Sei sulla pagina 1di 376

https://twitter.com/yudapearl (updated 5.5.

2020)

5.5.2020 2:24am - (Replying to @ilan_sinelnikov and @SSI_Movement) Congratulations!!

5.4.2020 10:22pm - #TheodorHerzl. And I almost missed his birthday! Here is his picture from a balcony in
Basel, a picture that ignited hearts, moved masses, made deserts bloom, created a state, and saved my family
from two infernos. What a balcony!!!

5.4.2020 4:43pm - (Replying to @zeemo_n and @AndrsMontealegr) The more I think about it the less certain I
am about the boundaries between science, philosophy and artificial intelligence. Especially if the science you
are doing is not conventional and the philosophy you are doing is driven by the puzzle: "So, how do people do
it?"

5.4.2020 5:10am - Philip Dawid has written a comprehensive overview of a his approach to CI
https://arxiv.org/pdf/2004.12493.pdf The do-operator is simulated by a decision variable in a Bayesian
Network. The paper illuminates what can be done without counterfactuals, a topic of my paper
https://ucla.in/2N92rBM

5.3.2020 8:41pm - (Replying to @EinatWilf) A tested way to uncover Palestinians' agenda to undo Israel is to
give their intellectuals a stage and let them rant freely about Zionism and its quest for coexistence. The truth
will come out immediately, on the first page, in perfect English, and in no ambiguous terms.

5.3.2020 5:27am - World leaders: Release imprisoned journalists worldwide - Sign the Petition!
http://chng.it/dfy75vvb via @Change

5.3.2020 5:22am - (Replying to @eyad_nawar and @eliasbareinboim) Determinism disappears if the action is
disjunctive, as in "paint the wall either green or red",see https://ucla.in/2N9nSCV. From the painter viewpoint
the action is deterministic (one choice) but from the policy maker, it is stochastic, not knowing what color will
be chosen.

5.3.2020 3:12am - There is a bit more to it than a language/calculus, it is a tractable algorithm that tells us,
not merely if a query is estimable, but also how to estimate it and, if it's not estimable, what kind of additional
measurements or experiments will turn it estimable. It was undersold.

5.3.2020 1:01am - (Replying to @ExogenyKarl and @eliasbareinboim) The paper you sited is one of my favorites,
but I think you missed the punch-line: "economics students should now be able to solve the eight toy problems I
posed in Pearl, 2013 (see Appendix A, Section A.2)." Do you think it has happened?

5.2.2020 10:30pm - Readers who have been shunning the do-operator for being "too deterministic" can now
embrace its stochastic cousin, "soft intervention", formalized in a recent paper by Correa and @eliasbareinboim
: https://aaai.org/Papers/AAAI/2020GB/AAAI-CorreaJ.8051.pdf

5.2.2020 5:32pm - (Replying to @agpatriota @guilhermejd1 and 4 others) It sure can be captured. E(Y|do(X=1))
= E(Y|do(X=0)) and E(Y|do(X=1),Z=0) =/= E(Y|do(X=0), Z=0) See Causality page 35-36

5 2 2020 4 57 (R l i t @R Bil10485167 @A N i d @U Ch 1) J ti t D i lP l
5.2.2020 4:57pm - (Replying to @RanaBil10485167 @AsraNomani and @UmarCheema1) Justice to Daniel Pearl
will usher justice to all innocent victims of brutality and extremism.

5.2.2020 4:50pm - (Replying to @Taks87) We hope that justice to Daniel will usher justice to all innocent
victims of extremism

5.2.2020 4:41pm - The parents of murdered Wall Street Journal reporter Daniel Pearl asked Pakistan’s
Supreme Court to prevent the men convicted of his abduction and murder from going free
https://wsj.com/articles/parents-of-slain-journalist-daniel-pearl-appeal-ruling-that-would-free-accused-
11588441160 via @WSJ

5.2.2020 5:30am - We wish to thank The Pearl Project @AsraNomani for leading the effort to realize justice for
Danny, the Committee to Protect Journalists @pressfreedom for their spirited support of this effort, & the Wall
Street Journal, for contributing half the legal fees needed for our appeal.

5.2.2020 4:52am - (Replying to @rameshfilms) Remember you very well, and the documentary too, with
Christiane Amanpour @camanpour . Glad you are going to screen it again soon; the world needs a reminder.

5.2.2020 3:54am - Today, my wife and I filed an appeal to the Supreme Court of Pakistan to overturn the
"acquittal" of our son's murderers. We ask all people of conscience to support our efforts to protect journalists
from imitators of Omar Sheikh and his ilk. For details https://medium.com/@dpearlproject/daniel-pearls-
parents-appeal-ruling-that-would-free-the-men-convicted-in-their-545f225014b8

5.2.2020 1:33am - (Replying to @ExogenyKarl and @RayDalio) Thanks, I'll try to refresh:

5.2.2020 12:44am - (Replying to @mribeirodantas @tdietterich and @StanfordHAI) I do not think the good
people at Microsoft and Google have the institutional clout to overturn the deeply entrenched thinking of the
data-fitting culture. Recall, it is quite traumatic to grasp the idea that "interpretation" means going beyond the
data and that it is necessary.

5.2.2020 12:22am - Could @RayDalio be the visionary executive who will save ML from itself by establishing a
"Center of Data-Interpretation" at the University of XYZ, to counter the "data fitting" hype? see
https://twitter.com/yudapearl/status/1255761208860237824Perhaps. If Microsoft and Amazon won't, why not
Wall Street? Dead serious!!

5.2.2020 12:03am - (Replying to @matt_vowels) Strangely, I had a similar experience, and so did many readers.
Doubly strangely, the authors of the right book write: "we have not found this approach to aid drawing of causal
inferences." You can't aid those who refuse your aid, but you can aim to understand of the refusal.

5.1.2020 7:53pm - (Replying to @kerstingAIML @mitbrainandcog and 4 others) Which "we" are we disagreeing on?

5.1.2020 7:33pm - (Replying to @bjh_ip) Aggree, and I am working on it.

4.30.2020 5:55am - A great article by David Suissa which, although does not target Dianne Lob explicitly,
provides an in-depth analysis of her mentality, and the mentality of other "Jews of discomfort", of whom I
wrote here: https://tinyurl.com/y5ryrgru

4.29.2020 11:12pm - (Replying to @KyleCranmer @tdietterich and @StanfordHAI) Simulators are causal, and
p ( py g y )
every causal model (completely specified) can be used as a simulator. But I can't see how DL can act as
surrogates for these models, since the inputs to DL are data, not the ropes behind the data, and the inverse
mapping data-->ropes is not unique.

4.29.2020 1:00am - A fairly harsh indictment of AI: https://1businessworld.com/2020/04/business/ai-cant-


solve-this-the-coronavirus-could-be-highlighting-just-how-overhyped-the-industry-is/ Why hasn't AI had more
impact? I have asked my colleagues at @StanfordHAI the same question: https://ucla.in/2JEhGyv, but I am not
sure they took notice. People blame the noisy data; shouldn't AI outsmart the noise-makers?

4.29.2020 12:15am - (1/ ) (Replying to @mribeirodantas @tdietterich and @StanfordHAI) It will happen when an
enlightened executive of Microsoft or Amazon, etc., will realize that data-fitting in addictive, and building
another "data-science center" will only worsen the addiction, and by the time industry will need to switch to
"data-interpretive science"
4.29.2020 12:29am - (2/ ) (Replying to @yudapearl @mribeirodantas and 2 others) there will be no one to hire
for the job. Nadda. The enlightened executive will then say: Let's establish a "Data-Interpretation Center" at
the University of xyz, and endow it with 0.01% of the resources now pouring into data-fitting centers. Get
ready, it will happen soon.

4.29.2020 12:10am - And the eyes of the world turn again to the Supreme Court of Pakistan and ask: What
message will it send to their sons and daughters?

4.28.2020 6:39pm - To readers who keep on asking: what's the secret of looking so good at 72, and of feeling so
happy despite 72 years of siege, the answer is simple: Its plain 1978 years of waiting, that's all it takes, just
1978 years.

4.28.2020 2:23pm - (Replying to @YFeyman @raj_mehta and 4 others) ignorability is "as if randomized, given Z"
and exclusion is "going only through X". Two independent restrictions that even seasoned IV experts can't
conceptualize w/o graphs, see why: https://ucla.in/2WX5T69

4.28.2020 2:07pm - (Replying to @YFeyman @raj_mehta and 3 others) Honestly, this is the first time I hear the
adjectives "clear" and "tractable" attached to the PO framework. It makes me extremely curious: can you
perhaps show us how you judge the plausibility of a simple ignorability assumption, like @stuartbuck1 :

4.28.2020 1:47pm - (Replying to @raj_mehta @YFeyman and 3 others) Good point. I could imagine conversations
like those we are witnessing on Twitter going on w/o the language of DAGs to facilitate them. Even devout
antaginsts of DAGs are using expressions such as "there are many paths from X to Y" "going only through W" etc.
It speaks a lot.

4.28.2020 3:46am - Another celebration starts today at sundown - Israel's Independence Day. Join me in
celebrating the most inspirational miracle of the 20th century: A tribe of beggars and peddlers lifting
themselves from the margin of history to become a world center of science, art and business.

4.28.2020 4:24am - (Replying to @richarddorset) Many oppressed peoples, Yazidis, Asyrians, Kurds, even
Tibetians are inspired by the Israeli experiment and studying the secret of its creation. The secret: Jews were
not a "nomadic tribe" after all, but a dormant nation, driven daily by the dream of returning to their homeland.

4.28.2020 4:46am - (Replying to @yudapearl and @richarddorset) For a good book on the creation of Israel, I'd
recommend Benny Morris "1948", though it may not meet your criterion of "new state based on ancient claims".
For my grandfather it was "old state based on daily dreams". He prayed 3 times a day "return us in sovereignty
to our land "
to our land.

4.28.2020 3:16am - (Replying to @maximananyev @analisereal and 6 others) I think we can generalize your
observation to read: applied econ use graphical metaphors quite often, albeit shyly and tacitly (eg see textbook
descriptions of IV), and the language of diagrams can improve their practice by encouraging them to make those
intuitions explicit.

4.28.2020 1:05am - Congratulations to @PHuenermund who will be joining the Editorial Board of the Journal of
Causal Inference (JCI). Paul's addition is an open invitation to all economists to embrace JCI as an effective
publication for disseminating ideas across disciplines.

4.27.2020 9:25pm - (Replying to @amt_shrma @autoregress and 8 others) The assignment need not be "random".
It can be by student's birthdate, or by results of an eyesight test, as long as it is not correlated with U and e_y.

4.27.2020 9:25pm -

4.27.2020 2:08am - Update on Israel Memorial Day. The sirens will sound today (Monday) at 8:00 pm Israel time
(1 pm EST,10 am PST). Join me in freezing (1 min) for Dani Balkore, who went to fight five invading armies in
May of 1948, and made it possible for my generation to live a semi-normal life.

4.26.2020 9:22pm - (Replying to @patrickkloesel and @PHuenermund) Based on my many conversations, the
percentage is around 95%. And, more importantly, it is not that they do not understand DAGs, they are
committed to not understanding them, expressed by: "we have not found them useful in our type of
applications".

4.26.2020 4:40pm - (1/ ) (Replying to @stuartbuck1 @PHuenermund and 2 others) I confess to not knowing
everything about "how economists think and reason about ignorability" because my knowledge comes primarily
from published articles, which may not reveal the entire process through which authors "think". I would
therefore be grateful to you if you could
4.26.2020 4:52pm - (2/ ) (Replying to @yudapearl @stuartbuck1 and 3 others) help me, and many other readers,
fill in the key principles of that process, by sharing with us how YOU think about ignorability. I capitalized YOU
because, economists may vary in their intepretations of the term. So, if you are willing, let us start with the
simplest
4.26.2020 5:00pm - (3/ ) (Replying to @yudapearl @stuartbuck1 and 3 others) 3/ "conditional ignorability"
statement, Y(X) || X | Z, where X is a treatment, Y the outcome, and Z a set of covariates that are candidates
for making the statement true. Can you walk us through the process by which you would judge whether the
statement holds in a given problem
4.26.2020 5:07pm - (4/4) (Replying to @yudapearl @stuartbuck1 and 3 others) or scenario, or a model, or a
story, or whatever helps you discern whether the statement holds true or not. No need to be pedantic on issues
such as "strong ignorability" vs. "weak ignorability", or how Z was chosen. Just examining the problem and
deciding true of false.

4.26.2020 6:52am - Good catch. This link should work I hope: https://arxiv.org/pdf/2004.09044.pdf

4.26.2020 6:27am - I am in receipt of a comprehensive paper on the role of causality in vision


https://arxiv.org/pdf/2004.09044.pdfAlthout Although I find the section on "causal Bayes Nets" a bit
incoherent, there is a lot I can learn from it, especially the interplay with naive physics and object-oriented
reasoning.

4.26.2020 3:41am - Caught me by surprise! But we must add that, with the exception of a few dragonic islands,
h d ll b l f
the dragons were eventually overcome by angels of commonsense.

4.26.2020 5:52am - (Replying to @PHuenermund) Not at all. It was in perfect timing; just a few days after the
angry dragons have issued a non-ignorable ultimatum: dragonize or perish.

4.26.2020 3:47am - 100th anniversary !!! And I almost missed it !! To readers who have not heard of the San
Remo Resolution, or the Balfour Declaration that triggered it, I wrote an oped on the latter
https://tinyurl.com/y4l2emnf which would be illuminating to read on this historic day.

4.26.2020 2:28am - ECOL VS. ECON While economists are moving from models to field experiments, ecologists seem to be going the other way. This paper:
https://researchgate.net/profile/James_Grace/publication/338163927_Scientist's_guide_to_developing_explanatory_statistical_models_using_causal_analysis_principles/links/5e8e72c9a6fdcca78901f151/Scientists-
guide-to-developing-explanatory-statistical-models-using-causal-analysis-principles.pdf seeks the advice of causal models to "guide scientists in their quest to develop explanatory hypotheses for evaluation."

4.25.2020 4:09pm - At sundown today, siren sounds freeze Israel into a 2-minute silence in remembrance of its
fallen - https://youtube.com/watch?v=R6rFMUw-pOk. In May 1948, our next-door neighbor, Dani Balkore (19),
kissed his family, smiled to us, and returned in a coffin two weeks later. I freeze for you Dani.

4.25.2020 2:46pm - (Replying to @LennyVds) I would not phrase it exactly as you did. I would say: The power of
SCM is that you can compute the likelihood of EVERY counterfactual sentence. For example, Y would be less
likely had X not happened, given that X actually did happen.

4.25.2020 2:27pm - (Replying to @aulderic @Liv_Boeree and @robertwiblin) Agree. The birth-weight paradox
should not be classified as an example of "Simpson's paradox". The surprise in the two paradoxes is of different
nature. Someone should fix Wiki, and add "not" to: "it can NOT be explained statistically".

4.25.2020 6:13am - Israel Yom Hazikaron National Online Memorial Ceremony 8:30EDT
https://israeliamerican.org/home/yhzceremony via @israeliamerican

4.25.2020 1:15am - (1/3) Here is a golden opportunity to learn something of lasting value during our COVID-19
hibernation. There have been dozens of articles, dialogues, discussions, even tantrum debates, laboring to
compare DAGs and PO frameworks, mostly written by distant commentators, who could
4.25.2020 1:15am - (2/3) not expose first-hand what it takes to solve a problem from beginning to end using
one framework or another. This link https://ucla.in/2WX5T69 will do it for you, in just a few paragraphs, using
a simple chain X-->Y-->Z. Once you read these paragraphs you will never
4.25.2020 1:15am - (3/3) be intimidated by an invitation to compare the two frameworks. You might even
become curious as to why some people would prefer one framework over another.

4.24.2020 10:21pm - (Replying to @maximananyev @Jabaluck and 7 others) I am not a bit surprised. "front-door"
is one of many models involving 3 variables and one latent confounder, so i am sure it came up naturally in
many applications. The fact that it permits nonparametric identification, came as a shock to statisticians
(narrated in #Bookofwhy)

4.24.2020 2:29pm - (Replying to @HannesMalmberg1 @Jabaluck and @maximananyev) You are beginning to see
why I called it "undersold". But the gains are not merely educational, they are methodological and
computational. Tasks that are intractable in PO (consistency, redundancy, testability, etc) suddenly become as
easy as writing down an SEM. Sadly undersold.

4.24.2020 2:17pm - (Replying to @HannesMalmberg1 @Jabaluck and @maximananyev) When I see the word
"design", I see a potential for an exciting PhD project: Let's take the mental model that gives the "designer" a
f li f "d i " f li it d h th d i h t d i thi b tt Thi d i l
feeling of "design", formalize it, and show the designer how to design things better. This was done in several
key areas, and science progresses.

4.24.2020 2:07pm - (Replying to @autoregress @PACESconsulting and 4 others) Thanks for adding a new acronym
to my arsenal. The missing-data literature refers to your CMAR as MAR. But there, "missing" means unobserved
variabls in some individuals, not missing individuals.

4.24.2020 1:56pm - (Replying to @autoregress @PACESconsulting and 4 others) What is CMAR? Is it Missing
Completely at Random MCAR, as used here https://ftp.cs.ucla.edu/pub/stat_ser/r473-L.pdf? If so, my answer is
probably YES. Most non-parametric problems in economics now have solutions. But it was undersold. Eco. was
not fortunate to have a leader saying: Look guys, Its Great!

4.24.2020 1:29pm - (1/ ) (Replying to @The_RickMc @PHuenermund and 3 others) This is an important point.
What make us judge a treatment as "as-if random" when we have no coins nor lotteries? It is our subjective
inability to conceive of a confounding mechanism. Take the price of beans in China and the traffic in LA. None
is random.
4.24.2020 1:40pm - (2/ ) (Replying to @yudapearl @The_RickMc and 4 others) Each varies systematically in its
own dynamics. Yet, unless you are a smart alec and argue that some future-broker in LA may panic and call his
clients to rush to their cars, the two variables are judged to be unconfounded. Randomization is sufficient, but
not necessary.

4.24.2020 1:20pm - (Replying to @PACESconsulting @gelbach and 4 others) My first sentence says that for
decades Heckman's selection-bias correction was tied to strong parametric assumptions. Nothing wrong with it.
But if we want to snap out of such assumptions we need the logic of non-parametric models. That logic was a
"novelty", highly undersold.

4.24.2020 11:00am - (Replying to @PACESconsulting @gelbach and 4 others) New results after decades of
stagnation do demand new tools. Specifically, the idea that selection bias can be reduced or eliminated without
making parametric assumptions demands tools to handle non parametric models. Why the resistance to novelty?

4.24.2020 10:49am - Measurement errors are represented as proxy variables, as shown here
https://ucla.in/2N5GNOK . The do-calculus just operates on the graph after these proxy nodes are added. In
most cases identification requires auxiliary machinery, as shown here https://ucla.in/2N9icIX.

4.24.2020 2:24am - For the many who inquired on the latest in the movement against the "acquittal" of our son's
murderer, this article offers a fairly comprehensive report. https://wsj.com/articles/pakistani-authorities-
seek-reinstatement-of-death-penalty-in-killing-of-wall-street-journal-reporter-daniel-pearl-11587582816
Thanking all readers for their help and empathy.

4.24.2020 2:21am - To our readers in Israel, The Israeli Association for Artificial Intelligence has announced a
Doctoral Dissertation Award. Details are here:

4.24.2020 1:37am - (Replying to @teemu_roos and @UusitaloLaura) You are so right, skipped my mind and my
search engine. It is indeed in https://ftp.cs.ucla.edu/pub/stat_ser/r69-reprint.pdf. Which, BTW, re-reading it
32 years later, is a damn good paper!. I don't recall if it was folklore. The firing squad was folklore (mentioned
in I. J. Good)

4.24.2020 1:25am - From the day I was told that RCT's require "careful design", I suspected that RCT's experts
must be using a mental causal model. Here is a paper that supports that suspicion:
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3331772 and shows experts what they can gain by making
that mental model explicit.

4.24.2020 12:32am - (Replying to @Jabaluck @PHuenermund and 2 others) Luckily, I have not met a serious
economist who thinks I am a charlatan. They know what you know, that my "guilt" is only in calling to make
assumptions explicit in a language that is most meaningful to researchers (SCM) but became unfashionable in
1980, yet producing new results

4.23.2020 10:51pm - (Replying to @yudapearl @PHuenermund and 3 others) And there is a good reason why
justification is avoided. Justification boils down to saying: "I cannot think of any factor that would affect both X
and Y." It is tabooed on 2 counts: (1) It's dreadfully subjective, (2) It invokes"factors affecting," almost like
graphical criteria

4.23.2020 10:43pm - (Replying to @PHuenermund @Jabaluck and 2 others) Moreover, students (and their
advisors) reading the final papers get the impression that as soon as you say: "we assumed ignorability" and cite
revered authors who also did so, you are a scientist. The ratio of such papers to justifying paper is 100:1 (based
on mental counting).

4.23.2020 10:28pm - It is always a thought provoking exercise to imagine how the general public would react to
issues that we have been debating on Twitter with such intensity. One thing is clear, causality is gaining stature
as an independent species of scientific inquiry.

4.23.2020 9:49pm - (Replying to @maximananyev and @Jabaluck) You are almost right. Applied econ. will
challenge every assumption that is explicit, but would rarely challenge those that hide under fashionable
technical titles such as "ignorable", which hardly anyone knows what it means. It's the old peril of being honest
and transparent.

4.23.2020 9:26pm - (Replying to @Jabaluck @autoregress and 3 others) What's the point? Creative people can
always find a link between the price of beans in China and traffic in LA. If they believe the link is strong, put an
arrow and continue. If its weak enough to ignore, remove the arrow and continue. W&C did the latter and failed
to continue.

4.23.2020 9:10pm - (Replying to @Jabaluck @autoregress and 3 others) The only economist I met who thought
"carefully about DAGs" was the late Hal White. If you know more, they were not here: https://ucla.in/2mhxKdO
when I asked them to answer a few elementary questions, like which parameter can be identified by OLS, or
which model is testable.

4.23.2020 8:53pm - (Replying to @Jabaluck @autoregress and 3 others) There are many ways to justify absence
of an arrow. For me, it is enough to say XWY are medical factors and Z's behavior is a bureaucratic one. Done. If
we must cheer Angrist for his insight, fine, but we still need to decide if the query is identified.This is where
WC fumbled.

4.23.2020 7:47pm - Great initiative, and a very informative window into the enterprise of "Evidence Based",
which I never understood. Speaking of evidence from population to individual, the idea of combining
experimental &observational data should illuminate the discussion: https://ucla.in/39Ey8sU

4.23.2020 4:07pm - (Replying to @autoregress @gelbach and 3 others) If treatment Z is administered by a clerk
who follows some protocol on how to respond to symptom W, then this clerk is shielded from the confounder U,
which affects patients, not clerks. I'm sure you have clerks in economics?
4.23.2020 3:57pm - (Replying to @Jabaluck @autoregress and 5 others) But to assure consistency among a set of
dependence-independence assertions, we must assume that they emanate from some, possibly unknown
distribution.

4.23.2020 2:35pm - (Replying to @Jabaluck @autoregress and 5 others) This is the beauty of DAGs. You do not
commit to a specific structural model. You commit only to the class that is compatible with your knowledge,
namely, the source of variation of each variable. This is many time more meaningful than "which PO is ignorable
conditional on whom"

4.23.2020 2:21pm - I surely read it, and commented on it here: https://ucla.in/36EoNzO

4.23.2020 2:12pm - (Replying to @Jabaluck @autoregress and 5 others) I have a friend who "understands" the
connection between independence and probabilities but he would never put the definition X||Y iff
P(X,Y)=P(X)P(Y) on paper, or in words. His papers are full of X||Y and X||Y|Z sentences but no mention of P.
His disciples claim he "understands"

4.23.2020 1:25pm - (Replying to @autoregress @guilhermejd1 and 5 others) If "Yup" means "Yes" then we
disagree on the 2nd "Yup". Imbens-Rubin-Angrist not only do not mention this connection in their books, they
resist the connection like a plague. But I welcome your first "Yup", and hope that we can agree on its logical
implications

4.23.2020 1:09pm - (Replying to @autoregress @guilhermejd1 and 5 others) I did not say "ARE" causal models. I
said "emanate from" a structural model, or, "can all be derived from" a structural model. Do you buy that? Do
Imbens-Rubin-Angrist-Peschke buy that? This is the 1st Fundamental Law of causal inference.

4.23.2020 12:48pm - (Replying to @guilhermejd1 @autoregress and 5 others) "Embrace"? You must be kidding. I
guess the idea that potential outcomes are intrinsic properties of SCM has not been sunk in yet. Every SCM
model assigns a probability to every conceivable potential outcome. Potential outcomes are rooted in SCM, and
trees w/o roots wither.

4.23.2020 4:53am - 1/ I am retweeting this post, knowing that hundreds among our readers are intimidated by
colleagues on the issue of "real-world" example. As if going to Somalia and conducting field experiments
elevates you to a new level of understanding. The post demonstrates a "real-world"
4.23.2020 4:53am - 2/ problem that was mishandled by prominent investigators precisely because they
dismissed the tools of causal inference. In my follow-up Tweet:
https://twitter.com/yudapearl/status/1253244362584031232 I ask: How many "real world" problems are
mishandled by those who talk "real world" to dismiss CI tools?

4.23.2020 1:48am - (Replying to @yudapearl @autoregress and 4 others) Now, Wermuth and Cox are not DAG
fanatics. They posed this problem b/c it came up in their investigation of sequential treatments. They wrote 3
articles on this problem, which involves only 5 variables. What does it tell us about economists and their "real
world" arguments?

4.23.2020 1:29am - (Replying to @yudapearl @autoregress and 4 others) More on "Real-world" herrings. I was
asked repeatedly to give one "Real-world" example which do-calculus can solve, and which "real-world" experts

could not. Here is one, https://ucla.in/2UX4jii, called "indirect confounding" by Wermuth and Cox, also
discussed in #Bookofwhy p.241
4.23.2020 12:59am - (Replying to @autoregress @gelbach and 3 others) Every setting that you would be willing
to call "real-world" can be solved by these problems, or through the general theorems that these problems
illustrate, or labeled "unsolvable" by the completeness theorems that follow (in later papers)."Real-world"
boasting is a red herring.

4.23.2020 12:45am - (Replying to @SylvainCF @pa_chevalier and 17 others) I am embarrassed to be on the same
list as the great Hume. But, given what I say about him and about counterfactuals on page 266-9 of
#Bookofwhy, perhaps he would accommodate my presence on the list.

4.23.2020 12:27am - (Replying to @Jabaluck) The trouble comes when you assume what you need BECAUSE you
need it, not because you know it. And this indeed is the common case when PO folks assume conditional
ignorability, save for RCT's, and some IV's settings, when natural lotteries are available.

4.22.2020 11:43pm - (Replying to @autoregress @analisereal and 3 others) How can you say that Pearl ignore
monotonicity? See my discussion on monotonicity with Imbens https://ucla.in/36EoNzO. Not only can SCM
articulate monotonicity when it is plausible, but we can also test it when it is in doubt (See Causality). We do
not shun tools, we use them.

4.22.2020 11:02pm - I like your crisp distinction between stating what one "needs" vs. stating what one "knows".
It really summarizes the entire debate between the CI and PO frameworks. In CI we use structure to encode
what we know, PO to encode what we wish to know, and logic to connect the two.

4.22.2020 10:47pm - (Replying to @analisereal @autoregress and 3 others) I have seen some. Imbens and Rubin
(2015) call these identifying assumptions "The Science."

4.22.2020 10:35pm - (Replying to @autoregress @gelbach and 2 others) Puzzled. How could economists possibly
have fought this or a similar war, if they did not have a 3-rung causal hierarchy in mind, nor complete tools for
managing each rung. Can you point me to any similar war? Was "moving on" a resolution or "give up, lets do
what we do best???"

4.22.2020 9:08pm - (Replying to @gelbach @guilhermejd1 and 3 others) Can you use Heckman's correction to
solve any of the siimple problems solved here: https://ucla.in/2OgMF4g ???. As I stated before. CONCERN - yes.
Results - ??? And I dont blame you, new results demand new tools, which your students will have to acquire by
independent reading.

4.22.2020 8:56pm - (Replying to @jeffreywatumull) No, this kind of assumptions have no testable implications
and are needed as input to causal calculus.

4.22.2020 8:43pm - (Replying to @maximananyev @autoregress and 5 others) Can you summarize how
@ecnomeager represents disparities between two populations.

4.22.2020 8:38pm - (Replying to @guilhermejd1 @autoregress and 3 others) Agree with you completely.
Economists have indeed been "concerned" with these questions, as were most scientists since Campbell etal.
But when you ask them to solve a tiny problem, eg https://ucla.in/2N7S0K9 you see how far behind they have
allowed their "concerns" to take them.

4.22.2020 8:14pm - (Replying to @Jabaluck @gelbach and 2 others) Can we put aside me and my failures to
engage. Let's return to economics "embracing" machine learning, and examine what precisely has economics
" b d" i th t3d d
"embraced" in the past 3 decades.

4.22.2020 8:09pm - (Replying to @autoregress @Jabaluck and 2 others) See my reply to Imbens
https://ucla.in/36EoNzO who made similar points.

4.22.2020 8:06pm - (Replying to @Jabaluck @gelbach and 2 others) It is more than a question "easier". It is one
of "intractability" which in practical terms means "impossible". And let's leave DAGs aside. I am insisting only on
transparency, and would welcome a transparent alternative to graphical models.

4.22.2020 8:00pm - (Replying to @Jabaluck @gelbach and 2 others) I seriously doubt that you will that that
flexibility. Your students will, but not those who grew up in the insular green house of 1990's econometrics.

4.22.2020 7:51pm - (Replying to @Jabaluck @gelbach and 2 others) Forget discovery. I want economists to do
what they have been advertising to be doing since Haavelmo, before being hijacked by statisticians who
believed that conditional independencies among counterfactuals is the "science" that researchers carry in their
minds while modeling.

4.22.2020 7:41pm - (Replying to @Jabaluck @gelbach and 2 others) Now you are generalizing. I never insist on
"my work", this is your hang up. I insist only on starting with cognitively meaningful and defensible encoding of
what we know, combining it with data, and formally deriving what we wish to know. Has any of your model
papers done that?

4.22.2020 7:13pm - Thanks for this positive and informative reviews. I found a short link here
https://tinyurl.com/ycjn3hg3. This is, if I am not mistaken, the first review of #Bookofwhy in a flagship
scholarly journal. I hope to see more soon, so that readers of BOW will begin to feel mainstream.

4.22.2020 6:52pm - (Replying to @Jabaluck @gelbach and 2 others) Can we get to the causal part?. These
papers are still in the data-fitting paradigm, taking "conditional ignorability" as a God given gift.

4.22.2020 6:37pm - Thanks for the feedback. And, now that you are at it, please read some of the antagonistic
critics, as many as you can, to truly appreciate the stifling paradigms that #Bookofwhy has attempted to shift.

4.22.2020 6:28pm - (Replying to @Jabaluck @gelbach and 2 others) Tell me more on how economics has
embraced machine learning, in substance please, not in hype. And please focus on that part of machine
learning which is not curve-fitting, namely computer-aided statistics. FYI, CI folks consider CI to be part of ML,
albeit above Rung-1

4.22.2020 6:22pm - (Replying to @Jabaluck @gelbach and 2 others) Beg to differ. The resistance generated was
not to overselling novelty but to my insistence that new tools are needed. Challenge: show me one oversold
novelty in https://ucla.in/2mhxKdO, or in: https://ucla.in/2WX5T69

4.22.2020 12:47pm - (Replying to @gelbach @Jabaluck and 2 others) The stagnation we see in certain
(unnamed) disciplines calls for courageous and rebellious spirit, not for professional humility (or timidity),
which has led to the insular stagnation.

4.22.2020 12:36pm - (Replying to @jeffreywatumull) Yet, unless we suspect that subjects response was affected
(positively or adversely) by the very idea of being selected for treatment, it is reasonable to take the un-
selected subjects as control group.
4.22.2020 12:30pm - Kudos to Sacks et al for creating the CRAN program, which should be used together with
their Biometrika paper https://sachsmc.github.io/causaloptim/articles/CausalBoundsMethods.pdf - A
comprehensive account on computing tight bounds for causal effects.

4.22.2020 12:28am - (Replying to @RemyLevin @HeerJeet and 2 others) I am not familiar with a lot of cross-
disciplinary critiques, except the ones I have articulated here https://ucla.in/2mhxKdO and here #Bookofwhy.
Speaking of these critiques, I see concrete examples, formal proofs and powerful tools - not one shred of
dishonesty or X-measuring!

4.21.2020 11:13pm - (Replying to @RemyLevin @HeerJeet and 2 others) Nice and rosy. Still, you would agree, I
hope, that some fields, at some points in time, could benefit from some jolt of self reflection. And the jolt may
not always come from inside, given the self-perpetuating structure of our academic culture.

4.21.2020 11:00pm - (Replying to @Jabaluck @autoregress and 2 others) Sorry, I did not realize that. So, back to
waiting for a prominent economist to take a sober look at the education and culture dominating the field.

4.21.2020 7:11pm - Responding to critical yet constructive reviewers, we have revised our "generalizing
experimental results.." paper: https://ucla.in/2FqikNZ, and we are happy to report new discoveries concerning
bounds and new Bayesian estimation of real data on Vitamin A supplement.

4.21.2020 4:21am - (Replying to @RWerpachowski and @Spinozasrose) I did not think you said it in a mean way.
My point still is that, in order for the memory of the Holocaust to be long lasting we must tie it with revival and
rebirth. And I do not think the directors of our 60 Museums understand it; you can hardly see a glimpse of Israel
in any.

4.20.2020 4:18pm - (Replying to @RWerpachowski) My grandparents failed to escape from the train, and I know
that they wanted to be remembered not by the train, but through their grandchildren lifting themselves from
the ashes and carving a future of hope and dignity for themselves, their people, and their descendants.

4.20.2020 3:52pm - This picture, not the cattle trains nor the gates of Auschwitz, should decorate the entrance
hall to each one of the 60 Holocaust Museums in the US.

4.20.2020 3:27pm - (Replying to @HeerJeet @guacamolebio and @beyerstein) I've never dreamed to hear this
echoed in the middle of the "credibility revolution." #Bookofwhy

4.20.2020 10:12pm - (Replying to @VUspenskiy and @andrewheiss) Actually, the sequences of do-calculus rules
have been systematized into an algorithm that gives us the result without going through the rules:
https://ucla.in/2N9JIWS But when I look at a diagram, I first ask if it contains one of the recognizable patterns
stored in my mind.

4.20.2020 10:01pm - Naive questions are everyone's questions. Ans. The backdoor and frontdoor are logical
consequences of do-calculus. So why do we decorate them with names? Because they are easily recognizable in
the DAG, so we store them explicitly in our arsenal of tools, so skip re-deriving them

4.20.2020 4:53pm - (Replying to @EinatWilf) Does anyone understand why the American Jewish Committee
would not acknowledge the dreams and ideas that inspired the Warsaw Ghetto Uprising?

4 20 2020 4:18pm (Replying to @RWerpachowski) My grandparents failed to escape from the train and I know
4.20.2020 4:18pm - (Replying to @RWerpachowski) My grandparents failed to escape from the train, and I know
that they wanted to be remembered not by the train, but through their grandchildren lifting themselves from
the ashes and carving a future of hope and dignity for themselves, their people, and their descendants.

4.20.2020 3:52pm - This picture, not the cattle trains nor the gates of Auschwitz, should decorate the entrance
hall to each one of the 60 Holocaust Museums in the US.

4.20.2020 2:25pm - (Replying to @danilobzdok @shakir_za and 3 others) Agree. It's an excellent Sunday read.
But the message of this paper https://ucla.in/2Jfl2VS was "a reminder of how easy it is to fall into a web of
paradoxical conclusions when relying solely on intuition, OR solely on statistics". We must snap out of statistics
to get it right.

4.19.2020 2:15pm - Conditioning just gives us a more specific class of individuals, with all the problems of
sparse data; it is still not the individual itself. This paradox https://ucla.in/2N6x36Q highlights the difference,
and the role of counterfactual reasoning in getting to the individual.

4.19.2020 12:41am - (Replying to @Physical_Prep) Not surprised that the #sportscience literature suffers from
similar myths. ACE is merely one evidence for what we wish to know about an individual. See eg our recent blog
https://ucla.in/39Ey8sU on which individual is "in greater need" for a hospital bed.

4.19.2020 12:32am - (Replying to @omaclaren and @unsorsodicorda) I am genuinely trying to learn from the vast
literature of DE. Can you give me the simplest example you have, time invariant, where the task of inferring
effects can benefit from the teachings of DE? I've just given you one from which qualitative DE folks can
benefit.

4.19.2020 12:12am - (Replying to @omaclaren and @unsorsodicorda) This sounds exactly like CI. In which case
can DE folks get the causal effect of X on Y, given data taken from Pr(X,Y,Z) if the input is three unknown
functions, x=f1(z,u1) y=f2(z,x,u2) z=f3(u2) and u1,u2,u3 are independent yet arbitrarily distributed random
variables? How?

4.18.2020 11:57pm - My current interest in "personalized medicine" compelled me to read this new paper
https://tinyurl.com/y7c7ngry. I must conclude though that, despite 3 decades of CI, "personalized nutritionists"
still have to accept that there can be no "personalized" inference w/o counterfactuals.

4.18.2020 11:33pm - (Replying to @unsorsodicorda and @omaclaren) Can you summarize some of this literature
in input-output language, as we folks do with CI? What must the input be like + what kind of data is taken and
what kind of answers we get from the analysis?

4.18.2020 11:19pm - (Replying to @seema_econ) Yes, there sure is "a little virus going around"!

4.18.2020 7:14pm - (Replying to @omaclaren) I would be very happy to enlist the work done in CI as a modest
contribution to the vast umbrella of DE, assuming the latter is open-minded. For example, the adjustment
formula can easily be viewed as an exercise in qualitative DE, so is the Instrumental Inequality etc. etc. etc.

4.18.2020 4:09pm - Readers suggest that this tutorial video https://tinyurl.com/y88fhnd9 may well serve as an
introduction to an online course in CI. Unfortunately, the videographer was sloppy on the slides, so I am posting
the slides here: http://web.cs.ucla.edu/~kaoru/nips-dec2013-with-eb-bw.pdf Feel free to use them in any
class.
4.18.2020 3:30pm - I am hearing great passion for ODE, so I ask myself: "Why not use ODE for causal inference?"
One answer: We don't have the functions that make up ODE. All we have are some qualitative properties, eg.
who are the arguments of each function, and we still venture to estimate effects

4.18.2020 4:20am - Another exciting paper arriving at my desk reads: Causal Relational Learning:
https://arxiv.org/pdf/2004.03644.pdf which promises to revolutionize causal inference the same way first-
order predicate logic has transformed Boolean logic.

4.18.2020 3:36am - I am always happy to see "completeness" results in CI, but these results
https://arxiv.org/pdf/2004.04872.pdf are especially gratifying, because graphical modeling of missing data
problems (https://ucla.in/2LdEjZW) have totally been ignored by MD practitioners. I hope things will change
soon.

4.18.2020 3:04am - On the basis of the first 2 accessible sections, this is the best on line course on the market.
Highly recommended. Especially for economists!!!

4.18.2020 1:17am - Here is a fun question that came up in conversation with my grandson. If it is bouncing off
the moving piston that speeds up molecules as gas is compressed, how come slow-moving and fast-moving
pistons result in the same temperature increase? A question for AI reasoning systems.

4.18.2020 1:04am - (Replying to @attilacsordas) What about "music breaks the silence". For every thing that is
created the absence of that thing is destroyed.

4.18.2020 1:00am - (Replying to @shaul_ido) Trying to find one

4.18.2020 12:56am - (Replying to @mribeirodantas) I love it too. And this paper https://ucla.in/2Jfl2VS
describes a machine that generate such data on demand.

4.18.2020 12:49am - (Replying to @firstmn59) Good for Rung 1 and Rung 2 (interventions), weak on Rung 3
(counterfactuals and SCM), so I would supplement it with chapter 4 of Primer: https://ucla.in/2G2rWBv

4.18.2020 12:46am - (Replying to @juli_schuess and @Jean_DeCarli) Highly recommended.

4.18.2020 12:40am - Your question is a good one, and it reminds me to remind readers of this link
https://ucla.in/2Kz0FoY which would provide them with a searchable file with all my past tweets (now
numbered 46K). Please search for "cyclic" or "feedback".

4.17.2020 5:29am - I am retweeting your query to all our readers, I hope some can point you to a good on-line
course based on Primer https://ucla.in/2KYYviP. I lost track.

4.17.2020 3:19am - The more I listen to it, the more I feel I can speak Mandarin. Perhaps because cause and
effect are universal or because (no, this is impossible!!) I can really understand Mandarin.

4.16.2020 10:13pm - Amazing photo. I believe it is the first public demonstration complying with COVID-19
"social distance" rules. Israelis protested yesterday government corruption and the erosion of democratic
institutions.
4.16.2020 5:46am - Here is how Richard Feynman explains why molecules speed up when gas is compressed.
About 5-6 minutes into this video https://youtube.com/watch?v=P1ww1IXRfTA&feature=emb_rel_end he
explains how a gas molecule, fairly stupid one, knows that it must increase it kinetic energy to make up for the
work done on the gas.

4.16.2020 5:12am - Replying to @omaclaren) The choice of the word "governing" may be misleading. He meant
the law is never violated, which does not imply that it is sufficient for, or that it cognitively "explains" all
phenomena.

4.16.2020 4:42am - This is incredible! Straight in the Journal of Statistics Education?! If it's true that "The Book
of Why (Pearl and Mackenzie, 2018) increased the motivation to teach concepts that Pearl calls the causal
revolution" then writing the book was worth every drop of ink (key stroke.)

4.16.2020 4:18am - Conservation laws do not satisfy our quest for causal explanation. Does a molecule know
that it needs to conserve energy? So what makes it speed up in a Venturi pipe? Or when gas is compressed?
Feynman always preferred putting himself "in the shoes" of those molecules. Me too.

4.15.2020 7:05pm - (Replying to @weissiam) A theory that explains everything, from quantum mechanics to
cosmology, humbles the wise to remain silent ("Yidom" Amos, 5). I must remain silent till I learn how to use this
theory to resolve Simpson's paradox, or to spot imperfect experiments. Will be back if wised.

4.15.2020 3:25am - Very pleased to find a powerful completeness result in the theory of transportability:
https://aaai.org/Papers/AAAI/2020GB/AAAI-LeeS.6006.pdf This time the heterogeneous data sets can have
missing components and the targets can be group-specific causal effect.

4.15.2020 1:05am - (Replying to @rodakker) Thanks for the link. Again, the abstract raises my objection. The
link is MORE than just interaction, its confounding too. Traditional mediation analysis attempts to cast a causal
problem in statistical language - an impossibility, as shown here: https://ucla.in/2N5Gdk2

4.14.2020 10:40pm - A pay-wall prevents me from reading this whole paper


https://ncbi.nlm.nih.gov/pubmed/32175754 But the abstract raises a huge red flag: "In psychology, the causal
process between 2 variables can be studied with statistical mediation analysis." IMHO, there is no such thing as
"Stat. Med. Analysis."

4.14.2020 9:04pm - Happy birthday, Tel Aviv I've never seen your beaches so sad Your sons and daughters alone,
yet playful and glad, Your true face, unmasked, smiling openly, wide Chosen by tomorrow, a city in white

4.14.2020 4:37am - Malki and Danny, my two fallen souls, Will the darkness of the night remember Two shooting
stars went by? Shining their most brilliant lights, their noblest, In their last dance through the sky?

4.14.2020 12:10am - (Replying to @aminsaadou) Of course, provided the editor of Econometrica commits to
publishing it. But why do we need a paper? Readers and students will benefit more from seeing economic

solutions to the questions I posed to Heckman here: https://ucla.in/2mhxKdO. A toy problem is worth 100
papers.

4.13.2020 7:43pm - Econonists are trained to find flaws in models, however plausible, so much so that they
p p y
have lost the capacity to ask: "Suppose the model was correct, what do I do with it?" As a result, they lost the
skill to repair the flaws they labor to discover -- others (epi?) are doing it.

4.13.2020 5:59pm - (Replying to @stephensenn @pash22 and 2 others) Good try, but causal calculus was never
proposed to represent "what we mean by causal effects" as did RCTs. "Causal chains" (eg DAGs) and
"counterfactuals" were proposed as the building blocks of causal thinking. And they are!

4.13.2020 5:56am - (Replying to @thosjleeper) Oh, thanks, I missed the beginning of the chatter; got there only
when the question was put for a vote: who is better? Eager to see the initial trigger.

4.13.2020 3:16am - Funny, we asked this question six years ago: https://ucla.in/2FwdsGV and received 38 juicy
comments. The summary https://ucla.in/2WX5T69 then describes sympathetically what it feels like being an
economist, denied the guidance of graphs, and pretend you don't need any.

4.13.2020 1:28am - Anyone concerned with inferring causes from effects as well as predicting individual
behavior from group data has probably stumbled on the questions raised in this blog entry
https://tinyurl.com/rvt6ry5, and would surely benefit from the answers provided by Carlos Cinelli.

4.12.2020 2:11pm - We are grateful to Congressmen Adam Schiff and Steve Chabot for taking a strong,
bipartisan stand on behalf of justice and press freedom . https://schiff.house.gov/news/press-
releases/statement-from-freedom-of-the-press-caucus-co-chairs-schiff-and-chabot-on_commutation-of-
sentences-for-those-convicted-of-murdering-daniel-pearl @justicefordanny

4.12.2020 2:32am - "Walking among alligators" means seducing scientific leaders to undergo a paradigm shift
under hypnosis, thinking they knew it all along. No one will forgive he/she who finds your wallet and proves
that the time you spent under the lamppost - an entire career - was a waste.

4.12.2020 1:55am - (Replying to @phyecon1) I was not aware of this incident. My Pakistani friends always depict
Jinnah as a champion of modernity and moderation.

4.11.2020 6:12pm - (Replying to @quantadan) I think I owe this affliction to my Kibbutz, in upper Galilee, for
sending me (1954) to study choir conducting in Haifa, see https://youtube.com/watch?v=PvByNKs1bmw, and to
my son Danny, who was the copy-editor of my first book, and kept on complaining about my writing: "It doesn't
sing, Dad!"

4.11.2020 5:06pm - (Replying to @vthorrf) Brilliant! I wish he lived in the digital age. He would have become an
AI pioneer; he wouldn't have allowed philosophy to linger in philosophy departments.

4.11.2020 2:30pm - (Replying to @vthorrf) Is it Hume, or Locke?

4.11.2020 4:20am - (Replying to @ClaudeAGarcia) At the risk of provoking more anger I would just say that,
understandably, the naked is angry at the sunlight, though we are all naked under our pajamas, and we all
know it.

4.11.2020 12:31am - (Replying to @yakir_bella) If you can find a translator, I would waive my royalties and
donate a free copy to every college library. Why? Because Israel has limited resources, and can't afford to see
its science education taken over by the data-fitting style of machine learning - it's addictive & blinding.
4.11.2020 12:16am - If we look carefully at his hand movement, we see that the little one is testing the
"mediation formula"; how the spiral wheel turns and stops with and without the linkage in between. He is surely
climbing rung-3, toward creative scientific thinking.

4.10.2020 3:56pm - (Replying to @11kilobytes) Two reasons for anger: (1) "We have been doing it already in
19xx" (2) "There is more to it than #Bookofwhy"

4.10.2020 3:16pm - I find this tweet impossible to disagree with, though I am suffering from an illusion that you
won't find a scientist who hasn't read #Bookofwhy, barring those who are angered by it.

4.10.2020 2:29am - (Replying to @Michael_D_Moor) Hillarious! No wonder economists refuse to solve "toy
problems" -- everyone would be able to see how their wheels turn.

4.9.2020 8:16pm - (Replying to @tylerlu) And the chemist will say: Why study connectionism if we can simulate
the chemistry of proteins and have machines solve the human brain mystery.

4.9.2020 7:37pm - I have discovered a spec of truth in Bajwa's fable: The order to keep Shaikh in custody on the
grounds of ‘public safety’ was taken INDEED under pressure from external agencies -- I was one of those
agencies when I reminded the world who Shaikh is and what he did to my son Danny.

4.9.2020 6:27pm - This is my favorite photo of Danny https://cjr.org/analysis/daniel-pearl-omar-saeed-sheikh-


justice.php . The article echos our hopes for justice, moderation and press freedom.

4.9.2020 3:42am - I was about to send warm wishes to all readers celebrating Passover, when this came in.
Obama is a much better orator, so I'll ride his words and second his wishes: Happy Passover !!!

4.9.2020 3:31am - Only those who have travelled this road in normal times would appreciate what it takes to
bring it to this state of Covid-19 paralysis. It is the road connecting Haifa and Tel-Aviv on Passover eve.

4.8.2020 7:35pm - You go with Sarsour and Omar, Senator, not me. You have brought shame and immeasurable
harm to your (former) people. Tonight, as Jews say: "Keivan Sh'Hotsi et Atsmo Min Haklal" ("estranged himself to
the community") we will be thinking of you, among other painful disappointments.

4.8.2020 1:48pm - Anyone who wishes to obtain the solution manual for the beautiful examples in
https://ucla.in/2n3VkuU should write to kaoru@cs.ucla.edu and indicate that it will be used for self-study and
will not undermine instructors who assign these questions as class homework #Bookofwhy

4.8.2020 2:51am - (Replying to @y2silence) Best compliment an author can dream of. #Bookofwhy

4.8.2020 1:09am - (1/ ) (Replying to @bariweiss) My memorable Seder was on April 23, 1948, 3 weeks before
Israel was created and attacked. We, kids, heard all the threats of "monumental genocide" and "rivers of blood"
that our neighbors broadcast from Cairo and Beirut, and were curious to see if our "adults" would dare
4.8.2020 1:16am - (2/ ) (Replying to @yudapearl and @bariweiss) celebrate Passover under such conditions. My
grandfather just read the Haggada, as usual, got to the story of the five Rabbis who defied Roman's rule and
chanted the exodus in Bnai-Brak (my home town), stopped, smiled, looked us in the eye and said: "So, what's
new?"
4.7.2020 11:12pm - (Replying to @fuzzydunlop123 @omaclaren and 7 others) Thanks for posting. I was not
aware of this insightful paper on the connection between the two models.

4.7.2020 10:45pm - For the life of me, why should hypocrisy and stupidity evoke more twitter chat than being a
Zionophobic racist? Who said inconsistency is a greater perversion than immorality.

4.7.2020 7:03pm - Another juicy proverb, just provoked by a discussion with a "missing data" expert: "When it
comes to perplexity, nothing tops the science of inertia, except perhaps the inertia of science."

4.7.2020 5:36pm - Any reason you did not like the ending?: "Data science is not a mirror through which data look
at themselves from different angles under different makeups"

4.7.2020 4:02am - (Replying to @Razorwindsg) We do not have an organized movement per se. But we have a
commitment to education (on-line) and a strong conviction that commonsense will prevail. I would start with
spreading the intellectual content of CI in whatever circles you dance; most circles would benefit from it.

4.7.2020 3:22am - A more effective way to write Pakistani Prime Minister Imran Khan in support of justice for
Daniel Pearl would be to use these three email addresses: ambassador@embassyofpakistanusa.org
rtbaldoz@gmail.com Cc: justicefordanny@danielpearlproject.org Thanks, and may justice prevail.

4.7.2020 2:16am - (Replying to @nickchk @Jacobb_Douglas and @Fhanksalot) Discussants on this thread might
enjoy a glimpse at the general condition for Z to be a valid IV. The condition goes: 1) There is an unblocked
path between Z and X, and 2) Every unblocked path between Z and Y contains an arrow into X. This condition
confirms of course the cases

4.6.2020 3:08pm - Responding to readers comments, we have improved our discussion of "Which Patients are in
Greater Need?..with reflections on COVID-19" https://ucla.in/39Ey8sU. Same with the related post, that argues
for "data-interpreting technology" https://ucla.in/2JEhGyv.

4.6.2020 3:56am - (Replying to @maliniw90th) Your father is lucky. I wish my daughter would ask me to explain
Fourier transform or integration by parts. Fun topics.

4.6.2020 1:55am - Speaking of DL and causal inference (CI) and keeping with our commitment to on-line
education, I am retweeting here a video-ed lecture on the subject https://tinyurl.com/weygyo6. It's more than
a year old, but covers many of the questions raised here, squarely and transparently.

4.5.2020 1:36am - (Replying to @desai_pratik) I haven't kept ups with Soar type of projects. From what I
remember this was a "production system" and we, expert systems folks, moved from rules to causal models. But
the cog. science people still use Causal Bayesian Networks as a model of cognition and "understanding".

4.5.2020 9:05pm - Agree. Yoshua Benjio is one of the few DL leaders who understands the role of causality in
AI, the rest are either avoiding the issue or paying curve-fitting lip-service to it. We, AI and society, will be
paying dearly for this neglect, as I note here: https://ucla.in/2JEhGyv

4.5.2020 8:44pm - (Replying to @VincentAB @causalinf and 2 others) I agree of course with your conclusions.
What I do not understand is your self-doubt. There is only one definition of IV and one way of verifying it. Only
economists are confused today about IV's because, having vowed to shun graphs, they are still debating what
"exogeneity" is.
4.5.2020 8:29pm - (Replying to @adepstein1 @RLong_Bailey and @Keir_Starmer) And I hope UK Jews get the
courage to talk honestly to @Keir_Starmer and tell him what his litmus test is: How he talks to Labor members
about Israel and Zionism and lasting peace in ME, not how he flatters UK Jews.

4.5.2020 3:16am - Readers who asked if they can help, please write to @ImranKhanPTI , the prime minister of
Pakistan, and express your hopes to see his government appeal the acquittal ruling, thus reaffirming Pakistan's
commitment to universal values of justice and the sanctity of human life.

4.5.2020 2:55am - (Replying to @RonKenett and @mario_angst_sci) I tend to side with @mario_angst_sci on this
issue. There is a profound difference between a tool that makes you ask: "How come no one told me?" and one
that leaves you blank even after being told. The former lets you do things you always wanted to do and
couldn't!

4.4.2020 11:14pm - (Replying to @DocGTLBrown and @mario_angst_sci) Extremely pleased and, if anyone was
busy preparing a "guide for students seeking enlightenment" I would recommend adding your course to the
White List. One needed ingredient: Add CI to the title, to show that your department is not bowing to pressures
or taboos.

4.4.2020 5:54pm - My thoughts on Imben's ideology are here: https://ucla.in/36EoNzO. Indeed, the econometric
leadership will have to answer some tough questions to the court of history. Econ. students will be summoned
as witnesses, and econ. textbooks as material evidence. Wish I live to see it!

4.4.2020 3:28pm - Likewise, I've been advising schools that advertize for students, postdocs and faculty to state
explicitly their interest in CI education and research, else they would not get the students and faculty they are
looking for. "Times they are achangin" (Dylan 1964).

4.4.2020 3:14pm - This past week, the Stanford's HAI Institute has organized a virtual conference on "AI and
COVID-19," a video of which is now available: https://hai.stanford.edu/events/covid-19-and-ai-virtual-
conference/video-archive. I have asked the organizers to share the following note with the participants:
https://ucla.in/2JEhGyv

4.4.2020 2:41pm - And how on earth is it that prominent leaders and educators in respectable scientific
disciplines do not ask themselves the same question to lift their fields from the margins of the causal
revolution? Historians of 21st century science will ask this question.

4.4.2020 7:21am - To all readers who shared hopeful thoughts with us, we are grateful and happy to inform you
that the government of Sindh has ordered that Daniel's murder suspects will be kept in detention for another 90
days, pending an appeal. Thanks for being with us.

4.2.2020 8:13pm - A new blog-page, carrying a humble contribution of causal inference to the fight against
Covid-19 was just posted here: https://ucla.in/39Ey8sU. It asks: "Which Patients are in Greater Need?" and
displays the options in vivid colors.

4.2.2020 5:27am - It is a mockery of justice. Anyone with a minimal sense of right and wrong now expects Faiz
Shah, prosecutor general of Sindh to do his duty and appeal this reprehensible decision to the Supreme Court of
Pakistan.

4.2.2020 3:21am - (Replying to @Abel_TorresM @dileeplearning and 3 others) Agree on first, disagree on
( py g @ _ @ p g ) g , g
second. Structural Causal Models exhibits true understanding of a domain (defined by the variables in the
model) in that it answers all causal and counterfactual questions about that domain, as shown in #Bookofwhy
chapter 1, "The Mini-Turing Test."

4.1.2020 4:19pm - (Replying to @Dr_Cuspy @tdietterich and @MarcioMinicz) I would hate to use the word
"autonomously" in a definition, b/c it is hard to tell if a program does things by choice, agency, or by being
programmed to do it. "Predicting consequences" is easier, we can just ask the program to predict, and test the
answer against outcomes.

4.1.2020 3:57pm - (Replying to @HundredthIdiot and @deaneckles) Really? I was told many readers filter their
tweets by subject matter, and will get only Tweets containing those hash tags. If it ain't true, I'll be more than
happy to gain 10 precious characters not using #Bookofwhy (I just did!!).

4.1.2020 3:47pm - (1/ ) In re-reading Holland (1986) I noted that he was the first to define CI as a missing data
problem, through his "Fundamental Problem of Causal Inference", which so many PO folks love to quote. Today,
in contrast, we are classifying "missing data" as a causal inference problem
4.1.2020 3:47pm - (2/ ) (see for example https://ucla.in/3423VTr) Wow! How "the times they are a changin"
(Bob Dylan, 1964). #Bookofwhy

4.1.2020 3:20pm - For readers who are re-raising @deaneckles question, I am retweeting my reply (below) and
add that I also praise DAGs for reminding us that it is not the end of the world, and guiding us to augment our
knowledge to get identification: IV, front-door, do-cal. etc. #Bookofwhy

4.1.2020 1:37am - Glad we have a causal model for COVID19 testing, so that we can accumulate facts and
judgement into one arena, and talk one language. #Bookofwhy

4.1.2020 1:23am - Thanks for sweet nostalgia. I often play with the idea of writing: "Statistics and CI - 30 yrs
later", but not sure JASA would publish it. Holland's paper has shaped the mindset of most living statisticians
working on CI, some irreversibly.#Bookofwhy

3.31.2020 8:43pm - Missing data is when Chinese journalists disappear, eg


https://nytimes.com/2020/02/14/business/wuhan-coronavirus-journalists.html and statisticians later fix with
multiple imputation. #Bookofwhy

3.31.2020 7:15pm - For coordinating, comparing and making sense of pandemic data sources, corrupted by local
idiosyncrasies we need causal meta-analysis. The theory is available: https://ucla.in/2Jc1kdD and is awaiting
an elite force of CI PhD's to teach & train practitioners. @EpiEllie #Bookofwhy

3.31.2020 4:56pm - This incredible scene comes straight from the twilight zone, where "The wolf lives with the
lamb, the leopard lies down with the goat, the calf and the lion and the yearling together; and little child leads
them." (Isaia 11:6). No child in sight, unfortunately.

3.30.20 11:28pm - The warm reception of my conversations with N Jewell tempts me to post another ancient
video: https://tinyurl.com/sbdtnj3. It took place 10 years ago, on the foundations of causal reasoning; a topic
more alive now than ever, especially while fighting for explanations #Bookofwhy

3.30.20 8:04pm - (Replying to @MEMRIReports and @HananyaNaftali) I know a few US Universities who would vie
for his scholarship in the department of Mid East Studies.
3.30.20 3:25pm - The crucial methodological part of the debate rests with fusing coherently all the data coming
from so many diverse and noisy sources. We have a calculus+algorithms for doing it, awaiting experts trained in
CI and listen; see https://ucla.in/2Jc1kdD. #Bookofwhy

3.30.20 3:07pm - I'm flattered, but don't stop here. Continue to Primer https://ucla.in/2KYYviP, and join the
common sense revolution - every rebel counts. #Bookofwhy

3.30.20 12:09am - The chloroquine debate should be settled by domain experts. As a methodologist, I can only
state that algorithms have been developed that can look at data and estimate the probability that a drug is
harmful to a given individual. I assume the experts are reading #Bookofwhy

3.29.20 6:00pm - Written especially for revolutionary-minded people like you. Who said epi is not poetry?
#Bookofwhy

3.29.20 3:40am - Another interesting paper: https://arxiv.org/pdf/2003.10341.pdf which provides an insightful


analysis of cross-world assumptions in mediation analysis. See #Bookofwhy chap. 9, and https://ucla.in/2N9IA5A
for nice plots of the mediation formula in linear, logistic, and probit models #Bookofwhy

3.29.20 3:08am - Retweeting an interesting paper by a Microsoft team on one of the most demanding task
performed by humans: system debugging https://arxiv.org/pdf/2003.09539.pdf. It combines causal and
counterfactual logic, as opposed standard debugging softwares which are statistically driven. #Bookofway.

3.29.20 2:17am - (Replying to @mgaldino) This is only part of the answer. Let's not forget that formalization
serves both: automation and coherence. Applied researchers, following their informal intuition might find
themselves producing contradictory recommendations (as in Simpson's paradox) #Bookofwhy

3.28.20 9:17pm - (Replying to @awmercer) Thanks for your honest and informative answers. Recall however
that for us, AI-ers, "makes the most sense" is just the beginning of a new challenge: How do we represent (on a
machine) the knowledge that tells you that one framework "makes more sense" than another. #Bookofwhy

3.28.20 6:08pm - (1/ ) (Replying to @awmercer) This is the first time I hear of structural equations in the
service of the survey field, which makes me super curious: Do these equations enjoy different notation than
regressions? Different treatment? Surely, the model-free approach misses something; it misses an opportunity
3.28.20 6:16pm - (2/ ) (Replying to @yudapearl and @awmercer) to benefit from valuable information coming
from the model, when such is available to the analyst. What is the simplest survey example in which we can see
the tradeoffs between the two approaches? A toy example is worth tons of scholarly words. #Bookofwhy

3.28.20 4:08pm - Another thought evoked by your informative description of survey methods. Since these
methods rely on the "no-confounding" assumption, how did survey experts articulate this assumption
mathematically? Or they just carried it in the head? What's in the tech literature? #Bookofwhy

3.28.20 3:44pm - (Replying to @shirokuriwaki and @deaneckles) This paper is not very helpful because,

following the PO tradition, it assumes away the hard part of the problem. Quoting: "The conventional solution
to this problem is to assume ignorable treatment assignment and overlap (Rosenbaum and Rubin 1983)"
#Bookofwhy

3.28.20 3:41am - (1/ ) I'm glad, Sean, that our brief exchange has resulted in your great clarification of the
3.28.20 3:41am (1/ ) I m glad, Sean, that our brief exchange has resulted in your great clarification of the
issues, from which I have learned a lot. Two thoughts come immediately to mind: (1) It is a blessing that we can
enjoy a division of labor between CI and statistics, the former generates
3.28.20 3:41am - (2/ ) causal estimands, the latter estimate them. Note though that the former is not totally
oblivious to the type of data available. Different types of data will result in different estimands.
eg.,experimental vs. observational, corrupted by missingness or by proxies or by
3.28.20 3:41am - (3/ ) differential selection etc. (2) I don't buy the mystification of "collecting adequate data". I
am in the business of automating a scientist, so, if there is human judgement involved in the data collection
task, we do not stop here and surrender it to humans. We see it as an
3.28.20 3:41am - (4/ ) invitation to ask: what knowledge allows you to decide that some data are "more
adequate" than others. We then model that knowledge and automate the process. I strongly suspect that behind
this piece of knowledge you'll find a causal model. I am willing to bet!! #Bookofwhy

3.28.20 3:29pm - Truly appreciate your substantive reply. So, generalization across population is the common
theme here. In https://ucla.in/2N7S0K9 we solved such problems using graphs. Have we missed something by
not listening to the survey literature and to the tools developed there?#Bookofwhy

3.28.20 3:19pm - (Replying to @awmercer and @deaneckles) Great. Now suppose we have good understanding
of confounders. What formal language is available to us to articulate our understanding, so that we can process
it coherently when the number of confounders exceeds our mental capacity? Is Meng's language sufficient?
#Bookofwhy

3.28.20 3:11pm - It is a great paper, and I retweeted it in Dec. 2020 with a warm blessing. However, I am still in
a learning mood: Survey methods have been around for 2 centuries. What tools have been developed that we,
CI folks, can use to speed up our research agenda? #bookofwhy

3.28.20 2:46pm - (Replying to @agostbiro and @deaneckles) No one doubt it. The entire success of DL rests on
this surprise. But we, CI students, have certain problems on our plate for which we cannot find (yet) salvation
in understanding correlation alone. And we are begging for help. #Bookofwhy

3.28.20 1:39pm - I am retweeting with the hope that perhaps some other readers beside @deaneckles can help
us on this question, which has come up again and again. In a way, ML/DL folks are operating as survey samplers,
so they too could help us: What can we learn from your experience? #Bookofwhy

3.28.20 12:45pm - There is another point to this pseudo-critique. Suppose you launch a very smart discovery
algorithm and get a causal structure, what then? If you do not teach your students how to leverage it to answer
causal questions we are back to where model-free analysts are today.#Bookofwhy

3.28.20 12:29pm - (Replying to @ewerlopes and @Muenchner_Junge) Thanks for posting this course. I have seen
many advertised under the catchy title "causal inference" which are doing ONLY estimation, leaving the
assumptions to divine intervention. This course has Primer in its reading list - a faithful sign of enlightenment.
#Bookofwhy

3.28.20 12:11pm - (Replying to @AngeloDalli) There could be some other type of knowledge guiding statisticians
in what they call "design". But the fact that they treat it as divine wisdom prevents us, AI researchers, from
automating it. We will, eventually. #Bookofwhy

3.27.20 8:32pm - (1/ ) I beg to differ. For readers interested in my unbiased opinion, this review is ill-informed,
ill-motivated and misleading. Bent on showing that "there is more to it than #Bookofwhy", sweating hard and
unsuccessfully to find any such "more to it", the reviewers neglect to tell
3.27.20 8:32pm - (2/ ) readers what they would miss not reading the book. The quote: "causal inquiry cannot be
reduced to a mathematical exercise nor automatized." is truly indicative of the occult "more to it" agenda of
the review, and its failure to appreciate how much has been automated already.
3.27.20 8:32pm - (3/ ) Another unbiased opinion I have regarding "the particular value of randomized
experiments". I challenge anyone to show me a clearer/deeper exposition of RCTs and their "particular value"
than that given in #Bookofwhy ch. 4, The Skillful Interrogation of Nature, Why RCTs Work?

3.27.20 7:53pm - (Replying to @blake_camp_1 @kchonyc and 2 others) I am late on the philosophy of SSL, so I
commend you for giving us a glimpse of what it does. Can you elaborate on how fillings blanks can get you a
causal effect without some causal assumptions encoded somewhere in the system? #Bookofwhy

3.27.20 7:42pm - (Replying to @omaclaren and @analisereal) In which case these undergrads should be able to
prove that back-door condition is sufficient for identifiability in a simpler, static system, with no derivatives.
They should also be able to run DAISY and decide yes/no on any of the models described in #Bookofwhy. Eager!

3.27.20 7:27pm - (Replying to @omaclaren) I would qualify it a bit. ‘Modernity’ is the sum total of things
that have been accomplished in the past 3 decades regardless of any particular formalism. This is how
#Bookofwhy defines the "causal revolution", and it lists those things explicitly, task after task.

3.27.20 7:16pm - (Replying to @analisereal and @omaclaren) I would add that structural models are causal
entities and should be annotated by arrows Y<--f(X,e), not equality signs. Once the distinction is made
explicitly, I embrace the paper as "relevant", else, "suspect of generating confusion" as in
https://ucla.in/2NRFn7e #Bookofwhy

3.27.20 2:51pm - (1/ ) (Replying to @seanjtaylor) "Useful" is a rich concept. Statistics papers, formal logic,
cognitive science, even history books and Greek mythology have been useful to my work. But when it comes to
causal questions, a paper that does not articulate the question in mathematically tells me a lot about what
3.27.20 3:19pm - (2/ ) (Replying to @yudapearl and @seanjtaylor) to expect from reading it. It tells me that
either (1) the paper deals with stat. estimation after assuming away all causal considerations, or (2) it has
nothing to do with causation despite the catchy title, or (3) the author is confused - like many regression
analysts -
3.27.20 3:19pm - (3/ ) (Replying to @yudapearl and @seanjtaylor) writing regression and thinking causation. It
does not make the paper useless, but it warns you that it ain't going to be easy to excavate for the relevant
material if the has any. So, if your time is valuable, you already saved 1000 hours, all for want of notation.
#Bookofwhy

3.27.20 1:49pm - (Replying to @blake_camp_1 @tyrell_turing and @ylecun) No need to toy. Just look at the
input information. If it contains a causal element (or interventional data) then your toying may succeed, if not,
you can't create it from data alone. It's like the conservation of energy (Helmholtz, 1847.) #Bookofwhy

3.27.20 1:42pm - (1/ ) You almost got it correctly. You can read 1000 papers, for 2000 reasons, but if you want
to know ahead of time whether it deals with causal questions formally, then it is not a matter of "preference" or
"taste", it is a matter of distinguishing causal from statistical
3.27.20 1:42pm - (2/ ) quantities and, then, the use of do(x), Y(x), or DAGs is a good indicator whether a
distinction was attempted. I've saved literally thousands of hours by this filter and I would be curious to know of
any causally-relevant paper that I've missed. Notation reveals! #Bookofwhy

3.27.20 1:56am - (Replying to @tobytfriend @davidpapineau and 3 others) Good catch, thanks. There are indeed
3 notational indicators to causally-minded articles: do(x), Y_x, or good old fashion DAGs. Thanks. #Bookofwhy

3.27.20 11:44pm - (Replying to @ewerlopes) It is very kind of you to trust me for giving such a course. I'll do
some checking around and see if it can be realized given everyone's constraints #Bookofwhy
some checking around and see if it can be realized given everyone s constraints. #Bookofwhy

3.27.20 1:45am - Responding to interests in the foundations of causal inference, and lacking
resources/army/time to marshal a full Zoom course, I am posting a video interview with N Jewell: Introduction
to Causality, Part I, https://ucla.in/3brEMnO, Part II, https://ucla.in/2vXrlwY, #Bookofwhy

3.26.20 7:33pm - (Replying to @btolgao @tvladeck and 3 others) I've been in this playground since 1990, see
https://ucla.in/2pKgKyu. The PC algorithm (Verma's IC) may discover beautiful structures and we know its
limitations. Jona's-type algorithms rest on non-scrutinizable assumptions (eg. additive noise), may turn out
promising.#Bookofwhy

3.26.20 4:16pm - (Replying to @btolgao @tvladeck and 3 others) When the DAG is not known, it means the
knowledge is not available; we need to acquire that knowledge by experimentation guided by analogies and
metaphors, like the Greeks did. Not like the Babylonians. #Bookofwhy

3.26.20 4:09pm - (Replying to @Jabaluck @VincentAB and 2 others) No, I have not seen those "other approaches"
arrive at those insights. Let's take the simplest of all: backdoor. Can you illuminate us on how a DAG-less
"approach" would arrive there. But, please, let's use input-output, no ignorability acrobatics. #Bookofwhy

3.26.20 3:58pm - (Replying to @deaneckles @tvladeck and 2 others) Yes, we often do not have the knowledge
required to answer our questions - I praise DAGs for telling us. We often have 3 equations with 4 unknowns - I
praise algebra for telling me that the solution is not unique. Who else would tell us? Statistics? PO? #Bookofwhy

3.26.20 3:50pm - (Replying to @peterlorentzen @ladder2babel and 3 others) ?????????????????

3.26.20 2:35pm - (Replying to @DanielHGill @davidpapineau and 3 others) No, orthogonal is you have left from
an orthodentist who failed dentistry school. Same as what's left from big-data lacking a causal model.
#Bookofwhy

3.26.20 2:14pm - (Replying to @davidpapineau @totteh and 2 others) The nice thing about modern thinking is
that you can tell orthogonality w/o even reading the paper. Just look at the equations, if all they have is
probabilistic expressions, no do(x) and no counterfactual subscripts - forget it, its traditional statistics, no
causation#Bookofwhy

3.26.20 1:46pm - (Replying to @totteh @davidpapineau and 2 others) Beg to disagree. Meng's paper
https://statistics.fas.harvard.edu/files/statistics-2/files/statistical_paradises_and_paradoxes.pdf is orthogonal
to the issue of randomization. It is written in the best tradition of statistical orthodoxy - avoid the C-word, and
avoid causal thinking. #Bookofwhy

3.26.20 1:21pm - (1/ ) (Replying to @deaneckles @rabois and @Jabaluck) You have assumed correctly. The
#Bookofwhy is all about the insufficiency of "big data" to get us over from association (rung 1) to causation (rung
2 and 3), no matter how skillful we are in processing the data. Large n only gets us from samples to
distributions. Another push
3.26.20 1:32pm - (2/ ) (Replying to @yudapearl @deaneckles and 2 others) is needed to get us from distributions
to causal conclusions. The push may come from interventional studies, or from causal models. Note that RCT's
must invoke interventions -- randomization alone is as useless as large n.

3.26.20 1:17am - (Replying to @Jabaluck @VincentAB and 2 others) The best way to address such complaints is
to ask: Do you acknowledge the existence of some inference tasks which DAGs perform well, and which the
alternatives find intractable, that is, close to impossible? #Bookofwhy

3.25.20 9:22pm - (Replying to @peterlorentzen @rabois and 2 others) Sorry, I haven't been around to catch what
the dispute is about, or what @rabois said that triggered @jabaluck anger. I am not sure #Bookofwhy can help
but I have found it a good start for any discussion about cause and effect.

3.25.20 1:27pm - I'll be back!!! I promise!!!

3.25.20 1:03pm - This is the first Certificat I have seen in "Causal Data Science"! Congratulations!!! I am anxious
to see the first PhD Diploma in Causal Data Science, perhaps 4 years from now. #Bookofwhy

3.25.20 6:34am - (1/ ) I used to think that a journal editor can help refresh a field by inviting "survey articles"
from outsiders. I no longer think so, b/c the questions that controvertial papers evoke need be answered by
direct interaction with the author. My recommendation: Publish a dialogue
3.25.20 6:34am - (2/2) (a technical one) between the outsider and the top leader in the field that seeks fresh
air. Anecdotally, I have just finished such a dialogue (for a prominent journal) and the editor who appointed the
interviewer rejected the outcome. Lesson: that's what's needed #Bookofwhy.

3.25.20 5:34am - (1/ ) Feldman's observations are insightful and thought provoking. I think we, scientists,
should learn from politicians (don't choke!). They knew that power corrupts and instituted 3 branches of Gvt to
check on each other. We do not have a "checking" branch to tell any of our
3.25.20 5:34am - (2/ ) self-sustaining "fields": Hey we are X years behind the time! We are waiting for outsiders
to do it, which is too slow a process. Here is a counterfactual thought: Vice Presidents in charge of "Fresh Air",
that is, innovation, openness, cross-field communication etc.#Bookofwhy

3.24.20 10:44pm - (1/ ) To assess where quantitative psychologists stand on the road up Mt. Causation, I
examined the citations in the "taboo paper" https://tinyurl.com/r5a79qa and noticed a dearth of causally sound
papers in psych.. [And I thought economists have the hardest time with modernity.] While
3.24.20 10:44pm - (2/ ) the survey paper by Foster https://gwern.net/docs/statistics/causality/2010-foster.pdf
is pretty good (for 2010), not much else from a field hopelessly sunk in SEM culture. I knew that Psychometrika
and JSEM can't shake stagnation, but I thought the rest of the field is more progressive. Cognitive Psychology,
3.24.20 10:44pm - (3/ ) for example, is a model of progressive thinking. It would be helpful to hear from
readers active in this field. Am I off in this gloomy assessment of quantitative psychology? #Bookofwhy

3.24.20 6:36pm - I have an even shorter link https://tinyurl.com/r5a79qa and note that the authors have not
yet used the heavy guns against the supremacy of RCT's: Selection bias (as discussed here:
https://ucla.in/2JfdQcv ) #Bookofwhy

3.24.20 6:04pm - I hope this link works: http://scholar.google.com/scholar_url?


url=https://psyarxiv.com/8hr7n/download%3Fformat%3Dpdf&hl=en&sa=X&d=18372133015727723954&scisig=AAGBfm1x5v0bDbmWowXA02DRGsG6cUz-

2A&nossl=1&oi=scholaralrt

3.24.20 5:56pm - The Taboo Against Explicit Causal Inference in Nonexperimental Psychology is undergoing a harsh examination here
file:///C:/Users/Judea/Downloads/Grosz%20et%20al.%20(2020,%20preprint)%20The%20Taboo%20Against%20Explicit%20Causal%20Inference%20in%20Nonexperimental%20Psychology.pdf

3.23.20 10:36pm - (Replying to @AdanZBecerra1 @mayfer and @_MiguelHernan) Of course. Likewise, doing real
world dpi is "doubly-super-hared," if not impossible, without SCM. I have not heard about "other causal
frameworks" that distinguish "cause of death" from "observed and death" and estimate the former, not the
latter. #Bookofwhy
y

3.23.20 10:24pm - (Replying to @AdanZBecerra1 @mayfer and @_MiguelHernan) In the case of CV, it is
reasonable to assume: "CV caused death" = "CV verified & death." Counterfactual logic is needed when we doubt
this identity. Note that "verified" is still different from "tested positive" which is behind the statistics we are
getting.#Bookofwhy

3.23.20 4:27pm - (Replying to @djnavarro) My personal apology to you, Danielle. Sorry.

3.23.20 3:02pm - (Replying to @AshbyHarper) Had I been in agreement with the ideology of the writer, I would
not have apologized. But I object to everything he stands for, so I apologize for having created an unintended
false impression. Sorry!

3.23.20 3:02pm - (Replying to @AshbyHarper) Had I been in agreement with the ideology of the writer, I would
not have apologized. But I object to everything he stands for, so I apologize for having created an unintended
false impression. Sorry!

3.23.20 2:52pm - Apologizing to the many readers who were offended by that miguided article that I posted,
not realizing it comes from CBN (which I misread as NBC) and which is deleted now. Next time I should read
what the author advocates, not merely what the title sings. Good lesson!

3.23.20 5:26am - (Replying to @Satoohn and @ChrisAdamsEcon) The weather predictor is observing two
different needles, goes through some mental process and makes a prediction. Someone asks "what is the effect
of Needle-1 on the prediction, given data from the past 100 days. This is where collinearity hurts. #Bookofwhy

3.23.20 3:38am - (Replying to @athos_damiani and @_rchaves_) I give publicity to many articles that invite
humor, criticism and, God forbid, even ridicule.

3.23.20 3:22am - (Replying to @athos_damiani and @_rchaves_) Sorry, the title was 2+2=4. The paper may have
jumped to 2+2=5, but the title is innocent. BTW, Who is God?

3.23.20 2:31am - (Replying to @DVBartram) Reconsidered. Retracting. Havn't foreseen the side effects or the
implications. Asking for forgiveness. Which, btw, is a counterfactual notion demanding formal explication. But
I'd settle for ordinary, pre-scientific forgiveness.

3.23.20 1:17am - (Replying to @_rchaves_) What's wtf? What's transphobic? What have I done wrong this time?
Can't we enjoy a title saying 2+2=4 w/o offending a mathematician (or an economist)?

3.22.20 12:48pm - I heed to your proposal and offer our Crash Course to all data-intensive professions:
https://ftp.cs.ucla.edu/pub/stat_ser/r493.pdf. I mentioned economists only because they coined the name
"bad control" and have been most obstinate in refusing the science of its remedies. #Bookofwhy

3.22.20 12:25am - I see that you singled out "psychologists" as potential beneficiaries of our "Crash Course on
Good and Bad Controls". What circles in psychology deserve this honor? #Bookofwhy

3.21.20 2:45pm - (Replying to @hangingnoodles and @aminsaadou) Causal modeling tools beat and escape the
Lars quote better than unaided "thinking", which some (eg., Keynes) see as a superior tool to rescue science and
policy makers from the complexities of social systems. #Bookofwhy
3.21.20 2:31pm - (Replying to @MehboobFarah) The distinction between "good and bad controls" is not used for
"constructing models" but for "utilizing models" once constructed and deemed plausible. #Bookofwhy

3.21.20 12:55pm - (Replying to @aminsaadou and @hangingnoodles) The causation tools available nowadays are
EASY for everyone (with the exception of some outdated old-guards), see https://ucla.in/2KYYviP, and they
also make clear when old tools are adequate. #Bookofwhy

3.21.20 12:36pm - For the many students of economics who express frustration with the way a problem known
as “bad control” is evaded, if not mishandled in econometrics, our "Crash Course" in now accessible as
Technical Report: https://ftp.cs.ucla.edu/pub/stat_ser/r493.pdf Enjoy and teach your professor. #Bookofwhy

3.21.20 5:10am - Proud to congratulate my colleague-in-law, professor Furstenberg, on winning the Abel Prize
in Mathematics: https://jpost.com/Israel-News/Hebrew-University-Professor-wins-Abel-Prize-the-Nobel-of-
Maths-621414 Wishing El-Al Airline a quick recovery, so he can fly to Norway to receive the Prize in person.
#Bookofwhy

3.21.20 1:28am - "Causal Transfer" and "Imitation Learning" is what called my attention to this new paper:
https://arxiv.org/pdf/2003.00806.pdf, plus the fact that it approaches the problem scientifically, as opposed
to data-fittingly. A prerequisite to data-science. #Bookofwhy

3.21.20 12:26am - To clarify, my comment below relates to @LarsPSyll statement "Econometric modelling
should never be a substitute for thinking." which some take as a license to substitute modelling with "thinking".
Modeling is a necessary amplification of the solid part of our thinking #Bookofwhy

3.20.20 7:31am - (Replying to @hangingnoodles) I am willing to accept the unmanageable nature of social
systems. What I find hard to accept is that an unaided and fallible machine called "commonsense economics" is
better equipped to manage those unmanageable systems.

3.20.20 1:38am - (Replying to @jules__shen) Impossible! How can a skill that governs so much of our every day
thinking and emotions be "very hard"? Impossible! Indeed, the 3-step process of computing counterfactuals is
easy. #Bookofwhy

3.19.20 7:15pm - (Replying to @tdietterich and @zacharylipton) Not familiar with the specifics, but would like
to mention that society paid dearly for brilliant perpetual motion machines, proposed, pursued, funded, argued
in courts etc. many decades after Helmholtz (1847) paper on conservation of energy. Seek balance & principles!
#Bookofwhy

3.18.20 8:21am - (Replying to @100trillionUSD @pierre_rochard and 5 others) No. I said: IF cointegration
analysis aimes at distinguishing spurious and non-spurious relationships, then it must be based on causal
assumptions and, thus far, I have not seen the literature on cointegration explicate such assumptions, nor use
causal vocabulary.#Bookofwhy

3.17.20 7:34pm - (Replying to @pierre_rochard @ercwl and 2 others) But what IS this remarkable relationship?
Can you show it in just 3-4 variables and 3 time slots? I would like to add it to my list of "remarkable"
relationships of correlated variables. If it is truly remarkable, it should be demonstrable in 3-4 variables.
Curious #Bookofwhy

3 17 20 7:26pm (Replying to @JohnMarkoff2 @ercwl and @100trillionUSD) If cointegration is a statistical


3.17.20 7:26pm - (Replying to @JohnMarkoff2 @ercwl and @100trillionUSD) If cointegration is a statistical
property of time series then @btconometrics is right, it does not imply causation. See
https://twitter.com/yudapearl/status/1240099100546052097 Whether it "smells" causation depends on the
assumptions made by the smeller; they may be plausible, but we need to see them #Bookofwhy

3.17.20 7:14pm - (Replying to @pierre_rochard @dopamine_uptake and 5 others) I am yet to find a signal that a
certain "correlation is not spurious" which is not based on causal assumptions. So, has any of its advocates been
able to explicate those causal assumptions? I know that Granger was not able, see
https://twitter.com/yudapearl/status/1220857382261030912 Who was? #Bookofwhy

3.17.20 4:03pm - (1/ ) Recent exchange with theory-averse readers compels me to retweet a slogan: "I would
rather follow Eratosthenes and measure the radius of a legendary turtle than fit curves in a Babylonian
observatory." It refers to Greek science in Toulmin's book "Forecast and Understanding"
3.17.20 4:03pm - (2/ ) (see https://ucla.in/2wj4pox, last page) and his observation that it takes a theory to
design a meaningful experiment. Moreover, we are fortunate to live in an era where the "understanding" part of
Toulmin's title can be emulated on a computer and tested against its rivals.
3.17.20 4:03pm - (3/ ) I am speaking of course about SCM which, to the best of knowledge, is the first system to
deserve the title "understanding", due to its ability to pass the mini-Turing test on all three levels of human
understanding (#Bookofwhy Ch. 1): Seeing, Acting and Imagining.
3.17.20 4:03pm - (4/ ) Its opponents do not usually offer computational alternatives, preferring instead to speak
in generalities about the superiority of data over "false narratives." Still, by the time they are ready, we should
be able to run the first empirical test of the "Scientific Method."

3.17.20 3:16pm - (Replying to @Saravanan_CU) The main disparities that I would include are disparities among
gv't policies. Still, transportability should be used whenever we generalize across populations.Unfortunately the
theory is only one decade old and analysts in power take 3 decades to shift paradigms. #Bookofwhy

3.17.20 9:40am - (Replying to @sean_a_mcclure and @HarryDCrane) The opaque culture you prefer reminds me
of Babylonian astronomy in Toulmin's "Forecast and Understanding" (see https://ucla.in/2wj4pox, last page) It
may get us to Babylon, but not to Athens; I prefer the footsteps of Eratosthenes and measure the turtle's radius
#Bookofwhy

3.17.20 12:57am - (Replying to @raymondshpeley) I like it too. And am going to keep it. I truly feel for those
who judge #Bookofwhy by how it confirms what they knew already, instead of rejoicing the tools it provides for
doing today what they wanted and could not do yesterday. Boy! how much fun they are missing!

3.16.20 9:16pm - On the contrary! I tell students: "formality amplifies common sense." Look at common sense
logic before and after Boole. Look at reasoning with uncertainty before and after Bernoulli + LaPlace.
#Bookofwhy

3.16.20 8:49pm - Commiserating with BDS cronies - blazing beacons of deceit. They smile while lying and smile
when caught. Do they really know the difference?

3.16.20 6:29pm - (Replying to @DKedmey and @HarryDCrane) One important caveat. If you know nothing about
your past behavior except that you are like anyone else in a given subpopulation S, then the personal question:

"Would I recover if I take this drug" can be answered from population-level data: find P(recover|do(drug),S).
#Bookofwhy

3.16.20 6:00pm - (1/ ) Back to students of CI and other sapiens fascinated by logic, counterfactuals, and
p ( ) p y g , ,
everyday language. Here is a commonsensical, everyday notion called "NEED", as in "lets first test people in
greatest NEED". We take it for granted that, if the need arises, analysts would
3.16.20 6:00pm - (2/ ) be able to define what "being in NEED" is, and how to quantify the degree of NEED. It
now downed on me that this simple notion is intrinsically counterfactual, hence only those familiar with PNS
would be able to agree on how to quantify it, given data and commonsense knowledge
3.16.20 6:00pm - (3/ ) I have nothing exciting to say to those who get angry at #Bookofwhy. Knowing everything
they are not fascinated by anything, and miss so much of the fun of discovery.

3.16.20 5:09pm - (Replying to @MalikMashail and @McDonalds) I tried not to get too personal. Oh, almost forgot
my logo: #Bookofwhy, sorry.

3.16.20 4:54pm - (Replying to @DKedmey) Well put! I could not express it more clearly. In addition, this
difference translates into (i) individual level vs. (ii) population level inferences. To guess whether "Joe would
have reacted differently" I need to undo what I know about Joe. #Bookofwhy [no off @HarryDCrane etal]

3.16.20 4:32pm - (Replying to @HarryDCrane) FYI, and for all those offended by my tweet:
https://twitter.com/yudapearl/status/1239691201986441217 please note: The tweet was addressed to
"students of CI" who have been estimating PNS for e-commerce and drug approval. Identifying "those in greatest
need" is another application we should keep in mind.

3.16.20 4:13pm - I end my tweets with #Bookofwhy because, so I am told, many readers filter their tweets by
this tag, and I do not want them to be left out of the discussion. It is not meant to suggest that the book "solves
every problem"; an epsilon of commonsense would rule out such suggestions

3.16.20 3:40pm - (Replying to @DavidPoe223) Right on the nail. It is now 30 years since I started asking
regression analysts: "What puzzle a regression is supposed to solve?" Some say:"Given information X, we want to
predict Y". Fine. But some say "We want to understand, to decompose, to make sense". Dont get it.#Bookofwhy

3.16.20 3:25pm - To students of CI, Coronavirus is a reminder that "people in greatest NEED" is a counterfactual
notion; the degree of "need" is the probability of "helped if treated & doomed if not", dubbed PNS in Causality
p.286, and tightly bounded in https://ucla.in/2NbER7C #Bookofwhy

3.16.20 2:57pm - (Replying to @DavidPoe223) In Simpson's paradox the puzzle is simple: We cant conceive of a
drug that is good for men, good for women, and bad for a person. What is the puzzle that "multilevel and fix
effects" models are trying to resolve? #Bookofwhy

3.16.20 7:02am - (Replying to @DavidPoe223) I am still curious: What is being sought? What is to be estimated?
What is it that does not make sense before you start writing down equations? #Bookofwhy

3.16.20 6:18am - (Replying to @DavidPoe223) May I ask: What is the research question that triggers this theory?
#Bookofwhy

3.16.20 5:59am - More from Michael Levitt on why the spread of Coronavirus is slowing https://jpost.com/HEALTH-SCIENCE/Israeli-nobel-
laureate-Coronavirus-spread-is-slowing-621145?

utm_source=ActiveCampaign&utm_medium=email&utm_content=Coronavirus%3A+Senior+scientist+at+Israel+s+largest+testing+lab+has+COVID-
19&utm_campaign=March+16+Day #Bookofwhy
3.16.20 4:03am - (Replying to @JRubinBlogger) And we can safely assume she won't be Linda Sarsour. If I were
Biden, I would make this point loud and punchy.

3.16.20 3:54am - (Replying to @rodakker @tudorcodes and @imleslahdin) I hope what I mean by "causal"
coincides with yours, causality is a human language. As to Jaynes, he was sold on probabilities, so I do not
believe he can define "causal arrows" better than other Bayesians, who are semi-alien to causation:
https://ucla.in/2nZN7IH #Bookofwhy

3.16.20 12:10am - (Replying to @rodakker @tudorcodes and @imleslahdin) This happened to be a Causal
Network, even though we are only interested in probabilities. For a non-causal network of the same physics,
consider AND--->D1---->D2 plus AND--->D2. It contains non-causal arrows which are necessary to get the
probabilities right. #Bookofwhy

3.15.20 3:14pm - (Replying to @tudorcodes and @imleslahdin) All data are generated by some causal processes,
hence causal diagrams can always be used to capture the statistical features of the data, provided we allow for
hidden variables (& selection B). To save the labor of dealing with hiddens, you may use undirected graphs.
#Bookofwhy

3.14.20 10:43pm - Of all the statistical analyses of the C-Vir outbreak, the most optimistic evaluation was given
by (Nobel) Michael Levitt, in an interview to China Daily:
https://global.chinadaily.com.cn/a/202002/25/WS5e545c76a310128217279f34.html His technical paper,
tables, curves and figures can be viewed here https://ucla.in/3aXNRog #Bookofwhy

3.14.20 6:04pm - (Replying to @svonava @Joshyblogz1 and @lexfridman) This wikipedia article on
counterfactual thinking was written before the algorithmization of counterfactuals. Someone ought to tell the
authors to reexamine those theories because much of what they describe can now be done by machines, eg.
https://ucla.in/2NceWNb, or #Bookofwhy

3.14.20 1:58pm - Why soap? And why it matters? Statistical analysis of the Coronavirus outbreak is illuminating,
but the most useful article I've read in the past two weeks is this:
https://theguardian.com/commentisfree/2020/mar/12/science-soap-kills-coronavirus-alcohol-based-
disinfectants And it has a touch of #Bookofwhy in it.

3.14.20 11:37am - Causal diagrams capture a small but important chunk of human understanding. To their
credit, they give us a friendly and formal object from which we can derive what Keynes probably meant by "the
broad reasons why past experience was what it was". #Bookofwhy

3.14.20 10:48am - (Replying to @mgaldino and @mendel_random) Barring a few exceptions, of course.

3.13.20 12:59pm - (Replying to @ganesh__s and @eliasbareinboim) Sure! Even static physical equations (eg
Hooke's Law) convey counterfactual information. But you have to be explicit how to simulate "undoing", e.g.,
"what if event E did not occur?" #Bookofwhy

3.13.20 11:39pm - (1/ ) On curve fitting vs. science, @eliasbareinboim showed me a beautiful quote from John
Locke (1960, Essay VI): "By the color, figure, taste, and smell and other sensible qualities, ...we cannot tell
what effects they will produce; nor when we see those effects can we so much as
3.13.20 11:39pm - (2/ ) (Replying to @yudapearl) guess, much less know, their manner of production." Locke
understood in 1690 that the key to human understanding, knowing "their manner of production," cannot be
inferred from data. #Bookofwhy
3.12.20 1:47pm - The March issue of the Philosophical Transaction of the Royal Society celebrates 50 years to
Price Equation, which describes how genes change over time. I was pleased to find two papers on causal
interpretations of the equation. https://royalsocietypublishing.org/doi/abs/10.1098/rstb.2019.0354
https://royalsocietypublishing.org/doi/abs/10.1098/rstb.2019.0365

3.12.20 1:28pm - (Replying to @ewerlopes) Yes. I purchased a new kindle version and, aside from their usual
sloppiness in handling equations, they have implemented almost all the errata. #Bookofwhy

3.12.20 5:36am - (Replying to @Stebbing_Heuer @regev_nir and @BernieSanders) My general philosophy is that
every experience, however subjective, is a piece of information that should be integrated with other sources to
forge theories. I don't know anyone among my colleagues who is not disgusted at the way @BernieSanders is
selling his soul to angry voters

3.12.20 5:21am - (Replying to @NNrehman) Sure. But please consult the Errata list here:
https://ucla.in/2KYYviP

3.12.20 4:46am - The day we were all been waiting for has arrived!!! Our publisher informs us that the new,
revised and corrected printing of Primer https://ucla.in/2KYYviP is now out at all US outlets. To make sure your
copy is "revised", check that the "about the authors" page cites #Bookofwhy

3.12.20 4:39am - (Replying to @Stebbing_Heuer @regev_nir and @BernieSanders) My personal, first-hand


acquaintance with my compatriots, the American people, and their deeply ingrained respect for leaders who do
not sell their heritage for a few angry votes.

3.12.20 4:31am - (Replying to @jasndoc and @Farzad_MD) A highly motivated reason to examine how modern
causal analysis can be harnessed to assist a task which appears purely associational but involves fusing together
diverse sources of information from diverse populations. #Bookofwhy

3.12.20 1:34am - (1/ ) (Replying to @regev_nir and @BernieSanders) "Think not with thyself that thou shalt
escape in the king's house, more than all the Jews. For if thou keep thy silence at this time, relief will arise to
the Jews from another place, but thou and thy father's house will perish". Mordechai would'nt use such harsh
words, unless
3.12.20 1:53am - (2/ ) (Replying to @yudapearl @regev_nir and @BernieSanders) he sensed reluctance on
Esther's part to plead with the King despite timing considerations ("I wan't called for 30 days"). I think Michigan
voters felt the same about Bernie's selling Israel, the culmination of his people's history, for a few angry votes,
promised by Rashida.

3.10.20 11:26pm - (Replying to @AnneHerzberg14 @richard_landes and @stan_state) B. Franklin's words keep
ringing in my ears: "He who sleeps with dogs will rise up with fleas." And the Mishna says: "Leaders, watch your

words! Lest your followers misinterpret them!" Poor Bernie, will he ever understand the dangers of bad
company?

3.10.20 10:57pm - (Replying to @AnneHerzberg14 @richard_landes and @stan_state) It is not funding that the
faculty of @stan_state should worry about, but their reputation. As much as I am telling myself "dont blame the

grass for one bad weed", it does reflect on their mentality, what they tolerate as "scholarship" and what they
feed students. STANISLAUS
3.10.20 10:32pm - (Replying to @regev_nir and @BernieSanders) Esther did need this. She was about to brush it
off with: "The timing is not right, etc etc." Mordechai reminded her: "Perhaps history elevated you to this
position for a purpose". Bernie had no one to remind him. I tried, but I doubt he can hear the bells of history.
Poor Bernie

3.9.20 1:21am - Poor Bernie, he thought by abandoning Israel he would win the hearts of his Zionophobic
followers. Alas, they now reject him as "rich Jew": https://tinyurl.com/sa7c3mc. He forgot what Mordechai told
Queen Esther: "You wont save your skin by abandoning your people". @BernieSanders

3.9.20 10:11pm - (1/ ) Columbia President Lee Bollinger has issued a statement https://tinyurl.com/wybo7f4
condemning antisemitism on campus. Kudos! But he evidently has not read my advice
https://tinyurl.com/uxwy9fn: To stop BDS hostilities it ain't enough to decry anti-semitism; you've got to
expose what
3.9.20 10:11pm - (2/ ) BDS stands for: the elimination of Israel, and the expulsion of Israel-inspired students
(and faculty) from campus life. And you've got to express how you, personally, feel about what they stand for,
as President Martha Pollack of Cornell did: https://tinyurl.com/rresc8a @Columbia

3.9.20 2:03pm - (Replying to @charleendadams) Never abandon babies, with or w/0 waters. It's good for science
to bemoan what we need: triangulation, external validity, selection bias, missing data, etc etc. Bemoaning loud
enough and long enough may get someone attention, lest it becomes a habit that avert us from doing it.

3.9.20 5:28am - (Replying to @kurtosis0 and @BernieSanders) Repeating: "Abandoning Israel" means supporting,
endorsing, and giving platform to those who seek Israel's destruction. You can abandon Israel being an American
or a Hottentot, all you need is moral frailty and Zionophobic voter base

3.9.20 5:10am - (Replying to @jmourabarbosa and @BernieSanders) Abandoning Israel means supporting,
endorsing and giving platform to those who seek Israel's destruction. Simple! Simpler even than causality.

3.9.20 3:34am - The word "triangulation" has been used by many to express dissatisfaction with established
methodologies and romantic aspirations for something new. This paper https://tinyurl.com/ufklebf however,
recognizes formal ways of triangulating information: https://ucla.in/2L6yTzE #Bookofwhy

3.9.20 3:06am - (Replying to @Stebbing_Heuer @kurtosis0 and @BernieSanders) You got it perfect. Being Jewish
is not genetic, nor religious, nor being a victim of the holocaust. It is identifying yourself with the history and
destiny of a collective calling itself Jewish. Anyone can acquire it. E.g., Large segment of former Soviet citizens
now in Israel.

3.9.20 1:32am - (Replying to @kurtosis0 and @BernieSanders) And Queen Esther was Persian. Truman and
Kennedy were American too, but would not admire Zionophobic racists like Linda Sarsour and Ilhan Omar.

3.8.20 6:27pm - (Replying to @GeoffWinestock) I did not realize you have a question. I agree that the British
could have done many things (eg. allow movement on the Danube to the Black Sea) but this does not absolve
the Mufti from the crime of strangling European Jewry by the anti-British anti-Jewish riots of 1936-39,

3.8.20 4:59pm - Today is international women's day, coinciding with the holiday of Purim, the day that Queen
Esther said: I must do it, whatever comes. I salute the Israeli female soldiers who served with me in the army
(1953) and kept my hopes high until this very day: https://tinyurl.com/tknzbqv
3.8.20 4:33pm - It is the first time that I hear this music and I see the charming smile of Al Husseini, the man
who prevented my grandfather from escaping Hitler, the father of BDS (1936 boycott of Jewish products) and
the creator of Zionophobia. May history remember him for his deeds.

3.7.20 2:01pm - (Replying to @RokoMijicUK and @BonbonFork) No qualms about "big jumps", but when I
characterize a tool, I say what the tool delivers, not how it operates. #Bookofwhy

3.7.20 9:03am - (Replying to @ChrisAdamsEcon) McFadden's random utility model is one model in which we can
compute certain POs. RCT is another. Speaking in general, however, given a structural model M, which PO's can
be computed, and how? #Bookofwhy

3.7.20 8:04am - (Replying to @ChrisAdamsEcon) That is why I prefer "first person" comments, rather than
hearsay. Can you state: "To me, Phil's views resolved several questions that I had: (1), (2)...." Or: "If you havn't
read Phil's slides you might be tempted to believe (1),(2).. Not so!" #Bookofwhy

3.7.20 7:44am - (Replying to @FriendlySceptic) Speaking of intellectual honesty, the word "equating" is out of
place. I said: When you endorse Zionophobic racists like Linda Sarsour and Ilhan Omar, you should expect other
racists to smell blood and yell: "Free season!" Havn't heard Sanders' defenders defend these endorsements

3.6.20 9:41am - (Replying to @TheMongoose4) Never mind what right-wing Jews are saying. Listen to what left-
wing Jews are saying: When you endorse Zionophobic racists like Linda Sarsour and Ilhan Omar, you should
expect other racists to smell blood and yell: " free season!". Signed: a disappointed Bernie sympathizer.

3.6.20 9:12am - (Replying to @roieki) Chas V'Chalila!! Never "BaMachane", and never "Nazism". But when you
endorse Zionophobic racists like Linda Sarsour and Ilhan Omar, you should expect other racists to smell blood
and yell: " free season!"

3.6.20 8:38am - Schoking? Yes, Surprised? No. I believe it was Benjamin Franklin who said: "He that lieth down
with dogs shall rise up with fleas". When you open your house to haters you can't screen them for table
manners.

3.6.20 6:18am - (Replying to @ChrisAdamsEcon) We discussed Haile's slides here, on Twitter, and I did not get
the impression that economists agree with him. Correct me if I am wrong. Also, help us find where economists
define PO's from structural equations. The closer I got was Heckman: https://ucla.in/2NnfGPQ#page=44.
#Bookofwhy

3.5.20 11:51am - Econometrics has two camps: Structural and quasi-experimental. The former say: PO comes
from structure, but don't tell us how. The latter resist structure, and pretend PO resides already in one's mind.
Both camps needs a shake, I tried here https://ucla.in/2mhxKdO, & #Bookofwhy

2.5.20 11:44pm - (Replying to @JDHaltigan) Thanks for educating me on this new measure. I'm still trying to
figure out when I would feel that strange need to hold Z constant for just X. BTW, the expression "hold Z
constant" is misleading, it should be "selecting only cases for which Z=z". #Bookofwhy

3.5.20 11:27am - (Replying to @jiaming_mao and @causalinf) Symmetric? No. Each can be separately derived
from a completely specified SCM (a set of structural equations). It is so nicely described in Primer, and it is
available for free: https://ucla.in/2G2rWBv, Why mystify? #Bookofwhy

3.4.20 11:34am - (Replying to @MrTeshnizi) Good question. I think we are born with a template for gathering
and using causal knowledge, which DL agents do not. If they do, they should be able to pass the mini-Turing test
of #Bookofwhy chapter 1.

3.4.20 4:49am - I'm absolutely delighted to be elected Honorary Fellow of the Royal Statistical Society, an
institute that has been rather lukewarm towards causality in the past. I see it as an invitation to all of us,
causal enthusiasts, to meet the challenge of becoming mainstream. #Bookofwhy

3.4.20 2:36am - Better in the sense of having lower asymptotic variance.

3.4.20 12:30am - Explanations of what? Examples of what? Sorry, I am a foreigner, please speak slowly to me.

3.4.20 12:04am - A reader just resurrected this yr-old tweet, and reminded me that my question is still
unanswered: Which estimator is better? The ML estimator of the product or the product of the ML estimators?

3.3.20 11:53pm - My heart goes to MN voters. Who will protect their children from the Zionophobic hatred
spewed by @BettyMcCollum04 ? Who will protect their reputation from the thought they might share her
mentality? My heart goes to Minnesota voters.

3.3.20 9:29am - (Replying to @mariabloec @jonasobleser and 5 others) I have a conciliatory ending for this
debate: The notion of "Well-definedness" is ten times more ambiguous than the causes it tries to disambiguate.
#Bookofwhy

3.3.20 8:46am - (Replying to @mariabloec @jonasobleser and 5 others) I cannot think of a well-defined
interventions for earthquakes and, still, their causal effects on property damage is not vague. I discuss it here
https://ucla.in/2LY8M2X and here https://ucla.in/2EpxcNU "Well definedness" suppresses causal thoughts.
#Bookofwhy

3.3.20 8:36am - I love the ending of this old post: "Once tried - always used". Thanks for re-posting. It applies to
so many tools of causal inference. #Bookofwhy

3.3.20 6:22am - An arrow between A-B would be a different story, which we could try later as the hard part of
the test. #Bookofwhy

3.3.20 12:22am - It is an honor to be in the company of these stellar authors, Thanks

3.2.20 11:23pm - My advice to readers who are asking me to help them recruit students and faculty in ML: State
explicitly that you seek candidates in the CI area. This would signal to candidates that your department is
serious about changing the causality-free tradition of ML education.#Bookofwhy

3.2.20 12:37pm - (Replying to @BlueManifold and @dmonett) Indeed, now that I re-read it, they do emphasize
the importance of causal and counterfactual reasoning. I did not pay attention to it in 1969, being deep into
heuristic search. It is only when I saw causality as the force behind Bayes networks that I asked: what is it?
#Bookofwhy
3.2.20 7:46am - (Replying to @mktopuz) It's easy to check using the mini-Turing test on the firing squad
(#Bookofwhy chapter 1). "Would the prisoner still be dead had soldier-1 refrained from shooting?"

3.2.20 7:23am - The next step is to ask: what kept the 1912 guy from recognizing the importance and/or
generality of his/her observation, and turn it into a "theory of 1912"? Language has lots to do with it.
#Bookofwhy

3.2.20 7:14am - (Replying to @BlueManifold and @dmonett) Computer science is the only discipline I know, in
which people understand that languages are here to serve a purpose, to be extended and replaced, not to be
religiously wedded to, till death do us part. #Bookofwhy

3.1.20 11:41pm - For curious readers, the intrasitivity of correlations, treated as an "issue" in @nntaleb paper
"fooled by correlation," is an invariant feature of colliders. In X-->Z<--Y we have Cov(XZ)>0, Cov(ZY)>0 yet
Cov(XY)=0 regardless of what numbers we spray on the arrows. However,

3.1.20 11:41pm - DAGs enjoy a weaker form of transitivity: If X and Y are each dependent on Z, then they must
be dependent on each other, either marginally or conditional on Z. See "Probabilistic Reasoning.." (1988,
p.129). It is not an immutable property of probabilistic dependence #Bookofwhy

3.1.20 3:29pm - "English" is great to describe a phenomenon and show how it is reflected in data. We need a
mathematical language to fully understand a phenomena, so as to meet the criteria proposed here
https://ucla.in/2Jfl2VS vis a vis Simpson's paradox. (English description=1899) #Bookofwhy

3.1.20 3:13pm - I like the way you put it: "formalize the intuition that you take for granted". But I dont foresee
"a big blunder" for those who resign to taking certain intuitions for granted. The only problem might be
repeating analyses on intuitions that come from the same source. #Bookofwhy

3.1.20 2:36pm - This is precisely the problem of answering counterfactual queries from group's Data, see
https://ucla.in/2G2rWBv for a gentle introduction and a tool kit. The "threats" paper you cite should help us
motivate modern "remedies", thanks. #Bookofwhy

3.1.20 2:12pm - (Replying to @nntaleb) Your paper shows how certain expectations (eg additivity, transitivity)
are often violated. I ask: whence those expectations come from? The logic of causal intuition provides insights
into this question. BTW, MI is also intransitive; is it bad? good? surprising? Why? #Bookofwhy.

3.1.20 8:07am - (Replying to @AngeloDalli @AdanZBecerra1 and 2 others) It depends on the granularity of our
variables. If I measure the movement of my coin-flipping finger in mm, the outcome is causally independent of
that movement, but if I measure it in Angstroms, it's a different story. #Bookofwhy

3.1.20 6:16am - It is fascinating to watch how traditional probabilists handle puzzles that emerge from causal
reasoning (in this case the collider structure) lacking a language to capture causal relations. How do we tell
them about the causal language w/o offending them? @nntaleb #Bookofwhy

2.29.20 2:03pm - Indeed our conception of cause and effect has only ONE point of contact with do-calculus -
the starting point. Do-calculus insists on starting with human conception of reality, faithful to the way it is
stored in the mind. The rest is mechanical, like algebra. #Bookofwhy

2.29.20 1:30pm - (Replying to @tmorris_mrc and @y2silence) We can't allow any reader to be left with "low
points". The quote is about human conception of cause effect relationships. While external tools for answering
certain causal questions may be absolutely necessary, this does not make them part of the conception. In the
case of RCT,
2.29.20 1:41pm - we can read text after text of scientific explanations of biological phenomena - what leads to
what, what prevents what etc etc - with not a single mention of RCT. See eg the Appendix to
https://ucla.in/2EpxcNU. RCT is a tool for interrogating, not explaining nature. #Bookofwhy

2.29.20 1:16pm - Why does it matter if Zionophobia IS antisemitism or just another form of racism. The
question is how genocidal it is in its aim and how deceptive it is in its advocacy.

2.29.20 5:41am - Our conception of cause and effect is almost always non-deterministic (eg,. careless driving
causes accidents), so what? It still can be reasoned about ,formally and informally, without RCT's. #Bookofwhy

2.29.20 5:02am - Got a neat brain teaser from a reader of #Bookofwhy regarding Game 2, page 159: Controlling
for D would constrain the mediator E, so why is it legitimate? It's good for a homework, for it ain't trivial. The
answer is in the Appendix of https://ucla.in/2LcpmHz. Enjoy.

2.29.20 4:18am - Thanks for the compliment. All we need to do now is to convince the establishment.

2.29.20 4:09am - Agree. The computational approaches rangesfrom "simple statistical models like linear
regression, to... [more statistical models] like convolutional neural networks". Hoping the 2nd edition snaps out
of Rung-1 to methods that interpret data. @DShaywitz #Bookofwhy

2.29.20 3:42am - Zionophobe is one who does not believe in the right of the Jewish people to a homeland, in
the context of "2 states for 2 peoples, equally legitimate and equally indigenous". Examples are: Palestinian
leadership, educators, clerics, BDS activists and their European supporters.

2.29.20 3:32am - I spread messages of peace and equity, summarized as: "2 states for 2 peoples, equally
legitimate and EQUALLY INDIGENOUS." I am risking my academic status to spread this message, because others
don't, and many of my like-minded colleagues are prevented from openly doing so.

2.28.20 7:18pm - (Replying to @jos_b_mahoney) No. Apology. I do not know this work.

2.28.20 6:28am - I have always thought that "object-property" relationships (ge, birds fly) require new
semantics, not captured by causation. This paper proves me wrong: https://academic.oup.com/jos/advance-
article/doi/10.1093/jos/ffz023/5721111 Time to preserve old beliefs and time to revise them (said King
Solomn). #Bookofwhy

2.28.20 5:42am - I would not call a statistical equality "constraint". It is more a "happening", or "coincidence" in
the observed data than a constraint imposed by the model's assumptions. #Bookofwhy.

2.28.20 5:37am - Of course! If the outcome of a random coin makes me win a dollar, then it is certainly causal.
You are probably asking whether the dependence of the coin outcome on the way I flip it is causal. Classical
mechanics says: Yes, despite the absence of prediction/control #Bookofwhy

2.28.20 3:00am - Sure! I would lump all such factors under "metabolism" and add it as a common cause of W_I
and W_F, on top of the arrow W_I ---> W_F. #Bookofwhy
2.28.20 12:37am - A Zionophobe is he who says that Israelis, as biological objects "have the right to live in
peace" , but will never say that Israel, as a sovereign state of the Jewish people, has the right to live in peace.

2.28.20 12:25am - (1/ ) On Dec. 9 1666, the Rabbis of Constantinople excommunicated Shabbetai Zvi, a self-
proclaimed messiah who led a sect of X-Jews to speak against the greater Jewish community. In Feb. 24 2020,
347 Rabbis rebuked Bernie Sanders for speaking against the greater Jewish community .
2.28.20 12:25am - (2/2) If I were a Rabbi, I would move to pronounce Bernie Sanders a X-Jew. This would not
diminish his standing as a Messiah, but it would revoke his credentials to speak as a Jew.

2.28.20 12:08am - Three days ago, 347 Jewish Rabbis rebuked Bernie Sanders for maligning AIPAC and Israel. I
have not heard any of them say: "Hey, Bernie you either are, or act like a Zionophobe, and this does not reflect
well on your moral character!"

2.27.20 2:49pm - In Lord's (1967) paradox, one constraint in his own plot is, cov(S, WI)=cov(S, WF), where WI &
WF are two repeated weights. To embrace it in his DAG, @yudapearl had to claim that i) S->WI (a) equals the
sum of ii) S->WI->WF (ac) & iii) S->WF (b): a=ac+b. Not super satisfying 2 me

2.27.20 9:13am - I do not usually engage in futuristic speculations, but this interview with Oren Etzioni on "How
to know if artificial intelligence is about to destroy civilization"
https://technologyreview.com/s/615264/artificial-intelligence-destroy-civilization-canaries-robot-overlords-
take-over-world-ai/ strikes me as a serious thought provoking session. #Bookofwhy

2.26.20 5:00pm - Sure. The best place is this personal story from the Jewish Journal
https://tinyurl.com/y5ko2nvs. I also have some photos that go with it: https://ucla.in/3caTC3l Note the
neighborhood bully, it's me, no messing around.

2.26.20 9:32am - Who is a proud Jew? S/he who calls Bernie Sanders a ZIONOPHOBE, in public. As far as I know,
no one has done it yet! May I have the honor? Proudly so!

2.26.20 8:19am - For readers who are asking about software tools to implement do-calculus, data-fusion and
other goodies mentioned in @Bookofwhy, here is http://bit.ly/36qUz4y

2.26.20 7:42am - No Jew is a proud Jew who fails to call @SenSander a Zionophobe, in public. And, as far as I
know, no one did. May I have the honor? Proudly so!

2.26.20 7:35am - Sharing a rare and breathtaking view of the Dead Sea coastline, which I, as a boy, knew only
as a sub-burned yellow desert.

2.26.20 7:09am - (Replying to @yudapearl @AngeloDalli and 3 others) Halpern's complete axiomatization of
counterfactuals, and Spirtes proof that d-separation holds in linear feedback systems. But we lack an
algorithmic method, like do-calculus to tell us, in the non-parametric case, when a query is identifiable and
when it is not. #Bookofwhy

2.26.20 7:04am - (Replying to @AngeloDalli @ylecun and 2 others) Depends on what we mean by "solution". We
have a satisfactory semantics, given by the 3-step process, once the model is fully specified. This is the logical
basis for next step: identification. It isn't as easy as in recursive models, but some nice results do exist, among
them
2.26.20 12:59am - Readers of #Bookofwhy who are in the healthcare practice or research will find this
comprehensive survey of Bayesian Networks to be most informative:
https://arxiv.org/ftp/arxiv/papers/2002/2002.08627.pdf It covers predictive, diagnostic as well interventional
applications.

2.25.20 9:27pm - Don't under-rate yourself. If you read #Bookofwhy you already know more than the millions
who have not read it, and that's quite a lot.

2.25.20 9:22pm - (Replying to @AngeloDalli @ylecun and 2 others) There is a paper by Richardson and Lauritzen
which gives cyclic interpretation to chain graphs. #Bookofwhy

2.25.20 9:18pm - I had to write them in this order because, lacking courage and audacity, I could not face the
public till I was sure I was right - a personality weakness. But now that I am sure, I do not buy the wimpy
"multiple approaches" myth - another personality weakness. #Bookofwhy

2.25.20 8:59pm - (Replying to @yudapearl @mribeirodantas and @diomavro) on the "linear microscope"
https://ucla.in/2LcpmHz, and https://ucla.in/2L8jUFg, which clearly distinguishes the two tasks and shows
that the partial regression BETA(yx|z) is independent of the value z of the variable Z. This is no longer true in
non-Gaussian models #Bookofwhy

2.25.20 8:49pm - Bias in predictive tasks occurs in finite-sample estimation, but not asymptotically, as in causal
tasks. Unfortunately, most texts on regression analysis confuse the two, as do texts in Economics and Social
Science. I must therefore overcome modesty and refer you to my papers

2.25.20 8:31pm - (Replying to @mribeirodantas and @diomavro) Bias due to bad-control only hurts us in causal
tasks, where our target is a causal effect. In purely predictive tasks there is no such notion as "bias", and
traditional wisdom says: the more information you have the better the prediction. True in linear Gaussian
model #Bookofwhy

2.25.20 12:39pm - In my blog comment, in case you could not find it, I provided three relevant links:
1. https://ucla.in/2EihVyD
2. "Bayesianism and Causality, or, Why I am Only a Half-Bayesian" -- https://ucla.in/2nZN7IH
3. Causality, section 11.1.1 https://ucla.in/2NnfGPQ#page=331 #Bookofwhy

2.25.20 10:36am - O.A. Arah has called my attention to a lengthy discussion on Gelman's blog:
https://statmodeling.stat.columbia.edu/2020/01/27/causal-inference-in-ai-expressing-potential-outcomes-in-
a-graphical-modeling-framework-that-can-be-fit-using-stan/ concerning a paper on "causal inference with Bayes
Rule" https://gradientinstitute.org/blog/6/. I have added a comment to the former, and will examine the
latter. #Bookofwhy

2.25.20 8:35am - The channel you are now using, by following @yudapearl , is an educational channel,
dedicated to the scientification, algorithmization and democratization of causal inference. #Bookofwhy

2.25.20 3:23am - (Replying to @yudapearl) On this occasion, let me also recount the availability of two
resources. (1) A searchable file of all our 4,263 Tweets. https://ucla.in/2Kz0FoY, and (2) A listing of all UCLA-
CSL papers and technical reports, dating back to 1965. https://ucla.in/2NWm9xu. Happy sailing #Bookofwhy

2.25.20 3:23am - My faithful Twitter counter says that, today, our educational channel has garnered the
attention of 30K followers. This encourages me to continue our journey towards the scientification,
algorithmization and democratization of causal inference. Thanks for the encouragement.

2.25.20 3:07am - (Replying to @AngeloDalli @ylecun and 2 others) Do you mean cyclic causal paths? (ie.
feedback loops)? If so, please examine causality p. 215-217. There is a robust solution, though the identification
phase is trickier. #Bookofwhy

2.24.20 5:48pm - (Replying to @jim_savage_ @steventberry and 5 others) For structural economists, an
agreement on "non-parametric" should come easy, it is a fancy surrogate to "structural", namely, properties that
emerge from the bare structure of the model (excluding for example shape restrictions, or distributional
assumptions) #Bookofwhy

2.24.20 5:20pm - (Replying to @steventberry @jim_savage_ and 5 others) The "honor" of structural models is
currently being restored through (1) clear definitions (which precede identification) and (2) tools (eg graphical)
for liberating models from their non-structural burdens and unleashing truly non-parametric identification
properties #Bookofwhy

2.24.20 4:52pm - (Replying to @steventberry @jim_savage_ and 5 others) As a student of philosophy I see it as a
challenge to track down sources of confusions in various disciplines and, although econ. is still an enigma to me,
I can see interesting patterns in econometric thought, going from inadequate notation to inadequate
definitions. #Bookofwhy

2.24.20 1:26pm - (Replying to @steventberry @jim_savage_ and 5 others) It feels great to be in agreement: "The
answer is: structural." Perhaps you also share with me another conjecture, that much of the confusion in Econ
emanates from not having a crisp, undisputed definition of Y_x, in terms of a (fully parametrized) structural
model. #Bookofwhy

2.24.20 12:17pm - (Replying to @jim_savage_ @steventberry and 5 others) Thanks for re-posting. The word
"counterfactual" appears dozens of times in these slides. Perhaps they can clarify how traditional econ defined
Y_x(u) or, even more ambitiously, which of "structural vs. reduced" form is more convenient for evaluating
E[Y_x|X=x', Y=y']? #Bookofwhy

2.24.20 12:16am - I am still not clear if there is anything new in p. 97 that econ students can use, or is it all
3rd-week freshman class. I was ready to offer a big apology for wrongly believing in the former, but I am not
getting a definitive affirmation of the freshman model. Help! #Bookofwhy

2.23.20 5:58pm - (Replying to @steventberry) I am counting the steps needed AFTER identification. How do we
compute the expected salary of Joe who now has skill level Z=1 had he had one more year of college (see Fig.
4.3 in https://ucla.in/2G2rWBv). I do not know how to skip the 3-steps of p.97, please teach me.#Bookofwhy

2.23.20 3:21pm - (Replying to @steventberry and @PHuenermund) Sorry this tweet got out of thread

2.23.20 3:18pm - Yes, I recall a lively discussion with Heckman on the definition of "effects":
https://ucla.in/2NnfGPQ#page=44 But counterfactuals like E(Y_x|x',y')are a bit more complicated, needing a 3-
steps computation, which I have not seen in the econ. lit. Perhaps you do it in one step? #Bookofwhy

2.23.20 1:25pm - I am happy to hear that counterfactual questions are now solved in freshmen econ. classes,
and it was only my bad luck that I could not find anyone who could solve it in 1998. I am still curious, though,
what definition of the counterfactual Y_x is used ? #Bookofwhy

2.23.20 5:51am - I hate to disagree, but I must note my disagreement with the analogy. Econometrics starts
with a (structural) model that every economist can comprehend. PO starts with ignorability assumptions that
almost NO epidemiologist can comprehend. See 3 bullets https://ucla.in/2wbH488

2.23.20 12:16am - I am still not clear if there is anything new in p. 97 that econ students can us e, or is it all
3rd-week freshman class. I was ready to offer a big apology for wrongly believing in the former, but I am not
getting a definitive affirmation of the freshman model. Help! #Bookofwhy

2.22.20 6:19pm - We are in receipt of a course on "Superforcasting", https://edge.org/event/a-short-course-in-


superforecasting%E2%80%94philip-tetlock-an-edge-mas ter-class, part IV of which says: "Superforcasting requires
countefactualizing". This means that an algorithm equipped with a causal model, should out-predict one based
on data alone. True, but how? #Bookofwhy

2.22.20 1:32pm - Commenting on my "Two new weapons.." https://tinyurl.com/uxwy9fn a reader claimed:


"Protestantism led the way in Zionism, not Jews". Reply: Not so; my grandpas prayed 3 times a day "Restore us
in sovereignty to our land" (Hamazon)

2.22.20 11:41am - (Replying to @pierre_rochard @dopamine_uptake and 5 others) I am eager to agree, but I do
not know what "s2f shills" is. Culturally deprived.

2.22.20 11:27am - (Replying to @dopamine_uptake @TheStalwart and 5 others) If by "useful" you mean
"causative", then the question need be rephrased to read: What causal assumptions are needed to make Granger
causality "causative"? I recall (vaguely)

2.22.20 11:18am - (Replying to @pierre_rochard @dopamine_uptake and 5 others) Another mathematical result
says that there is no way to distinguish spurious from non spurious correlations without causal assumptions. See
https://ucla.in/2NnfGPQ#page=5. These ought to have been part of econ. education, but see what happened:
https://ucla.in/2mhxKdO. #Bookofwhy

2.22.20 9:41am - (Replying to @TheStalwart @dopamine_uptake and 5 others) It is not exactly a mathematical
dispute as it is a mathematical fact that one cannot define causal relations in the language of probabilities, nor
can one claim causal relations without causal assumptions. The latter are missing from Granger's causality.
#Bookofwhy

2.22.20 7:30am - The shameful complicity of the journalistic fraternity will always embitter my memories. Like
those "journalists" who cooked up the story of Mohamed Al Dura, whose picture was used by Danny's murderers.

2.22.20 7:04am - Well put. "facelift to Sharansky's 3 Ds"; so badly needed. The 3 D's served their purpose, but
they also reinforced the outdated fixation that anti-Zionism is evil only in as much as it contributes to
antisemitism, which has spawned and legitimized Zionophobes like Linda Sarsour.

2.21.20 9:07am - "How much the system IMPROVES the likelihood of a good outcome" is precisely the
counterfactual term that PNS captures. #Bookofwhy

2.21.20 5:21am - And let's not forget the Passover Hagada: Whoever did not say the following three things in the
past week has not fulfilled his obligations to students on campus: Zionophobe Zionophobic and Zionophobia
past week, has not fulfilled his obligations to students on campus: Zionophobe, Zionophobic and Zionophobia.
Words count, and silence corrupts.

2.21.20 5:00am - Cornell's President, Martha Pollack, should serve as a role model to all academic leaders,
especially to Berkeley's Chancellor, Carol Christ, who said last week: "Jewish students have the right to feel
dismay and concern" instead of: "I, personally, feel dismayed and concerned."

2.21.20 3:28am - (1/ ) Legality aside, I would assess how likely I am to encounter a complication never seen by
the AI surgeon, which a human surgeon can handle through her knowledge of anatomy. It actually happened to
me, when a famous surgeon quoted us 90% success record but, when asked
2.21.20 3:28am - (Replying to @yudapearl) (2/ ) how many cases like THAT have you seen? He said: "None. I
have never seen a lung as bad a lung as this." We cancelled the surgery, and my wife is still alive. Remarkably,
counterfactual logic permits us to bound the probability that a given patient will benefit from a given
2.21.20 3:28am - (3/ ) treatment, once we have a combination of both experimental and observational data.
See tight bounds on PNS (prob. of necessary and sufficient causation)

2.20.20 3:01pm - (Replying to @omaclaren @AdanZBecerra1 and 2 others) Who would care about identifiability
if is were not a necessary condition for estimability? And who would care about "a particular class of models" if
it were not a faithful representation of your very own understanding of the world? "Estimability" reminds us or
that. #Bookofwhy

2.20.20 12:56pm - I define an expression to be "estimable" if it is made up of sums and products of conditional
probabilities, and the underlying distributions are non-pathological. #Bookofwhy

2.20.20 12:44pm - It was an honor to speak to Alums for Campus Fairness (not 'algorithmic fairness') and to
share with them my two defensive weapons. I hope these weapons will be used effectively by all colleagues and
students who are subjected to Zionophobic harassments on their campuses.

2.20.20 10:36am - I can forgive them all, including the last one, is he only did not sell his Jewish soul to Linda
Sarsour and other Zionophobes.

2.14.20 6:40am - Today we will say Kaddish in memory of our son, Daniel
https://algemeiner.com/2020/02/14/remembering-daniel-pearl/ If you happen to be in Encino, CA, join us at
5pm at Valley Beth Shalom. Else, please join us spiritually. After the Kaddish, please sing with us: Tsadik
Katamar Yifrach (Like an Immortal Palm Tree)

2.20.20 6:25am - The http://fairmlbook.org gives indeed a good overview of causal inference, albeit from a
fairness perspective, especially the CI chapter https://fairmlbook.org/pdf/causal.pdf. I'm glad they took
Berkeley's sex-discrimination debate as a test case, and built new ideas around it. #Bookofwhy

2.20.20 3:15am - (Replying to @ChengSoonOng @mrityunjay_99 and @suzatweet) Beg to differ. The test is
fruitful, necessary, available and has led to only few debates, mostly among traditional statisticians, who could
not reconcile to the fact that the most powerful language they knew, prob. theory, cannot do what they always
wanted to do (CI). #Bookofwhy

2.20.20 3:08am - (Replying to @ChengSoonOng @mrityunjay_99 and @suzatweet) The test is as formal and crisp
as mathematics Herself. See https://ucla.in/2NnfGPQ#page=5. Any inference defined in terms of probability
distributions of observable variables is Statistical, not causal. #Bookofwhy
2.20.20 2:51am - Bergstein's article raises all the important issues. I would only wish it was titled "what AI can
and will do" instead of "what AI cannot do", which reminds me of the old speculations of Penrose, Searle and
Dreyfus. #Bookofwhy

2.20.20 2:25am - The key point is "have the architecture" which is an innate "template" to store and interrogate
what we learn through experience. What we cannot do having this template we surely cannot do having to
acquire it from scratch. #Bookofwhy

2.20.20 2:14am - Glad you asked me, the least biased arbiter. The @Bookofwhy will equip you with elementary
tools to solve elementary problems that even the most sophisticated theories in statistics, logic and NN cannot
solve. It's "low-level" on the one hand and super "high-level" on the other.

2.20.20 1:58am - (Replying to @ChengSoonOng @mrityunjay_99 and @suzatweet) What I meant is a chapter on
the mathematical framework (called CSM) that unifies all these seemingly different "approaches" or "styles".
(BTW Granger causality is not causality). #Bookofwhy

2.19.20 4:47pm - Statisticians are not to be undervalued. Rather, they need to be awakened to the age of
causation and internalize the fact that the language of statistics need be extended to adapt to the new age.
#Bookofwhy

2.19.20 7:07am - With the number of enemies I have on Twitter, I am not sure my retweeting would serve you
well. Watch your steps carefully and ...(almost forgotten).. take a look at the Sequential Backdoor
https://ucla.in/2IyJqTJ for more pleasant "time varying" journeys in the future.#Bookofwhy

2.19.20 5:47am - If do-calculus is a special case of X, I would be thrilled to see how X derives the front-door
formula. It reminds me how, not too long ago, epi folks considered Rung 1 and 2 and 3 as the same rung, see
#Bookofwhy page 152. Should epi students wait another 25 years?

2.19.20 4:38am - Replying with my retweet https://twitter.com/yudapearl/status/1230104263042334722 to


keep the thread unbroken, and to add an iota: I am willing to help an interested author put together a
comprehensive textbook on MLL that includes a CI chapter. Will do so anonymously and delightfully.
#Bookofwhy

2.19.20 4:18am - Where is MML heading? It seems to be heading towards another period of Rung-1 existence,
awaiting textbook authors to discover that a unifying mathematics of causality does exist, within which all
seemingly different "approaches" have converged. A golden opportunity! #Bookofwhy
https://twitter.com/ChengSoonOng/status/1230009666458935296

2.18.20 12:24pm - When I see political scientists learning from epidemiologists, I see signs of progress,
especially when they're using graphical models, like here:
https://openscholarship.wustl.edu/art_sci_etds/1757/ What I do not understand is why they aren't using do-
calculus, to get identification with no tears #Bookofwhy

2.18.20 10:52am - Israel celebrates the disappearance of an island in the Sea of Galilee:
https://youtu.be/0Dl87iG_bmU. On a personal note, this island was completely covered (a latent variable)
during my army-service days, 1953-1955.

2.18.20 2:30am - My point is that Jews were not actively "criticizing" revealed truths. "Debates" were forced
upon them What provoked hostility was not the persistence of their beliefs as much as the persistence of their
upon them. What provoked hostility was not the persistence of their beliefs as much as the persistence of their

biological existence; Wrong beliefs had to show up as abnormal existence.

2.18.20 2:06am - This new paper, "Conditional Path analysis in singly-connected path diagrams"
https://arxiv.org/pdf/2002.05226.pdf provides interesting factorization results to the "Linear Microscope" of
https://ucla.in/2LcpmHz. #Bookofwhy

2.18.20 12:21am - There is some truth to it, as told in #Bookofwhy p.244, but not exactly as Clark describes it.
What I learned from Spirtes (aug. 1991) was arrow-cutting for intervention, but the causal interpretation of
DAGs came already in 1987, eg https://ucla.in/2qb2yyU and others, 88-90.

2.17.20 1:35pm - The protest was legally tolerable yet morally despicable. Berkeley is waiting now for the
Chancellor's personal condemnation of such Zionophobic disruptions on moral, not legal grounds.`

2.17.20 12:18pm - Debra is right in NOT equating the two. By saying: "anti-Zionism IS ANTISEMITISM" one implies
that the latter is more despicable than the former and, indirectly, bestows some leniency onto an eliminationist
ideology, which is more dangerous than its emotional rival.

2.17.20 7:25am - (Replying to @ghost_of_roger @deborahlipstadt and @IzaTabaro) The only way to cure this
misconception is to use the term Zionophobia, which makes people jump to their feet and ask: What do you
mean? Tell me more! The gullible don't understand that being anti-Zionist means being ugly and immoral.

2.17.20 7:12am - anti-Zionism is crisply defined, unlike anti-Semitism which invites infinite discussions and
disputations. Perhaps because the latter fell into ill-repute after the holocaust, for a short period of time.

2.17.20 7:06am - Really? I am truly delighted. It is such a fitting metaphor, that I thought would help
statisticians understand why causal notions exist only in the 3rd dimension, and what they need to do to
capture it. Still waiting. #Bookofwhy

2.17.20 6:59am - You sound like a reasonable person, but you speak with Zionophobic slogans like "apartheid
policies". I refrain from using such terms for fear of being identified as gullible and misinformed, who buys into
BDS propaganda.

2.17.20 4:35am - Logic does dictate that, if you read my elaboration on why the logic of Zionophobes makes
most Jews suspects of criminal activity, unless proven "innocent," like those fringe X-Jews (the new Maranos),
who think distancing themselves from Israel would protect them from harassment.

2.17.20 4:22am - Good point. We should really make clear it's not "probabilistic models" but "models of reality",
as opposed to models of data. I hate to say "causal models", b/c ML folks would say "It must be awfully
complicated". "Reality" on the other hand is supposed to be simple. #Bookofwhy

2.17.20 4:03am - Emil Coman just sent me a new toy, a Stata module to do-the-DAGs:
https://ideas.repec.org/c/boc/bocode/s458467.html Some of our readers might find it useful to have a '1 stop
shop' to do things they couldn't do before. #Bookofwhy

2.17.20 1:04am - I heard people say "he was a spy", or "he had no business investigating Pakistan" This is the
first time I hear someone saying it was "propaganda". The capacity of certain cultures to deny reality exceeds
our capacity to be surprised
our capacity to be surprised.

2.17.20 12:31am - The reason I said "insurmountable obstacle" is that it cannot be solved by model-free ML
methods. The examples in https://ucla.in/2N7S0K9 show that the same shift in probability may require two
different repairs, depending on which structure caused the shift.#Bookofwhy

2.16.20 9:17pm - Mike, you made my day! Finally, a Jewish leader with a spine! Who recognizes that we must
change the conversation from whether anti-Zionism is or is not antisemitism to whether it is genocidal or just
disgustingly ugly.

2.16.20 8:03pm - (Replying to @MisterSquirrel2 @LomaahhMore and 4 others) What's the point? That distancing
ourselves from Israel should protect us? They (Zionophobes) disrespect Jews who trade their values for social
acceptance, something they wouldn't do. They will never stop hating us just because "hate" is socially
unacceptable.

2.16.20 7:52pm - Dear @BDSreport . I just tweeted this


https://twitter.com/yudapearl/status/1229242370341793792 Q. Why don't you use the word "Zionophobia"? and
turn it into the ugliest word on campus? When we use the word "antisemitism" we bestow some legitimacy onto
anti-Zionism, legitimacy that it does not deserve.

2.16.20 7:29pm - In my entire life I havn't heard anyone calling for the elimination of Iran or China from the
surface of the earth, or calling whoever opposes the elimination "criminal" or subhuman. Linda Sarsour just said:
Any one who humanizes Israelis is on the side of the "oppressor".

2.16.20 7:13pm - This harassment will continue as long as we play their game that, unlike antisemitism, anti-
Zionism has forgivable qualities - it doesn't. The harassment will stop when we raise counter-signs saying: "Its
worse!" or "Zionophobia kills," perhaps even in lectures on Islamophobia.

2.16.20 11:12am - (Replying to @Prof_Livengood @deborahlipstadt and @IzaTabaro) I am not inventing those
terms, I am just quoting my Zionophobic colleagues: "Yes, every Israeli who does not support the dismantling of
Israel and turning it into a Palestinian-majority state is complicit with the crime of the century." I read what
they write.

2.16.20 10:57am - Both intention and material support are a statistical fact. Recall however that the crime is in
Israel's very existence, so are the intention and the support -- both are aimed to help Israel survive.

2.16.20 10:09am - I would say "structure" - everything then follows.

2.16.20 9:23am - I don't think we can afford to wait for them to realize that they are committing suicide, for
they will drag all of us into the abyss. It's time to close ranks, and save American Jewry from itself -- Israel and
Zionism have the inspirational power to do it.

2.15.20 10:38pm - (1/ ) The book you mentioned


https://amazon.com/dp/1580232590/ref=cm_sw_r_sms_awdo_t1_oX.rEbXEZHHE is indeed a penetrating and
unparalleled window to the mystery of Jewish identity. It was created at an emotional moment of history, when
one was able to convince 150 Jews, from all walks of life, to take time off their busy
2.15.20 10:41pm - (2/ ) (Replying to @yudapearl @kisjdmls and @nprscottsimon) schedules and address a
question they would normally push aside: "When you say "I am Jewish" what do you really mean?" The result was
an authentic panoramic view of how Jews see themselves in the 21st century. We were gratified to know that,
17 years after publication a book
17 years after publication, a book

2.15.20 10:42pm - (3/3) inspired by Danny's last words is still informing and empowering curious minds; quite a
few schools are giving it as a gift to their graduating youngsters.

2.15.20 2:42pm - (Replying to @Ostrov_A) Your tweet has made my day, thanks. And may that, from now on,
the inspirational story of Israel will be told through its miracles, not through sacrifice.

2.14.20 2:56pm - Sharing a poem I wrote on the tenth anniversary of Daniel's death: The Lion's Den
https://jewishjournal.com/opinion/101390/the-lions-den-to-daniel-pearl-on-the-anniversary-of-his-death/

2.14.20 6:40am - Today we will say Kaddish in memory of our son, Daniel
https://algemeiner.com/2020/02/14/remembering-daniel-pearl/ If you happen to be in Encino, CA, join us at
5pm at Valley Beth Shalom. Else, please join us spiritually. After the Kaddish, please sing with us: Tsadik
Katamar Yifrach (Like an Immortal Palm Tree)

2.14.20 5:22am - I am surprised that, with all their wisdom and resources, the authors do not consider
combining the two, as advocated here: https://ucla.in/2JfdQcv Can anyone inform them that the tools of
combination are available. #Bookofwhy

2.14.20 4:07am - For ML folks, "domain adaptation" connotes an insurmountable obstacle. For CI folks it is a
causal graphs problem embraced under "transportability" theory. This paper
https://arxiv.org/pdf/2002.03278.pdf views the problem as Bayes inference on graphical models. #Bookofwhy

2.13.20 2:20pm - Thanks for posting. I was not aware of this paper https://papers.ssrn.com/sol3/papers.cfm?
abstract_id=2692274, which is the most comprehensive I've seen on instrumental inequalities and their
extensions. Highly recommended to those who wish to test the assumptions behind instrumental variables
#Bookofwhy

2.13.20 12:51pm - It's great to see the principles of causal inference propagate through multiple eyes, multiple
lenses and multiple wavelengths. At the end, I believe we will have a super bright sun beam that even blind
heathens would notice. #Bookofwhy

2.13.20 6:31am - From my recollection, the Instrumental Inequality is tight for binary treatments, see
https://ucla.in/2pFa9pc, and we also have tight bounds for monotonicity , see https://ucla.in/2NbER7C. Thus,
if we find violation, its good to know if its structural or functional #Bookofwhy

2.12.20 8:59pm - The insight of Yossi Klein Halevy and his careful choice of words have always swept my
admiration. This piece in particular deserves careful attention, it is half prophesy and half calling.

2.12.20 6:40am - A new paper has reached my desk "Revisiting Sequential Attributable fractions" http://cholar.google.com/scholar_url?
url=https://assets.researchsquare.com/files/9994855e-5543-44f8-a01f-
f6e5a21f588e/v1/AAF_extensions_23Jan.pdf&hl=en&sa=X&d=1348223756497013068&scisig=AAGBfm0Cx6PdSh2H_mpNoYTBmettGwnQOw&nossl=1&oi=scholaralr
. "Attributable fractions" is an old epidemiological concept that has been treated badly when statistics ruled epi. Interesting to see it in causal dressing.
#Bookofwhy

2.12.20 7:23pm - Epi was "ruled by statistics" when it allowed causal concepts such as "attributable fractions"
and even "confounding" to be defined in statistical terms, and no one complained. This was "ruling" by language
and thought #Bookofwhy
and thought. #Bookofwhy

2.12.20 1:10pm - Sample complexity has been both an enigma and Achilles heel of causal, non-parametric
estimation. This paper establishes sample and time complexity bounds on non-parametric identifiability, thus
bringing CI closer to the general ML fold. #Bookofwhy.

2.12.20 6:19am - At the request of readers, I am sharing a transcript of my ACF talk on "Two New Weapons for
Reclaiming Israel's Posture on Campus" https://ucla.in/37iAMni Use them while you can in defense of equity,
diversity and inclusion.

2.12.20 12:51am - For those of us still interested in Lord's paradox, and his question: "how to allow for
uncontrolled pre-treatment differences between groups", I've added a comment https://ucla.in/2SiNLkn to my
post, examining why statisticians have given up the challenge prematurely #Bookofwhy

2.11.20 1:26pm - As Israel celebrates #WomenInScienceDay I can't help recalling how, in 1960, my EE class at
the Technion had 4 female students among 120 males (1 of the 4 was my wife). Today the percentage of women
is 40% ( 20% of them Arab). A more inviting culture for a male student.

2.11.20 6:26am - I used "head-to-head" in the Bayesian Network paper of 1985 and thereafter. But "collider"
became more popular later, and I believe it was coined by the CMU group of Spirtes etal in 1994. #Bookofwhy

2.11.20 6:14am - I wish someone rewrites this wiki entry. I have never heard of a book that is "laying the
foundations of the modern debate on causal inference". The writer tried so hard not to take sides that naive
readers may take CI to be just one big "debate". #Bookofwhy

2.11.20 5:58am - Warning! Good paper, but I would not recommend expressions like: "A DAG is one method that
investigators may use to clarify which variables have causal interpretations in statistical models."1st, NO
method can clarify that. 2nd, DAGs are causal, not statistical models. #Bookofwhy

2.11.20 2:17am - It was a pleasure to present the logic of "Emancipating our Identity" and to see how the word
Zionophobia gives people the power to express what they have been yearning to express all along. I am now
convinced, Zionophobia WILL become the ugliest word on campus.

2.11.20 1:47am - (1/ ) This is one of the most perplexing question in Jewish history: to embrace or to shun. King
Solomon said: To every thing there is a season... a time to embrace, and a time to refrain from embracing
(Proverbs 3). I feel we are approaching a tipping point in our existence where
2.11.20 1:47am - (2/ ) we should CLOSE RANKS rather than EXTEND TENTS. Those who trade core values (eg,
Israel) for social acceptance are becoming existential threats, thus making a painful community split or a
unilateral secession almost inevitable.

2.11.20 1:26am - (Replying to @nicidob @maximananyev and @TradeandMoney) Recalling that "what's occurring"
in not the data, but the ropes behind the data, "a sketch of what's occurring" is precisely what a causal model
gives you, and what MHE pretends to ignore. #Bookofwhy

2.11.20 12:16am - I am not after definitions. I am intrigued by your accurate description of day-to-day practice
of social scientists, a confession I could not obtain from practicing social scientists, or from authors of econ.
texts. No qualms. #Bookofwhy

2 11 20 11:03pm - I think you have summarized the current practice of "natural experiment" fairly well though
2.11.20 11:03pm - I think you have summarized the current practice of natural experiment fairly well, though
MHE gurus would prefer to dress it up in more socially accepted terms, e.g., https://ucla.in/36EoNzO. Note,
matching cannot undo model-blindness. See #Bookofwhy page 274.

2.10.20 10:40pm - I could not let this phrase pass w/o a comment: "a pretty standard regression analysis. It's
causal in that the coefficients show predictive power". Now I do not know where to begin. Perhaps here
@Bookofwhy? Perhaps Wright (1920)? Perhaps the "microscope" https://ucla.in/2L8jUFg?

2.10.20 2:21am - Your tweet has enticed me to share a nostalgic piece from my ancient past. It is a popular
exposition of the hopes that superconductivity inspired in the 1960's:
https://ftp.cs.ucla.edu/pub/stat_ser/pearl-1966-electronics.pdf . Enjoy. Nothing like #Bookofwhy though.

2.9.20 11:48am - Agree with @eliasbareinboim . This paper https://causalai.net/r384-lnai.pdf gives a more
extensive and formal characterization of notions such as autonomy/locality/invariance that are usually
attributed to causal relations. #Bookofwhy

2.9.20 6:50am - (Replying to @yudapearl and @attilacsordas) Sorry, I meant do(slippery), not do(wet).

2.9.20 6:48am - It also contains answers to: "Who needs causality if all you want is prediction". As I once
tweeted, ask any pollster what the secret is to good prediction and the answer will soaked with causal
vocabulary. Why? Integrating data from multiple sources (Fusion) is a causal exercise

2.9.20 6:24am - I watched this Ted talk in 2016, impressed by the capability of her system. Last month she gave
an interview at the Technion, Israel, (our alma mater) and, when I watched it again I noticed that this is the
best promotion for causal networks in existence (not counting #Bookofwhy)

2.9.20 6:36am - (Replying to @BenjaminShender) You mean conspiracy theories that emanate from hate are
genuinely believed by the victim (eg., Roger Waters) and cannot be labeled "dishonesty", despite his rational
knowledge of the truth. Interesting. My heart goes out to him.

2.9.20 6:24am - I watched this Ted talk in 2016, impressed by the capability of her system. Last month she gave
an interview at the Technion, Israel, (our alma mater) and, when I watched it again I noticed that this is the
best promotion for causal networks in existence (not counting #Bookofwhy)

2.9.20 6:05am - The journey to predict the future: Kira Radinsky at TEDxHiriya https://youtu.be/gAifa_CVGCY
via @YouTube

2.9.20 3:55am - (Replying to @attilacsordas) Banana peel is a good instantiation of do(Wet), but the DAG does
not need such specific external forces. Just tell it do(Wet) and it will imagine one on its own. On page 71 we
will indeed represent do(Wet) by a force F(Wet), but why load the DAGs with redundant nodes? #Bookofwhy

2.9.20 2:45am - (Replying to @EpiEllie) Sorry, this tweet


https://twitter.com/yudapearl/status/1226420822211878913 was meant for @EpiEllie and the toddler who was
so impressed by what he thought he should have known. #Bookofwhy

2.9.20 1:55am - (Replying to @Irishchutzpah) Why give Roger Water the benefit of a "thinking disorder" or
"hereditary disease". He knows the truth about Israel and still acts as if he believes the slogans against it. That
makes him "dishonest", not "sick".
2.9.20 1:21am - (Replying to @Spinozasrose and @m_pezzolla) It is not "hatred" nor "antisemitism" but a
straightforward continuation of the culture of deceit, as conveyed to The Guardian by Ehud Barak

https://theguardian.com/world/2002/may/23/israel. The more of it, the more its character is exposed and
understood.

2.9.20 12:33am - (Replying to @Claire_Voltaire) Bernie and @ewarren have not learned from Corbyn's defeat,
they think those grand "coalitions" represent anyone but their creative slogan writers. Like Corbyn, they
underestimate voters' intelligence and voters' sense of right and wrong.

2.9.20 12:21am - (Replying to @PHuenermund) If you read the "linear microscopes", here
https://ucla.in/2LcpmHz and here https://ucla.in/2L8jUFg, you will know more than what most Econ&PoliSci
will ever know. And that's a lot. @Bookofwhy

2.8.20 11:45pm - Which raises a deep cognitive science question: Can dishonesty be localized? Can an artist be
dishonest in his moral makeup and, simultaneously, honest in his artistry?

2.8.20 11:26pm - From my current reading of the article I do not see the super-breakthrough advertised. I see a
serious attempt to narrow the set of data-compatible structures, using recently proposed (2006-) asymmetry-
producing ASSUMPTIONS. Worth watching & reprimanding the advertiser #Bookofwhy

2.8.20 10:44pm - I cannot overemphasize the importance of page 24 for understanding: "intervention",
"parents", "invariance" and other notions, often treated informally. Chapter 3 later explicates the ramifications
of this one page, and gives us backdoor, do-calculus and much more, #Bookofwhy

2.8.20 10:04pm - My understanding of "measurement errors" is summarized in two papers on "effect


restoration", linked via https://ucla.in/2N5GNOK and https://ucla.in/2N9icIX. I have the hunch the answers can
be found there. #Bookofwhy

2.8.20 9:56pm - The advertisement for this super-breakthrough in causal inference is unfortunately too sketchy
to comprehend. I hope someone attends the AAAI-2020 talk and tell us, in technical language, how they
overcome the First Law: No causes in - No causes out. #Bookofwhy

2.8.20 10:20am - This is something (some) American Jews do not understand (e.g., New Israel Fund, post-Zionist
Synagogues) as they accelerate the disappearance of American Jewry by weakening the moral basis for Israel's
sovereignty.

2.8.20 9:46am - (Replying to @AngeloDalli) It is surely possible, because we can always fit a function to data.
The question is whether the fitted function (w/o causally accounting for daily temperature) would correspond
to our understanding of what banning ice cream will do to the world. #Bookofwhy

2.8.20 3:09am - (Replying to @ZionessMovement @JewishCJustice and 2 others) Anyone knows whose Bernie's
surrogate is? Who makes up the questions? Whether anyone there is prepared to ask the hard questions? And
where we can find the program?

2.7.20 8:20pm - It's final! Eden Alene will represent Israel in the Eurovision contest 2020. When I saw her
perform: https://youtube.com/watch?v=iRFiHNa8wsw&list=PLxHkdEXvO_4wsUouFBH6J9xK5CmkEtkB- I could
not control my goose pimples. Thinking how her parents walked the vast deserts of Ethiopia to give their
hild h ii h h
children the opportunities they now have.

2.7.20 7:53pm - Thanks for retweeting this exercise. It was proposed in Sep 2019 but, in light of our recent
discussion on DL and explainability, it remains doubly relevant. Let's do it together. #Bookofwhy

2.7.20 7:28pm - (Replying to @oacarah @JohannesTextor and @ProfMattFox) Agree, the "microscope" is a useful
paper https://ucla.in/2LcpmHz, I frequently consult it when linear models come up. I wonder why the dont
teach it in Stat 101, since it can be taught as an innocent exercise in linear regression, not mentioning the C-
word. #Bookofwhy

2.7.20 7:16pm - (Replying to @memosisland and @ylecun) Not so. Loops are allowed. See Causality page 315,
and its complete solution on page 316. Dont miss footnote 10, on my funny experience with solving a problem
that economists themselves posed. #Bookofwhy

2.7.20 8:34am - (Replying to @EpiEllie and @PHuenermund) It is alien in two senses: (1) The founding fathers of
econ, Haavelmo, Koopman and Marschak would rather quit academia than teach from MHE. (2)Thinking about
an economic reality as an RCT is convoluted and constraining, which is why they can't use DAGs, only IV.
#Bookofwhy

2.7.20 8:22am - (Replying to @EpiEllie and @PHuenermund) This is the secret of Neo-econometrics. They quote
and quote and requote themselves, which helps create an illusion of progress. But see what this echo-chamber
quoting amounts to in my answer to Imbens https://ucla.in/36EoNzO; #Bookofwhy

2.7.20 8:07am - (Replying to @nano_unanue) It does not need to be a DAG. SCM may have a cyclic structure.
But when the structure is acyclic certain properties come into being what facilitate analysis and identification
and do-calculus. #Bookofwhy

2.7.20 8:03am - (Replying to @EpiEllie and @PHuenermund) The other way around (remember cause and
effect?). I do not like Target Trial framework (do you still call it "framework"?) because I can see which
direction it takes, not the other way. #Bookofwhy

2.7.20 7:47am - (Replying to @PHuenermund and @EpiEllie) I am not sure we agreed here, on tweeter, which
direction it goes. Remind me, but I am fairly sure it was not the "right" direction. #Bookofwhy

2.7.20 7:15am - (1/ ) It is refreshing to read about "Causal diagrams: pitfalls an tips"
https://jstage.jst.go.jp/article/jea/advpub/0/advpub_JE20190192/_pdf by folks who actually use DAGs, as
opposed to distant commentators. I recommend however that the paper be read in the context of (1) my
critique of Imbens https://ucla.in/36EoNzO
2.7.20 7:15am - (2/2) and (2) the examples in https://ucla.in/2wbH488 which deconstruct myths about TWIGs.
#Bookofwhy

2.6.20 4:38pm - If you are planning to attend AAAI-20 in NYC https://aaai.org/Conferences/AAAI-20/invited-


speakers/ dont miss our humble contribution to causal discovery https://ucla.in/2LyMbHo, Monday 2/10

Technical session 10: Reasoning under uncertainty. "For the want of a nail... a victory was won/lost" (BF)
#Bookofwhy

2.6.20 8:21am - More on Fairness, this time in Nature https://nature.com/articles/d41586-020-00274-3


accepting and explaining the necessity of causal models to define and quantify what we mean by Fairness or
absence of Fairness. @Bookofwhy

2.6.20 7:46am - (Replying to @TheAngrySemite) Mena is where your language incubated, where your heroes

sang their songs, and where your mind wanders when you say "homeland".

2.6.20 5:04am - (Replying to @joedotfaith @ylecun and @federalreserve) Very interesting! I wasn't aware of
those regs. What backgrounds did the authors have?

2.6.20 2:39am - (1/n) Two comments on @ylecun thread, on doing engineering w/o causal explanations. True,
flight engineers can improve wing design with Navier Stokes eq. But naive users need naive explanations on why
a loan application was denied, gradient-based optimization in NN will not satisfy
2.6.20 2:39am - (2/ ) this need. I touched on this aspect of man-machine communication here
https://ucla.in/2wj4pox. 2nd comment: Every heat engineer knows the 2 laws of thermodynamics, albeit not
their atomic explanations. These 2 laws have saved humanity many hours of searching for perpetual
2.6.20 2:39am - (3/ ) motion machines, and many many hours reviewing and rejecting proposals for such
machines. In machine learning we also have two conservation laws: 1. Conservation of interventional
information (the barrier on Rung 1) and 2. Conservation of counterfactual information (Rung 2
2.6.20 2:39am - (4/4) barrier). Taking seriously the analogy with engineering, one would insist on teaching ML
thermodynamics to every ML class in the country. Would Facebook join UCLA in initiating this huge educational
project? Just think of the savings in perpetual motion machines. #Bookofwhy

2.5.20 4:23am - (Replying to @shamoons) No! Jihadi music is Jihadi music. "From the river..." is a death chant
to human beings living in the neighborhood. Not to mention the students whose meeting was silenced by the
saintly "simple chant".

2.5.20 2:02am (Replying to @arxter and @Claire_Voltaire) Cultures void of history claim "indigenous is a silly
notion" Others embrace it to heighten creativity.

2.4.20 10:58pm - I've heard this Jihadi music since I was 3 yrs old. What's new is the participation of gullible
American students. We, academics, are partly responsible, by begging protection from anti-semitism instead of
demanding a stop to anti-Zionism - the more lethal of the two racist cults

2.4.20 3:44pm - I do not give up and claim (as did Kruskal and Lord) that the solution lies beyond statistics.
Rather, I am asking: (3.1) What information is needed for a solution? (3.2) What notation would this information
be cast in? (3.3) Would I be able to read this notation and judge the
2.4.20 3:44pm - plausibility of the claims? (3.4) What would I do with it, once I agree with the plausibility? (3.5)
Does it have any testable implications? (3.6) Unlike my semi-revitalized colleagues, I will begin critiquing
specific models only when satisfied with (3.1)-(3.5). #Bookofwhy

2.4.20 2:02pm - The last question is answered fairly well in the conclusion section of the Lord Paradox posting
https://ucla.in/2YZjVFL. A revitalized mainstreamer recognizes that (1) The two clashing intuitions are deeply
entrenched in statistical thinking and should not be brushed off , and
2.4.20 2:02pm - (2) Both intuitions are causal, hence, to reconcile the apparent clash between them we need a
causal language; statistics alone won’t do. A fully revitalized mainstreamer goes further: (3) accepting that
every causal assertions must invoke untested causal assumptions I do not

2.3.20 5:40am - Replying to @Plinz and @ArcusCoTangens) I would like to believe that defenders of mainstream
listen to me, not because I have an authority in any subject, but because my arguments for revitalizing
mainstream make good sense and science would benefit from the revitalizatioin. #Bookofwhy
2.3.20 1:05am - Any disdain or conspiratorial undertones on my side are imaginary at best. In fact, as I
articulate here: https://ucla.in/2v72QK5 I respectfully invite mainstreamers to join me in the effort, by
temporarily halting the question "What if the model was wrong?" and attend to another
2.3.20 1:05am - , equally important question: "Suppose it was right, what would we do with it?". Attending to
this question is pre-requisite to resolving causal problems such as Lord's Paradox. (or, more generally: should we
adjust for base-line conditions?) #Bookofwhy

2.3.20 12:07am - (Replying to @jtrecenti @athos_damiani and @agpatriota) Perhaps you can explain "the
situation" w/o DAGs. But I have not seen an explanation of why we should come to a different decision, with
the same data, depending on the story. #Bookofwhy

2.2.20 11:57pm - (Replying to @MarcusCrede) First, the data show that initial weight is correlated with diet.
Second it stands to reason that over-weight students would choose a dining room differently than under-weight.
#Bookofwhy

2.2.20 9:55pm - (Replying to @ArcusCoTangens) The question was: what information do we need to decide
correctly. This can be answered independently of the question "do we have this information?" or "do we have
sufficient evidence to support the needed information?" Separating tasks does not mean neglecting tasks
#Bookofwhy

2.2.20 8:34pm - Correcting a link to the Lord Paradox posting. The correct link is
http://causality.cs.ucla.edu/blog/index.php/2019/08/13/lords-paradox-the-power-of-causal-thinking/ and it
should go to: Lord Paradox and the Power of Causal Thinking. (Thanks to Stephen Leroy for noting). #Bookofwhy

2.2.20 8:25pm - (1/ ) About a dozen or so readers have offered creative proposals for resolving Simpson's
paradox in the X,Y,Color scatter plot example. I can't comments on each of the proposals, but I would beg the
discussants to focus on my humble proposal: A causal model is both necessary and
https://twitter.com/agpatriota/status/1224145232952025088
2.2.20 8:25pm - (2/ ) sufficient for resolving the paradox, namely, for deciding if X increases Y for a person with
unknown color. I would be happy to respond to anyone who thinks this statement is in some way incomplete.
#Bookofwhy

2.2.20 8:08pm - (Replying to @jtrecenti @athos_damiani and @agpatriota) What explanation is "equivalent" to
DAGs which does not use DAGs? perhaps PO? or "exchangeability"? or "higher resolution?" or "context-sensitive"? I
am yet to see one. #Bookofwhy

2.2.20 5:22pm - (Replying to @jtrecenti @athos_damiani and @agpatriota) Why do you say "Causal models are
useful" instead of "are necessary"? "Useful" is what economists use to justify not using (see
https://ucla.in/36EoNzO). When we have same data demanding two different conclusions depending on the
model, we say "necessary" #Bookofwhy

2.2.20 5:08pm - (Replying to @agpatriota @eduardohorta and @michael_nielsen) Explanations are man-made,
but we are facing a decision that has true/false value: Will Joe gain a higher Y with more X ? Philosophy aside;
do we or dont we have enough information in the scatter plot to decide correctly? #Bookofwhy

2.2.20 4:59pm - (Replying to @DiogoFerrari) I would say: It is impossible to "deal with Simpson's Paradox"
without a causal model, and it is impossible to specify a causal model in the language of probability
distributions, however intricate. See https://ucla.in/2Jfl2VS which you've cited but not taken
p y
seriously.#Bookofwhy

2.2.20 4:37pm - (Replying to @eduardohorta @agpatriota and @michael_nielsen) I interpret your hesitation to
mean: It depends on the causal relationships between X, Y and Color. Agree. Absent these relationships we

cannot decide if X would increase or decrease Y for that person. No need to bemoan science, we need only
glance at the causal graph. #Bookofwhy

2.2.20 7:33am - (Replying to @eduardohorta and @michael_nielsen) It surely helps. And the real puzzle begins
here: A person with unknown color comes in, would X increase its Y or decrease it? #Bookofwhy

2.1.20 5:49pm - (Replying to @ArcusCoTangens) Too bad for the standard. Graphs are man made. Better to
define things in their organic habitat - probabilities.

2.1.20 5:41pm - (Replying to @attilacsordas and @d_spiegel) It is not the "style" that gets things in order, but
the introduction of the 3rd-dimension: causation. Hopefully a new chapter in @d_spiegel next edition.

2.1.20 5:31pm - (Replying to @agpatriota) It's the appendix to a prepublication report. It proves that reversal
cannot occur in causal logic, in which the notion of "good for men" is expressed in do-operator. And it claims our
intuition comes from this calculus, not from statistics, where reversal does occur. #Bookofwhy

2.1.20 11:45am - Cook & Campbell spent years enumerating threats to validity -- they could not do any better,
and this paper explains why: https://ucla.in/2N7S0K9. Today, that we know how to establish external validity,
it is a pleasure to see all the threats that are circumvented. #Bookofwhy

2.1.20 9:55am - (Replying to @agpatriota) No! do(x) does not appear in


https://ftp.cs.ucla.edu/pub/stat_ser/r414-reprint.pdf I just identified where the paradox comes from -- i.e.,

clash between causal intuition and statistical logic -- and resolved it by replacing stat logic with causal logic.
Plain commonsense. #Bookofwhy

2.1.20 8:56am - I do not believe any of these countries thinks this "peace plan" was meant to be a peace plan. It
was meant to be a sober reminder to Palestinians that time may no longer be on their side, to stop the tantrum
and accept their neighbors as permanent, equally indigenous neighbors.

2.1.20 5:10am - A colleague alerted me to a new wikipedia entry on Market Blanket


https://en.wikipedia.org/wiki/Markov_blanket It is badly written, defining MB as a property of a graph, instead
of a probability distribution. The most astonishing feature, uniqueness under positivity, is not mentioned.
#Bookofwhy

2.1.20 5:04am - (1/ ) Curious me took a glimpse at Michael Nielsen's blog, which triggered Gelman's discussion
of Simpson's paradox http://michaelnielsen.org/reinventing_explanation/index.html. Michael is blunt: "[The
paradox] shows that some of our ingrained intuitions about statistics are not just wrong, but spectacularly
wrong." Yet
2.1.20 5:04am - (2/ ) Yet 90% of living statisticians still believe it can be "resolved" by re-visualizing data. My
esteemed discussants on The American Statistician, https://ucla.in/2Jfl2VS (eg, Xiao-Li Meng) wouldn't even
utter the word "causation". And it is 2020, and Statisticians get angry
2.1.20 5:04am - (3/3) when I remind them of the date, and their students still can't cope with a paradox that
has haunted statistics for the past 120 years. Pearl unfairly bashes statistics, they say. #Bookofwhy
2.1.20 4:05am - (Replying to @zakkohane and @StatModeling) I am no longer convinced this is true, after seeing
readers comments on "compatibility among the experts". The only way to grasp differences and commonalities
is to take ONE toy problem and try to solve it using competing "approaches". The rest is Hollywood. #Bookofwhy

2.1.20 3:56am - (Replying to @dalmiaman) Don't let them "rehabilitate" you back into correlations.

1.31.20 11:04pm - (Replying to @KordingLab @danilobzdok and 2 others) 90% of living statisticians say the
same:"I am interested in non-causal questions". Fine! But when a scholar titles a blog "causal inference" you
expect to see some interest in causal questions on that blog, not "you do yours and I'll do mine" #Bookofwhy

1.31.20 7:56am - My disagreement with Gelman is fundamental, because his views represent an attitude that
paralyzes wide circles of statistical researchers. My initial reaction was posted on https://bit.ly/2H3BH3b
Related posts: https://ucla.in/2sgzkPZ and https://ucla.in/2v72QK5 #Bookofwhy

1.31.20 7:28am - (Replying to @KordingLab @danilobzdok and 2 others) That's one of them, thanks. As you can
see, the causal character of Simpson's paradox is avoided like a plague. Why? I am asking you as one who knows
the solution: Is it possible to solve w/o graphs? Those who can't solve a problem, avoid it, and cover up the
void..#Bookofwhy

1.31.20 6:13am - (Replying to @danilobzdok @blei_lab and @StatModeling) Does anyone in 2020 believes that
linear regr can be used for causal discovery?

1.31.20 6:10am - (Replying to @KordingLab @danilobzdok and 2 others) In person? You must be kidding. If you
knew the number of hours I spent on his blog, you would not say "different language". The example that brings
everything to the surface is Simpson's paradox. I dont have the strength to dig it again from his blog. If you can,
please share.

1.31.20 5:47am - (Replying to @KordingLab @danilobzdok and 2 others) I am not so sure. Why? Because all my
attempts to convince @statmodeling to show us how he solves a simple causal problem (one whose solution is
known in advance) ended up in failure. I lost my charm. Simpson's would be a good example #Bookofwhy

1.31.20 5:22am - (Replying to @mf_schomaker) One argument is, indeed, that it should have been encoded in
the SCM. Another, that it should be treated as "disjuctive action", and extrapolated by imaging:
https://ftp.cs.ucla.edu/pub/stat_ser/r359-reprint-forthcoming.pdf #Bookofwhy

1.31.20 5:11am - (Replying to @sbuhai) But look at the cost of this "standard". Invited papers tell readers: "Here
is a body of work that has been neglected, catching up requires that we waive the "originality" requirement." If
it were not for such editorial leadership, my work would be unknown outside CS. #Bookofwhy

1.31.20 3:59am - (Replying to @sbuhai) I beg to differ. An "invited survey" has a totally different status than
ordinary submission. The requirements of innovation and importance are waived, b/c the editor decides on the
latter. There remains only the requirement of "representation." #Bookofwhy

1.31.20 3:37am - Retweeting my rebuttal to Imbens's paper, because readers complained about bad @@ signs in
the original post. The link is still the same: https://ucla.in/36EoNzO #Bookofwhy
1.31.20 1:43am - (1/ ) Another important clarification. In what way does SCM embrace the logic of PO? Ans. It
supports the consistency rule Y_x = Y if X=x, which is the main inference engine of PO. Consistency a theorem in
SCM https://ucla.in/2MlB5Utand and an assumption in PO (part of SUTVA, which
1.31.20 1:43am - (2/ ) deals with side effects in experimental settings). In this sense one can safely say that
SCM provides legitimization for the logic of PO, but rejects PO as a "framework" or "approach" #Bookofwhy

1.31.20 1:08am - (Replying to @ThomSeaton) Don't give up so easily. "Ein Lecha Adam SheEin Lo Shaa" says the
Mishna (Every person has untapped potential). Even "y=mx+b" is not trivial. I once asked a statistician what Y_x
is, when b is correlated with x. Repeating this story would get me into another trouble. #Bookofwhy

1.31.20 12:30am - (1/ ) Important clarification of this point: SCM embraces the counterfactual notation Y_x and
its logic, but not the "PO approach" which is a research methodology built around Y_x. The fundamental
difference: SCM starts with what you believe (eg DAG), PO starts with what you need (e
1.31.20 12:30am - (2/ ) (eg., ignorability conditions). Why can't PO start with what we believe? Because it
insists on expressing everything in the language of Y_x, no structure, no DAGs, while SCM starts with the
language in which scientific knowledge is stored: "who listens to whom?"#Bookofwhy

1.30.20 11:47pm - A chapter in history. Readers asked when the relationship between PO and DAGs was first
reported. My earliest record is this '93 Stat Conference in Florence https://ftp.cs.ucla.edu/pub/stat_ser/R195-
LL.pdf (Section 6) As I recall, the audience had NO IDEA what I was talking about - blank eyed.#Bookofwhy

1.30.20 8:27pm - (Replying to @Jabaluck) Can we skip the "unfortunate" - the fault is all mine. And can we start
rejoicing what we can do today that we couldn't yesterday? Rejoice man! Show your students how to do it. And
write to Imbens to open Econometrica to those who want to learn. #Bookofwhy.

1.30.20 5:18pm - (Replying to @GregZ_MD @ipam_ucla and @GlockerBen) Would have loved to hear the talk, or
get the paper, if available.

1.30.20 5:14pm - (Replying to @wagonomics @causalinf and 2 others) Clarification! @yudapearl has embraced

the counterfactual notation Y_x, as did PO. But he could not embrace the "PO approach", which permits
researchers to assume what they need, instead of what they believe. Not their fault, PO could not express what
they believed @Bookofwhy

1.30.20 4:27pm (Replying to @btshapir and @Jabaluck) Research answer: Economist: Citation for me? Citation
for whom? Non-economist: Rejoice man! Look what we can do today that we couldn't yesterday! Rejoice!

1.30.20 1:51pm - For the benefit of all readers, I have compiled my "Causal, Casual and Curious" articles in one
searchable page, now posted here: https://ucla.in/3aWaQkf . Enjoy. #Bookofwhy

1.30.20 1:38pm - (Replying to @Jabaluck) You are right. I was saying it more excitedly: "Rejoice! ALL
(quantitative) Economists should find it useful" . I say the same today: Rejoice!

1.30.20 5:26am - (1/2) It is FAT2020 week, so I'll not spoil the fun. But I am not convinced by Issa's argument
that counterfactual logic cannot capture fairness because it presumes autonomy which breaks down in social
systems. Heckman made similar arguments for economic systems: "Shutting down one
1.30.20 5:26am - (2/2) equation might also affect the parameters of the other equations in the system and
violate the requirements of parameter stability." (Heckman and Vytlacil 2007). I countered here
https://ucla.in/2NnfGPQ#page=44, as well as Cartwright's critique of autonomy, earlier.#Bookofwhy
1.30.20 1:32am - @causalinf , @PHuenermund , @EconBookClub, @jeremyskog Speaking of heathens, converts
and imposters in these last days of DAG-deprived economics, I've posted a brief rebuttal to Imbens paper on PO
and DAGs: https://ucla.in/36EoNzO. Comments are welcome. #Bookofwhy.

1.30.20 1:06am - (Replying to @attilacsordas) Can you rephrase the question in terms of information relevance,
given that we "know". ???

1.29.20 3:37pm - (Replying to @shamoons) The question was addressed to colleagues who are dancing to the
eliminationist tune of "from the river to the sea", not to Governments, organizations, or phenomena. I assume
you are not one of those dancers.

1.29.20 7:20am - (Replying to @TheShubhanshu and @fatconference) It is still only one page long; is that all you
have?

1.29.20 7:14am - (Replying to @causalinf) Remind me what our bonus agreement was. But for the life of me,
where did find a "skeptical friend"? In 2020? In economics? You must have been searching really hard.
#Bookofwhy

1.29.20 6:25am - (Replying to @TheShubhanshu and @fatconference) Thanks for posting. Does anyone has her
compete paper? I saw it someplace, and it seems to challenge the capacity of counterfactual logic to capture
"fairness", so I am wondering what logic can capture it. #Bookofwhy

1.29.20 6:08am - (Replying to @EmraniMd) I would be careful before I give a cold-blooded Zionophobe the
benefits of hereditary disease.

1.29.20 5:18am - What is Causal Cognition? I thought this article will help me find the answer:
https://frontiersin.org/articles/10.3389/fpsyg.2020.00003/full. But not seeing computational models of the key
concepts I feel unable understand the goals and methodology of this intriguing discipline. My weakness.
#Bookofwhy.

1.29.20 3:50am - For goodness sake, she really thinks apologies were made for self promotion.

1.28.20 10:35pm - Any peace plan that does not address the elephant in the room (i.e., the religion of "From
the river to the sea") will be remembered, not for what it says but for the eulogies in its funeral.

1.28.20 8:36pm - (1/2) I had the thrill of discussing this counterfactual world with @DavidDeutschOxf yesterday.
He viewed it as a swift allies victory in WW-II, assisted by a sizable army of dedicated Jewish volunteers from
Palestine/Israel. I would add to it: a true Palestine-Israel co-existence
1.28.20 8:36pm - (2/2) un-impeded by traumatic memories of the Arab 1948 attack on the Yishuv, and the
massive displacement of Arab population in that attack.

1.28.20 5:51pm - (Replying to @EinatWilf) Palestinian detriments are well documented, but their victories are
not widely known. One such "victory" was the entrapment of European Jewry, with the British Navy, and its
genocidal consequences. The memory of this successful "victory" makes "From the river..." appear feasible.

1.28.20 4:57pm - Everyone of my esteemed colleagues who is dancing and prancing to the tune of "From the
River to the Sea" will have to answer to Lady History when she asks: What have you done for peace?
y y y p

1.28.20 4:57am - (Replying to @RussInMtl @rlmcelreath and @Lester_Domes) Will follow your lead, as long as
end up with the latter.

1.28.20 2:46am - (Replying to @rlmcelreath and @Lester_Domes) Why do we assume that students are "scared"
by equations? I do not think this was on our minds in 1999. If it was, perhaps we should go back and sprinkle a
few more equations there. "A formula is a baked idea. Words are ideas in the oven." #Bookofwhy page. 335.

1.28.20 1:54am - As some of you heard, UCLA is under investigation by the Education Department on charges of
"discrimination and harassment," filed by two Jewish students. Here is a LA Times letter written by another
student, explaining the background:

1.27.20 5:52am - I am compelled to retweet your pleadings @RealSarahIdan , partly b/c you grew up across the
street from where my wife did, in Baghdad, partly b/c you sang Hatikvah with me, loud and clear, in the SWU
Conference, but mainly b/c the Iraqi people deserve freedom we take for granted.

1.27.20 5:10am - Sharing exciting nanotechnology news from the Technion (my alma mater), Prince Charles,
and Winston Churchill's Auditorium. The connection: https://jpost.com/Israel-News/Prince-Charles-meets-with-
Israeli-scientists-at-Technion-University-615355. Nostalgia: I was present at the 1958 dedication of this
auditorium, and later attended lectures there. #Bookofwhy

1.27.20 1:33am - (Replying to @mohomran) And I thought I was the first to name this on-going pattern of lies as
"CULTURAL" (not racial). Where did I get the idea of being first? On my campus, UCLA, where the word
"Palestinian" means "always right" "holy of holy" "never to be criticized". Anti-Palestinian??? Come to UCLA

1.27.20 1:14am - (Replying to @mohomran) It is not "acceptable" at all. I just coined this expression today,
seeing this culture grow from tiny sprinkles of denying Jewish history, to denying the Holocaust, to blood liable,
to denying yesterday's snow. But now, having stained the US Congress, someone has to name it.

1.27.20 12:53am - (Replying to @SachaBaronCohen @EinatWilf and 2 others) @SachaBaronCohen , you are so
right, and so is @JGreenblattADL . But don't those who delegitimize Israel also aim to encourage another one?
What else do they aim to encourage? And watch how close they are getting to their aim when tolerated as "just
something people get wrong".

1.27.20 12:30am - Why does everyone want Rashida to apologize? If she apologizes she would surely repeat the
lies in different dressings. Let her stew in her own soup as a monument to Palestinian culture of deceit. Some
Palestinians will eventually recognize what she does to their reputation.

1.26.20 8:04pm - Speaking about AI, machine learning, what DL can and cannot do? This analysis of 2800 yr old
Hebrew inscription tells it all. Can CI be of help? Perhaps, in case we want to generate the explanations
automatically. #Bookofwhy

1.26.20 3:43pm - (Replying to @jeffreywatumull @talyarkoni and 4 others) What humans do, we observe; what
NN can't do we can prove. Eg, humans shout: "Impossible" when told of a drug that is good for men, good for
women and bad for a person. What would NNs do if given Simpson's data? What would a statistician do? See
https://ucla.in/2Jfl2VS #Bookofwhy

1.26.20 5:21am - Be careful. Hannan Ashrawi is the Darling of the West. She always tells reporters what they
tt h E Wh k d if PA i I l i ht t i t h l "A f t i dI l
want to hear. E.g., When asked if PA recognizes Israel right to exist, she always says: "Arafat recognized Israel
in 1988". None ever asks her: "Do you?" They know she can't say YES.

1.26.20 4:45am - Replying to @attilacsordas) I agree that Mackie's INUS is inadequate and full of inconsistencies.
I mentioned them in a Tweet discussion on Rothman's "sufficient cause PIE", but Harvard Epis were so wedded to
it that I felt like I'm depriving them of a seat in heaven. See https://ucla.in/2lEOJtO #Bookofwhy

1.26.20 4:17am - (Replying to @DaveBrady72) This discussion intrigued me b/c it challenges the capacity of
counterfactuals to define "race", but I could not find the paper. Is it available? It raises the question: Are human
capable of ever agreeing on the distinction between "racist" and "non-racist" practices? #Bookofwhy

1.26.20 12:53am - (Replying to @talyarkoni @dileeplearning and 3 others) Should we really cast them as "open
empirical questions" when we have mathematical results on what is doable and undoable irrespective of NN

architectures ?? Not sure. Empirical claims are subject to interpretation when we get to higher level AI tasks
#Bookofwhy

1.26.20 12:41am - We spoke about innate templates that help reasoners do prediction intervertions and
counterfactuals. This new paper in Developmental Psychology explores how young children identify "causally
relevant variables" https://doi.org/10.1037/dev0000872 #Bookofwhy https://tinyurl.com/wg6b5og

1.25.20 10:52pm - Not entirely surprising. A culture rooted in deceit spawns leaders of like character.

1.25.20 9:57pm - (Replying to @HannesMalmberg1 and @HarryDCrane) I think of SCM in econ as reclaiming
commonsense in economics (hesitating betwee foreign invasion or home-grown abberation).

1.25.20 7:10pm - The more requests I get to participate in live debates with critics of CI, the more I wish to ask
if there is anything missing from the "Dialogue with A Hostile Examiner"( p. 369,
https://ucla.in/2NnfGPQ#page=38 that we can learn from a live debate or a staged fist-fight. #Bookofwhy

1.25.20 6:36pm - Nor will I get tired of commending those who recognized the educational power of Primer
https://ucla.in/2KYYviP While educators debate where to start - estimation, correlation, PO, or ML - Primer
starts where knowledge resides: Structure. The rest follows organically. #Bookofwhy

1.25.20 2:36pm - Replying to @HannesMalmberg1 and @HarryDCrane) Yet despite common ancestry, the new
culture of "natural experimentalists" does view "structural economics", or even mere "structure," as an
existential threat. If you think this is an exaggeration, I'll dig out some juicy quotes. #Bookofwhy

1.25.20 7:01am - (Replying to @cecilejanssens) Sorry if I misunderstood. My question was also simple: To decide
if a {hypothesis + data} supports a conclusion takes more than "reasoning 101". We need a non-standard logic to
do it if the conclusion is causal. Will wait patiently for a written summary when available. #Bookofwhy

1.25.20 6:27am - (Replying to @dr_benner) Hats off!!! Still, the entrapment of 1936-1940 needs to be taught,
especially in Rashida Tlaib's school district, in Michigan.

1.25.20 6:17am - (Replying to @trumanfrancis) Great question!!! Textbooks tell us: No way! correlation is all
that's needed for prediction. Butask any pollster what the secret is to good prediction and the answer will
invoke causal vocabulary. Why? Data Fusion is a causal exercise. See https://ucla.in/2Jc1kdD #Bookofwhy
1.25.20 5:55am - The most memorable moment I carry from this somber Holocaust memorial is listening to the
President of Germany recite the Hebrew prayer "Shehrchianu" ("to the living") instead of "in memory of the
dead". So clearly telling the world what Israel is all about. I wish the hundreds,
1.25.20 5:55am - perhaps thousands of Holocaust Museums around the world will follow his insight and dedicate
one exhibit room to celebrate the "to the living" message that Israel sends to the rest of the world.

1.25.20 5:25am - Replying to @cecilejanssens) What kind of logic have you used? My reading into the philosophy
of science says that, if the hypothesis is causal, classical logic will not do, and we need some sort of causal
logic. Which did you find most useful? #Bookofwhy

1.25.20 5:12am - Replying to @MaartenvSmeden and @trumanfrancis) Agree. It holds for any statistical model,
any ML algorithm of any form, regardless how smart, and regardless how big the data. It's hard for ML folks to
swallow, but I thought statisticians got over the trauma in 2002, when Denis Lindley confessed to this effect.
#Bookofwhy

1.25.20 4:34am - (Replying to @trumanfrancis) Indeed. Logistic regression in itself, without extra-statistical
causal assumptions, is as helpless as linear regression. It reveals NOTHING about causality. Please alert the
thousands of "logistic regressionists" of this fact; textbooks lure them to think differently. #Bookofwhy

1.25.20 4:19am - (Replying to @btolgao @nntaleb and 2 others) What element of irrational human behavior do
you think the introduction to #Bookofwhy should have emphasized more?

1.25.20 4:12am - Same goes for readers who ask "could you please explain what causal inference is about?".
Please read: "A Dialogue with a Hostile Examiner", page 369, in https://ucla.in/2NnfGPQ#page=38. The entire
miscommunication between statistics and CI is encapsulated here in 3 entertaining pages

1.25.20 3:55am - To readers craving to watch a fist-fight: What's the point? You can learn so much more from
solving one toy problem in CI. Still, for those who are incurably blood-thirsty, I got one for you. "A Dialogue with
a Hostile Examiner", page 369 here https://ucla.in/2NnfGPQ#page=38. #Bookofwhy

1.25.20 3:30am - (Replying to @BenjaminMenashe @nntaleb and @HarryDCrane) I believe @nntaleb had in mind
causal inference in probabilistic setting which, of course, #Bookofwhy and each of my writings is advertising in
big megaphones. Not to be confused with "probabilistic causality" as an outdated branch of philosophy. See
https://ucla.in/2ugpzTg

1.25.20 2:11am - (Replying to @CasualBrady) Sensitivity analysis with graphical models has not made it to
textbooks yet." But see this paper https://tinyurl.com/txcyan3 for a solid introduction to the framework.
#Bookofwhy

1.25.20 1:55am - (Replying to @CasualBrady and @Peninha_13) Agree. "can respond to" is as good. I happened to
like anthropomorphic metaphors, which irritates statisticians, but they dig it better. "can respond to changes in"
is also useful, unless you talk to a mathematician, for whom "non-trivial function of" is sufficient. #Bookofwhy

1.25.20 1:14am - (Replying to @HannesMalmberg1 and @HarryDCrane) If natural experimentalists where


interested in methodology they would rejoice finding their work perfectly aligned with SCM, enriched with more
powerful machinery. Unfortunately, they see "structure" as an existential threat. #Bookofwhy

1.25.20 12:40am - (Replying to @avicenna @joedotfaith and 2 others) Moreover, this assumption is "causal", not
statistical, so it cannot be expressed as some restriction on parametric family of distribution functions, as we
normally find in the p-value literature. Its an assumption of a new dimension. #Bookofwhy

1.24.20 9:02am - (Replying to @mathalytics @BatteryHorse and @HarryDCrane) No shame in self promotion,
especially when institutions expected to promote new ideas prefer to perpetuate old ones. I have done a few
tutorials at JSM, but a 1-day course should be more effective. Good luck. #Bookofwhy

1.24.20 8:26pm - (Replying to @joe_shipman @nntaleb and @HarryDCrane) Sorry if I misinterpret you meaning
of "meaning", it reminded me how I used to argue with Stats who kept insisting on "SEM has no meaning" or "the
only meaning of y=ax+eps is E[Y|x]=ax" [Holland 1988] etc. etc. The influx of Tweets from @nntaleb followers
reminded me of old days.

1.24.20 8:16pm - (Replying to @HannesMalmberg1 and @HarryDCrane) The partial derivative definition is a good
alternative to do(x). It has obvious problems with discrete x but, more importantly, we do not really have a
calculus based on partials that can deliver something like the backdoor/frontdor adjustment formulas. Do we?
#Bookofwhy

1.24.20 8:04pm - (Replying to @joe_shipman @nntaleb and @HarryDCrane) CI is orthogonal to the Bayesian vs.
Frequentiss debate. See here https://ucla.in/2nZN7IH Why I consider myself only half Bayesian. #Bookofwhy

1.24.20 8:00pm - (Replying to @joe_shipman @nntaleb and @HarryDCrane) No offense, but I can tell you are
statistician. How? For my stat colleagues things has "meaning" only if they can be expressed in the language of
statistics. Causal questions cannot be expressed thus, and I presume this is what is so frustrating to @nntaleb
and @HarryDCrane .

1.24.20 7:51pm - (Replying to @analisereal and @HarryDCrane @analisereal) Thanks for reminding us of this
ancient thread. Today I would answer it differently: "Bashing Statistics? I was generous! Count 2020 stat
textbooks that mention CI, stat dpts offering CI courses, PhD's who can solve a toy problem in CI (eg, Simpson
prdx?) #Bookofwhy

1.24.20 7:03pm - (Replying to @HarryDCrane) You are not missing a thing, if you don't have a causal question to
answer. If you have such a question we can examine if you can answer it in some other mathematical
framework, alternative to CI. I do not know of any, but am always open to learn. How about
Simpson's.#Bookofwhy

1.24.20 6:54pm - (Replying to @pastramimachine and @HarryDCrane) I will never get tired of recommending
Primer https://ucla.in/2KYYviP

1.24.20 6:52pm - I concur with @eddericu recommendation, though it depends on readers background. Some
are impatient with Primer's gentleness and prefer Causality with all its proofs. #Bookofwhy

1.24.20 6:49pm - Many readers ask me this question about Primer https://ucla.in/2KYYviP. But publishers are
driven by greed, and this is one secret they won't divulge. They want you to buy an uncorrected edition by
hiding all marks identifying the corrected one. Future authors, Beware! #Bookofwhy

1.24.20 6:37pm - (Replying to @HannesMalmberg1 and @HarryDCrane) Almost! It is logically equivalent to PO,
but infinitely more transparent. Similar to Haavelmo? Yes, see the parallels in https://ucla.in/2mhxKdO. As to
"econ defn" I have not seen one since Strotz & Wold (1960); it has been buried by the "natural experimental"
craze #Bookofwhy
craze. #Bookofwhy

1.24.20 6:26pm - (Replying to @HarryDCrane) You told me you are curious, not antagonistic, so I am searching
for your questions, and all I see are statements about what causality is, and what do(x) is good for. So, what are
you curious about? I am curious myself #Bookofwhy

1.24.20 6:15pm - (Replying to @nntaleb and @HarryDCrane) I assume you are asking these questions as a curious
observer, and you are not in antagonistic mood, like I've mistaken Harry to be. If so, permit me to ask how
familiar you are with the CI literature (it is all in the probabilistic dimension) & I will start from
there.#Bookofwhy

1.24.20 6:05pm - (Replying to @RotemEren @ShekatkarSnehal and 2 others) Who is blocking? Who is great? Why
block? I thought I mentioned Granger causality in #Bookofwhy, but was not sure. Thanks for quoting.

1.24.20 6:02pm - (Replying to @quantadan @ShekatkarSnehal and 3 others) I dont understand. Who is deleting?
who is blocking? who is dismissive? I just got assurance that this thread is motivated by curiosity, not
antagonism.

1.24.20 5:49pm - (Replying to @agpatriota @nntaleb and @HarryDCrane) Beautiful example. I use it often to
demonstrate the difference between Rung-2 and Rung-3 (see Causality, Section 1.4.4). Glad you brought it up.
#Bookofwhy

1.24.20 4:33pm - (1/ ) (Replying to @HarryDCrane) Since you are in such antagonistic mood, I'll answer your
question using "CI for the Infidel". You are unhappy with the Economist's article on minimum wage, and you
want to expose its weaknesses by pointing out how vulnerable the conclusions are to unwarranted assumptions.
1.24.20 4:44pm - (2/ ) (Replying to @yudapearl and @HarryDCrane) So far you are in good company. But now
you want to pick the critiqued assumptions effectively, making sure that they are CRUCIAL, not tangential, to
the conclusions. CI logic helps you decide whether a given assumption is necessary (aka sufficient) for the
conclusion. Nice!
1.24.20 4:51pm - (3/ ) (Replying to @yudapearl and @HarryDCrane) I say "Nice!" b/c no branch of statistics or
classical logic can give you such powerful machinery to criticize papers you don't like. But, to be more
constructive, here is another chapter in "CI for the infidel" https://ucla.in/2NnfGPQ#page=38 I called it a
"survival kit". #Bookofwhy

1.24.20 4:16pm - Anyone with a determination to read Primer is already "up to the task", and will soon join CI.
But make sure you've got the "corrected printing". How? Page ix should cite #Bookofwhy. This is the litmus test.

1.24.20 3:54pm - (1/ ) (Replying to @nntaleb and @HarryDCrane) This one is easy. In 1991, I had a quiet dinner
with Clive Granger in Uppsala, Sweden. Between the 2nd and 3rd glass of wine, he confessed to me that he
feels embarrassed by the name: "Granger causality", since it has nothing to do with causality, but he can't stop
people from
1.24.20 3:59pm - (2/3) (Replying to @yudapearl @nntaleb and @HarryDCrane) using it; they need some way to
express what they wish to estimate. I think we should honor him by echoing his understanding. An easy way to
see that GC has nothing to do with causality is to look at the defining equations and note that they comprise
only conditional
1.24.20 4:04pm - (3/3) (Replying to @yudapearl @nntaleb and @HarryDCrane) probabilities, no do(x)
expressions, nor counterfactual terms Y_x. Bingo! We are done! Whenever a concept is defined in terms of a
distribution of observable variables it can't be "causal". No causes in - no causes out (N. Cartwright) #Bookofwhy

1.24.20 3:36pm - (Replying to @littlebode) I wrote #Bookofwhy and opened a Twitter account so that each one
f 28K f ll h ld b bl d i d MOOC C li A h ?
of our 28K followers should be able to, and excited to create a MOOC on Causality. Are you there?

1.24.20 3:23pm - (Replying to @hubertpaulo and @Peninha_13) The invariance of the axioms under changing
definitions is important to mathematics, but the question was why not accept ONE of those definitions and work
with it as THE definition?

1.24.20 6:45am - The cover looks very familiar. I hope the translators fixed all our errors. Oh, BTW, we just
commissioned a Spanish translation. #Bookofwhy

1.24.20 6:39am - (Replying to @ShekatkarSnehal @HarryDCrane and 2 others) People think I'm kidding when I
say that most statistically-trained folks are still unaware of what CI is all about. Glad I have this Twitter post
from which to watch the glacial progress of science. #Bookofwhy

1.24.20 6:22am - nother forgotten chapter of history. Glad it sees the light of day. My grandparents were
stranded in Poland as a result of this Arab-British-Nazi circle of appeasement. It was "Genocide by entrapment",
still not taught in Palestinian schools.

1.24.20 3:41am - Nice example of how SCM can serve as a laboratory to test various interpretations of familiar
and colloquially used terms, in this case "incentives". #Bookofwhy

1.24.20 3:13am - (Replying to @Meetasengupta) Hilarious, if it wasn't so true. But don't blame AI. Blame big
business for enslaving us to premature AI. #Bookofwhy

1.24.20 3:01am - (Replying to @Peninha_13) Euclid "defined" lines and points in terms of other undefined
notions, "part of" "length" "breadth" etc. This reductive definition must stop at some point where we choose the
irreducible primitives and declare them "self evident". #Bookofwhy chose "listens to" as primitive.

1.24.20 1:48am - (Replying to @tdietterich @BethCarey12 and @GaryMarcus) Let me explain. Interventional
data cannot distinguish between two competing Rung-3 hypotheses. Eg, "No effect" in RCT cannot be
distinguished from "treatment kills some and cures others". See gentle introduction here:
https://ucla.in/2G2rWBv A true Rung-3 Primer. #Bookofwhy

1.23.20 6:44am - Your dream is my dream. Just signed. #Bookofwhy

1.23.20 6:31am - A new dawn in sensitivity analysis! A systematic approach based on meaningful assumptions.
#Bookofwhy

1.23.20 6:16am - The brightest light I see in this somber Holocaust memorial
https://jpost.com/Diaspora/Antisemitism/The-Fifth-World-Holocaust-Forum-Watch-Live-615063 is its taking
place in Israel, a country that symbolizes a metamorphosis from tragedy to hope, death to renewal and
historical injustice to inspiration for all oppressed minorities

1.23.20 2:52am - Clarification: Although the paper referenced does not mention do-calculus explicitly,
explainability criteria are defined in terms of interventions on a structural model M meeting condition C. So
readers can see what data+assumptions would enable do-calculus to prove C.#Bookofwhy

1.23.20 2:10am - Explainability is ML's Achilles heel. This paper https://google.com/search?


1.23.20 2:10am Explainability is ML s Achilles heel. This paper https://google.com/search?
q=%5BPDF%5D+The+Explanation+Game%3A+A+Formal+Framework+for+Interpretable+Machine+Learning&oq=%5BPDF%5D+The+Explanation+Game%3A+A+Formal+Framework+for+Interpretable+Machine+Learning&aqs=ch
8 presents formal criteria & measures for various aspects of explainability. Its use of do-calculus gives readers immediate view of the data and assumptions needed to meet each of the criteria. #Bookofwhy

1.22.20 1:48am - (Replying to @udansk) The issue is, if I am not mistaken, whether DAGs could be taken
seriously as computational models of human causal inference, as well as of mental store of scientific
knowledge. This is a stronger claim than the correctness of the math. #Bookofwhy

1.22.20 1:06am - (Replying to @kerstingAIML @mitbrainandcog and 4 others) Agree. I used "we" figuratively,
referring to the trendy community dominated by DL. I am still under the influence of the DL vs. CI discussion we
had here a week ago. #Bookofwhy

1.21.20 11:37pm - Speaking for myself, I've gained a lot of insight about myself by playing around with various
computational models. Example, I'd never guess that I had a causal diagram in my mind's eye until realizing that
no other representation can account for my swiftness and versatility.

1.21.20 10:49pm - In the 1970-80's, AI was all about "computational models of mental processes". With the
advent of ML, many now view AI as "algorithmic replacement of mental processes". We lost something in this
transition, b/c the fun part of doing AI is gaining an understanding of ourselves.

1.21.20 2:31pm - (Replying to @ArcusCoTangens @bariweiss) and 3 others Enormous difference! The former is
emotional, the latter calculated and genocidal. See https://ucla.in/2SxjgWH

1.21.20 1:50pm - (Replying to @udansk) I said DAGs (SCMs) make causal reasoning POSSIBLE. To refute it, you
need to present another mathematical object that allows the derivation of predictions interventions and
counterfactuals simultaneously. I am not aware of any. Is anyone? #Bookofwhy .

1.21.20 1:40pm - (Replying to @WiringTheBrain) Do we have computational models of that "deep knowledge",
ie, how it is represented in the mind and how it is accessed and transformed into a "causal diagram"? Is it not
possible that causal diagram are already the deepest form of knowledge accessible to an agent? #Bookofwhy

1.21.20 1:31pm - (Replying to @andrewthesmart) Disagree. "Computational models" can be emulated on a


computer. "familiarity" or "fieldwork", even "design", cannot!

1.21.20 1:27pm - (Replying to @SMBrocklehurst and @WiringTheBrain) The SCM (structural causal model)
supports feedback loops (see Causality chapter 7). People use the term DAG because acyclicity gives us many
computational advantages. #Bookofwhy

1.21.20 6:15am - (1/ ) Has AI done that? To a limited extent YES. Not through DL, sadly. But if you agree that
"understanding" means having a representation that allows you to predict and control a phenomenon then DAGs
do a fairly amazing job. Where else have you seen a compact mathematical object
1.21.20 6:15am - (2/2) from which you can derive all four: predictions, redrodictions, interventions and
counterfactuals? It's primitive, yes, but it offers us a laboratory in which we can test ideas about understanding.
Refinements? By all means! Knitpicking? Don't slow us down, please. #Bookofwhy

1.21.20 3:18am - (Replying to @WiringTheBrain @mendel_random and @causalinf) It is science indeed. But AI
(at least me) has endeavored to ask a slightly more ambitious question: What has changed in our mental
representation of a phenomenon that makes us feel: "We understand something well enough to define a
hypothesis that can be tested.". #Bookofwhy
1.20.20 7:07am - I summarize this discussion thus: Progress in our century amounts to building computational
models of mental processes that have escaped scrutiny under slippery terms such as "design" "discovery"
"background knowledge" "familiarity" "fieldwork". etc. Saying "DAGs are not enough"
1.20.20 10:25pm - (2/3) produces no progress until one is prepared to propose a mathematical model for what
IS enough. The IV criterion I presented is necessary for justification, so keep it in mind and teach it to natural
experimentalists who should be jubilant seeing their IV's justified.

1.20.20 10:25pm - (3/3) We should always be open to extensions and refinements, but mystification holds us
back. We had enough of that in the 20th century. #Bookofwhy

1.20.20 7:07am - (Replying to @deaneckles @thosjleeper and @aecoppock) What makes a promising candidate
differ from an unpromising candidate if not features that would enable it to pass the IV criterion against beliefs
about the world? #Bookofwhy

1.20.20 7:03am - (Replying to @ryancbriggs @causalinf and @mendel_random) I dont see any debate here. I see
an attempt to demystify what some people describe as a mysterious process of obtaining new domain
knowledge and spotting a "natural experiment" in the knowledge obtained. #Bookofwhy

1.20.20 6:55am - (Replying to @thosjleeper @deaneckles and @aecoppock) Sure. And then? Eventually, after
forming the new simplified DAG, someone has to say: "Hey I got a new IV, and it's a good one too!" #Bookofwhy

1.20.20 6:51am - (Replying to @causalinf and @ez_angus) Agree. I would only equate DAGS with "theory, model,
intuition, and deep institutional knowledge". Is there something in the last four that is missing in the DAG? If
there is, lets add it, after deciding of course how DEEP we wish to go. #Bookofwhy

1.20.20 6:43am - (Replying to @causalinf and @ez_angus) All circularities disappear if we equate DAGs with "an
encoding of what we know". Difficulties may surface if you can postulate a more refined, or more natural
encoding of knowledge than DAGs. I can't, but am open to suggestions. #Bookofwhy

1.20.20 6:30am - (Replying to @thosjleeper and @deaneckles) Talking with people is always enlightening, but I
am trying to demystify the process of "finding new variables, arrows etc. etc." and see how it leads us at the
end to "Hey, I got a new IV". Reaching such a conclusion requires checking an IV criterion on some model.
#Bookofwhy

1.20.20 3:49am - (Replying to @deaneckles) I do not see any disconnect. I see a quest to understand (semi
formally) what "fieldwork" gives us that we did not have before, and where a "fieldworker" puts the newly
learned variables if not in a some revised model. #Bookofwhy

1.20.20 3:07am - Let us try to give "a situation" a somewhat technical interpretation. You do "fieldwork" for 3
months, 24/7, you come back to your office and reflect: "Have I seen a "situation" that satisfies a definition?"
What's in memory are you interrogating with this question? #Bookofwhy

1.20.20 12:43am - To focus the discussion: Z is a good IV if there exists a set S of measured variables such that:
(1) S does not separate Z from X, and (2) S separates Z from Y after removing all arrows entering X (Causality
p.248). We see that the criterion does not require data. #Bookofwhy

1.19.20 11:57pm - Preparing myself for the #WorldHolocaustForum. My grandparents were murdered in
Auschwitz in 1942, my son has not been spared the wrath of this hate, and my colleagues in Israel are still under
the threats of Ayatollahs designs and Palestinian deligitimization. Preparing myself.

1.19.20 11:23pm - Replying to @PHuenermund Best news from econ. since Haavelmo.

1.19.20 11:20pm - Replying to @maximananyev But you are unfair to people who can only talk "natural
experiments", for whom the whole economy is a huge salad of natural experiments, and he who chooses one
and ignores the others is called an "experimentalist". #Bookofwhy

1.19.20 9:13pm - (Replying to @DrMikeH49 @StandWithUs and 5 others) Amazing day indeed. My highlights: 1)
Singing Hatikva with Sarah Idan (Miss Iraq 2017) and discussing the psychology of Palestinian Rejectionism with
Hussein Aboubakr (Egyptian scholar, CCA)

1.19.20 8:35pm - (Replying to @AdanZBecerra1 @PHuenermund and @laura_tastic) Good detective work,
thanks. The thing to do now is to petition Susan R. Bailey, AMA President Elect, to bring up this issue for public
discussion. My banner would be: "Censor the claims, not the language. Causes have produced effects Myrs
before RCT was invented." #Bookofwhy

1.19.20 5:45pm - (Replying to @yudapearl and @EinatWilf) "Culture or mentality?" Sadly, it holds the key to ME
peace: If Palestinian Mentality becomes the center of conversation, chances are the embarrassment will sober
up minds and plans. Here is a beginning:

1.19.20 5:23pm - (Replying to @EinatWilf) Agree. That's why Omar Barghouti started his lecture at UCLA (Jan
2014) with "They (Jews) are not a people..." Strange, Edward Said insisted on Arabs' exclusive right to define
themselves; his disciples now insist on exclusive right to define others. Is this culture or mentality?

1.19.20 6:35am - (Replying to @EinatWilf) If EU has any obligation at all it is to ask Dr. Saeb Erekat if he ever
thought whether Israel's has a right to self-determination. Last I heard from him https://tinyurl.com/uaoer8t he
said: NEVER! Can a people seek nationhood on its neighbor's tomb?

1.19.20 4:15am - (Replying to @yudapearl and @causalinf) For readers asking how one can spot a natural
experiment in a model, the answer is very simple and is given here https://ucla.in/2mhxKdO among the toy

problems I presented to economists in 2015. To the best of my understanding, they are still working on it. Are
they? #Bookofwhy

1.19.20 3:46am - (Replying to @attilacsordas @sciencescanner and @freesci) Discrepancies between RCT and
observational studies usually point to confounding (ie spurious correlation) in the latter. However, certain kind
of discrepancies may indicate selection bias in the RCT. See Causality page 294. #Bookofwhy

1.19.20 1:03am - (Replying to @aceyuan) Admitting ignorance and overwhelmed by acronyms: What is BERT?

1.19.20 12:58am - (Replying to @causalinf) I am not an economist, nor social scientist. But if I were one, I
would first learn how to spot a natural experiment in a well specified model of the phenomenon. Only then I'll
try to apply the method to the mental model I have of the phenomenon, however familiar it is.#Bookofwhy

1.19.20 12:07am - (Replying to @HenMazzig) The antisemites are poor victims of hereditary disease. It's the
Zionophobes that are doing it in cold blood, destroying every movement that gives them a nod of approval.
1.17.20 10:15pm - (Replying to @AdanZBecerra1 @Andrew___Baker and 3 others) Are you sure the PO guardians
of "well definedness" (say @_MiguelHernan ) would agree with your suggestion (and my position in
https://ucla.in/2LY8M2X) that if you can find it in the DAG it is automatically "well defined"? #Bookofwhy

1.17.20 9:52pm - (Replying to @Andrew___Baker @PHuenermund and 2 others) The notion of "potential
outcome" in the PO framework requires "treatment assignment" w/o which you cannot define Y_x. That is why

PO folks are debating whether non-manipulative x (say an earthquake) can have counterfactuals -- the RCT
roots are still strong & stifling.#Bookofwhy

1.17.20 9:28pm - (Replying to @Andrew___Baker @PHuenermund and 2 others) Many authors are still using "as if
randomized" instead of "unconfounded", and the whole enterprise called "target trial" is based on the idea that
if you try to emulate some RCT you would do better than trying to think about the underlying cause-effect
relations. #Bookofwhy

1.17.20 6:27pm - (Replying to @IzaTabaro) We are partly responsible for fermenting this aberration, by exposing
it ONLY when it crosses into anti-Semitism. Aren't they ugly enough, and racist enough, just being anti-Zionists?

1.17.20 6:07pm - (Replying to @PHuenermund @laura_tastic) , Where is this page proof taken from?

1.17.20 6:05pm - (Replying to @christopherruhm @PHuenermund and @causalinf) RCT draws its legitimacy from
a proof that, if conducted properly, it delivers a sought-after quantity called "causal effect". This means that
causal language must exist BEFORE RCT. The #Bookofwhy makes this point on pages 146-150.

1.17.20 12:02pm - (Replying to @IntuitMachine and @jm_alexia) Examples?

1.17.20 11:55am - (Replying to @MaxALittle and @jm_alexia) Who said it is hard? Only those who have not tried
it. Those who have, can’t understand the hesitation of the others. #Bookofwhy

1.17.20 4:58am - Readers who, like me, believe that symbols make history should understand why i felt
compelled to retweet this photo. One of the few hopeful signs in the past week, after Omar’s congress
refused to support iranian protestors, another history shaping symbol.

1.17.20 4:25am - (Replying to @DKedmey) “listening” is a good and accurate NL metaphor, and so is
“source of variation”. But these are good for pre-scientific discourse, or for education. Are we in a similar
pressing need for NL definition of “correlation”? #Bookofwhy

1.16.20 5:32pm - (Replying to @PhilosopherMD1) Keep up informed either way. Its been a long time since we
heard from philosophers of statistics.

1.16.20 5:27pm - (Replying to @PHuenermund @Lester_Domes and 6 others) The question was not about path-
specific effects (for which we have the Avin etal result) but about the Total Effect, ie, E[Stroke|do(Age)]. The
miracle of DAGs (not widely acknowledged) is that expressions of type E(Y|do(x), z)] do not need summation
over paths.#Bookofwhy

1.16.20 12:40am - Do we need a NL definition of a “cause”? We are blessed with a compact and meaningful
mathematical object from which we can derive NL utterances on all three Rungs, association, intervention and
counterfactuals! Why chase the unreachable? #Bookofwhy

1.15.20 11:30pm - (Replying to @Lester_Domes @RUBENSARO and 6 others) If we insist on doing path tracing
Wright style then, yes, all paths from AGE to STROKE should be counted. Fortunately, we have DAGS to save us:
The effect is simply E[STROKE|AGE] , no paths, no multiplication, no additions. Welcome to the magic world of
DAGs-land #Bookofwhy

1.15.20 4:58am - (Replying to @PHuenermund and @Andrew___Baker) Indeed, what’s in “design”?

Pressing hard, l’ve found that “design” is a word used by economists to cover decisions they prefer to do
informally, away from piers scrutiny, thinking old-fashionally that they cannot be made formally. #Bookofwhy

1.15.20 3:47am - Highly recommended watch. And while watching, worth keeping in mind what #Bookofwhy
says about Kahaneman’s dictum, and his “undoing project”.

1.14.20 2:48pm - (Replying to @JDHaltigan) If you can be specific, I'll be glad to demonstrate "no malice" in
#Bookofwhy

1.13.20 11:48pm - (Replying to @Physical_Prep and @Alan_Couzens) I have great faith in AI, and I have read
#Bookofwhy long time ago. The book, too, has great faith in AI.

1.13.20 11:26pm - (Replying to @GodfreySnorgyrs) Why dont you write one for anthropology? Epidemiology owes
a lot to that pioneering article, and I am sure anthropology will be grateful to you, if you write one.
#Bookofwhy

1.13.20 11:11pm - (Replying to @yudapearl @ShalitUri and @omaclaren) expect a "methodology" to deliver, not
guesswork about the dark side of the moon, which we can generate absent a "methodology", still consistent
with the standard mathematical definition of "solution. #Bookofwhy

1.13.20 11:06pm - (Replying to @ShalitUri and @omaclaren) That's exactly my point. Before we submit our
causal questions to the mercy of DL we need to ask whether we possess sufficient knowledge to prevent
contradictory solutions, or in prediction, for which curve-fitting has unique solution. This notion of "unique
solution" is what we.

1.13.20 10:51pm - (Replying to @richard_landes @DKedmey and @EinatWilf) I am thinking of fake feminists like
Linda Sarsour https://tinyurl.com/y4rjozub who is probably sincere in her admiration of Jews like Sen. Sanders,
and fake liberals like Corbyn who hate Jews because they support Israel more than he hate them for rejecting
Jesus.

1.13.20 9:49pm - (Replying to @omaclaren) A "solution" and a contradiction:


X,Y,Z = 1,2,2
X,Y,Z = 0,3,2
You just told me X=1, now you telling me X=0, and you want me to call it a "solution"? Are DL folks willing to
accept both: "the drug is helpful" and "the drug is dangerous" as a legitimate "solution" of DL methodology

1.13.20 9:05pm - (Replying to @richard_landes @DKedmey and @EinatWilf) Your video is very convincing,
though I've been exploring another theory, that antisemites like Richard Wagner and Mel Gibson can be forgiven
for sucking it with their mother milk, not so Zionophobes, like BDS activists, for whom Israel elimination is cold
g p
and calculated agenda.

1.13.20 8:31pm - (Replying to @DKedmey and @EinatWilf) I was not familiar with @DavidDeutsch pattern,
probably because I grew up in Israel, with a firm conviction that all past patterns are a thing of the past -- the
New Jew will put an end to irrationality and inhumanity. It is 2020, and I am anxious to study the "pattern".
Thanks.

1.13.20 8:21pm - (Replying to @DonBeham @RandomlyWalking and @raamana_) Yes, the diagram imposes
constraints on the data, so data can refute or satisfy a diagram. This the 2nd fundamental law of causal
inference which, fortunately, ML folks are beginning to internalize (unlike economists) .#Bookofwhy

1.13.20 1:13am - (Replying to @TheLeanAcademic @tdietterich and 2 others) The classical (and easiest) case is
d-separation; it tells us if some members impose conditional independence constraint on the data. But speaking
of Rung-3, things are more involved, and Section 8.5.2 of Causality (2009) has beautiful examples how data
shape priors. #Bookofwhy

1.13.20 12:48am - (Replying to @tdietterich @BethCarey12 and @RaajeevVerma) The right way to think about it
is, first, to ask if the "evidence" is capable of ruling out members of the model space. The hardest thing for
statisticians to swallow is that data can be equally compatible with each member of the space. So, we must act
as if we've NO evidence.

1.13.20 12:34am - (Replying to @nitalon and @ylecun) Chibuk L'Katin

1.13.20 12:31am - (Replying to @InnovatArt and @ylecun) This still begs the question: "mathematical models of
WHAT?", and what do we do with it once we have it? #Bookofwhy

1.13.20 12:25am - (Replying to @SinghalApoorve) I sometimes feel I need to re-read my own fables to reinforce
basic truth. That's what months in the trenches do to you. #Bookofwhy

1.12.20 10:25pm - (Replying to @ylecun) I am the last person to doubt your definition of DL, but I'll have very
hard time explaining it to my students; they will rebuke me if I ever say "construct a model" without saying a
model of WHAT. Is it a model of the dark side of the moon? Any data from that side? #Bookofwhy

1.12.20 10:05pm - (Replying to @roydanroy @suchisaria and 2 others) Last time (1990) I examined Dempster-
Shafer theory, it couldn't even do Monty Hall correctly. See https://ucla.in/2RetYPS and
https://ucla.in/2Ranb9U Anything new to consider? #Bookofwhy

1.12.20 9:54pm - (Replying to @roydanroy @suchisaria and 2 others) We, in the trenches, need more gentle
instructions on how you implement SEM/SCM as POMDP when you dont have the state transitions, not even
actions, just passive observations. Have you tried it? Has anyone you personally know? #Bookofwhy

1.12.20 9:46pm - (Replying to @InfoSymmetries) Explanation: This is Amazon announcement of the paperback
edition, which our publisher decided to release in August 2020. Good news, it will contain all the errata that
readers labored to find. Thanks. #Bookofwhy

1.12.20 9:39pm - (Replying to @BethCarey12 @RaajeevVerma and @tdietterich) In my little corner of the
i th i t fd t lift f R 1t R 2 f R 2t R 3
universe the answer is: no amount of data can lift us from Rung-1 to Rung-2, or from Rung-2 to Rung-3, no
matter how smart we are, and no matter if we call it DL or NN or AI or stochastic optimization. #Bookofwhy

1.12.20 4:44pm - (Replying to @roydanroy @suchisaria and 2 others) Please walk us by the hand through some
simplified issues. Take us with POMDP to decide which drug is better, the one that is good for men and good for
women, or the one that is good for people. Let's start with the model-free brethern. #Bookofwhy

1.12.20 4:33pm - (Replying to @suchisaria @RandomlyWalking and 2 others) How do we provide them with more
talk space than twitter with occasional encouragement to "Please rebel!"

1.12.20 3:29pm - (Replying to @simonbchen) The implicit assumption you are talking about, has hardly been
touched in DL circles, which is an interesting social phenomenon: A whole civilization discussing how to get a
better fit to data, rather than "what can we do once we get a perfect fit" #Bookofwh

1.12.20 3:16pm - (Replying to @VeronicaEpi) I've not met anyone saying "causal inference is still confusing" after
reading #Bookofwhy How can anything be confusing if it merely reflects human thought?

1.12.20 3:08pm - (Replying to @suchisaria @RandomlyWalking and @raamana_) I like your distinctions, for they
are cast in terms of the questions we aim to answer, rather than the mathematics used to answer them. I hope
they bring DL students to realize: "Gee! We never dealt with questions which we could not answer! The world IS
3-dimensional" @Bookofwhy

1.12.20 2:56pm - (Replying to @RandomlyWalking and @raamana_) "Snark" is not in my vocabulary. The high-
school algebra example simply demonstrates that, if we do not have the data to answer certain questions, we
should not just guess at random. It has nothing to do with inductive bias, which is a Rung-1 (function-fitting)
concept #Bookofwhy

1.12.20 7:03am - (Replying to @davidcnorrismd @venkmurthy and 5 others) I would not call it "maligns". It is a
very accurate description that most authors agree with, and only few dare share, which helps perpetuate the
mystery. #Bookofwhy

1.12.20 1:50am - (1/ ) (Replying to @raamana_) Hilarious! But the more we listen to DL talks, the more it
sounds like "everything". I am still hoping one of the faithfuls would remember how the idea that we need n
equations to solve for n unknowns changed how we did algebra in high-school. We first asked ourselves:
1.12.20 2:02am - (2/ ) (Replying to @yudapearl and @raamana_) "Do we have n equations?" If not: "Have we
forgotten one?", if not: "Perhaps we can impose one?" and then, only then, we submitted our equations to the
algebraic machinery that we could trust. The analogous process for CI is spelled out in https://ucla.in/2HI2yyx
#Bookofwhy

1.11.20 10:59pm - (1/ ) (Replying to @raamana_) What I did learn from the discussion of "What is DL" is that,
though no definition is in sight, DL folks are utterly intoxicated by the seemingly unlimited potentials of their
programs, techniques and vocabulary. I remember our second or third week of high-school algebra,
1.11.20 10:59pm - (2/ ) (Replying to @raamana_) utterly intoxicated by its unlimited solution-finding
capabilities, the teacher told us that you can't solve two equations with three unknowns. Humbled and
disappointed, we were nevertheless excited by this revelation; it saved us hours upon hours of chasing after
solutions
1.11.20 10:59pm - (3/3) (Replying to @raamana_) that do not exist. But these were high-school days.
#Bookofwhy
1.11.20 2:34am - The intense discussion on "What is DL" has evidently not converged on a consensus among DL
practitioners. Nor has it convinced me to reconsider any of the impediments reported in
https://ucla.in/2umzd65 or the ML/CI division of labor recommended. Ready to be educated. #Bookofwhy

1.11.20 1:27am - (Replying to @DKedmey and @tangled_zans) Reply to your tweet removed: @EinatWilf is my
hero. I wish Israeli leaders will be as clear as she is on the kind of peace Israel society strives to achieve and
what kind of "peace talks" are a waist of time, because they do not entail "equally legitimate and equally
indigenous".

1.10.20 4:58am - (Replying to @tyrell_turing) Not really. The other reason for the Q "what is DL" was that good
ppl identified DL with everything that AI can ever hope to achieve, which naturally led to the question: "What is
NOT DL?" #Bookofwhy

1.10.20 4:54am - (Replying to @witbrock @zaffama and @ylecun) The problem is not what classes of
computable functions can be effectively learned by a DL system? But rather, whether the answer you expect
your algorithm to deliver is a computable function of your data. #Bookofwhy

1.10.20 4:41am - I am re-tweeting with the hope of reaching one or two econ. students and enticing them to
ask some hard questions. #Bookofwhy

1.10.20 4:02am - (Replying to @csilviavr) Your paper makes the point very clear. I was first hesitant seeing the
term Causal Bayesian Network CBN, which is defined on interventional distributions (Causality p. 24). But you
corrected for it through path-specific, ie. counterfactual notation. #Bookofwhy

1.10.20 1:50am - (Replying to @javisamo) I couldn't access Lily Hu's paper. Can you summarize the limitations of
counterfactual reasoning? A punchy example, perhaps? #Bookofwhy

1.10.20 1:41am - (Replying to @PHuenermund @causalinf and @Dunkin_Donuts_2) I think viewing these
anecdotes as "personality cult culture" misses the real crisis of econ -- the only data-driven discipline the
majority of whose students are still deprived of 21st-century tools such as graphical models, and are not
rebelling, not even striking. #Bookofwhy

1.10.20 1:01am - I'm reading this extensive review of Fairness in Machine Learning https://phil-fak.uni-
duesseldorf.de/fileadmin/Redaktion/Institute/Sozialwissenschaften/Kommunikations-
_und_Medienwissenschaft/KMW_I/Working_Paper/Dunkelau___Leuschel__2019__Fairness-
Aware_Machine_Learning.pdf and am happy to see that it acknowledges (though not strongly enough imo) that
"fairness" is a causal notion, and that DAGs plus counterfactuals are needed to make sense of it. #Bookofwhy

1.9.20 7:28pm - As a Special Anointer of saints, I hereby commit the next sainthood to the Editor of the first
econometric journal to invite a survey paper on "graphical models in econometric". Spread the word. Candidates
may include previous saints. #Bookofwhy

1.9.20 7:03pm - The discussion on "What is DL?" reminds me of identical discussion (among Bayesians) on "Who is
a Bayesian?" ending up with a dead-end: "Do it! Dont ask!". Good for justifying what you have been doing, not so
good for science, as I confess here https://ucla.in/2nZN7IH #Bookofwhy

1.9.20 3:27am - (Replying to @AndrewLBeam and @ylecun) The division of labor between CI and DL is
mathematically defined by every CI exercise. See https://ucla.in/2HI2yyx. Each probabilistic expression that
appears in the derived CI estimand calls for a DL exercise Data Fusion entails a similar philosophy
appears in the derived CI estimand calls for a DL exercise. Data Fusion entails a similar philosophy
https://ucla.in/2Jc1kdD

1.9.20 2:33am - (Replying to @_mb46_ and @ylecun) Your simple definition of DL "using ANNs to do stuff" is too
broad for me. Is a computer chip an ANN? It is certainly a "network" of wires and gates. What does it take to
qualify as "neural" once we allow "artificial". #Bookofwhy

1.9.20 2:27am - (Replying to @tangled_zans) What for you may seem "Politics" is for me "moral imperative". My
take: I am a student of Middle East history and a long-time peace activist towards "two equally legitimate
states, for two equally indigenous peoples".

1.9.20 2:06am - (Replying to @tangled_zans) The silencing takes the form of protecting this woman from
criticism, for fear of being called "Islamophobe". I feel I have earned credibility as a fighter against
Islamophobia to articulate what others see but are afraid to condemn.

1.9.20 1:46am - (Replying to @_mb46_ and @ylecun) Surely "considerable differences" exist. But are they
judged to be fundamental, or merely temporary limitations of current DL techniques? If the former, then we
need indeed to better understand what the the definition is of DL. #Bookofwhy

1.9.20 1:31am - (Replying to @tangled_zans) Random strangers have immense influence on me, but I can't
understand what you see wrong in me alerting readers to the danger that this woman poses, through her hate
speeches. It is my responsibility to share my analysis with readers who, silenced by PC, do not dare speak out.

1.9.20 12:58am - (Replying to @zaffama and @ylecun) I am not sure DL advocates would be satisfied with your
modest definition. They would probably wish to extend it to include "every success story in the next century".
But I should not speak for them. Curious: "What is NOT DL?" #Bookofwhy

1.9.20 12:47am - (Replying to @_mb46_ and @ylecun) I believe it was #tdietterich who defined DL as "the
science of intelligence systems". https://twitter.com/yudapearl/status/1210552584680796161 And the answer
to "Are ANN enough" is trivial: "Of course! Because we know that Organic NN are enough." I am waiting for a less
ambitious clarification of what DL is.

1.8.20 11:56pm - (Replying to @tangled_zans) The "bigger picture": To assure descent readers that I share their
disgust with the racism of certain Zionophobes, that the educational origin of this aberration is explained, and
that it is OK to speak against it. No INSULT here, since Zionophobes are proud of their disorder.

1.8.20 11:32pm - I wish I could comment on the question "Is DL enough?" But I can't, because we do not know
what DL is. Some define it as an aspiration to emulate intelligence. Others, eg @ylecun , define it as a specific
NN implementation. We are waiting for clarification. #Bookofwhy

1.8.20 11:10pm - Asking a Zionophobe for logical consistency is like asking a snake to walk on two. According to
Mogadishu-born Ayan Hirsi Ali, the only semblance of consistency in her (& Omar's) school curriculum in
Mogadishu was the consistent blaming of Jews for bad grades and water shortage.

1.8.20 10:46pm - (Replying to @sherrirose and @RhubbBstat) Muchos Congratulationes. #Bookofwhy

1.8.20 10:40pm - OOPPS. I've just answered it here:


https://twitter.com/yudapearl/status/1215157258939596800
ttps: tw tte .co yudapea l status 5 5 58939596800

1.8.20 10:24pm - It's a bit more that "offsetting paths". Consider Y=f(X,U)= 1 iff X=U, and let X and U be two fair
coins. No offsetting paths in X-->Y<--U. Yet X||Y and Y||U. Statistical magic. #Bookofwhy

1.8.20 12:28am - Just bumped into this wonderful Ted Talk of @harari_yuval https://youtube.com/watch?
v=nzj7Wg4DAbs from whom I've learned about the origins of counterfactuals and imagination, and where I've
found a connection between the structural theory of counterfactuals and futuristic robotics. #Bookofwhy

1.7.20 11:00pm - (Replying to @EpiEllie @ruschenpohler and @AndrewJDBell) I would dare say that to ensure
generalizability under changing location or time we need to postulate a data-generating-model of some kind. I
don't think we can get by with frequency information alone. #Bookofwhy

1.7.20 7:26am - (Replying to @carlgieringer) And how is opposing "killing frogs" a victory for commonsense? The
charge varies from year to year, and is becoming more "benign" and "universal". Who does not love frogs? BDS
scores it victory from the music "Israel is on trial" and the megaphone, not the libretto.

1.7.20 6:55am - Am. Hist. Assoc. scored victory for commonsense, but the BDS circus will continue. Next year
the resolution will call on Israel to stop killing frogs on Tuesdays. Do they? Questions later! First listen to the
sweet music of "Israel killing" - BDS symphony #99 in E-flat major.

1.7.20 4:44am - Clarification. MR, IV, and conditional-IV methods belong to bullet 5 (More elaborate policy
evaluation.) These are available in DAGGITY, if I recall correctly. It is important to emphasize that DAGs should
be used to bestow defensibility upon these methods. #Bookofwhy

1.7.20 12:26am - (1/ ) Comments on "Use of DAGs". Commending you on an important survey. I was worried that
DAGs will be introduced too narrowly, but the route of "data generating mechanism" avoids misconception that
gets Epi people into trouble when talking "treatment-assignment" jargon.
1.7.20 12:26am - (2/ ) Readers noticed that the papers selection criterion was not clearly defined. E.g., did you
include papers where DAGs were used merely to communicate context? or was the causal component necessary?
I would suggest the following taxonomy of DAG usages:
1.7.20 12:26am - (3/ )
1. Communicate contextual information
2. Communicate and defend causal assumptions
3. Selection of covariates for adjustment
4. identify testable implications.
5. Elaborate policy evaluatioin
6. Transportability and selection bias.
7. Missing data
1.7.20 12:26am - (4/4) I was surprised to find no papers in bullets 4-7 Dont people ever test their models? Is Epi
still in the adjustment era ? Much work for educators.
Overall, a useful panoramic view of a field gone scientific. #Bookofwhy

1.6.20 6:29pm - (1/ ) I can't wait to see the faces of our two smearing experts when they find that the person
whom they labeled "conservative" "religious" "blinded by hatred" etc. etc. is in fact a (1) devout liberal, (2) a
voting registered Democrat, (3) a devout atheist, (4)award winner for
1.6.20 6:29pm - (2/ ) anti-hate programs.(5) active writer on Middle East history. What these smear peddlers
could not stomach was my criticism of the untouchable saintly Rashida Tlaib. I therefore repeat what I know
about her. She has been spewing hate, deceit and outright racism since being
1.6.20 6:29pm - (3/ ) elected. For readers who respect my sincerity, domain knowledge and ability to read
beneath the surface, I have also commented on her sweet anti-war statement after soleimani death, and noted
that she cannot afford to tell her audience what kind of mass-murderer Suleimani was
1.6.20 6:29pm - (4/4) because a large segment of her constituency sympathizes with Suleimani's agenda of
weakening the US and wiping out Israel -- two holy aims that justify all means.

1.6.20 7:25am - (Replying to @timminglab) Moreover, the more you get to know colleagues at Ivy league
universities, the more you appreciate the tribal games you did not have to play being yourself.

1.6.20 7:15am - Or, at the very least, @SenSanders , stand up like a Mench and answer questions about what
chunks of Sarsour's ideology you DON'T share. I am willing to do the interview, since none of the reporters
behind the microphones is willing to touch on this issue.

1.6.20 5:45am - (Replying to @KordingLab) Thanks for the update. Interesting interpretation. The sequence of
presenting a method first and asking for veracity second is what triggered my simplistic interpretation.

1.6.20 4:51am - (Replying to @KordingLab) Is there a new "Athey tutorial" in circulation? The last I have seen
said essentially: "Whatever economists do is automatically causal" (my humble interpretation). Is there a re-
freshener? #Bookofwhy

1.6.20 1:50am - New sounds: "naive, ignorant, unschooled, pasty". What makes some academics so allergic to
new information from a new perspective? Put differently, what makes Rashida Tlaib more knowledgeable,
educated, revered or morally compelling than humble me? Ready for the first test? is automatically causal" (my
humble interpretation). Is there a re-freshener? #Bookofwhy

1.5.20 9:13pm - (Replying to @KordingLab) Agree! Most of the discussed issues are rung-1. Still, they all
eventually lead to questions of policy and/or explanation, and the paper gives the impression that DL is well
prepared to tackle these issues too. Your honest opinion: Is it? #Bookofwhy

1.5.20 8:32pm - (1/ ) Fair question! My aim was not to "interpret" Rashida's statement, but to express my
reaction to her rhetoric in light of what she has been doing in the past. Specifically, in light of the hate, deceit
and outright racism she has been spewing since being elected.
1.5.20 8:32pm - (2/ ) 3/3 which do not tell us who Suleimani was and why Rashida can't speak about it. She
can't because she would lose her support base, a big chunk of which sympathizes with Suleimani's agenda of
weakening America and wiping out Israel. Sad. But rhetoric won't change reality.
1.5.20 8:32pm - (3/3) which do not tell us who Suleimani was and why Rashida can't speak about it. She can't
because she would lose her support base, a big chunk of which sympathizes with Suleimani's agenda of
weakening America and wiping out Israel. Sad. But rhetoric won't change reality.

1.5.20 8:01pm - (Replying to @bariweiss) Kol Hakavod

1.5.20 6:21pm - (Replying to @fayyazhere)


1. The limitations of rung-3 reasoning are that it relies on information about the functions behind the arrows.
Lacking it, gives us bounds, not point estimates.
2. It is useful to assume that my ML colleagues can extract the optimal predictive information
available.#Bookofwhy

1.5.20 8:29pm - Beautiful 1-pager. Now I see my math teacher winking to me: Should I be pedantic and spoil
the beauty? No. I'll let it go. Just a tiny spoiler: All "statistically dependents" in bullet 5 should read "likely
statistically dependents". Why? Some other time. #Bookofwhy
1.5.20 4:52am - (Replying to @ktmud) What you are saying is: It's TOUGH to see someone who has spent lots of
time studying the middle east conflict, and worked day and night spreading love and understanding for whole
mankind express ideas that differ than MINE. So I'll dismiss him as "blinded by born identity".EASY!

1.5.20 3:48am - This paper https://arxiv.org/pdf/1906.05433.pdf encourages ML folks to tackle problems of


Climate Change, yet neglects to mention the unique causal and counterfactual tools needed for the challenge.
In particular the evaluation of sufficient and necessary causes, eg: https://ucla.in/2wAMXeX

1.5.20 3:27am - (Replying to @TheDavidSJ) If I were the murderer of Rina Shernav (17) I'd say: "Job worth
repeating! Even when we kill a girl in the heart of Tel Aviv, we now have someone in the US Congress to blame
"Israeli occupation" for the crime. It's a new dawn!" [Note how she did not say "equal rights to Israelis"

1.5.20 1:02am - (Replying to @TheDavidSJ) If she truly cares for "loss of innocent lives", including those
hundreds of thousands of innocent lives destroyed by Suleimani, Congresswoman Rashida Tlaib has had ample
opportunities to weave it into her speeches since his departure. She has not done so & we know why. She cant!

1.4.20 11:41pm - Congresswoman Rashida Tlaib "cannot stay silent". She is deeply shaken by the unexpected
loss of an ideological idol, mass murderer Qasem Suleimani. Her sudden interest in "innocent lives...at home
and across the globe" now humors her American voters in Michigan.

1.4.20 8:55pm - This is one of the most futuristic, yet sensible futuristic conversations I have heard:
https://futureoflife.org/2019/12/31/on-consciousness-morality-effective-altruism-myth-with-yuval-noah-
harari-max-tegmark/ I'm only surprised they treat consciousness as a mysterious inaccessible concept, laden
with pain and suffering. #Bookofwhy

1.4.20 4:30pm - Glad we agree. I believe, however, that it goes deeper than "origin". The refusal of the PO
community (including Econs and Harvard) to adopt the intervention/counterfactual distinction represents
refusal to accept structural models as a basis of scientific thinking.#Bookofwhy

1.4.20 3:29pm - (Replying to @jrgptrs @3ieNews and 13 others) If I were able to contribute to your blog, I would
ask: "Why quote Deaton who laments careless handling of #externalvalidity instead of promoting methods that
properly handle the problem?" For example: http://cs.ucla.edu/~kaoru/misc.html or https://ucla.in/2JfdQcv.
#Bookofwhy

1.4.20 2:56pm - I highly recommend Primer: https://ucla.in/2KYYviP, which has a wealth of beautiful
examples. Plus, the recent survey by Paul and Elias, geared to economists:
http://cs.ucla.edu/~kaoru/misc.html #Bookofwhy

1.4.20 6:02am - (Replying to @attilacsordas) Let's discuss by email.

1.4.20 5:54am - (1/ ) What is "counterfactual prediction"? The quantity estimated in:
https://pdfs.semanticscholar.org/94d7/652764a6cce9cd25c8c5bf8c6ed6785c42b0.pdf turns out not a Rung-3,
but a Rung-2 interventional expression. To avoid confusion & false expectations, I recommend that ML folks
adopt the hierarchy in #Bookofwhy. We need to distinguish
1.4.20 5:54am - (2/2) "effects of causes" from "causes of effects" (see https://ucla.in/2BTa9Lc and
https://ucla.in/2N92rBM) else ML students will come to believe that DL has found a magic way of inferring
counterfactuals from data, thus risking an inevitable disappointment.
1.4.20 4:02am - I welcome this new review of #Bookofwhy in American Mathematical Monthly
http://bayes.cs.ucla.edu/WHY/Pearl-Mackenzie_American-Mathematical-Monthly.pdf. It is written by a hard
core statistician unafraid to confess the nature of statistics discomfort with causation.

1.4.20 3:34am - Rabbi Sacks is always insightful and to the point. I would only add that the irrational animosity
towards Israel originates from the same swamp: failure to accept a cohesive thriving society not based on class
struggle. For elaboration, see https://ucla.in/2SxjgWH

1.4.20 12:12am - (Replying to @attilacsordas) What did you find incomplete in the formal definitions of
"sufficient cause" and "necessary cause" as given here: https://ucla.in/2lEOJtO or here:
https://ucla.in/2G2rWBv I find them quite satisfactory. Plato.stanford is outdated. #Bookofwhy

1.3.20 6:01am - (Replying to @jonathanborows2) What is "clickbait marketing"? I love the sound of it.

1.3.20 5:58am - (Replying to @mendel_random) I noted, with sadness and humor, that Johnson will have hard
time finding qualified applicants for those jobs. (ps. where did you dream up "faithfulness"?) #Bookofwhy

1.3.20 12:50am - I see a much more immediate risk, that @downingstreet will barely find any applicants. I can
hardly name a dozen who would qualify, can you? Not because it is hard to explain, but b/c DS folks are too
busy doing what they are so good in doing. #Bookofwhy

1.3.20 12:39am - (Replying to @eliasbareinboim) I am trying to interpret the expression "using another DL
model". I thought DL prides itself on being "model free". Perhaps you mean "using the estimation prescribed by
the CI model" ? #Bookofwhy

1.2.20 11:19pm - (Replying to @UlrichJunker and @sahilsingla47) Thanks

1.2.20 11:18pm - The wrong link, to the "counterfactual prediction" paper was not meant to be an
endorsement. I am still not sure the authors are doing counterfactuals in the Rung-3 sense. There is a tendency
among OP folks to call interventions counterfactuals; too bad ML folks follow #Bookofwhy

1.2.20 10:31pm - Correcting link to the video: https://israelunwired.com/san-diego-state-university-invites-


jew-hating-muslim-to-speak/ Though it will never correct the mentality of the department that invited this
pervert to speak in an "Institute of higher learning". We ARE responsible.

1.2.20 10:21pm - A must watch VIDEO, from San-Diego State University:


https://pdfs.semanticscholar.org/94d7/652764a6cce9cd25c8c5bf8c6ed6785c42b0.pdf And what are we doing,
as refined academicians, to prevent this poison from igniting our campuses? Have you called out the Chair of
that idiotic department? History records both action and apathy.

1.2.20 3:01am - (1/ ) Reflecting back on our DL/CI discussion with @tdietterich, @ylecun, @GaryMarcus etal, we
should note that CI folks are invoking the same scientific argument that has brought DL its fame and success.
Back propagation and subsequently gradient-based optimization were inspired by
1.2.20 3:01am - (2/ ) machine known in the 1970's. As we go to higher levels of cognition, eg causal inference,
it would make sense to ask (and imitate), how the mind preforms this mode of inference, that is, what
cognitive
1.2.20 3:01am - (3/ ) templates enable it to manage association, interventions and counterfactuals
simultanewously, compactly and swiftly. I know of only one such template: Structural Causal Model. Thus,
instead of hoping that CI templates would pop up spontaneously from DL, let us ask first
p g p p p p p y
1.2.20 3:01am - (4/4) what should they look like? How should we represent them, and what we would be able to
do once we get them. #Bookofwhy

1.2.20 12:36am - Glad to see that subtitles were added to my podcast with Lex Fridman
https://youtube.com/watch?v=pEBI0vF45ic I can now understand my Hebrew accent. #Bookofwhy

1.1.20 3:53pm - (Replying to @AdanZBecerra1 @Lester_Domes and @nature) I was called "disruptive" in 1st
grade of school, sent home, and was back only three weeks later, after seeing an oppressive psychiatrist who
vowed that I was ready. Evidently, I wasn't. #Bookofwhy

1.1.20 2:49am - (Replying to @maximananyev) How many Masters does it take to correct what some Metrics
refuse to do? #Bookofwhy

12.31.19 11:29pm - (Replying to @BariFaisal) I assume you refer to #Bookofwhy? My thoughts on empirical
research are summarized here: https://ucla.in/2mhxKdO, and on RCT, here: https://ucla.in/2JfdQcv. Warning:
Graph-shunning economists are depriving their students of ever combining experimental and observational
studies.

12.31.19 7:49pm - (Replying to @Isalomaki) Yes, "Be" means "We", not "Me, Me, Me".

12.31.19 4:56am - (Replying to @HenMazzig)` Isn't it what congress-woman Tlaib has been howling all along?

12.31.19 4:45am - (1/ ) My New-Year greeting this year is again in a form of a poem/song, through which I wish
all readers:
"Make it be, Make it be, Pleading: "Make it be", All for which we pray Make it be."
In 1973, Nomi Shemer's, set out to translate the Beatle's song "Let it be" into Hebrew. But
12.31.19 4:45am - (2/ ) then the Yom Kippur war broke out, and what came out her pen was a prayer that
swept the country with immense yearnings for peace. It still resonates today in every street and on every
occasion:
The summer ends, the journey's over, Let them please at last come home...
12.31.19 4:45am - (3/3) All for which we pray: "Make it be".
You can hear the entire song about 7:10 minutes into this video. https://youtube.com/watch?v=PvByNKs1bmw
Happy New Year! and "Make it be!"
#Bookofwhy

12.30.19 1:55pm - (Replying to @AwokeKnowing @ylecun and @LeCun) Sure. The math of woodwork is not too
complex, so I dont see why it cannot be expressed as output of the math of metallurgy, given the unique
metallurgical properties of wood. #Bookofwhy

12.30.19 1:39pm - (Replying to @vatsal_maru @ylecun and @lexfridman) Solving ONE toy problem in causal
inference is worth 1000 conversations, interviews, debates & what-have-you about what others do. Take, say,
Simpson's paradox and analyze it through the math of "gradient-based optimization", but stick to the toy!
#Bookofwhy.

12.30.19 1:06pm - (Replying to @ylecun) Please listen carefully to what the carpenter is saying:
* Carpenter: Sure, you can do the nails, I'll do the woodwork, together we can build a house.
He should also add that the math of metallurgy is not very helpful in carpentry, a new math is needed.
#bookofwhy
12.30.19 6:38am - (Replying to @ElliotMalin) Wishing you and your lucky bride many years of happiness. And
may your commitment to Israel be a constant inspiration to you and to your children.

12.30.19 4:48am - (Replying to @PolandCherieM and @oralassila) In #Bookofwhy I was just summarizing half a
century of zero-progress in external-validity research, stuck for lack of language. DL, in contrast, has shown
immense progress, but lack of language is still a fundamental impediment, and is beginning to show its face.

12.30.19 4:06am - A strange thought on what DL can and cannot do.


` If "X can do Y" whenever we need X to do Y, then transistor experts "can do" AI, and nail-makers "can do"
furniture.

12.30.19 3:33am - (Replying to @ArnaudMegret) And naive me assumed that every ML researcher read
#Bookofwhy, with just a few hard cases who remained unconvinced. No wonder @ylecun 's followers keep
asking: "Can't we just "extend" DL a bit to cover causality?" I don't believe any reader of #Bookofwhy would ask
such a question.

12.30.19 3:13am - (Replying to @ArnaudMegret and @LeCun) This paper summarizes our latest findings on how
much information one can get out of missing data, both MAR and non-MAR https://ucla.in/2LdEjZW. As you
point out, the question is causal, and it is hard to imagine how any statistical analysis can obtain these results.
#Bookofwhy

12.30.19 2:05am - (Replying to @AngeloDalli @ylecun and @LeCun)


* Metallurgist: My nails just did a table, they can do a house?"
* Carpenter: Sure, you can do the nails, I'll do the woodwork, together we can build a house.
* Metallurgist: No! The nails should do it. They can perhaps be extended to handle the woodwork.
#Bookofwhy

12.29.19 4:13pm - (Replying to @ylecun and @LeCun) If we have interventions, we do not need CI, we just fit
the observed data (eg by DL) and we're done. CI is needed to answer questions that cannot be answered directly
from data, eg "Find the effect of X, when I only have interventions on Z". A new math is needed here.
#Bookofwhy

12.29.19 6:48am - (Replying to @ngutten @jsusskin and @LeCun) Sure! Everyone (except perhaps Yann)
understands that the goal of DL is to fit a function to 10K+vars, and that CI has a totally different goal
(requiring totally different math): to answer causal questions about reality ASSUMING that DL folks succeed in
their fitting efforts.

12.29.19 6:36am - Today's madness could not be more vividly portrayed than through this record of @KenRoth ,
the Orwellian watchman of (the world's) Human Rights.

12.29.19 6:27am - (Replying to @SussexFriends) Is this moral pervert still talking? After all that he did to
embolden hate against Israel? May my last words to him be: Nes Gaddol Haya Po (in UK).

12.29.19 5:33am - (1/ ) (Replying to @ngutten @jsusskin and @LeCun) We should discourage researchers from
thinking that 'to do CI with regression requires an extra step'. This would lead them into believing that doing CI
with regression is not an oxymoron, that the extra step is trivial and would pop up spontaneously if only they
study more
12.29.19 5:48am - (2/ ) (Replying to @yudapearl @ngutten and 2 others) and more regression. It won't. On the
contrary. Today, the best experts on regression are the least likely to understand what is needed to do CI. I
co t a y. oday, t e best e pe ts o eg ess o a e t e least l ely to u de sta d w at s eeded to do C .
foresee the same fate awaiting students of DL. My advice: start afresh, with CI, and resort to DL only if needed.
#Bookofwhy

12.29.19 3:59am - I've hoped to celebrate the 8th night of Chanukka on a joyful note. In the wake of the Monsey
stabbing, I must retreat to what I wrote in 2009 (in the WSJ): The normalization of evil begins with the
normalization of hate, and our leaders are not watchful: https://tw.forumosa.com/t/daniel-pearl-and-the-
normalization-of-evil/49441

12.29.19 2:32am - (Replying to @Imamofpeace) That's essentially what Linda Sarsour told Bernie Sander. He
believed her, assuming she represents the next wave of votes. The consequences are in the streets of NYC.

12.28.19 5:38pm - (Replying to @KordingLab @GaryMarcus and 7 others) As you can see here:
https://ucla.in/2mhxKdO, I love everything about the economists tradition, except their latest betrayal of that
tradition, and their cultish avoidance of tools that operationalize that tradition. See eg
https://ucla.in/2ZcRpRq. #Bookofwhy

12.28.19 5:14pm - (Replying to @KordingLab @GaryMarcus and 7 others) I am sincerely trying to understand
what it IS, or what it means to YOU, and you are slapping me with "denying the validity". Unfair! #Bookofwhy

12.28.19 5:08pm - (Replying to @KordingLab @GaryMarcus and 7 others) For the sake of amazing clarity I have
been trying to avoid the term "different flavor", because it has been misused to create the impression that
there are differences in substance where there are none. @Bookofwhy

12.28.19 4:06pm - (Replying to @KordingLab @GaryMarcus and 7 others) The arguments are perfectly solid. I am
just not clear what econ. practitioners call "OV bias equation", and how they use the equation to reach new
heights. #Bookofwhy

12.28.19 3:54pm - (Replying to @edwardsjk @EpidByDesign and 11 others) I take it that the idea is to "combine"
observational and experimental studies, rather than exclude one in favor of the other. Are the results of
"combining" different than those of "fusing"? As in eg https://ucla.in/2Jc1kdD ?#Bookofwhy

12.28.19 3:39pm - (Replying to @KordingLab @GaryMarcus and 7 others) A sincere question: What is the
"omitted variable bias equation"? How does one realistically "apply it to real world reasoning?" I honestly havn't
seen it applied, but I might have missed the equation title. #Bookofwhy

12.28.19 3:24pm - (Replying to @KordingLab @GaryMarcus and 7 others) I would love to join you in admiration
of "econ style CI" if I only knew what it is, what it is "complementary" to, and how it "beautifully complements"
pieces that are missing elsewhere. Curious. #Bookofwhy

12.28.19 6:14am - (Replying to @oralassila) It is not an implicit prediction as much as it is an explicit concern. If
we continue to invest most of our resources in the science of nuts and bolts, engrossed and intoxicated by its
successes, our ability to build engines will be delayed. #Bookofwhy

12.28.19 4:16am - (Replying to @PolandCherieM) Thanks for posting this illuminating paper on the history of
regression. Biased by the lens of causation we emphasized the odd fact that Galton sought a causal explanation
to Darwin's theory and ended up abandoning causation to the mercy of correlation. What a miss! #Bookofwhy
12.28.19 3:55am - In addition to introducing new methods of generalization, this paper also identifies and
solves a new and important problem of learning causal models from both observations and interventions. Highly
recommended. #Bookofwhy

12.28.19 3:14am - To all who celebrate the 6th night of Hanukkah with my friends and colleagues in Israel, I am
sharing a song that we used to sing since Kindergarten:
A miracle did not happen to us, We have not found a vessel of oil, We carved the rock till we bled, And there
was LIGHT!

12.28.19 3:03am - Variation on a painful, left-leaning theme: The only thing less politically convenient for
leftists to recognize than left-wing antisemitism is left-wing Zionophobia and Palestinian elimination-ism.

12.28.19 2:47am - I am retweeting this reply, for it crystallizes my position in the latest conversation on the
relationships between DL (deep learning) and CI (causal inference) with @tdietterich , @ylecun , @GaryMarcus ,
@rodneyabrooks and significant others. #Bookofwhy

12.28.19 2:06am - (Replying to @moultano @bradpwyble and @GaryMarcus) It is extremely difficult to make
precision nuts and bolts, and it is by no means "trivial". But the science of making nuts and bolts is not the
science of making engines, and a community engrossed in the former will be limping on the latter. #Bookofwhy

12.27.19 8:08am - (Replying to @kudkudakpl @GaryMarcus and @moultano) A different approach is needed. Now
it is my turn to ask "out of curiosity": Why do you keep saying "doing DL"? Is it because you enjoy approximating
functions? Or because you want to ride the DL hype? Seriously, I am truly curious. #Bookofwhy

12.27.19 8:01am - (Replying to @databoydg and @tdietterich) Same can be said about linear algebra. Current
"approaches" are insufficient but, eventually, everything AI will achieve in the future will use linear algebra to
some extent, So, linear algebra will spontaneously spawn intelligence. Tough? Don't bet against DL. #Bookofwhy

12.27.19 7:39am - (Replying to @kudkudakpl @GaryMarcus and @moultano) Good question. What I mean by
"mathematical impossibility" is w/o causal (extra-data) assumptions of some kind. The paper u cite makes such
causal assumptions and the study of what assumptions are needed has been the central focus of CI research in
the past 3 decades. #Bookofwhy

12.27.19 7:29am - (Replying to @databoydg and @tdietterich) Why assume anyone pretends they can't exist
together? I just wrote that I have always assumed DL is almighty and can perfectly approximate any function
and, so, it can exist together with CI, approximating galore all functions that CI says should be approximated.
#Bookofwhy

12.27.19 7:05am - (Replying to @kishkushkay)` When was that, Kay?

12.27.19 6:55am - (Replying to @memosisland @GaryMarcus and 2 others) @tdietterich wrote that DL's aim was
to develop "science and engineering of intelligent systems." I removed "engineering" from his quote, because I
want to focus on the "science", where CI has made some modest contributions. #Bookofwhy

12.27.19 6:41am - (Replying to @moultano and @GaryMarcus) My expectations about DL never ended up wrong,
because I have always assumed DL is almighty, and can approximate perfectly every function it is given,
regardless how complex. But for DL to spontaneously produce CI, it's not implausible, its a mathematical
impossibility #Bookofwhy
12.27.19 5:26am - (1/ ) There is more to @GaryMarcus argument than taking credit. If the goal of DL was indeed
as ambitious as you describe it ("science of intelligent systems") than it failed miserably by not realizing the
barriers to rising above Rung-1 (ie., function fitting). If, however, the
12.27.19 5:26am - (2/ ) goal was more modest, say to perfect the art of function fitting, then there is no room
for claims that CI will eventually become part of the DL umbrella. The distinction between "CI will be under the
DL umbrella" vs. "DL is part of the CI umbrella" is not about credit. It's
12.27.19 5:26am - (3/ ) about the agenda of AI research in the next decade. Saying "CI is merely an extension of
DL methodology" amounts to telling educators, funders and the public: "Let us continue doing what we have
been doing all along, CI will emerge organically." It won't! And the result:
12.27.19 5:26am - (4/ ) AI progress will be slowed, and the disappointment will be painful. In contrast, saying:
"DL is a component of CI" amounts to telling funders and educators: "We must educate a new breed of DL
researchers, capable of tackling new challenges. Things as usual just won't work!"
12.27.19 5:26am - (5/5) I believe the second alternative is more constructive. "Things as usual" leads to "Things
as usual" unless there is a paradigm change & a community awakening to the limitations of "Things as usual".
Luckily, we can formalize those limitations. Let's overcome them! #Bookofwhy

12.27.19 3:46am - Happy new year! To all readers who are hoping for one, me included. Glad Israel remains the
world's beacon of moral clarity, something the world has forgotten exists, and finds it hard to stomach. Happy
new year!!

12.27.19 3:15am - Thanks for posting this interesting paper. I've never imagined that causal graphs will find
applications in such esoteric fields as Visual Dialogue. But I am known to suffer from occasional lapse of
imagination. #Bookofwhy

12.27.19 2:38am - (1/ ) (Replying to @tdietterich) Disparity in levels is one source of the confusion. You
congratulate DL engineers for their achievement in approximating such complex functions, and I say: it is ONLY
"curve fitting". We are both right! In the past 3 decades I took function approximation to be a done deal.
12.27.19 2:45am - (2/ ) (Replying to @yudapearl and @tdietterich) How? Every time my math ended up with a
probability expression, say E(Y|x,z), I labeled it "solved", namely, DL folks will find a way of approximating this
function from data. Done deal! Let's look at the more challenging task of reducing our causal questions to
probabilities!
12.27.19 2:56am - (3/ ) (Replying to @yudapearl and @tdietterich) Today, that we know almost everything
about what can or cannot be reduced (meaning what DL can or cannot do) we find DL folks awakening to, and
dismissing the main challenge with: "We can fit everything!" It's true, they only need someone to tell them what
to fit.#Bookofwhy

12.27.19 2:19am - (Replying to @eyad_nawar) If we have 1&3 you do not need controlled experiment. If we
don't, we can't move beyond "curve fitting". This is something DL folks are beginning to realize, in their own
unique pace. #Bookofwhy

12.27.19 2:05am - (Replying to @steventberry @PHuenermund and @ChrisAdamsEcon) The confusion about
"what to control for" is symptomatic of a fundamental methodological neglect econometric is currently
recovering from. I compiled a few more symptoms in: https://ucla.in/2mhxKdO. For survey and remedy, see:
https://arxiv.org/pdf/1912.09104.pdf. #Bookofwhy, @ben_golub

12.26.19 6:54am - Arnold Roth is both a friend and a grieving comrade. His daughter Malki and my son Daniel
were victims of the same wave of hate that swept our planet in 2001-2002. Please listen to his plea. Judea
Pearl

12.26.19 6:17am - (Replying to @guillefix @maier ak and 5 others) What is not available to DL training is not
12.26.19 6:17am (Replying to @guillefix @maier_ak and 5 others) What is not available to DL training is not
"data" but "information" on how to get the appropriate data. For example, that "age" is a confounder, hence,
estimate the age-specific effects, not aggregate effects. See Simpson's paradox, #Bookofwhy p.211, or
https://ucla.in/2Jfl2VS

12.26.19 6:05am - (Replying to @bioinfochat @tdietterich and 3 others) Agree, "regression is here to stay" does
not imply "DL is here to stay", I should have said "Some form of function approximation is here to stay", which is
tautological and elevates DL to a special status it rightly deserves. #Bookofwhy

12.26.19 5:44am - The Middle East Studies Association (MESA) sent a letter to President Trump in which they
decide how Jews should define themselves. The verdict: Jews are not a people of common national origin.
https://mesana.org/advocacy/committee-on-academic-freedom/2019/12/12/letter-criticizing-president-
trumps-executive-order-on-combating-anti-semitism Their logic betrays their wisdom --a true academic gem.
12.26.19 5:44am - (2/ ) The logic goes: Since not ALL Jews share a common national origin, therefore, NO Jew,
including the vast majority who DO, should be treated as such. May the 5th candle of Hanuka remind those
academic "experts" how Jews have been defining themselves for the past 2200 years.

12.26.19 3:06am - (Replying to @LuisMateusRocha and @LeCun) Thanks, but this is over my head. Think about a
naive AI fellow who wants to implement one of those "non-inductive" machine. Where should she start?

12.26.19 2:55am - Who said John Searle will never pass the Chinese Room Turing Test? See #Bookofwhy page 38-
39.

12.26.19 2:35am - (Replying to @LuisMateusRocha and @LeCun) I'm not familiar with the technical details of
this literature. Can you Tweet what information "anticipatory system" has that #induction machines do not,
where it gets it, and how it represents it. Thanks #Bookofwhy.

12.26.19 2:02am - (1/ ) (Replying to @tdietterich @Plinz and @GaryMarcus) The point is that the term
"functional approximation" is more informative to the engineer than an account of flexibility. It tell the
engineer to stop asking how good the approximation and start asking whether she approximated the right
function, and how to get the correct
12.26.19 2:12am - (2/ ) (Replying to @yudapearl @tdietterich and 2 others) input-output pairs. It also means
that @ylecun statement "DL is here to stay" is tautological, akin to "programming is here to stay". What counts is
whether we write the right program. And when we do, we don't label it "Turing Machine" #Bookofwhy

12.25.19 4:39pm - (Replying to @ylecun @BethCarey12 and 2 others) Just for the music: Replace DL with
"regression", and see how things still stick, and the music is still flowing soundly and harmoniously, even more

so, for we know so much about "regression" and its limitations. BUT: Is "regression" is the right scientific model?
NO! #Bookofwhy

12.25.19 2:30pm - (Replying to @tdietterich @Plinz and @GaryMarcus) You'r speaking from a function-builder
perspective. But from the end-user, input-output, what have we got? A function that approximates the actual
relationship between input and output. It need not capture the pain, labor, flexibility and sophistication of the
builder. #Bookofwhy

12.25.19 5:19am - (Replying to @maier_ak @hardmaru and 4 others) When we talk "algorithms" everything
seems to be doable in DL or NN etc. But when we talk about "information" the limitations become clear --
information cannot be created by processing data, it is either there or not. As Cartwright said: "No causes in, no
causes out." #Bookofwhy
12.25.19 3:25am - (Replying to @recursus @basorot and @LeCun) Fine, but "working out the entailments of a
model" requires a data structure to represent that model. Once you commit to such representation, Bingo! You
are doing CI! And if you still want to keep your memebership in the DL club, fine, we wont excommunicate you.
#Bookofwhy

12.25.19 2:15am - (Replying to @recursus @basorot and @LeCun) Sorry to disappoint, but the "model" cannot
unfortunately "be found through a generate-and-test process", b/c there is nothing in the training data that can
refute a bad model. It can only be tested when we make a bad policy decision and, then, it's too late.
#Bookofwhy

12.25.19 1:24am - To all readers who celebrate it, Merry Christmas from me and from my family in Jerusalem. I
fell in love with Christmas the moment we arrived to US, 1960, primarily b/c it reminded me so much of Hanuka
as it was celebrated in my home town, with songs and lights from every window.

12.25.19 12:56am - (Replying to @hardmaru @ngutten and 3 others) These proposals reinforce my conclusion
that some DL folks have not internalized the Ladder of Causation and its implications. One of these says: No
matter how you squeeze the data, no matter how smart the squeezing, you can't get causal information out of
it. Sorry #Bookofwhy

12.25.19 12:11am - (Replying to @ngutten and @jsusskin) Interesting viewpoint. Trouble is, regression is such a
small part of CI that to say "To do CI with regression requires an extra step" is almost like saying: "To do CI with
algebra requires an extra step". Its better to say "Do CI first, add NN if needed" @ylecun #Bookofwhy

12.25.19 12:11am - (Replying to @ngutten and @jsusskin) Interesting viewpoint. Trouble is, regression is such a
small part of CI that to say "To do CI with regression requires an extra step" is almost like saying: "To do CI with
algebra requires an extra step". Its better to say "Do CI first, add NN if needed" @ylecun #Bookofwhy

12.24.19 11:10pm - (Replying to @ylecun and @LeCun) And there is another reason why I want to convince YOU,
rather than Leon. I have noticed a day-and-night difference between folks who actually solved a toy problem in
CI and those who know all about how others do it. If you are among the former, Leon would do. #Bookofwhy

12.24.19 10:57pm - (Replying to @ylecun and @LeCun) And there is another reason why I want to convince YOU,
rather than Leon. I have noticed a day-and-night difference between folks who actually solved a toy problem in
CI and those who know all about how others do it. If you are among the former, Leon would do. #Bookofwhy

12.24.19 10:57pm - (Replying to @Plinz and @GaryMarcus) And what will be lost if we drop the "compositional"
and call it just "function approximation"?? or, God forbid "curve fitting"? (see footnote b in
https://ucla.in/2HI2yyx) #Bookofwhy

12.24.19 10:47pm - (Replying to @GaryMarcus) Great questions, @BaryMarcus! May I add a 5th question: What
will be lost to AI if we replace the term DL with "super-efficient way of doing regression" or simply "regression"?
A lot will be gained, b/c we know so much about what can and cannot be done w/ regression. #Bookofwhy

12.24.19 10:22pm - (1/ ) (Replying to @ngutten @jsusskin and @LeCun) Sure. A NN trained on specially prepared
data would capture the features that the preparation meant it to capture. How much preparation can RL do?
See Tweet: By deploying interventions in training, RL allows us to infer consequences of those interventions,
but ONLY those
12.24.19 10:22pm - (2/ ) (Replying to @yudapearl @ngutten and 2 others) interventions. A causal model is
needed to go BEYOND, ie, to predict consequences of actions not used in training, or even combinations of
needed to go BEYOND, ie, to predict consequences of actions not used in training, or even combinations of
actions used in training. Are NN useful in CI? Sure! Same as asking "is regression useful in CI?" Sure!. But is NN
different? #Bookofwhy

12.24.19 9:12pm - (Replying to @basorot and @LeCun) Wait, Wait! This is an interesting point!. You propose to
call the model "part of processing". Fine. Have you seen a DL system given a model and, after processing it with
the data, answer questions it could not answer before? This is a critical ingredient ML folks lack #Bookofwhy

12.24.19 9:03pm - (Replying to @jsusskin and @LeCun) If we go beyond the definition, we find Yann saying:
Don't say "DL doesn't do causal inference" when you really mean "a plain, supervised neural net does not
spontaneously discover causal relationships." The correct answer is: "DL just doesn't! Regardless of the NN!"
@Bookofwhy

12.24.19 8:51pm - (Replying to @jsusskin and @LeCun) If we go beyond the definition, we find Yann saying:
Don't say "DL doesn't do causal inference" when you really mean "a plain, supervised neural net does not
spontaneously discover causal relationships." The correct answer is: "DL just doesn't! Regardless of the NN!"
@Bookofwhy

12.24.19 8:23pm - (Replying to @basorot and @LeCun) If you cannot get answers for questions by processing the
data, and those questions ought to be answered by a sensible agent, it means that the sensible agent uses some
information that is not in the data itself. This extra-information is called "model", not "data."#Bookofwhy

12.24.19 7:01pm - This bold defense of deep learning says that some DL folks have yet to internalize the Ladder
of Causation, eg https://ucla.in/2HI2yyx. Can any DL-insider think of a good way to convince @Lecun that some
questions cannot be answered from data alone, no matter what? #Bookofwhy.

12.24.19 5:12am - (Replying to @HenMazzig) Well said. However, I was informed that Arafat started the erasure
of the Jewish history already in 2000, at Camp David. See https://jewishjournal.com/online/308874/chanukah-
our-trust-deed-to-history/fabricated . Also, the Palestinian claims are weakest when "indigeneity" is defined as
an intellectual-cultural state of mind.

12.24.19 4:43am - (Replying to @jrgptrs and @JustinSandefur) In 1987 David Freedman wrote "As others see us",
which had a soul-searching impact on social scientists. Econometrics is waiting for an insider to write such a
paper. I volunteer to play one of the "others", but the field is begging for a courageous insider. #Bookofwhy

12.24.19 2:27am - (1/ ) (Replying to @jrgptrs and @JustinSandefur) 1/ I perfectly understand the culture that
reinforces your arguments, but I beg to question its logic. The fact that "no referee requires us to use [the] Data
Fusion approach." does not negate the possibility that the methods used by that culture are grossly outdated. In
fact,
12.24.19 2:37am - (2/ ) (Replying to @yudapearl @jrgptrs and @JustinSandefur) readers on this Twitter post
have been decrying the insular, echo-chambered character of economics literature in general, and in the area
of external validity in particular. So, I am glad you are looking into this issue from a non-cultish perspective.
And, BTW, "Data Fusion"
12.24.19 2:42am - (3/ ) (Replying to @yudapearl @jrgptrs and @JustinSandefur) is not "my approach" nor is it
something you may choose to apply or ignore. The findings of "Data Fusion" analysis are universal, regardless of
how you choose to extrapolate across environments. The environments either permit extrapolations or not and,
ignoring the latter
12.24.19 2:49am - (4/ ) (Replying to @yudapearl @jrgptrs and @JustinSandefur) may lead to errors regardless of
what referees say or not say. I trust that, after looking into the state of art, you will join your colleagues'
efforts to elevate econometrics to the age of modernity. #Bookofwhy
12.24.19 12:14am - (1/ ) This paper introduces an interesting refinement of selection diagrams:
https://arxiv.org/ftp/arxiv/papers/1912/1912.04432.pdf If one is interested in transporting a single contrast
(say Risk Difference) and is willing to make parametric assumptions to match this aim, then selection diagrams
can be pruned
12.24.19 12:14am - (2/ ) so as to reduce the number of covariates involved. This refinement is similar in spirit
to that proposed in http://causality.cs.ucla.edu/blog/wp-
content/uploads/2019/12/cinelli_pearl_generalizing_experimental_results.pdf, which leverages knowledge of
how mechanisms interact to produce the outcome. #Bookofwhy
12.24.19 12:49am - (2/ ) (Replying to @yudapearl and @DavidHarrisAJC) This means that, seeing any such claim
opens up an opportunity to expose, with facts, figures and dates, the glaring recentness of the claimant's
historical heritage.

12.23.19 10:38pm - (Replying to @DavidHarrisAJC) And let's not forget that claims of "no Jewish historical tie w/
Israel & Jerusalem!" are fabricated for a reason; to justify an embarrassing 'no historical tie' on the side of the
claimants. As described here

12.23.19 7:31pm - We, who are bound to earth, salute you @Astro_Jessica for showing us how cosmic
imagination can enrich a humble tradition, of a tiny people, who did not have much to offer the world, except
learning, memory and imagination.

12.23.19 7:16pm - The legitimization of proxies is treated here https://ucla.in/2N5GNOK and here
https://ucla.in/2N9icIX We call it "Measurement Bias and Effect Restoration". BTW, the conditions established
are not unique to Pearl-stale analysis; they govern ALL analyses, no escape. #Bookofwhy

12.23.19 4:26pm - (Replying to @AngeloDalli @GaryMarcus and 2 others) I'll second the conclusions. Now, let's
do it. That's what the do-operator calls on us to do. #Bookofwhy

12.23.19 3:22pm - (Replying to @EpiEllie and @smueller) If I were a mentor to 5-yr old, I would say: Isn't it
amazing? With the same tools we used to determine backdoor conditions you can now determine how to re-
weight data to estimate target effects? This possibility will not occur to an economist in the next 17 years!
#Bookofwhy

12.23.19 3:09pm - (Replying to @emzanotti) She doesn't. Hanuka represents Jewish sovereignty in some part of
the middle east, https://twitter.com/yudapearl/status/1209103037991534592 an idea that @IlhanMN Omar has
vowed to eradicate, in blunt betrayal of her Minnesota voters.

12.23.19 6:31am - (Replying to @stuartbuck1 @Jabaluck and @metrics52) Inverting matrices is a good analogy. I
would use "solving equations algebraically" vs. trying out all possible solutions and see which satisfies the
equations. We can do it by hand, true, but where would science be today w/o algebra? #Bookofwhy

12.23.19 5:10am - (Replying to @charleendadams) Ocho Candelikas is my favorite Chanukah song. Inviting all
our Spanish speaking readers to join us in lighting the candles and celebrating the defeat of darkness. Here:
https://youtube.com/watch?v=1p5FBFdYufk

12.23.19 2:13am - Readers elevated by the Holidays spirit are invited to a sing-along in Hebrew, ending with a
Chanukah song we used to sing in kindergarten: Sevivon Sov Sov Sov, about 11 minutes into this video
https://youtube.com/watch?v=PvByNKs1bmw Hallelujah!

12 22 19 5:16am (Replying to @totteh @jttiehen and @keithfrankish) I wrote "sifting" not "shifting" because I
12.22.19 5:16am - (Replying to @totteh @jttiehen and @keithfrankish) I wrote sifting , not shifting , because I
am always hoping for some reader to show me a nugget that I have neglected in my combing of this literature.
#Bookofwhy

12.21.19 5:52pm - For readers baked in traditional philosophical theories of causation, who wish to see how
these gel with modern formalisms, this thesis provides a panoramic view: https://su.diva-
portal.org/smash/get/diva2:1373243/FULLTEXT01.pdf I am still sifting this tradition for a gold nugget worth
adding to SCM #Bookofwhy

12.21.19 12:49pm - (Replying to @mattshomepage and @Elias) More than that! The completeness of the do-
calculus is nice, but it does not provide us a procedure for finding an identifying pattern. The algorithm
(summarized here: https://twitter.com/PHuenermund/status/1207338778886574080) actually finds it,
whenever such a pattern exists. #Bookofwhy

12.21.19 11:05am - (Replying to @mattshomepage) The answer is YES. @Elias algorithm can be proven to be
"complete" namely, guaranteed to discover ALL doors, where by "doors" we mean patterns that allow
identification. #Bookofwhy

12.21.19 10:38am - (Replying to @mattshomepage) I do not buy the "very specific structure" limitation. The
class of structures satisfying the front-door condition is not smaller than those satisfying backdoor condition. It
is only that the latter is in more common use. #Bookofwhy

12.21.19 10:33am - (Replying to @mattshomepage) Well said, and it entices me to add: When you wish to
express WHAT you know, you better express it the WAY you know it, that is, the way it is stored in your mind,
not the way it would justify favorite statistical routines. #Bookofwhy

12.21.19 4:41am - (Replying to @PWGTennant and @medrxivpreprint) I am anxious to see the first paragraph,
where DAGs are introduced. eg. DAGs are .....

12.21.19 4:20am - (Replying to @ang_hermann) Good luck @ang_hermann on your French translation project. I
know it is a work of love, else you would not have undertaken it. I'm looking forward to see @Bookofwhy on
French book-shelves, and hear France communicate with Portugal in DAGs. PS. a Spanish one is in the making
too.

12.21.19 3:04am - (1/ ) (Replying to @ErakatSaeb) This is not a joke! It is in fact the cause, the root, and the
essence of the Palestinian tragedy. Unable to celebrate any holiday connected to the land to which they claim
sole ownership. Unable to chant a single hymn authored in the days of Jesus or Judas Maccabeus,
12.21.19 3:09am - (2/ ) (Replying to @yudapearl and @ErakatSaeb) lacking any cultural connection to those
days, Palestinians been laboring to fabricate such connection by molesting the heritage of their neighbors, and
hoping to be taken seriously and well-intentioned if/when they decide to come to the negotiating table again. If
only ...
12.21.19 3:14am - (3/3) (Replying to @yudapearl and @ErakatSaeb) If only @ErakatSaeb understood what this
circus acrobatic does to Palestinian posture.

12.21.19 2:04am - (1/ ) Halleluya, It's Hanuka! Sunday night we will be lighting the first candle to commemorate
the Maccabees revolt, Jerusalem, 161 BC. Join me in a holiday that inspires all who cherish the ideas of
freedom and self-determination. My grandson asked me why we make such a big fuss l
12.21.19 2:04am - (2/ ) about this holiday. I told him Hanuka is our TRUST DEED to the birthplace of our history,
more solid even than the ancient synagogues they excavate in Israel, or the arch of Titus in Rome. Why?
Because stones can be faked, not so a continuous celebration for 110 generations.
12.21.19 2:04am - (3/3) I wrote about this aspect of Hanuka when the LA Times asked me "what does Israel
. . 9 :0 a (3 3) w ote about t s aspect o a u a w e t e es as ed e w at does s ael
mean to you?" https://latimes.com/opinion/la-op-pearl-bisharat16-2008may16-story.html and why
understanding Hanuka is so essential for dreaming any peace prospect between these two equally indigenous
peoples.

12.21.19 12:36am - Halleluya!! Good tidings! Kindle has finally uploaded a revised edition of PRIMER
https://ucla.in/2KYYviP with all the errata smoothed to perfection. Enjoy. As to the hard copy, publishers try
to keep this information very very secret. But I'll Tweet when it happens #Bookofwhy

12.21.19 12:14am - (1/ ) Gratified to see researchers committed to decision making adopt modern tools of
causal inference. I can tell right away that the authors have not graduated from UCLA. How? (1) Consistency is
an "assumption" in PO, not so in SCM, where it is a "theorem", https://ucla.in/2MlB5Ut
12.21.19 12:14am - (2/2) and (2) Graphs/potential-outcomes symbiosis is not sacrilegious in SCM; it is in fact
the norm; the former to express what we know, the latter what we wish to know. See #Bookofwhy (though it
may be sacrilegious in some islands of PO, eg Lilliput and secret others)

12.20.19 11:41pm - Yes, worth keeping in mind, and add to it: Hypothesizing means unabashedly committing to
a representation (e.g., a model) from which we can deduce how things operate, when the need arises.
#Bookofwhy

12.20.19 3:33am - (Replying to @maximananyev @Jabaluck and @metrics52) Do you think this great paper gives
potential users any idea about how to select variables for valid reweighing? (even if we knew exactly how study-
participation was determined in MA and in NJ). #Bookofwhy

12.20.19 1:28am - In the island of Lilliput they don't teach addition. Why? Because the problems they truly truly
wish to solve involve multiplication; addition alone cannot handle. That's in the legendary island of Lilliput.
#Bookofwhy

12.19.19 8:25pm - (Replying to @Jabaluck and @metrics52) Ask not what I did wrong nor what I did right, ask
what YOU can do to elevate your field to new heights. #Bookofwhy

12.19.19 7:36am - (1/ ) True, Journals normally do not bid and do not solicit. However, what makes an
enlightened editor ENLIGHTENED is his/her ability to spot promising new develpments and invite authors to
write a review article for readers who have not been exposed to it. Phil Dawid, for example
12.19.19 7:36am - (2/ ) invited me in 1995 to write an article to Biometrika, because he felt that it was about
time that Stat readers will learn something about graphical models. He probably faced stiff resistance by
reviewers and the established elite. He did it, and his leadership made an impact.
12.19.19 7:36am - (3/ ) I wrote at least a dozen papers thus invited by enlightened editors. You don't have to be
a genius to see that your field lags behind in a given area, you need only be a leader to do something about it.
This is what I meant by "bidding." Perhaps econ. editors are still
12.19.19 7:36am - (4/ ) fighting for the honor to be the first? Or perhaps they have not been informed of the
paper. Alert them. Some editors need a jolt. Play on their patriotic duties to their readership. Econ. students
can't afford another decade of arrow-phobic prohibition. Can they? #Bookofwhy

12.19.19 3:49am - (Replying to @SadiqKhan) Commending you on your courage and good will. But to be
effective, please heed to an advise of someone who has studied the anatomy of anti-Semitism. Change your
statement to read: "a hostile environment for anti-Semites, Islamophobes and Zionophobes. @EinatWilf

12.19.19 2:42am - I bet you already received 4-5 invitations from top Econ journals. Seriously, one can argue till
dawn (as @jabaluck & @metrics52 ) that economists do not need new identification strategies. But I dont know
one who'd argue they don't need external validity or data fusion.#Bookofwhy
12.19.19 2:32am - (Replying to @Meetasengupta and @prem_k) Would love to discuss those possibilities with
you, but my family will be descending on us on Chrismass, so I can't make any plans in advance. Let's play it by
ear.

12.19.19 12:12am - (Replying to @Meetasengupta @prem_k and 3 others) I believe we can go a step beyond "try
to narrow the gap", since we are in the age of causation, where "trying" is just not good enough. We can at least
formalize when gaps are narrowable, what information is needed to narrow them, etc. We have the language to
do it! #Bookofwhy

12.18.19 11:40pm - (Replying to @prem_k @Meetasengupta and 3 others) I am not sure what "Evidence Based"
people count as "evidence". I know that N. Cartwright, one of the champions of EB has also complained bitterly

about external (non)validity of RCT's, and I've tried to introduce her (and Deaton) to S-diagrams:
https://ucla.in/2JfdQcv #Bookofwhy

12.18.19 11:28pm - Evidently, wines do taste better with age. As I re-read the 1995 paper on instrumental
inequalities, https://ucla.in/2pFa9pc, I can't help but wish that IV's would be introduced today with the same
clarity and freshness as they were treated then, in 1995. #Bookofwhy

12.18.19 10:47pm - To readers who asked whether I heard back from Miguel why he goes back to PO, given that
it is so easy to generalize in DAGs, https://twitter.com/yudapearl/status/1206522147998289922, the answer is
not yet. I think Harvard folks love DAGs, but still have problems swallowing do-calculus. Who knows?
#Bookofwhy

12.18.19 5:20pm - (1/2) Glad to see this ancient inequality finding useful applications in Mendelian
Randomization. Truthfully, I knew already in 1995 that it would be resurrected some day, but I wrote it for
economists, who loved IV and bemoaned: "it can't be tested". I thought they would jump
12.18.19 5:20pm - (2/2) from joy and celebrate the testability of their beloved IV. Little did I know that
economists have been waiting for epidemiologists to take the lead. #Bookofwhy

12.18.19 3:01pm - This paper may just be what econometrics has been waiting for in its effort to catch up with
the causal generation. I envision dozens of editors bidding on the right to publish it in their journals. Imagine
what such paper could do to the reputation of Econometrika. #Bookofwhy

12.18.19 1:24am - (1/ ) Our success of showing readers how re-weighing works in moving across populations
entices me to show how it works in correcting sampling-selection bias. Assume we preferentially select subjects
for a study based of an arbitrary set of characteristics (some measured & some not) 12.18.19 1:24am - (2/ ) ,
and we wish to estimate P(y|x) in the target (unsampled) population. Can the DAG tell us when we can do it?
Yes. Whenever we can measure a set Z of variables such that: (1) {X,Z} separates S from Y and (2) The
conditional distribution P(z|x) can be estimated for the target 12.18.19 1:24am - (3/ ) (unsampled) population.
If these hold, then the re-weighing formula P(y|x) = SEM_z P(y|x,z, S=1) P(z|x) gives us an unbiased estimate of
P(y|x) in the target population. Illuminating examples, generalizations and refinements are given in
https://ucla.in/2OgMF4g. 12.18.19 1:24am - (4/4) Note that the only difference between this re-weighing
formula and the one used in transportability is the weighing term P(z|x) vs. P(z). Note also that even @EpiEllie
5-yr old can enjoy and appreciate the power of selection diagrams; how else would one select Z?? #Bookofwhy

12.17.19 6:36pm - Our success of showing readers how re-weighing works in moving across populations entices
me to show how it works in correcting sampling-selection bias. Assume we preferentially select subjects for a
study based of an arbitrary set of characteristics (some measured & some not)
y y ( )

12.17.19 5:59pm - (1/ ) Great to see Uber joining the causal era. There is one glitch in this article which might
be misleading to the novice - the word "approach" appears 19 times. Why misleading? Because "approach"
connotes an "option" under the analyst control. Yet the 19 methods discussed
12.17.19 5:59pm - (2/ ) are not under our control but are "opportunities" made available to us by (what we
believe is true about) reality, ie, our model of reality (DAG). Once we have a model of reality we cannot speak
about the 19 methods as "approaches," ie, as if they were our options. #Bookofwhy

12.17.19 6:31am - (Replying to @jeremyphoward and @calabi_and_yau) Students of causality would read this
paper with a constant question in mind: How do you climb the Ladder of Causation, from predictions, to effect
of actions, to counterfactuals (eg would this customer be swayed by this discount). I guess it is all in the model
used. #bookofwhy

12.17.19 6:25am - (Replying to @HenningStrandin) Thanks for posting this interesting summary. Just in case you

find it relevant, I have found comfort in escaping the circularity trap through another route. It is in Section
33.5.3 of this paper: https://ucla.in/2ugpzTg. #Bookofwhy

12.17.19 12:54am - (1/3) I am retweeting Charlene's insightful and personal testimony from the holy city of
Bristol. The city where anti-semitism does not exist, and whatever does exist has nothing to do with Corbyn,
and whatever has to do with him is Jewish invention or Israel's fault.
12.17.19 12:54am - (2/3) Bristol is a city of highly devoted Corbynites who helped turned a true "Labour Party"
into a dangerous and un-electable cult of directionless excuse makers. A personal question to each member of
your faculty, including prof. Harvey Goldstein and those who like his report:
12.17.19 12:54am - (3/3) What have YOU personally done to fight BDS activities on YOUR campus? Have you too
decided that hostilities connected with Israel are kosher hostilities, not affecting students like me who are, or
are perceived to be, proud of Israel? Wasn't it YOUR academic duty to act?

12.16.19 11:55pm - (Replying to @calabi_and_yau and @jeremyphoward) I listened to @jeremyphoward talk and
kept asking myself: When was it given? It turned out it was given in 2012, before analysts realized that most of
his aspirations can be algorithmatized. So, yes, it is related to #Bookofwhy

12.16.19 11:01pm - (Replying to @jouni_helske) Thanks for saving me an n-th attempt to understand what
Gelman is talking about. It is not his fault, it is mainly mine. We live in two non-intersecting universes. I have
solved a few toy problems, Gelman refuses to try even one. #Bookofwhy

12.16.19 10:56pm - (Replying to @calabi_and_yau) It would be of interest to document the various contexts in
which business people yearn for "insight" and see if it confirms my theory that what they yearn for is "causal
understanding", or at least some chunk of a causal model compatible with the data. #Bookofwhy

12.16.19 2:58pm - (Replying to @jouni_helske) I will comment on Gelman's post if you tell us why you find it to
be an "interesting take." Or, how it could possibly change the way we/others are doing Causal Inference,
however minutely. #Bookofwhy

12.16.19 4:28am - I consider myself progressive, so some academic colleagues call on me, resentful of Israel's
courting right-wing politicians to pass anti-BDS laws. My answer to them: What have YOU personally done past
10 years to fight BDS activity on YOUR campus? Wasn't it YOUR academic duty?

12.16.19 3:34am - (Replying to @thosjleeper) We have considered, but S-admissible is reminiscent of backdoor
admissible for identification Moreover it is almost 9 years since its first publication What would you suggest?
admissible for identification. Moreover, it is almost 9 years since its first publication. What would you suggest?
Reweighing-admissible"? perhaps. S-separator? "S-Equalizer"? Suggestions? #Bookofwhy

12.16.19 2:31am - (1/4) In view of the dominant role that re-weighing plays in extrapolating effects across
populations, and the many Twitter requests for a concise graphical criterion that gives re-weighing its
legitimacy, I am retweeting the criterion (called "S-admissibility"), in next 4 tweets.
12.16.19 2:31am - (2/4) It works on a selection diagram in which S nodes represent disparities between the
target (*) population and study population (experimental). Z is a set of measurements. To test if Z is S-
admissible (1) Remove all arrows pointing to X (2) Check if {X, Z} d-separates S from Y
12.16.19 2:31am - (3/4) If Z passes this test, then the reweighing formula is valid: P*(y|do(x)) = SUM_z
P(y|do(x),z)P*(z) In words: Effect at target equals the Z-specific effects at study, averaged over Z, using the
target distribution P*(z) as weight. Warning, this is merely a sufficient test.
12.16.19 2:31am - (4/4) Many more opportunities are available for situations in which a S-admissible set cannot
be found. see https://ucla.in/2N7S0K9 and https://ucla.in/2Jc1kdD #Bookofwhy (And dont dismiss "trivial
transportability," described in https://ucla.in/2N7S0K9) Graphs are fun! #Bookofwhy

12.16.19 1:50am - I am also unsure of what "rainfall became to Economics", but I am retweeting ONE sentence,
because it is a GEM: "Mendelian Randomization is just another observational study, with a different set of
assumptions." Quasi-experimentalists shoud keep this in mind. #Bookofwhyk

12.16.19 1:38am - (Replying to @davidsirota) You may be Jewish by birth, but that does not prevent you from
conflating the issue, perhaps intentionally. Bernie Sanders is criticize for surrounding himself with Zionophobes
like Linda Sarsour, consumed by genocidal agenda to eliminate Israel. Jews do not escape issues.

12.15.19 1:35pm - (Replying to @EpiEllie) Yes. (I thought I tweeted it before, but can't find it.)This criterion
(called S-admissibility) is sufficient for transporting the effect by simple re-calibration on Z. But this is the
simplest kind of transport. We can do much more even if such a Z is not measured.#Bookofwhy

12.15.19 1:23am - (Replying to @DrMikeH49 @markshiffer00 and @YousefMunayyer) Mike, thanks for using
Zionophobia, our only fighting word which, for some reason, is still not "the ugliest word in town" but it's getting
there. See my reasons for preferring it over anti-semitism: https://tinyurl.com/y53hvecj

12.15.19 12:45pm - (Replying to @EpiEllie) Which diagram specifically gives you hard time to decide if {X,Z}
separates Y from S (in the revised DAG from which all arrows entering X are deleted)? This is a simple d-
separation test -- duck soup in your Epi class. What's the problem? #Bookofwhy

12.15.19 12:21pm - (Replying to @YousefMunayyer) Agree! Equating anti-zionism with antisemitism is


misguided; it gives the former a speck of legitimacy that racist ideologies do not deserve. Zionophobia is more
dangerous on its own character and genocidal aims. Here is why: https://tinyurl.com/ak4ccg It's the ugliest
word in town.

12.15.19 12:01pm - (Replying to @yudapearl and @wtgowers) This is another reason why I viewed Corbyn as a
danger to humanity: https://timesofisrael.com/former-uk-chief-rabbi-lord-sacks-jeremy-corbyn-is-a-dangerous-
anti-semite/ Rabbi Sacks is a champion of universal values. I know him personally. We went together to Muslim
schools in London and talked to kids on our common heritage. I weigh his word over Guardian

12.15.19 4:05am - (Replying to @Howard_Lovy and @lsarsour) Which proves that there IS a fundamental
difference between "left-wing" and "right-wing" anti-Semitism. The latter is emotional and mother-milked, the
former is calculated, strategic and eliminationist, ala @lsarsour .
12.15.19 3:50am - (Replying to @wtgowers) Admittedly, I was unaware of some these very concerning
allegations. However, this report https://twitter.com/simonwiesenthal/status/1204510314986754048 was
prepared by dedicated scholars whom I know for years. In contrast, I know NONE of the academic signatories on
the Guardian. Strange! I should know at least ONE.

12.15.19 3:21am - (Replying to @AndreaSaltelli @RonKenett and @fhuszar) I have read your Nature paper with
great interest, yet it is not clear to me where "causal models" reside? In the statistical or mathematical
category? My taxonomy of models (Ladder of Causation) is based on what language the assumptions must be
expressed. Thoughts? #Bookofwhy

12.15.19 3:06am - (Replying to @RonKenett @AndreaSaltelli and @fhuszar) Pearl has this habit of celebrating
what we can do today that we could not do yesterday. Ron has the habit of playing Solomon the wise: "Nothing
is new under the sun" relevant or not. I'd go with Pearl; celebrations are fun and yield new results. Solomon
doesn't #Bookofwhy

12.15.19 2:22am - (Replying to @james_r_lucas @roydanroy and 3 others) To predict a "wall" and overcome it,
you need to have some idea of what's behind it. ML per se has no language to describe "what's behind". But see
https://ucla.in/2HI2yyx #Bookofwhy @tdietterich

12.15.19 1:59am - Little Jack Horner, sat in the corner... Oh What a Good Boy Am I. !!! Singing Oh What a merry
land is England Singing Oh What a merry land is England

12.15.19 12:51am - (Replying to @mendel_random @stephensenn and 2 others) The #Bookofwhy unveils several
appearances of causal notions in the periphery of statistical analysis, it even celebrates with great fanfare
David Cox's (1958) explicit expression "quite unaffected". Those peripheral sparks can now be made explicit and
unleashed. R U unhappy?

12.14.19 11:46pm - (Replying to @RonKenett and @AndreaSaltelli) I am celebrating a new ability: To take 10
meaningful examples and solve them formally from A to Z, and you are inviting me to go back to the hand-
waving days of "gaining understanding" w/o solving any of those examples.What is gained by reminding people
of those days? #Bookofwhy

12.14.19 11:05pm - (1/2) (Replying to @RonKenett and @AndreaSaltelli) I am open to be convinced about the
existence of "different ways to generalise findings" as soon as you show me (not 'so and so', but YOU) how ONE
of those "ways" decides which of the ten simple examples in https://ucla.in/2N7S0K9 is generalizable.
#Bookofwhy
12.14.19 11:26pm - (2/2) (Replying to @yudapearl @RonKenett and @AndreaSaltelli) 2/2 Elias has compiled
dozens of quotes on "different ways" and "so and so", from Campbell etal, to David Cox, to Manski. I don't see
why the belief that "there r different ways" gives you more comfort than my conclusion (10 yr exploration) that
there r'nt such ways. #Bookofwhy

12.14.19 8:21pm - (1/ ) Good question: "So how do researchers generalize experimental results if they don't
know selection-diagrams or, worse, if they (eg econ.) don't even use DAGs?" Ans. They dont. They publish tons
of paper with "generalizing" in the titles but, inside, they assume the analyst can
12.14.19 9:11pm - (2/ ) do the S-admissibility test in her/his head and come out with some "conditional
ignorability" expression that justifies the result they want. Bingo! They did it! Reminiscent of the way folks did
adjustments before DAGs. In https://ucla.in/2wbH488 we decry this culture in
12.14.19 9:11pm - (3/3) no ambiguous terms. But to those in the cultural bubble, decrying amounts to non-
existing. And to economists, it may even amount to: "Don't show our students." Why the resistance? Selection
diagrams highlight the needs to start with what we know - too hard for PO #Bookofwhy
12.14.19 7:52pm - I echo Democritus (~400 BC) who said: "I would rather find one cause than be a King of
Persia". My bit: "You learn more by solving ONE toy problem than watching 10 panels"- even when everyone tells
you "it ain't REAL-DATA" or "its textbook econ." It ain't! Try one! #Bookofwhy

12.14.19 6:54pm - (Replying to @EpiEllie) As they say in the Mishna "A child that does't ask, lead him to a
question" (At Ptach Lo). So, which of the following can you read as "S-admissible"? (i.e., the x->y effect is
transportable by reweighing on Z) #Bookofwhy

12.14.19 6:42pm - (Replying to @vatsal_maru and @BernieSanders) Poor Bernie wants us to believe he is a "man
of principles," but is appointing advisers (eg Sarsour) that are consumed by unprincipled agenda. Why? Because
other advisers tell him "this is what young voters go for" -- an unprincipled "principle", proven a disaster by
Corbyn.

12.14.19 6:21pm - (Replying to @paulpharoah) You are absolutely right to demand English, but you need to say
more than just "gibberish". Tell us where you get stuck.

12.14.19 6:18pm - (Replying to @nathankallus @SusanMurphylab1 and @Susan_Athey) My take: (1) “read the
stats literature!” but only after taking "causal inference 101", eg https://ucla.in/2KYYviP (2) “pay SOME
attention to https://ucla.in/2HI2yyx (3) “read the empirical literature! but only after (1)&(2) (4) Solve one
"toy problem" in CI, say from #Bookofwhy

12.14.19 3:00pm - Seriously. What did the panelists think when the words "the ML community's approach to
causal inference" were spoken? Did they take the word "approach" to mean "aspirations" or "buzz words" or "pay
attention to" or (my favorite) "careful thinking" #Bookofwhy

12.14.19 2:42pm - (Replying to @analisereal @EpiEllie and 2 others) The Mishna says: "The shy cannot learn and
the strict cannot teach". We are facing a 5-yr old that is too shy to ask, and teachers that are open to any
question. Will the shy help with: "I would like to know how...?" or "What do we do in case..." or... or...
#Bookofwhy

12.14.19 2:30pm - Great thread! As an outside observer, the punchiest punch-line was: Contrast traditions that

start with *the data generating process* versus those starting with *the algorithm*. The most perplexed line was
"the ML community's *APPROACH* (???) to causal inference" #Bookofwhy

12.14.19 1:59pm - (Replying to @EpiEllie) What's the problem? Constructing a selection diagram? Or reading it to
determine how to transport things? We can go over it one at a time, assuming, of course, that your 5-yr old can
read DAGs. & while we are at it, other readers can enjoy its power on favorite probls.#Bookoofwhy

12.14.19 9:39am - (1/ ) Before parting from the beauty of selection diagrams and the S-admissibility criterion,
it is important to empower @EpiEllie 5-yr olds with the proper historical perspective. In the same way that
backdoor holds the key to the adjustment formula P(y|do(x)) = SUM_z P(y|x,z)P(z)
12.14.19 9:39am - (2/ ) which was used informally since Yule (1899), so does S-admissibility hold the key to the
re-calibration formula (or re-stratification, or re-weighing, or re-standardization) P*(y|do(x)) = SUM_z
P(y|do(x),z) P*(z) which was used informally by 18th century demographers
12.14.19 9:39am - (3/3) to transport mortality rates across populations (from PI to PI*). We make this historical
connection in footnote 13 of https://ucla.in/2N7S0K9 and https://ucla.in/2L6yTzE where readers can find
additional magics of the graphical S-admissibility criterion. Enjoy. #bookofwhy
12.14.19 2:41am - A new paper deserving our attention, on alternative CI methods:
https://sciencedirect.com/science/article/pii/S2352827319301545. My first reaction: Not clear why
triangulating IV with backdoor methods should give us more information than, say, one backdoor method with
two distinct admissible sets. #Bookofwhy

12.14.19 1:02am - Sincere apologies to my colleagues in the dpt of Archaeology at UCLA. It is the dpt of
Anthropology that deserves the shame, as described here
https://jewishjournal.com/news/los_angeles/298850/ucla-guest-lecturer-calls-zionists-white-supremacists/
and here: https://tinyurl.com/y66pnnsm Sorry for the confusion.

12.14.19 12:29am - Corbyn to Pellosi: Nancy, they all lied to us; about the angry mob, the shifting voting base,
the radical millenials of Omar, Tlaib and JVP --they are just not there! Voters can't stand liberal politicians
when they betray liberal values.

12.13.19 10:16pm - (Replying to @omaclaren) Where can we read more about Fraser and Bunke and what they
related the fiducial argument for?

12.13.19 7:13pm - (Replying to @charleendadams) And my only unhappiness with Corbyn's defeat is that
(according to analysts) he was not defeated for the reason he should have been: Racism, but for being wishy-
washy on Brexit. Still, his defeat removes fears of shame & darkness for all my colleagues in the UK -
Congratulations!

12.13.19 4:35pm - Personally, I have NOT encountered antisemitism, as confessed here


https://tinyurl.com/u4eyccd. But my students HAVE, especially in dpts of History and Archaeology, shamefully
mentioned in https://tinyurl.com/u4eyccd. It is the duty of every professor to expose those pockets of hate.

12.13.19 10:46am - (Replying to @charleendadams and @pablo_gps) University of Bristol!!! What a


coincidence!!! I believe George @mendel_random is a professor there, and I presume he spends hours fighting
the phenomenon you are describing. As I am doing in my university. Our reputation is at stakes, not to speak of
our moral commitments.

12.13.19 10:23am - (Replying to @mendel_random and @pablo_gps) The understatement of the year: "Corbyn
should have apologised for the behaviour of a smallish number of people mascerading as labour supporters."
Compare to a comprehensive, documented report, by professionals that I know personally:
https://twitter.com/simonwiesenthal/status/1204510314986754048 "Smallish Number"!!!

12.13.19 1:37am - (Replying to @pablo_gps) The disaster I feared was not Labor (I loved Blair), nor Brexit. I
feared seeing a racist like Corbyn as a PM of a European country that would be turned into a safe haven for
Hamas and other anti-Western movements.

12.13.19 1:24am - This is the normal controversy we find in the literature. Then we have the added dimension
of causality, see https://ucla.in/2nZN7IH, and the fact that Bayes was chasing causes. #Bookofwhy

12.13.19 12:37am - (Replying to @ulrichspeck On the nail. Hot-headed students hijacked the volume on social
media, and our politicians believe their voting base is really shifting. Look at what Linda Sarsour has done to
Sanders and what Omar and Tlaib are doing to Pelosi.

12 13 19 12:21am (Replying to @pablo gps) Upon facing a sure disaster and an unknown you dont "choose" the
12.13.19 12:21am - (Replying to @pablo_gps) Upon facing a sure disaster and an unknown, you dont choose the
unknown but you sure escape the disaster.

12.13.19 12:16am - The pleasure was mutual and I was thrilled by the "philosophical twist" of today's talk. First
time I speak about Bayes' pool table (p. 98 #Bookofwhy) realizing that most people (including Bayesians) dont
really know what Bayes did and what the controversy is all about. New book?

12.12.19 11:19pm - On such historical events, my grandpa used to bless: Oh God! Hagomel Chassadim Tovim
L'Amo Israel. As a secular Jew, but one who does believe in miracles, all my lips can say: Oh God! please give
me the illusion that my Tweet has swayed at least one voter to defeat this racist!

12.12.19 9:40pm - (1/ ) (Replying to @EpiEllie) If DAG rules take a 5-year old, then selection takes a 6-year old.
First, a tiny correction! Only a tiny fraction of DAG rules (ie, backdoor) is taught to 5-yr. The entire spectrum of
identification opportunities (ie, do-calculus) awaits to be embraced in Epitweet #Bookofwhy
12.12.19 10:08pm - (2/ ) (Replying to @yudapearl and @EpiEllie) You are probably hoping for a backdoor rule
when you say "5-year old". Don't worry, we've one for your 5-yr old in transportability. It is Def-8 & Cor-1 in
https://ucla.in/2N7S0K9: If Y can be separated from S in G_X, Bingo! The effect is transportable! (by Eq 3.1)
#Bookofwhy
12.12.19 10:19pm - (3/3) (Replying to @yudapearl and @EpiEllie) You see, It's duck-soup easy!! Even to a 5-yr
old. All it takes is a simple language to say how populations differ, and a deep sense of belief that, if things are
done the hard way in PO, they can be done the easy way in our natural language. #Bookofwhy

12.12.19 9:26pm - (Replying to @EpiEllie) The new name reminds users that the diagram stands for TWO
populations, not one, and that the selection nodes [S] just switch from one population to another. This in
mind,anyone who understand DAGs should understand 2-DAGs, ie, S-diagrams, as Elias shows
https://twitter.com/eliasbareinboim/status/1205140352061116416 .

12.12.19 7:27pm - (Replying to @EpiEllie) If you are happy with the functions behind the arrows, just follows
the beautiful slides by Carlos and Elias, here: https://twitter.com/yudapearl/status/1205083830584872961 and
tell me if/where their explanation of selection diagrams is less than simple. We will continue from there. Is it a
deal? #Bookofwhy

12.12.19 3:15am - (Replying to @eddericu @EpiEllie and 2 others) This are great slides indeed, but they still
require one to understand that behind the DAG's arrows there are FUNCTIONS and, not less important, that
writing y=f(x,u*) is another way of writing y=g(x,u), with g neq f. #Bookofwhy

12.12.19 1:20am - My My! I almost forgot I did this podcast. Now the counterfactuals are beginning to play tricks
on you: "You should have answered it this way, or that way." No, I shouldn't; we know that free will is just an
illusion #Bookofwhy

12.11.19 11:00pm - (Replying to @yudapearl and @Corey_Yanofsky) But, for havens sake, even if p-Jews
(People-hood defined Jews) are just a minority (they are a vast majority) aren't they entitled to dignity and
respect like other p-minorities? Especially by the Office of (don't smile) "Equity, Diversity, Inclusion" at my
university, UCLA.?

12.11.19 10:09pm - (Replying to @EpiEllie) I've tried to get inspired by your Dags Cartoons, to explain *selection
diagrams* but, OOPS, I got stuck on agreeing about what a DAG is. "A Research Tool"? True. But a DAG also
represents a collection of functions that defines a population. Can we agree, to continue? #Bookofwhy

12 11 19 9:28pm (Replying to @EpiEllie) You know that I am the last person to "change topic" on Tweeters For
12.11.19 9:28pm - (Replying to @EpiEllie) You know that I am the last person to change topic on Tweeters. For
me, a selection diagram is just a DAG's way of representing two populations that differ in some of their
underlying mechanisms (ie, equations). If you know DAGs, you know selection diagrams. #Bookofwhy

12.11.19 8:55pm - (Replying to @Corey_Yanofsky) And Ruth said: Your people will be my people, and your God
will be my God. People-hood first, religion second.

12.11.19 7:05pm - I am retweeting, because this question comes up again and again: Why are diagrams
"natural", "transparent" and "communicable"? Why? It also touches on the sanctity of RCT and the obsession with
"well defined intervention" that some brilliant scholars refuse to abandon #Bookofwhy

12.11.19 6:51pm - (Replying to @EpiEllie) For ordinary folks who represent knowledge the way it is stored in our
brain, in terms of FACTORS influencing one another, who have learned (eg. #Bookofwhy or
https://ucla.in/2NceWNb) that Potential Outcomes are DERIVED properties of that knowledge, not its
elementary particles

12.11.19 5:13pm - Which reminds me. Tomorrow I'll be speaking to History of Science Reading Group (3pm UCLA
Royce 314) where I will resurrect Bayes, Galton, Pearson, Wright and Haavelmo and ask again: Where have all
the causes gone? Will Haavelmo's legacy survive in Econometrics ?#Bookofwhy

12.11.19 5:01pm - Remember Ander's "Russian Roulette" and all the arguments why one needs to use cross-world
cryptic like P(Y(1)|Y(0)) to generalize experimental results? Well, Carlos and I took another look,
https://tinyurl.com/uzlknjj. Good news: You can do better in common language. #Bookofwhy.

12.11.19 4:37pm - It is a victory for BDS and an indictment of American universities that it should take a
Presidential Order to define "Jewishness" the same way as Jews are defining themselves today, and the same
way they have been defining themselves since the exodus from Egypt.

12.11.19 4:14pm - (Replying to @mstephens999 and @y2silence) Sure, because "salah" is not the only factor that
determines the outcome of the game. Recall, however, a "unit" u, in both SCM and in PO stands for the sum
total of all factors necessary to make the outcome 0 or 1, leaving nothing to uncertainty, thus no Pr(*).
#Bookofwhy

12.11.19 3:50pm - (Replying to @mstephens999 and @y2silence) What makes "u" different than "i" ? If there is no
difference, and both are just "anyone", we can remove the index. But if "u" have an allergy to zinc or some
specific experience with zinc, we say: Pr (cold|do(zinc), specific knowledge). Still no index. #Bookofwhy

12.11.19 5:11am - This thoroughly-documented report by the Simon Wiesenthal Center chases sleep from my
eyes. I imagine UK slipping into Corbynism, what it would do to my people, to my former country, Israel, and to
the moral compass of the civilized world. Extremely worried.

12.11.19 1:44am - (Replying to @mstephens999 and @y2silence) I haven't thought about what Pr(Y(u)=y |
do(X(u)=x)) could mean. But now that I think about it, the answer would be either 1 or 0. Because the Y(u)
takes some value, say y' , which is either equal to y or not. If u is picked up at random we get P(Y=y|do(X=x))
w/o u. #Bookofwhy

12.11.19 1:33am - (Replying to @patrickkloesel @ASSAMeeting and 2 others) If invited, I'll break my vacation
and go to San Diego to remind economists how Haavelmo defined "economic model", as I started to do here:
https://ucla in/2mhxKdO and of all the new things they can do by restoring his vision #Bookofwhy
https://ucla.in/2mhxKdO, and of all the new things they can do by restoring his vision. #Bookofwhy

12.10.19 12:08am - This is a powerful presentation of the message I am trying to send to ML folks: "You and I
know that we must climb the ladder of causation at some point. Therefore, to be safe, let's look at the rungs
before we climb, and surely before we make believe we are already up"#Bookofwhy

12.9.19 9:52pm - (Replying to @mstephens999 and @y2silence) I see it as an advantage, to be able to


distinguish between questions that can be answered by intervention, do-expressions, and those that cannot,
because they are Rung-3 counterfactuals. Look how important this distinction is: https://ucla.in/2G2rWBv ,
#Bookofwhy

12.9.19 4:31pm - (Replying to @mstephens999 and @y2silence) Do-expressions are probabilisitic, involving
populations. To go to individual level, we go to SCM (Stractural Causal Model), where U=u is the individual. All
counterfactuals, PO, ignorability etc etc can be derived from a fully specified SCM.(Including
"consistency")#Bookofwhy

12.9.19 3:22pm - (Replying to @VartanKA @haldaume3 and 3 others) Attempts to synthesize have been
successful. All is unified now and we fully understand where each concept stands in the scheme of things. eg.
"ignorability" is logically implied by 'backdoor." No "synthesis" in Mostly Harmless where graphs mean
disloyalty.#Bookofwhy

12.9.19 3:10pm - (Replying to @mstephens999 and @y2silence) When we write P(y|do(x), z) it means the full
distribution of Y (not "mean") for every intervention do(X=x), for every z-specific class of individuals. It is so
heterogeneous that calling it "heterogeneous" only diminishes its heterogeneity. #Bookofwhy

12.9.19 4:09pm - This is one merit of social networks: Reviewers and editors are no longer operating with
impunity, they know that history will judge them, albeit not personally. They do not want to shame their
Journal or their Alma Mater. #Bookofwhy

12.9.19 1:25am - (Replying to @RonKenett) Recall that transportability theory covers statistical transportability
as well. So, what exactly changed that needs to be "transferred"? Population? Environment? distribution?
#Bookofwhy

12.9.19 12:45am - (Replying to @RonKenett) The fact that they do not deal with causal effects does not make it
"much wider". What does? Can you name a problem area (not an algorithm) that makes the scope of this survey
"wider"? #Bookofwhy

12.9.19 12:27am - Where did you say this "editor" got his education? The quote is a vivid example of how
research is impeded by those who slap careless labels on other people works (yesterday it was NPSEM-IE). As to
content, I have not met a more heterogeneous object than a DAG. #Bookofwhy

12.8.19 11:30pm - I was also surprised to find the context worse. Modern statisticians are so embarrassed by his
text that some resist: No, he wan't one of the founders of statistics. Some even told me: "Well, he was a
socialist." Which proves that dogmas & politics still rule science. #Bookofwhy

12.8.19 5:24pm - (Replying to @AsraNomani) When Linda Sarsour says a Zionist can’t be a
feminist....#zioness #WW84 #GalGadot, she is just confessing to being a FAKE feminist (see
https://tinyurl.com/y4rjozub) unfit to lead any progressive movement. When will Bernie see the harm she is
causing him?
12.8.19 4:29pm - (Replying to @AndersHuitfeldt) Many people (unnamed) interpret the term "independent
errors" (IE) to mean "Causal Markov condition", which is not just "semantics", but a severe restriction. This,
IMHO, is a "gross misunderstanding," perhaps (?) unintended by the authors of the term "IE". #Bookofwhy

12.8.19 4:04pm - (Replying to @AndersHuitfeldt) The "misunderstanding" is on the part of readers who take "IE"
as a restriction on SCM. Giving researchers the OPTION to express independence should not be construed as a
restriction, it is LIBERATION when compared to systems that do not provide such option. #Bookofwhy

12.8.19 3:33pm - (Replying to @AndersHuitfeldt) That's a gross misunderstanding. The Causal Markov condition
is NOT assumed to hold a priori in nonparametric SCM. Labeling it NPSEM-IE (IE ="independent errors") is
misleading. A notational system that gives you the OPTION to assume IE should not be labeled IE. #Bookofwhy

12.8.19 3:08pm - (Replying to @hazqiyal) The bi-directed arrows denote correlated errors for some reasons,
e.g., latent variables with names, un-named latents, selection bias, etc. If you suspect correlation, use them.
Absence of those bi-directed arrows implies error-un-correlatedness for any reason. #Bookofwhy

12.8.19 8:06am - (Replying to @hazqiyal) Almost every diagram in Causality (2000) has bidirected arrows, X< - -
->Y, which stand for correlation between eps_X and eps_Y. #Bookofwhy

12.8.19 3:03am - "Transfer learning" (another word for "transportability") is what called my attention to this
paper http://ccc.inaoep.mx/archivos/CCC-19-004.pdf The title: "Learning subject-specific causal effects" made
me doubly curious: Is this possible? Isn't "class-specific" the most we can hope for?#Bookofwhy

12.8.19 2:09am - Causal discovery enthusiasts among our readers should be interested in this symbiotic
"discover-identify" approach, just posted on our website: https://ftp.cs.ucla.edu/pub/stat_ser/r491-L.pdf
#Bookofwhy

12.7.19 11:41pm - (Replying to @steppenjiff) Corbyn cannot be excused as lacking the brain to foresee the
genocidal consequences of his words and actions. Do you genuinely believe the call for "from the river to the
sea" is anything less than genocidal?

12.7.19 10:27pm - Agree with @GaryMarcus on the insufficiency of causal inference. But the "type-token
distinction" is IMHO well formalized in structural models as "unit level vs. population level distinction" (Causality
p. 310). The firing-squad example in #Bookofwhy shows token counterfactuals.

12.7.19 8:28pm - (Replying to @jon_y_huang) Thanks for posting @diedre_tobias . And I am happy to see that
Jamie and I agree on at least one thing: “We are in the midst of a Causal Revolution!". Our disagreements
dwarf when we remember what it took to get here and the huge work still ahead of us. #Bookofwhy

12.7.19 8:01pm - Luciana Berger is my UK heroine, for standing up to Corbynism.

12.7.19 7:44pm - (Replying to @aminjorati) "voices are not heard?" You must be kidding! Come to UCLA, my
campus, where Palestinian victim-hood is celebrated day and night with megaphones and marching bands, and
where BDS activists chant: “We don't want two states, we want 1948!" You must be kidding!

12.7.19 7:01pm - (Replying to @peder_isager @dingding_peng and 3 others) In cyclic equations, the d-
separation may breakdown when non-linearities are introduced. [It holds for linear cyclic systems]. For
"Identifying Independencies in Causal Graphs with Feedback" see https://ucla.in/2OeKvCe #Bookofwhy

12.7.19 6:49pm - "Zionophobic" is the only word I use for Zionophobes -- a much uglier word than "Anti-Semite".
The latter may often be a victim of misguided upbringing. The former has no excuse, it's a cold & calculated
call for genocide.

12.7.19 6:30pm - There are (at least) two inaccuracies in this paper:(1) Nonparametric SEM does not assume
error-independence (see<-->) therefore, no one should use the acronym NPSEM-IE. (2) NPSEM does NOT assume
that all variables (V ) can be intervened on. See https://ucla.in/2LY8M2X #Bookofwhy

12.7.19 12:13pm - (Replying to @wtgowers) She will always find a reason, or a paragraph, or an adjective to
prevent her from voting for a two-state solution. Why? Such a vote would turn her into a Zionist, thus losing her
voting base. Watch and see how next time she says: It does't include the "right of return".

12.7.19 5:30am - (Replying to @Kaniska_Mohanty) I do recommend Primer. Chapters are available here
https://ucla.in/2KYYviP, and it is a really good book!!! #Bookofwhy

12.7.19 5:26am - Omar, Tlaib, AOC & Pressley -- the four elephants in the room. American law makers have not
internalized a simple fact: Zionophobes cannot utter the words "two-states solution" and Israel's neighbors
suffer from the same speech impairment.

12.6.19 11:00pm - (Replying to @geomblog @L_badikho and 4 others) These people are wonderful, very broad,
and technically superb. But all the King's horses and all the King's men can't turn a mathematical impossibility
into a possibility. "Fairness" requires acknowledging certain impossibilities, else all King's men would be
wasted.#Bookofwhy

12.6.19 1:20am - (Replying to @markcannon5 @AnimaAnandkumar and 5 others) I hope you are not suggesting
that DL is beyond enrichment, or that DL+CI symbiosis is not a compelling way of overcoming DL's shortcomings.
see https://ucla.in/2HI2yyx. #Bookofwhy

12.6.19 1:06am - Continuing this track on Bernie Sanders & Linda Sarsour marriage. Poor Bernie thought
Zionophobes could be kept under leash for some semblance of progressive values. Now he is wondering who
owns the leash, who's "surrogate" to whom?

12.5.19 11:23pm - (Replying to @L_badikho @timnitGebru and 6 others) Agree totally. Logic too originated from
moral disputations. At the same time we should not ignore the fact that current ML proposals for dealing with
"fairness" are dominated by a model-blind thinking, while "fairness" demands a model-based approach.
#Bookofwhy.

12.5.19 11:08pm - (Replying to @L_badikho @timnitGebru and 2 others) I was explicit about "expand their
expertise" which is different from "lack technical expertise". The former is respectful and hopeful, so am I.
#Bookofwhy

12.5.19 10:51pm - Very well put. I am seeing this "ignorance above all" phenomenon first hand at UCLA,
Department of Archaeology. I, too, cast the conflict as a morality play, a simple choice between co-existence
and elimination. A very simple choice.
12.5.19 10:13pm - (Replying to @timelessdev) CI stands for "Causal Inference" as defined and described in
https://ucla.in/2KYYviP or https://ucla.in/2KYvzau

12.5.19 10:06pm - Agree! And the question is: Are ML folks flexible enough to understand that "fairness"
demands expansion of their expertise? Or will they insist on handling it the model-blind way, and waste the
funds and trust that society now pours on them. Same to "explainability #Bookofwhy

12.5.19 7:20pm - "Fairness" involves CI issues such as mediation (ie distinction between direct and indirect
effects), undesirable side effects, and protected variables. It is hard to see how these issues can be managed
without a causal model of reality. #Bookofwhys

12.5.19 7:14pm - Why are "fairness" and ML so often mentioned in the same breath? True, the former can be
compromised by irresponsible practices of the latter, but this does not mean that the tools available to ML folks
are adequate for preventing those compromises. #Bookofwhy

12.5.19 7:01pm - A glimpse at the NeurIPS 2019 program shows that CI is well represented. IBM, for example,
seems to be taking the DL+CI symbiosis quite seriously: https://ibm.com/blogs/research/2019/12/ibm-
research-ai-neurips-2019/. I fail to see though why "fairness" and ML are so often mentioned in the same breath.
#Bookofwhy

12.5.19 6:05pm - I can try and contribute a couple of papers, but I am afraid the title will invite "what
computers cannot do" type of speculating philosophers (eg Penrose, Searle) with whom I do not wish to be
identified. #Bookofwhy

12.5.19 2:41pm - (Replying to @IzaTabaro) Bernie can't decided between the Sarsourish crowd that flocks his
band wagon and the principles he learned at home.

12.5.19 1:57pm - Viewing history from a causal lens is an exciting adventure for me. I will be giving this talk
https://sites.google.com/g.ucla.edu/hpass/speakers to the History & Philosophy of Science Group, at UCLA.
December 12, 3 pm, Royce Hall Room 314. If you happen to be around that day, just drop in. #Bookofwhy

12.5.19 6:36am - (Replying to @BlakeFlayton) On the contrary. She just confessed to being an incurable racist,
hence unfit in any progressive movement. Poor Bernie, choosing "surrogates" turns out more hazardous than one
thinks. Has he given up on attracting non-racist voters. ???

12.4.19 11:29pm - Linda Sarsour (Bernie Sanders' surrogate) is my muse this week. Friday she said: "Israel is
based on supremacy," Yesterday, she apologized: "[No, No]...not the Jewish people. I apologize for the
confusion." I had to tweet her: Sister Linda! Sister Linda!

12.4.19 7:24pm - (Replying to @lsarsour) Sister Linda, Sister Linda! A decent way to take back what you said is
to say what you meant: e.g., "Zionism is NOT what I said! It's a noble dream, a people's homecoming journey,
one we should revere, respect and assist --That's what I meant." Then apologize for the confusion.

12.4.19 5:33pm - (Replying to @RabbiWolpe) Here is the context: "I love Jews, and Bernie is my proof, but I was
bought up to hate, and Zionophobia is safe, and so much chic"

12.3.19 10:03pm - I have thoroughly enjoyed re-reading @EinatWilf 's article a day after Linda Sarsour, the
darling of Fake Feminism (see https://tinyurl.com/y4rjozub), tried again to contaminate Zionism with her racist
g ( p y y j ), g
thoughts.

12.3.19 12:56pm - (Replying to @EpiEllie) I do not recall any paper rejected for having been pre-printed. But I
may be too young to recall all cases, I'll check with my buddies on the Editorial Board.

12.3.19 8:20am - (Replying to @nickchk) This may be a new IV, but the graph-void presentation makes me
commiserate with econ. readers who must make sense of cryptic sentences like: "an instrument correlated with
the confounders, but which itself is not causally related to the direct effect of the treatment." #Bookofwhy

12.3.19 1:44am - (Replying to @DemMaj4Israel) "Israel’s built on idea that every people is entitled to self
determination" - How true! But did you ask Bernie Sanders and Nancy Pelosi if they think about that truth when
they choose advisers and spokespersons?.

12.3.19 1:17am - Being part of a condemned minority is not new to Jewish history, but being a PROUD member
of such minority IS new and, strangely, you feel fortunate not sharing the embarrassment of the Sapiens
delegates voting tomorrow, having to face their families the next day.

12.2.19 3:25am - (Replying to @juli_schuess @thosjleeper and 6 others) On first scan, this is a very valuable
paper, connecting the huge survey-inference literature with causal graphs and showing how crucial design issues
in the former are rooted in causal thinking and can now be formalized and made transparent in graphical
models. Kudos! #Bookofwhy

12.2.19 1:29am - (Replying to @curtatkisson) I'm not familiar with Meng 2018, should I? And who are "they"?
Sorry, I can't guess.

12.1.19 10:40pm - (Replying to @JDHaltigan) I rarely agree with Gelman but, in this case, he is right that the
true effects can be higher or lower than what they found; the same holds under "natural experiments" which are
just observational studies with different assumptions, see https://ucla.in/2mhxKdO, #Bookofwhy

12.1.19 9:43pm - (Replying to @RonKenett) What is not clear to me about the "second dimension" in your paper
is whether the "matching of data resolution/aggregation and analysis goal" is a desirable feature to have, or an

algorithm that actually establishes this matching in some formal setting. #Bookofwhy

12.1.19 7:47pm - (Replying to @AviMayer) It's worse than pure antisemitism, its ugly Zionophobia. It's time we
stop treating the latter as the lesser of the two evils. See https://tinyurl.com/ak4ccg. And let's deny its
legitimacy using its real name, Zionophobia, not antizionism, which sounds almost as innocent as anti-tax

12.1.19 4:55pm - Granularity in Scientific Explanation has been a favorite topic of discussion for many
philosophers, but when cast in a specific representation (DAG), capable of producing explanations, it now
assumes a new level of clarity: http://etheses.lse.ac.uk/3996/1/Kinney__Problem-granularity-scientific-
explanation.pdf #Bookofwhy

12.1.19 4:16pm - A new paper on my desk (i.e., screen), calling for attention, combining linear or nonlinear
conditional independence tests with a causal discovery algorithm to estimate causal networks from real-world
data. https://advances.sciencemag.org/content/5/11/eaau4996 #Bookofwhy

12.1.19 3:22pm - (Replying to @brianoflondon @EinatWilf and 2 others) Fairly gloomy picture, I've suspected
thi ki d f d f ti f it ti A h lth i ti t i t th @B d fD ti
this kind of defeatism for quite some time. Any healthy organization trying to pressure the @BoardofDeputies
into some level of pride and conviction?

12.1.19 8:21am - (Replying to @yudapearl @XiXiDu and 2 others) And too often they treat Antizionism as a semi-
acceptable form of racism, less dangerous than Antisemitism. It isn't.

12.1.19 7:53am - (Replying to @XiXiDu @EinatWilf and @HananyaNaftali) Part of the reason Antizionism has
metastasized among the political left is that OUR leaders and OUR spokespersons so rarely challenge its
legitimacy and barely expose its moral stench.

12.1.19 7:39am - For heaven's sake, why is everyone questioning Corbyn on antisemitism, a question he is so
skillful in deflecting. Why not charge him squarely with incurable Zionophobia, a moral deformity he cannot
possibly hide. @EinatWilf @HananyaNaftali

12.1.19 1:31am - (Replying to @MaccormickIan and @ewenharrison) This is a step-change given current
approaches. I agree. And one way toward making this step-change is outlined here #Bookofwhy. If there are

others, bring them along.

12.1.19 1:13am - (Replying to @ArcusCoTangens @bariweiss and 3 others) Sure there is: The latter is twice
uglier and much more dangerous:

12.1.19 12:04am - The litmus question our spokespersons refrain from asking presidential candidates: "Are you
prepared to state that, as president, you will not support anyone running for office who denies Jews the right to
self-determination in their historical homeland?" So, I am asking instead.

11.30.19 10:26pm - (Replying to @bariweiss @EinatWilf and 2 others) Senator Sanders, Fighting antisemitism
requires courage, not populist talking-points. Are you ready for the Pearl's litmus test: "As President, I will not
support anyone running for office who denies Jews the right to self determination in their historical homeland."
Answer=?

11.30.19 8:35pm - I should not stay silent on this day, Nov. 30, marking the expulsion of Jews from Arab
countries. My wife was one of them, and her story is documented here
https://sfiaccess.usc.edu/Testimonies/ViewTestimony.aspx?RequestID=00893fd7-a7b8-428b-b3c3-449c2791a3ae
in an interview for the Shoah Foundation

11.30.19 1:42pm - I now realize that ML folks must be going through the same trauma. Armed with Logic and
Probability, the two most powerful languages in science, they find it hard to believe that they cannot climb
from Rung-1 to Rung-2 (&-3) in the Ladder of Causation. It's a scandal. #Bookofwhy

11.30.19 1:09pm - (Replying to @prem_k and @GaryMarcus) This was one of the most traumatic realization of
my life, that probability, my mother tongue, and the mother tongue of statistical sciences cannot deal with
cause & effect, the elementary particles of human thought. I do not blame statisticians for chocking over the
#Bookofwhy

11.30.19 5:45am - (Replying to @fdabl) This is a nice summary of CI at its core. Thank you @fdabl for
introducing the subject to a wider audience in such friendly and commonsensical way. You proved my point that
causality is simple if not purposely mystified #Bookofwhy
11.30.19 4:06am - Surprise!! #Bookofwhy is #1 bestseller in Biostatistics.
https://amazon.com/gp/bestsellers/books/227277/ref=zg_b_bs_227277_1 Reasoning statistically, if Bio
statisticians take it seriously, chances are statisticians too are not irredeemable.

11.30.19 3:56am - (Replying to @MarkusSchacher) A standard causal DAG presumes unknown (and
unrepresented) factors u to affect every node in the graph. The firing squad example, however, does not allow
such factors, to demonstrate how counterfactuals are computed in deterministic systems. #Bookofwhy

11.30.19 3:49am - (Replying to @yudapearl @omaclaren and 10 others) Correction. The query is Q: "There exists
a units u s.t. Y(x,u)>Y(x',u)" for some x and x' in the domain of X." The post at http://ucla.in/2qnaII1 estimates
the percentage of such units u. #Bookofwhy

11.30.19 3:19am - (Replying to @MarkusSchacher) You are absolutely right. All of them should be iff. "D only if A
or B" asserts that we rule out other causes of death, except (possibly) A or B". Thanks for pointing this out.
#Bookofwhy

11.30.19 3:13am - (Replying to @bschoelkopf and @ylecun) What Festschrift? I am not going to retire. I still
have 37 productive years (Jewish legend). And someone must convince the remaining 90% of academia that
students should take CI-101 before stat-101. Thanks for a great paper. #Bookofwhy

11.30.19 2:57am - (Replying to @omaclaren @Lester_Domes and 9 others) From seeing just one Y(x) we cannot
conclude that "X causes Y". To conclude, we need to define a new Boolean query Q: "There exist two units, u
and u' s.t. Y(x,u)>Y(x,u')" , and then estimate E[Q] . For one way of estimating E[Q], see
https://ucla.in/2qnaII1. #Bookofwhy

11.30.19 2:33am - (Replying to @yudapearl @omaclaren and 10 others) I expect the TT crowd to hit the roof:
"Yes, but how do you estimate E[Y(x=1)] without an RCT?" Me: "Relax, the question was whether E[Y(x=1)] would
satisfy me, not how I estimate it. I can do the estimation too (see #Bookofwhy), but lets not conflate two
separate questions."

11.30.19 2:23am - (Replying to @yudapearl @omaclaren and 10 others) Fisher, "You really want to "compare"
the new yield to last year's yield, don't you?" Me: "Never mind what I do with that information, it's my business,
just tell me E[Y(x=1)]". Fisher quits, and I remain offering 100 Euro to any statistician who can predict
E[Y(x=1)]. #Bookofwhy

11.30.19 2:11am - (Replying to @omaclaren @Lester_Domes and 9 others) In some cases I would be happy with
just E[Y(x)]. If I am a farmer in Rothhamstead 1924, and Fisher came to me with his RCT, I'd say: Just tell me
what Yield I'd get if I use the new Fertilizer on the entire field. I dont want to Latin Square my field, nor to
randomize. Aha! Says

11.30.19 1:31am - Here is another offering to Lady History in the wake of November 29th, laden with facts and
driven by a question: What kind of Palestine would our Lady wish to see in the Middle East?

11.30.19 12:53am - To close the gap between Big Data and Causality, a panel will be conducted at the IM DATA
conference in Pasadena, where 3 UCLA graduates will show how CI can make data interpretation robust and
meaningful See details: https://tinyurl.com/v34wct7, https://tinyurl.com/td682sx #Bookofwhy

11.29.19 11:18pm - To clarify, what is a "Causal Notion"? It's a notion that cannot be defined in terms of
p y,
distributions over observed variables, so Y(x) is certainly "causal". Harvard's folks would object: "If you find
E[Y(x)]=5, it doesn't tell you if X causes Y." True RCT thinking. #Bookofwhy

11.29.19 10:59pm - (Replying to @omaclaren @Lester_Domes and 9 others) If you are asking me whether I
ultimately want a function Y(x) -- not necessarily; in some problems I would be happy with just E[Y(23)/Y(7)]. In
SCM you get the whole function for free. In Harvard you can't talk about Y(x) w/o some Y(x') to feel part of the
RCT culture.#Bookofwhy

11.29.19 10:47pm - (Replying to @omaclaren @Lester_Domes and 9 others) Change in variance of Y(x) means
that there is at least one individual (labeled u, in Y=f(x,u)) for whom the change is between Y=0 and Y=1. Quite
dramatic. #Bookofwhy

11.29.19 10:42pm - (Replying to @omaclaren @Lester_Domes and 9 others) I can't really speak for what the
others want. I know that many expressed suspicion of a creature named Y(x), or even E[Y(x)], because (my
conjecture) Harvard's graduates are trained, born and conceived in RCT, all else is suspect. TT fashion is part of
this culture #Bookofwhy

11.29.19 7:12pm - (Replying to @kaz_yos) Thanks for posting. I was not aware of this Cookie blog before. And I
am always surprised to see articles on missing data containing no graphs. How do the authors check their
assumptions? Do they really believe in them? Beyond me! #Bookofwhy

11.29.19 6:59pm - (Replying to @omaclaren @Lester_Domes and 9 others) There is a subtlety regarding the
object of discussion. Is it the object Y(x)? Or the exercise of estimating Y(x) or the estimate E[Y(x)])? Do we
have a established standard criterion as to what is "causal" and what is not? I have proposed a crisp criterion.
Objection?#Bookofwhy

11.29.19 3:19pm - (Replying to @omaclaren @Lester_Domes and 9 others) The expression "X causes Y" is used in
scientific conversations, and it often stands for the formal condition "Y(x) is NOT constant in x." But some
people use it more narrowly, to mean E[Y(x)] is not constant, or even E[Y(1)]>E[Y(0)]. #Bookofwhy

11.29.19 2:21pm - (Replying to @omaclaren @Lester_Domes and 9 others) If I, as the author of the DAG, know
that Y(X=x) is a constant, then I have the license to remove the arrow. But putting an arrow there will not count
as "model misspecification". It will count only as under-utilizing one's knowledge. Recall, "May be" is not
knowledge #Bookofwhy

11.29.19 1:50pm - (Replying to @Lester_Domes @EpiEllie and 8 others) Great! No pathology. So we agree that it
is OK to call the function f(x,y) = P(Y=y| do(X=x)) "causal effect of X on Y", though in some extreme cases it may
be independent on x. And despite the fact that some people may confuse it with the difference E(x=1, y)-
E(x=0,y).#Bookofwhy

11.29.19 5:18am - (Replying to @Lester_Domes @EpiEllie and 8 others) Moreover, a null-effect is still a kind of
"effect", same as a flat-response being some kind of a "response", and a "constant function" is still a function,
albeit pathological. #Bookofwhy.

11.29.19 4:01am - (Replying to @Kweku_OA) @Kweku_OA I somehow overlooked your question, sorry. The best
introduction is Pirkei Avot, which stands for "Essays of our Fathers," a collection of ancient Tweets (100-200 AD)
on principles of ethical behavior https://sefaria.org/Pirkei_Avot.4.3?lang=bi&with=all&lang2=en Enjoy.

11.29.19 3:04am - (Replying to @elderofziyon) I usually add one more item: Final ownership to be determined
. 9. 9 3:0 a ( eply g to @elde o yo ) usually add o e o e te : al ow e s p to be dete ed
by Palestinians readiness to spell the word "co-existence".

11.29.19 2:53am - (Replying to @MarkusSchacher) Consider the Firing Squad example in #Bookofwhy page 39.
Bidirected logical implication exists between any two variables. Yes causal implication goes only along the
arrows.

11.29.19 2:49am - From Deep Learning to "Deep Feeling: AI and Emotions." This mind-blowing exhibition should
be of interest to our artistic readers: http://petachtikvamuseum.com/en/Exhibitions.aspx?
eid=4987&utm_source=InforuMail&utm_medium=email&utm_campaign=Culture+Newsletter+November+2019
Petach Tikvah is not far, just 10 km from Tel Aviv. Founded in 1878 by 5 orthodox Jewish cowboys. Worth a
visit. #Bookofwhy

11.28.19 9:35pm - (Replying to @MarkusSchacher) No relation. Logical implication is invariant under contra-
position, not so causal implication. #Bookofwhy

11.28.19 5:48am - (1/ ) Wishing all readers a joyous and meaningful Thanksgiving weekend, and adding a
personal note for Friday, November 29. I have a secret contract with Lady History: She forgives my moments of

weakness and I remember her moments of grace. One such moment was November 29, 1947
11.28.19 5:48am - (2/ ) when I, my family, our neighbors and the whole town heard the voice from the radio:
"Yugoslavia abstained! The resolution [UN 181] has passed!" And suddenly, a terrifying shout, 80 generations of
age, came out of the throats of everyone in town, including pets, fish, and
11.28.19 5:48am - (3/3) babies in their cribs: WE HAVE A STATE !!!!!!!! To fulfill my obligation to Lady History, I
am sharing a 2018 oped that further describes the significance of that moment: https://tinyurl.com/y8w2qcrg
Happy Thanksgiving!

11.27.19 11:40pm - (Replying to @ConiByera @EpiEllie and 8 others) Not only who are "we"?. What are the
"counterfactual framework?".

11.27.19 8:22pm - (Replying to @ConiByera @LucyStats and 8 others) What I called "causal effect" (in Causality
2000) was the entire sequence: Y(X=x) x=1, 2, 3, ... or E[Y(X=x)] x=1,2,3, .... From which you can form
differences eg E[Y(X=5)]-E[Y(X=3)] or play with ratios, eg E[Y(X=2)]/E[Y(X=1)], or follow your fancy. #Bookofwhy

11.27.19 3:27pm - (Replying to @EpiEllie) Eureka !!! Eureka!! With the help of another colleague I now
understand what Target Trial (TT) is! And I also understand why no TT expert could explain it to me. It is really
a "framework", a misguided one, but a "framework". I will soon post what I've found. Eureka! #Bookofwhy

11.27.19 8:08am - (Replying to @robertwplatt @rohitpojha and 5 others) I cannot work through these question
in the TT framework, because no one would tell me whether TT constrains me, and how. Perhaps someone
versed in TT can do it and tell us if there is anything wrong in assuming no constraints. IOW TT permits us to
forget about TT. #Bookofwhy

11.27.19 7:58am - (Replying to @LucyStats @AlexaYakubovich and 7 others) I can't believe this is the official
position of TT. Are you saying E[Y(X=0)] and E[Y(X=1)] may not be causal, and the difference E[Y(X=0)] -
E[Y(X=1)] is causal just because we have a comparison? We need a top TT expert to confirm. You see why I am
less than enthused? #Bookofwhy

11.27.19 7:49am - (Replying to @robertwplatt @rohitpojha and 5 others) I read the literature and I acknowledge
that many find it helpful as a way of thinking. The answer I do not get is "What is it?" New or old, what is the
idea? Is each of my aspirin questions legitimate though it invokes no comparison no trial and no randomization?
idea? Is each of my aspirin questions legitimate though it invokes no comparison, no trial and no randomization?
#Bookofwhy

11.27.19 7:38am - (Replying to @AlexaYakubovich @LucyStats and 7 others) I did not say they aren't, I merely
asked the adherents of TT to tell us if each of those questions IN ISOLATION, each invoking no comparison, no
trial and no randomization, would be a legitimate question to ask. Simple question. #Bookofwhy

11.27.19 7:02am - (Replying to @LucyStats @AnjaLeist and 6 others) Before I decide what to do with the
predictions under 1 tablet of aspirin, 2 tablets, 0 tablets etc etc, I want to make sure that each of these
predictions is a legitimate question to ask. (Regardless of how we establish them). Q:Are they legitimate
questions under TT? #Bookofwhy

11.27.19 5:46am - (1/ ) (Replying to @AnjaLeist @rohitpojha and 5 others) You are the first to tell me what
"target trial" is, thanks! But you are imposing a severe restriction on the kind of questions we can ask,
forbidding me from asking for the probability that my headache will subside in 1 hour if I, given my medical
history etc, decide to take
11.27.19 6:08am - (2/ ) (Replying to @yudapearl @AnjaLeist and 6 others) an aspirin, just because my question
invokes NO COMPARISON. I am talking about ASKING, not about establishing the answer. I now look with great
compasion at the oppressed adherents of the "target trial" framework, forbidden from asking the questions they
really wish to ask and
11.27.19 6:08am - (2/2) (Replying to @yudapearl @AnjaLeist and 6 others) forced to ask only questions that
someone put in their mouth. I can't join this constraining culture; would you? And all for the sake of emulating
RCT? I can't. A good framework is one that allows researchers to articulate what they truly wish to know.
#Bookofwhy

11.27.19 5:11am - (Replying to @luketrailrunner) My indicators would be: (1) Explicit mention of commitment to
causality research and acknowledgement of its crucial role, in the Institute's Statement of Purpose. (2) Inclusion
of genuine CI researchers on the institute's Advisory Board. Without the two, we're back to statistics.

11.27.19 4:59am - (Replying to @rohitpojha @EpiEllie and 4 others) No objection! I am genuinely trying to
understand what it is. Since it is marketed and hailed as a "new framework," I fear to be missing something by
not adopting it. Am I? Or can I and my students relax back to the framework that kept us happy and productive
thus far? #Bookofwhy

11.27.19 3:29am - (Replying to @rohitpojha @EpiEllie and 4 others) We agree on what is needed to predict the
consequences of taking an aspirin, we are both causal inference experts, so we have no differences here. My
humble question is "What would we lose if we purge the phrase "target trial" entirely from our language?" What?
#Bookofwhy

11.27.19 3:12am - (Replying to @_JakeHumphries @EpiEllie and 5 others) I am using metaphors when I wish to
describe someone who is committed to consistently violate the very essence of the "target trial" philosophy, and
I humbly ask "Where would this someone go wrong?". Well, where? #Bookofwhy

11.27.19 3:04am - (Replying to @_JakeHumphries @EpiEllie and 5 others) I need analysis (and data) to estimate
the probability that my headache will subside next half hour if I take aspirin. Note: I did not mention "trial". I
will mention "trial" if I use RCT to interrogate nature. But I would never use "target trial" in the same breath.
#Bookofwhy

11.27.19 2:43am - (Replying to @EpiEllie @rohitpojha and 4 others) I will bend to this nuance and talk only
about my target being "an intent to implement a policy". Where would I go wrong viewing this intent to be my
TARGET, not a TRIAL, vowing never to mention TRIAL when I define what I wish to estimate. Would such heresy
hurt me? #Bookofwhy

11.27.19 1:30am - (Replying to @EpiEllie @rohitpojha and 4 others) The podcast appears useful to many in the
audience. But humble me still wonder what error would I make if I continue to view RCT as in #Bookofwhy, ie, a
means to interrogate nature, not a target in itself, and I view my target to be a policy, not a trial. Where would
I go wrong?

11.26.19 11:11pm - (Replying to @rohitpojha @AbbyCScience and 2 others) I have asked epidemiologists
colleagues if they knew what "target trial" is. They smiled and admitted "not a clue". Can you explain it to the
uninitiated? What if one's ultimate goal is not a "target trial" but a "target policy? Would anything be missed?
#Bookofwhy

11.26.19 10:00pm - I am retweeting this reply because of its general concern to many CI researchers. Public
funds are pouring to many "Data Science Institutes" which are likely to continue the exclusive "data-centric"
focus of DL/ML unless an explicit commitment is made to snap out of it.#Bookofwhy

11.26.19 5:28pm - (Replying to @ylecun and @bschoelkopf) I see dozens of "Data Science Institutes" erected
across the country, I read their manifestos and I check their advisory boards. Causality does not seem to be on
their agenda. Which makes one doubt whether the Ladder has been internalized and where this hype will end.
#Bookofwhy

11.26.19 4:24pm - Now that we all know how to compute counterfactuals (e.g., https://ucla.in/2NceWNb), we
can take a step back and enjoy Silvia Karasu's article (Psychology Today) on the poetic side of Counterfactual
Thinking and Imagining: https://sylviakarasu.com/wp-content/uploads/counterfactual-thinking-imagining-what-
might-have-been.pdf. #Bookofwhy

11.26.19 3:56pm - (Replying to @EricTopol) Thanks for the kind words, Eric. I am usually against humility in
science, because so much CAN be accomplished. My point, however, is that to climb the Ladder one has to look
at the Rungs. #Bookofwhy

11.26.19 2:43pm - There's much truth to what you're saying. The idea that there are theoretical impediments to
ML methods is hard for ML folks to internalize.And repeated assurances that causal inference is just one aspect
of what ML has been doing all along do not encourage them to try.#Bookofwhy

11.26.19 1:41pm - (Replying to @ArielOrtizBobea and @causalinf) This is fair, with two additional touches. (1)
among those "complex relationships" rests this one: "is the effect estimable by any new method?" which DAGs
now automates. (2) "intractable" practically means "undoable", like coloring a map with 3 colors. #Bookofwhy

11.26.19 4:51am - A very comprehensive, delightful and inspiring paper. Recommended to ALL, not just MANY
ML/AI folks. Note also that @bschoelkopf does not perceive me as "polarizing the field" as suggested by @ylecun
here: https://twitter.com/ylecun/status/1198319448320548865. SCM unifies, invites and educates. #Bookofwhy

11.25.19 9:53pm - (Replying to @ArielOrtizBobea and @causalinf) Ed Leamer asked me the same question. Take
your favorite algebraic representation, say 3 structural eqs., and ask your students if it has any testable
implication, or, which parameter is estimable by OLS, or identified by some IV. Its doable, yes, but have they
been asked before?
11.25.19 9:40pm - (Replying to @causalinf and @ArielOrtizBobea) Indeed, visualization is only one tiny feature
of DAGs. The feature that is least appreciated by economists is computation. "So what" some say, "It takes me
longer, but I'll get it!" Wrong! Intractability induces helplessness! They just wont try! They try only tractable
problems.

11.25.19 6:59pm - (Replying to @EpiEllie) I've never met a "clinician", I only heard about them. What epi
principle are they missing?

11.25.19 1:47pm - (Replying to @ChrisCrandall16 @joel_saarelma and 3 others) My point is that the arguments
that convinced psych departments to include stat classes in their programs (1/2 century ago) are doubly more
compelling when applied to causal reasoning classes. #Bookofwhy

11.25.19 1:37pm - (Replying to @osazuwa @ang_hermann and 3 others) Conditional Policies and Stochastic
Policies are discussed in Chapter 4 of Causality. It shows that if we know P(y|do(x), z) then we can compute the
effect of a policy X=g(z) i.e., Act X in response to observation Z=z. See also Dynamic Process Control p. 74.
#Bookofwhy

11.25.19 1:18pm - (Replying to @learnfromerror) I've read what Spanos writes on Simpson's paradox. But do you,
as a philosopher of statistics find it satisfactory? e.g., sufficient for determining the correct action in any of the
examples discussed here https://ucla.in/2Jfl2VS ??? #Bookofwhy

11.25.19 5:44am - (Replying to @joel_saarelma @ChrisCrandall16 and 3 others) PRIMER is taught in high schools,
https://ucla.in/2KYYviP Are psych students less deserving? Introduction to causal inference is recommended
before stat 101. #Bookofwhy

11.25.19 3:13am - (Replying to @RonKenett) Let's start with the virtues of automation as a tactical effort, then,
if we are lucky and succeed to emulate Ron Kenett on a computer, I would not reject the extra benefit, i.e.,
understanding scientific thought, understanding ourselves, and finding a cure for cancer.#Bookofwhy

11.25.19 3:05am - (Replying to @RamiEmad1988 and @JustinTrudeau) "Not for the civilians" !!!. The credibility
of anyone reporting from the Middle East hinges upon one simple litmus test: Willingness to ascribe some human
qualities to Israelis, however feeble. Zionophobes never do.

11.25.19 1:27am - (Replying to @RonKenett) The goal of automation forces you to think about your thinking as
an object demanding understanding at the input-output level. Informal thinking, however decorated, "critical"
or not, is still thinking "from within". #Bookofwhy

11.25.19 12:21am - (Replying to @RonKenett) Aspirations to formalize then automate human thinking did more
for our understanding of human thinking than all the thinking that was done before. We might never get there,
but the fall outs of trying far exceed the dead-ends of old-fashion "thinking". #Bookofwhy

11.24.19 7:40pm - (Replying to @mdubowitz) Reform Khomeini-style makes me crave for Reza Shah.

11.24.19 9:25pm - (Replying to @learnfromerror) This has been my observation too, but isn't it the same when
statisticians say "this is a subject matter issue", thus avoiding a position on which claim is warranted and which
is not. Recall how statisticians have been avoiding Simpson's paradox https://ucla.in/2Jfl2VS #Bookofwhy
11.24.19 7:28pm - (Replying to @joe_miller0 @_pierreblanc and 8 others) Great question. The functions used in
the estimation stage (say a regression line) approximate a population. Those that operate behind the arrows
govern each individual unit. The finite sample accuracy of the former can be quantified, not so the latter.See
Cslty p. 35. #Bookofwhy

11.24.19 5:12pm - (Replying to @joe_miller0 @_pierreblanc and 8 others) Or with finite data combined with a
ML wizard.

11.24.19 5:06pm - (Replying to @joe_miller0 @_pierreblanc and 8 others) Given the X--->Y graph, no functions,
we get P(y|x) and P(y|do(x)) from the data. We need the function y=f(x, eps) if we want Y_x (eps). #Bookofwhy

11.24.19 4:49pm - @NicoleBarbaro . A most generous compliment, thanks! I only wish economists start to view
themselves as social scientists. #Bookofwhy

11.24.19 4:04pm - (Replying to @RamiEmad1988 and @JustinTrudeau) I wrote: "The truth about what Israel has
done for Syrian refugees", not Zionophobic propaganda.

11.24.19 4:00pm - (Replying to @joe_miller0 @_pierreblanc and 8 others) Depending on what "causal model" we
have. The graph in itself (with data) can only give us interventions. We need the functions behind the arrows to
get countefactuals. It is neatly explained and exemplified here: https://ucla.in/2G2rWBv #Bookofwhy

11.24.19 2:52pm - Twitter is a powerful provoker of sound bites. I wish this one gets inscribed on my tomb: "We
need to see the Rungs before we climb the Ladder" #Bookofwhy

11.24.19 2:43pm - (Replying to @prem_k and @Atmaprajnananda) It might have some connection, but it is too
spiritual for me. I am an input-output poet. #Bookofwhy

11.23.19 7:03pm - (Replying to @JakeSearcy) how incredibly useful curve fitting can be, and how incredibly
easy it is to see its limits.

11.23.19 4:47pm - (Replying to @dileeplearning) Do we have a Turing test for 'symbolism"?

11.23.19 2:40pm - What I am saying about ML/AI and "curve fitting" is not really aggressive, nor is it my saying;
it is an innocent message sent to us from the mathematics of cause and effect. We need to see its Rungs before
we climb the Ladder, and it is on p.28 of #Bookofwhy

11.23.19 6:09am - (1/ ) This is a key insight. This mysterious place holder got different names in different
disciplines. Statisticians called it "subject matter information", some used "Bayesian priors." Economists called it
"economic theory", etc. etc., Tell us more, about what pushed philosophers
11.23.19 6:09am - (2/ ) to postulate this animal. It comes up again and again, what was suppressed into
namelessness turns out to be none other but a causal model, mathematizable, well understood, and definitely
not mysterious. #Bookofwhy

11.23.19 2:07pm - (Replying to @Ace_Harsh) Kindly explain what you mean by "relative".

11.23.19 2:01am - (Replying to @JustinTrudeau) Respectfully, denunciation is what racists love to hear. The
way to neutralize this mob is for your office to echo the message that this speaker came to tell York University:
the truth about what Israel has done for Syrian refugees. They fear truth, not denunciations.

11.22.19 10:16pm - (Replying to @adamdedwards and @Yu_Ke__) If customers of difference races enter the
bank at random, then "race" becomes a "random variable" in the eyes of the bank teller. "Interaction" does not
make it "complicated" or "metaphysical". #Bookofwhy

11.22.19 10:08pm - (Replying to @analisereal) The analogy is not perfect. When you use a slide-rule the
assumptions are as clear as English, .... #Bookofwhy

11.22.19 2:35pm - (Replying to @Yu_Ke__ and @adamdedwards) Suppose you apply for a loan in a Nazi-own
bank and then you complain: "Why was my application denied?" Wouldn't the answer: "Because you are not
Arian" be an honest answer? (albeit not exactly "fair") #Bookofwhy

11.22.19 3:34am - I remember another quote related to Anderson's "correlation supersedes causation". It goes:
"Correlation can get us to Babylon, but not to Athens". I used it here: https://ucla.in/2wj4pox . #Bookofwhy

11.22.19 3:15am - (Replying to @Yu_Ke__ and @adamdedwards) Being a member of one race or another surely
CAUSES racially biased people to act differently. Why deny them that honor? #Bookofwhy

11.22.19 12:51am - (Replying to @ERMANigeria) I am not sure what Anderson meant by "supersedes" but, if we
judge by correlation-based statistics, it hasn't advanced much since Pearson if we count the number of
regression analysts who can resolve Simpson's paradox. eg., https://ucla.in/2Jfl2VS. #Bookofwhy.

11.22.19 12:15am - As I read this post today, 5 years after, I am amazed at how little things have changed in
the eco-bubble. Glad this Twitter post helps grad students in economics see what they need to rebel against.
#Bookofwhy

11.21.19 11:07pm - Sharing an interesting paper: http://papers.nips.cc/paper/8547-identifying-causal-effects-


via-context-specific-independence-relations.pdf. The need for context-specific independencies emerged in the
1980's to supplement d-separation. This paper extends it to causal inference using annotated graphs; it seems
more transparent than specifying PO's. #Bookofwhy

11.21.19 10:40pm - (Replying to @ma_macneil @rlmcelreath and 2 others) I would not use the phrase "causal
salad" because it connotes a complex and hotly contested topic. It isn't. Causal Inference is embarrassingly
simple and coherent, despite some who are trying to mystify it. Dont let them. #Bookofwhy

11.21.19 10:23pm - (Replying to @Yu_Ke__ and @adamdedwards) Race has effects, same with earthquake. And
these effects exist "beyond correlation". #Bookofwhy

11.21.19 10:03pm - (1/ ) (Replying to @JessGeraldYoung @WhitneyEpi and 9 others) 1/ I would separate
conceptual understanding of mediation from its identification. After all we would only labor to identify
something if it is meaningful once identified. My favorite conceptualization is: "The fraction of cases OWED to
mediation" (necessity) and, likewise,
11.21.19 10:07pm - (2/ ) (Replying to @yudapearl @JessGeraldYoung and 10 others) 2/ "The fraction EXPLAINED
BY mediation" (sufficiency). Both are defined, discussed and exemplified here https://ucla.in/2N9IA5A. Now,
given that our scientific knowledge (SCM) assigns values to ALL counterfactuals, I do not see any impediment to
conceptualize mediation in terms
11.21.19 10:17pm - (3/3) (Replying to @yudapearl @JessGeraldYoung and 10 others) of nested counterfactuals -
identification comes later. And if my meaningful effect is not identified, I would approximate, but I would not
de t cat o co es late . d y ea g ul e ect s ot de t ed, would app o ate, but would ot
sacrifice meaning on the alter of identification. @Bookofwhy ps. Check https://ucla.in/2N9IA5A, it's a good
paper, obscured by Wiley

11.21.19 12:35pm - Interesting connections between tractable versions of Bayesian networks and causal
diagrams https://kr2ml.github.io/2019/papers/KR2ML_2019_paper_38.pdf, despite the somewhat negative
conclusions. #Bookofwhy

11.21.19 3:32am - Gee, where did I get the time in July to answer Quora type questions? But, in retrospect, it's
a good answer http://bit.ly/2Wfxm0s, which has withstood the test of time. Glad to share it on Twitter.
#Bookofwhy

11.21.19 1:49am - I am not a health scientist, nor am I around NYC today, but I would surely love to hear the
latest in "causal data science". It sounds like a new building Columbia is constructing. #Bookofwhy

11.20.19 9:13pm - (Replying to @stewarthu and @mark_vdlaan) We do not have strict guidelines on this
distinction. Our main criterion is: Would other readers be able to learn some lesson (or a tool, or an
impossibility) that would help their next research problem. #Bookofwhy

11.20.19 8:42pm - The enthusiastic interest in Rebane's Polytree Algorithm (118 likes) reminds me of another

beautiful discovery algorithm. Suppose you have a tree, and observe only its leaf-nodes. Can you discover the
internal-structure of the tree? A miracle! https://ucla.in/2q6zmJj #Bookofwhy

11.20.19 8:26pm - Please welcome @mark_vdlaan who has just joined our education channel. Mark is a friend of
commonsense, and a co-founder of the Journal of Causal Inference, which is about to become a serious
competitor of Econometrika (for reasons we have discussed here yesterday) #Bookofwhy

11.20.19 7:55pm - Fully agree! The do-operator operates on variables in the context of a causal model of
reality. The recipient of the operator can be a state (eg temperature, race) or manipulable event (eg. take
drug). In https://ucla.in/2LY8M2X I explain the importance of the former. #Bookofwhy

11.20.19 5:35pm - (Replying to @renatrigiorese) Agree. "Deep associationism" is a more accurate term.

#Bookofwhy

11.20.19 3:53pm - (Replying to @DaniCMBelg @lgmoneda and 3 others) From what I can see, the paper by Mason
etal is not using DAGs and, as a result, the assumptions are opaque and theoretical guarantees are absent. By
DAGs I mean a model of the reasons for missingness, as is described here: https://ucla.in/2LdEjZW .
#Bookofwhy

11.19.19 11:00pm - Still another historic day: Our Twitter counter tells me that this station has reached 25K
followers. Not bad for an education channel that started 17 months ago, June 27, 2018. I thank all readers for
keeping me awake, and for giving me the illusion of being useful. #Bookofwhy

11.19.19 10:45pm - Another historical moment I must retweet, since I am old enough to vividly remember Sadat
speech at the Israeli Knesset in 1977, see https://youtube.com/watch?v=_tnQvgJX7OQ, it gave my countrymen
an existence proof to hope that, one day, a Palestinian leader will have such courage.
11.19.19 10:24pm - (Replying to @wagonomics and @amitabhchandra2) Interesting theory, but if we read
carefully what economists write about regression (see eg https://ucla.in/2NRFn7e) I am not sure the "long CI
traditions" have helped them much, especially these days, when those traditions are trashed the "natural
experimentalists" #Bookofwhy

11.19.19 9:58pm - (Replying to @wagonomics and @amitabhchandra2) Only history will judge if econ. editorial
leadership is endowed with the kind of insight you describe: "to see important papers using a new tool and
become convinced a review paper will be impactful". My job is to alert them to existing gaps and growing
insularity. #Bookofwhy

11.19.19 9:23pm - A proverb one can't help but retweet. I only wish it comes true in our lifetime, not like other
Biblical promises.

11.19.19 7:20pm - (Replying to @yudapearl @wagonomics and @amitabhchandra2) And, moreover, "pressuring"
the editorial establishment is simply not in econ. culture; academic survival means tribal allegiance, not
"pressure". #Bookofwhy

11.19.19 7:14pm - (Replying to @wagonomics and @amitabhchandra2) I hold fellow economists to a higher
esteem. I know they do not need a minimum-wage example to recognize the importance, in their own research,
of mediation analysis and external validity. Yet it would not occur to them that other fields are already 10 years
ahead in these tasks.

11.19.19 7:00pm - (Replying to @AlexAlvPerez) I will never get tired recommending Primer. It is essentially
available free, here https://ucla.in/2KYYviP. Enjoy! #Bookofwhy

11.19.19 4:45pm - @amitabhchandra2 What you are saying is very encouraging to me. Still, as an Editor, I would
feel it ain't right if my readers have to wait 10-20 years for an amazon book to keep up with the latest. I wonder
if other editors are aware of how econ. is perceived. #Bookofwhy.

11.19.19 4:01pm - (Replying to @chrisalbon and @databozo) Recall, every structural model confers an
unambiguous truth value to every conceivable counterfactual. Need a ruling? Unveil your model. #Bookofwhy

11.19.19 1:35pm - Discovering Polytrees. Rebane's algorithm (1987) was the first attempt to uncover causal
structures from raw data. I was gratified today to see that it was extended to include hidden variables:
http://papers.nips.cc/paper/9647-an-algorithm-to-learn-polytree-networks-with-hidden-nodes.pdf.
#Bookofwhy

11.19.19 5:13am - (Replying to @elderofziyon) Settlements are not THE obstacle to peace, because there is a
ten time taller obstacle there. Yet by diverting attention from the taller one they make it harder to remove, or
even to point to.

11.19.19 4:46am - (Replying to @mikejohansenmd @amitabhchandra2 and @Kweku_OA) Highly accessible review
articles are available in all disciplines (eg https://ucla.in/2N9f28c https://ucla.in/2MpjsT6) except economics.
We know for a fact that economists do not read such articles unless they are home blessed, and NBER is an
exclusive club. #Bookofwhya

11.19.19 4:15am - (Replying to @amitabhchandra2 and @Kweku_OA) I am more concerned about the fact that in
the past half a century the econometric literature has not invited/published any REVIEW ARTICLE on graphical
models, a methodology that other disciplines have adopted as a second language. This would worry me as an
Editor. #Bookofwhy

11.19.19 3:28am - (Replying to @Kweku_OA) Thanks for bestowing hope onto economics. I am still waiting to
hear from @amitabhchandra2 (former Editor of RES) if he agrees that editors should assume greater leadership
in de-insulating their respective fields. https://twitter.com/yudapearl/status/1195345400384720896.
#Bookofwhy

11.19.19 2:11am - (Replying to @arinbasu and @UCNZ) Spoiling the young is what makes book-writing such a
thrill.

11.19.19 1:53am - (Replying to @StephStammel) Or, at the very least, not spoil it with statistics.

11.18.19 6:21pm - Hillarious! I can't forgive my mother for denying me such Bayes'ic rights! But I took revenge in
this cartoon: https://britannica.com/biography/Judea-Pearl/images-videos

11.18.19 6:04pm - (Replying to @EinatWilf) As much as I think Israel’s settlement project is its greatest folly,
this much is true: People have a right to a land to the extent they bestow that right to their neighbors. And that
works both ways.

11.18.19 2:02pm - (Replying to @deboerk07) This is indeed interesting, the first "innovation" that came to her
mind was about causal inference. She will soon see more innovationis, now that @eliasbareinboim joined
Columbia. #Bookofwhy

11.18.19 4:01am - Sport tends to make miracles: First time I hear the Israeli anthem (Hatikva = "The Hope")
played by Arab musicians, sending shivers in my spine. It's fusing two hopes: (1) "to be a free nation" and (2) to
make peace with our neighbors. The first is 2,000 yrs old, the second 71.

11.18.19 3:41am - An interesting new paper just crossed my screen:


https://tandfonline.com/doi/abs/10.1080/07350015.2019.1624293, connecting graphical models to machine
learning routines in an economic journal. A word of caution: The black-box is not really black, it is back-door
admissible. #Bookofwhy

11.17.19 8:52pm - For our Argentinian scholars and international soccer fans, a rare picture of Leo Messi landing
in Israel, from a Jordanian Airline plane, despite rockets from Gaza and in defiance of BDS cacophony. More
soccer may be the treatment needed and good will the effect modifier.

11.17.19 3:37pm - (Replying to @stephensenn @RonKenett and 2 others) For AI-ers, "comes from experience" is
not a sufficient answer, because experiential knowledge must reside in one's mind to be usable. Notice that,
taking your description literally, a regression model is all we need to automate your thoughts. Not sure you
would agree. #Bookofwhy

11.17.19 3:26pm - (Replying to @yudapearl @RonKenett and 3 others) So, my next project is to read Cox (1958).
Give me a week to see if I can translate Cox's hints to encodable knowledge. This conversation reveals a basic
disconnect between "designers" and AI-ers. The former THINK, the latter WORRY: Can I automate my thoughts?
#Bookofwhy

11 17 19 3:16pm - (Replying to @RonKenett @f2harrell and 2 others) I am not referring to re-randomization I


11.17.19 3:16pm - (Replying to @RonKenett @f2harrell and 2 others) I am not referring to re-randomization. I
am referring to the knowledge that makes @stephensenn smarter than a novice or a coin flipper. I am unable to
extract from him the nature of that knowledge but, U'r right, Cox has lots of comments about where this
knowledge comes from.

11.17.19 2:26pm - (Replying to @RonKenett @f2harrell and 2 others) Of course blocking and randomization are
different. Still, compare an expert designer who assigns blocking scientifically to a novice, who assigns things
thoughtlessly, say by flipping a coin. If the former does better, why? What does he have? Knowledge?Calculator?
DAG? #Bookofwhy

11.17.19 2:15pm - (Replying to @stephensenn @f2harrell and 2 others) The reason I was changing from "hair
color" to "eye color" was to guess the general criterion for choosing a blocking factor. But if it is so
simple:"CONSIDERED PREDICTIVE" we can automate it, all we need is a model of predictiveness." Now you will
say NO & I'm back to guessing

11.17.19 1:53pm - (Replying to @RonKenett @f2harrell and 2 others) If human researchers can assign blocking
better than a random coin, and they have no more than a mind-DAG to guide them, then the answer to your
question is "Of course!". If they use more than a mind-DAG, what is the extra ingredient? Lets capture it on a
computer. #Bookofwhy

11.17.19 7:32am - (Replying to @f2harrell @stephensenn and 2 others) @f2harrell . If you understand
"experimental design" (I don't) perhaps you can help me unveil the mental process by which @stephensenn
decides, prior to seeing data, that "age" will be a better block than say "eye color" for reducing error in
predicting heart attacks. #Bookofwhy

11.17.19 5:31am - (Replying to @stephensenn @RonKenett and @learnfromerror) Whats the point of challenging
DAGs to do this or that when I am trying to understand what this or that is? I am coming to you as a student,
and you take a defensive position as if I was a threat. I am interested only in the mental process behind the
choice of blocks.#Bookofwhy

11.17.19 5:22am - (Replying to @yudapearl @RonKenett and 2 others) I am trying to learn what this mental
process is, so that you and me can decide if it is within statistics on not. So, please be patient with me and help
me understand this mental process. Lo Hakapdan Melamed (Hebrew for: "The uptight makes a bad teacher").
#Bookofwhy

11.17.19 5:02am - (Replying to @RonKenett @stephensenn and @learnfromerror) I did not say "statistics does
not have the tools to handle blocking." I said the phrase that you used: "this is one of the important areas of
statistics" was also used about causality, since 1975. And I now understand that it meant: "statistics cannot
handle causality" (cont.)
11.17.19 5:07am - (Replying to @yudapearl @RonKenett and 2 others) which was true, it did not have the
language. Whether statistics has the tools to handle blocking depends on whether the mental process by which
you decide (prior to examining the data) that one factor is a better block than another can be articulated in
statistical language.

11.17.19 2:44am - For the many readers of our education channel, this new post https://ucla.in/2qnaII1
introduces Fréchet inequalities using modern visualization techniques and discusses their applications and
their fascinating history. Enjoy, and thank @smueller for the movies. #Bookofwhy

11.17.19 2:04am - (Replying to @RonKenett @stephensenn and @learnfromerror) Funny. I have heard the phrase
" one of the important areas of statistics" reiterated since 1975, about causality. And I now understand that it
meant: "Sorry, statistics does not have the tools to handle it." Nowadays, with new tools, I hope to capture your
meant: Sorry, statistics does not have the tools to handle it. Nowadays, with new tools, I hope to capture your
intuition. #Bookofwhy

11.17.19 1:48am - (Replying to @stephensenn @RonKenett and @learnfromerror) This is precisely what we,
causal inference students, are trying to learn from you. I, for one, do not understand the vocabulary of "design",
"block" "within subject studies" "inherently predictive" etc. But once we learn your intuition, we will automate
it, promise. #Bookofwhy

11.17.19 1:37am - (Replying to @stephensenn @RonKenett and @learnfromerror) Is "inherently predictive"


different from "predictive" or "highly associated with"? Recall, this judgement is made BEFORE we collect any
data, so it cannot be defined circularly as e.g., "choose whatever improves precision". #Bookofwhy

11.17.19 1:28am - (Replying to @RonKenett @stephensenn and @learnfromerror) Your paragraph makes students
(me) believe that you need to try out both "age" and "hair colors" on the data before deciding. But you surely
advise them not to waste time trying out every possible measurement. What intuition guides you prior to taking
data? #Bookofwhy

11.17.19 1:16am - (Replying to @RonKenett @stephensenn and @learnfromerror) I look back at my Tweet and
find nothing there about causality, just an innocent question: "what makes 'age' a better block than say 'hair
color'?" . How do you explain to students in class, BEFORE seeing any data, that the former would be better for
estimation. #Bookofwhy

11.16.19 10:39pm - (Replying to @RonKenett @stephensenn and @learnfromerror) @RonKenett . Why is it that
design experts tend to refer me to authorities when I ask "what makes 'age' a better block than say 'hair color'"
Why cant I find one to tell me what a 'good block' means to him/her PERSONALLY, or how they explain it to
their students? Shy? #Bookofwhy

11.16.19 9:24pm - Another thing that would help us get out of the "well-defined intervention rut" is to recall
how the rut came about. In 1974 Don Rubin defined PO in the context of "treatment assignment", which
appealed to statisticians and epidemiologists. In 1973, however, David Lewis ...
11.16.19 9:24pm - defined counterfactuals (ie, PO) in terms of "alternative worlds", a metaphysical conception
that only philosophers could buy. The two conceptions are unified in SCM (https://ucla.in/2N9nSCV) but neither
field managed to cut its umbilical cord. Glad U'R "getting out". #Bookofwhy

11.16.19 7:19pm - Thanks for the original text. I was quoting from memory, and was sure there was some
reference to "Rabbi". But "honor" serves the same purpose in our secular world. "Who is honored? He who honors
his fellow human beings".(Ben Zoma).

11.16.19 6:28pm - (Replying to @ProfStiff and @marclamonthill) Marc Lamont Hill will continue to spew BDS-
hate with impunity until someone dares call him a "Zionophobic Racist", which may perhaps convince him to
look at himself in the mirror.

11.16.19 6:13pm - This Tweet helped me formulate an exciting agenda for philosophers of statistics: "Go over
the statistical literature of the 20th century and explicate, formalize and algorithmatize key concepts which
statisticians dismissed as "beliefs" "subjective judgement" or "too Bayesian"

11.16.19 6:01pm - Sharing a new refinement on "what is a cause", and "Is race a cause" and "why does it
matter?" https://academic.oup.com/ije/advance-article/doi/10.1093/ije/dyz229/5621469 #Bookofwhy
11.16.19 2:57pm - (Replying to @JBlevins0 @RonKenett and 2 others) Beautifully stated: "approaches the
experiment with some beliefs". The reason I am excited is that, nowadays, "beliefs" are not left to speculations
but are being explicated, formalized and algorithmitized. So, what are the beliefs that make blocking improve
precision? #Bookofwhy

11.16.19 2:40pm - (Replying to @JBlevins0 @RonKenett and 2 others) Unfortunately, I dont have Vol 1 of K&H.
Can any design expert summarize, conceptually, what makes one block better than another, w/o referring to
authority? In computer science we have no authorities, so we just confess: "For me, a good block is one which
...". #Bookofwhy

11.16.19 2:15pm - (Replying to @ben_golub @causalinf and 2 others) Oddly, my patience only ran thinner since
turning 83. And Jewish tradition now gives me only 37 years of hope to see econometric catching up to
modernity. Short time, considering the observed movement. #Bookofwhy

11.16.19 1:36pm - (Replying to @JBlevins0 @RonKenett and 2 others) All I could get my hands on was vol. 2 of
H&K. Do you remember how they show (or justify) that "age" is a good block, while "street address" is not so
good? Or, better yet, how does a good stat professor justify it in class? Thanks #Bookofwhy

11.16.19 1:06pm - (Replying to @JBlevins0 @RonKenett and 2 others) Thanks for pointing to this source which I
am eager to explore. But the label "Bayesian problem" must be taken with a grain of salt, because statisticians
tend to label any subjective judgement "Bayesian", even judgments about causation, see
https://ucla.in/2nZN7IH #Bookofwhy

11.16.19 12:54am - (Replying to @arinbasu) I am sorry you got stuck on a beautiful chapter
https://ucla.in/2G2rWBv, which some readers have found to make counterfactuals as clear as daylight. See if
Section 4.4 of https://ucla.in/2Nbq7W2 makes it clearer, it has more examples. #Bookofwhy

11.15.19 6:28am - (Replying to @stephensenn @RonKenett and @learnfromerror) I am trying to learn from you
about experimental design. I am not comparing it to causal calculus, nor do I claim anything. I'm just trying
humbly to learn how blocking was proved useful. Why not teach me? #Bookofwhy

11.15.19 6:18am - A LOADED question to @amitabhchandra2 . We have had a lively discussion here on Twitter
whether economics is more insular than other disciplines. Do you think econ. editors are doing enough to
expose readers to emerging methodologies that neighboring disciplines have found useful?

11.15.19 6:07am - I can't envision an AI PhD student that would miss this opportunity. #Bookofwhy

11.15.19 6:01am - (Replying to @stephensenn @RonKenett and @learnfromerror) May I assume that Nelder's
"allocation map" is the notational device that I was searching for? "Reading" will be situated differently than
"age", and he has the mathematics to tell us that "age" is a good block, while "reading" may be terrible. Did I get
it right? #Bookofwhy

11.15.19 5:33am - (Replying to @stephensenn @RonKenett and @learnfromerror) I still can't see how John
Nedler could distinguish a good block (say "age") from its proxy (say "reading level") lacking mathematical
notation in 1965. Moreover, how could he possibly PROVE that blocking on age improves precision, when
blocking on "reading" may not. #Bookofwhy

11.15.19 5:22am - A down-to-earth guide for the missing-data perplexed. Our tutorial has just been posted on
arxiv: http://arxiv.org/abs/1801.03583 #Bookofwhy
11.15.19 3:07am - (Replying to @stephensenn @RonKenett and @learnfromerror) It is hard for me to believe
that this intuition was made formal before 1975. Why? Because "determinant" is a causal notion, and statistics
did not have the notation to distinguish a "determinant" from its proxies; the latter could be bad blocks. Thus
my disbelief. #Bookofwhy

11.15.19 1:39am - (Replying to @RonKenett @stephensenn and @learnfromerror) I am stuck on something much
simpler. What is "block"? Is it what you DECIDE to block, or what you SHOULD block? Wikipedia says:
"Block=Where units are similar to each other". Do you buy it? I'm also looking for a proof that if you block on
some B you would do better. #Bookofwhy

11.15.19 12:47am - (Replying to @stephensenn @RonKenett and @learnfromerror) I could not find the history of
"block design" in Peter Goos writings. Does any reader know where block design was first proposed? And, more
importantly, where it was first proven (mathematically) to improve precision? History galore! #Bookofwhy

11.15.19 12:20am - (Replying to @cryptoecongames @kareem_carr and @EpiEllie) A DAG/tool forces more than
clarity. For example, it tells researchers if they chose a good covariate set for adjustment. As for
"implementation"? Here it is: "Choose a good set and continue the same way as you did with the bad set."
#Bookofwhy

11.15.19 12:07am - (Replying to @RonKenett @stephensenn and @learnfromerror) Ron - speaking of Lord's
paradox, @stephensenn and @learnfromerror , do you happen to know where "block design" was first used? first
defined? or first justified formally? #Bookofwhy

11.14.19 11:31pm - (Replying to @RonKenett) My point was/is that even what we took to be purely statistical
tasks (eg running RCT, optimizing predictions, handling missing data) require causal knowledge, hence DAGs.
Statisticians avoided DAGs successfully, because they used their mind-DAGs instead. Robots can't. #Bookofwhy

11.14.19 11:14pm - (Replying to @hoss_tbf and @doinkboy) 1. Are you sure this was a stats textbook? Mine
refrained from discussing mediation, because it requires causal thinking & vocabulary - strongly tabooed. 2.
Would you be surprised if some journals (JSEM, Psychometrika) still spoke the way your textbook did?
#Bookofwhy

11.14.19 10:09pm - (Replying to @YoniMichanie and @HananyaNaftali) Beg to disagree, Palestinian education is
a much greater obstacle to peace.

11.14.19 7:21pm - Even just to "go get reliable, replicable, experimental data..." we need to understand the
problem domain, ie, its DAG. How else can we decide if "what worked here will also work there"? How else can
we decide what nuisance factors to control for? No escape! #Bookofwhy

11.14.19 6:46pm - Nice exposition of transport methodology, unconflated by "target trials" and other RCT
imitations. But what is a "g-comp estimator"? Is it a special estimator of the transport formula, or just any
consistent estimator? #Bookofwhy

11.14.19 6:20pm - Thanks for re-posting and re-calling a quote: " tools that are indispensable in solving simple
problems are unlikely to become dispensable when problems become more complex." Which explains why some
researchers hate simple problems--they tend to expose one's tool set. #Bookofwhy
11.14.19 2:05pm - (Replying to @jonasobleser @stephensenn and 3 others) I would love to weigh in but can
someone explain (in Twitter friendly language) what "latent change score model" is? And how it is used by its
proponents? Thanks #Bookofwhy

11.14.19 1:46am - (Replying to @kareem_carr) Can't help asking: Did you say that "classical statistics knows how
to generalize to new data"? I am asking because I have learned that classical statistics has no clue on how to
generalize, eg., https://ucla.in/2N7S0K9 Are we speaking same "classics"? #Bookofwhy

11.14.19 1:24am - (Replying to @Statisticko and @JuhaKarvanen) Sounds like a wild party that I should enjoy. I
know @JuhaKarvanen since his 2014 paper on study design. Say hello for me. Now I know causality is in good
hands. Thanks. #Bookofwhy

11.14.19 12:18am - These are the best lectures (and course notes) I have seen in statistical education. Not so
much because they are DAGy, but b/c they are given by a "data scientist", as opposed to a "data analyst." We
need a new discipline: "scientific statistics" to replace "statistical science"

11.13.19 9:54pm - (Replying to @HananyaNaftali) California

11.13.19 3:04pm - I am not worried about the woolly socks, my heart goes to "causality" being squeezed
between two statisticians, top and bottom. Give her a chance!! Or, better yet, give them a chance to rethink
what they are doing. #Bookofwhy

11.13.19 7:02am - Any connection to the missingness graphs used here: https://ucla.in/2LdEjZW ??

11.13.19 6:30am - My hero! 6 years old Shoham. I wish I could tell him what I was told at my 6th birthday, that
when I grow up there would be no wars in the world.

11.13.19 5:44am - Several readers inquired on the relations between Markov processes and structural models.
This new paper https://arxiv.org/pdf/1911.02175.pdf translates continuous-time discrete-state Markov
processes to structural equations for the purpose of managing actions and counterfactuals. #Bookofwhy

11.13.19 5:37am - (Replying to @lgmoneda @Khipu_AI and 3 others) Anyone analyzing missing data using DAGs
must understand missing data. Kudos! #Bookofwhy

11.13.19 1:30am - Erika, had I not been inflicted with incurable modesty I would recommend the Primer
https://ucla.in/2KYYviP much more strongly than I do. It is geared for non-statisticians, and it starts where
intuition resides (not RCT) @smueller uses it for high school. The best! @Bookofwhy

11.12.19 10:46pm - (Replying to @GilTroy) "Fighting against hatred" is easy. Even hate-soaked Linda Sarsour says
she fights hate. What counts is "fighting Zionophobia", because no living Zionophobe can control this
pathological obsession.

11.12.19 8:59pm - (Replying to @jazchaz) Joe Biden made history, not by repeating what others say, but by the
way he stood up to JVP hecklers: "Show me one Arab leader willing to accept Israel, and I will talk to you about
the occupation" (paraphrased from a friendly memory).

11.12.19 7:18pm - To my brothers and sisters in Israel, I join all people of conscience in telling you that your
ili d t l f lif d j ti i i ti l t ll f h it
resilience and struggle for life, peace and justice are inspirational to all of humanity.

11.12.19 4:04pm - (Replying to @JadePinkSameera) The merits of DAGs lie not in helping the "most logical"
researchers, but in helping the lazy ones, those who want the arrows to do the thinking for them, and protect
them from errors and over-generalizations. That's why I am surprised when cultures slip into darkness.
#Bookofwhy

11.12.19 3:43pm - (Replying to @AlyssaHarlow) #epitwitter and @EpiEllie , I just tweeted this
https://twitter.com/yudapearl/status/1194384353381384193, because I see counter-examples to every
statement on this thread. I think there is an endemic misconception here, about selection bias, that can be
easily corrected with a quick glance at the DAG. #Bookofwhy

11.12.19 2:39pm - (Replying to @admiyoung and @AlyssaHarlow) More generally, questions about selection bias
cannot be answered in the language of "associations" or "prediction", no matter how hard we try. We must
consult a causal diagram and first find an unbiased "estimand", IPW is just an estimator. see
https://ucla.in/2OgMF4g #Bookofwhy

11.12.19 2:39pm - (Replying to @admiyoung and @AlyssaHarlow)

11.12.19 5:51am - (Replying to @hangingnoodles) Well put! Competent statisticians nowadays are grad students
who return their textbooks to the publisher with a note: "Same old stuff?" #Bookofwhy @rorysutherland

11.12.19 3:15am - The siren sound that I hear from Israel today has the same pitch as the one I heard in 1948,
when Egyptian war planes bombed Tel Aviv. But then, people did not say "Israel has a right to defend", it was
self evident. Today, a nation must beg public opinion for the right to live.

11.12.19 2:06am - (Replying to @AdanZBecerra1 @UChicago and @WestPoint_USMA) Did students find it "super
hard"? or "super fun"?

11.11.19 7:51pm - (Replying to @DrCatherineB) Did you say "super hard"?. The only complaints I have heard
about causal inference were: "How come no one told us it's so easy?" Seriously, have you read any of the "for
fun" literature? BTW, consistency is not an "assumption" but a simple corollary of counterfactuals. #Bookofwhy

11.11.19 7:36pm - Thank you @Carthica for re-posting this tutorial on graphical models. It should open many
eyes to the simplicity of causal inference, as it did in 2012, when most people still did "belief updating." Note
the 3 bullets on the last slide, how true! #Bookofwhy

11.11.19 4:54pm - (Replying to @GilTroy) For J Street to convince us that "pro-Israel pro-peace" are not empty
words, it must commit at least 1% of its efforts to fighting Zionophobia. It does'nt. Thus, J Street may not be
anti-Semitic in intent, but it sure aint "pro-Israel pro-peace". Is it, God forbids pro-deception?

11.11.19 4:28pm - The beauty of the graphical framework in https://ucla.in/2wbH488 is that all
generalizations, including subset to superset, are managed by the same methodology. Life would be utterly
impossible if we demand "a new approach" for every nuance of generalization. #bookofwhy

11.11.19 1:07pm - (Replying to @deaneckles and @juli_schuess) I think N. Cartwright was the first to use := .
The notation Y(x) = a + bx + ε is not sufficient to define a structural equation. You need a structural model to
define what Y(x) is. IOW, for y= a + bx + ε to be structural, X must be the only direct cause of Y. #Bookofwhy
11.11.19 4:11am - (Replying to @juli_schuess) Today we think it is "disturbing"; at the time it was standard
thinking. Even today, what do we know about the thousands upon thousands of "regression analysts"? We do not
know much about what they think, because they are silent. Dead silent. Scared to confess thoughts #Bookofwhy

11.11.19 1:47am - The thread itself, including Miguel's proposal, can be found here
https://twitter.com/yudapearl/status/1193773633245573122 #Bookofwhy

11.11.19 1:24am - "How to extend causal inferences from a RTC to a target population?" I am retweeting this
thread since many readers ask this question, and many know of the comprehensive solution described in
https://ucla.in/2wbH488 and in https://ucla.in/2Jc1kdD using graphical models. 1/2 #Bookofwhy.
11.11.19 1:24am - I was surprised therefore to read Miguel's proposal to recast the whole question in the
opaque language of PO, where assumptions are so far removed from scientific understanding. A face-to-face
comparison of the two approaches is given here: https://ucla.in/2L6yTzE

11.10.19 10:07pm - (1/2) (Replying to @_MiguelHernan @Biometrics_ibs and @EpidemiologyLWW) Miguel, I beg

to strongly disagree with your proposed approach to generalizing empirical results. Translating issues and
solutions from the language of graphical models to the language of potential outcomes blurs the issues and
hides the results. Such translations suffer from 11.10.19 10:13pm - (2/2) (Replying to @yudapearl
@_MiguelHernan and 2 others) three incurable shortcomings, summarized in 3 vivid bullets here:
https://ucla.in/2wbH488, also known as TPT (Transparency, Power and Testability). So, given that Epi folks
already speak graphs, what's the point of retreating from clarity? #Bookofwhy

11.10.19 9:35pm - For mediation analysts among us, here is an interesting article that hit my screen today:
https://rss.onlinelibrary.wiley.com/doi/abs/10.1111/rssb.12345. I still havn't absorbed the semantics of the
proposed Indirect Effect, but the fact that it generalizes the frontdoor tells me there's substance to
it.#Bookofwhy

11.10.19 1:33pm - (1/2) It's an excellent paper, agree, that is rarely cited, because it got published in a
"handbook", ie, books that only librarians read. The official publications of the SEM community, eg JSEM or
Psychometrika, are still buried in confusions over what SEM is about, awaiting,
11.10.19 1:33pm - (2/2) like most econometric journals, for a change of guards, and a speedy catch up with
modernity. The litmus test for modernity is Eq. (6), the definition of counterfactuals, of which most SEM folks
are unaware, and which is the basis of most works in causal inference.#Bookofwhy

11.10.19 12:59am - For machines to generate human-level explanations, we should listen to what philosophers
of language say about causative nuances in English. This paper
https://scholars.huji.ac.il/sites/default/files/elitzurbarashersiegal/files/baglini.bar-ashersiegal2019.pdf say it
in the language of SCM, so it should be accessible to readers of #Bookofwhy.

11.10.19 12:32am - (Replying to @tdietterich and @twimlai) I buy your main thesis that "understanding" is a
matter of degree. Would you buy mine, that "understanding a domain" means answering 'what if' questions
about hypothetical scenarios in that domain? #Bookofwhy

11.9.19 9:10pm - (Replying to @_MiguelHernan @MariaGlymour and @HarvardEpi) And I am waiting eagerly for
the day the teaching curiculum @HarvardEpi will be revamped forward, to the #targetQuery approach, with the
#TargetTrial approach being one way of answering queries of interest, including queries about counterfactuals
and explanations. #Bookofwhy
11.9.19 8:46pm - (Replying to @tdietterich and @twimlai) Thanks.

11.9.19 7:25pm - (Replying to @tdietterich and @twimlai @tdietterich) Would love to read you blog post on
"understanding", but could not find a link. Can you link-a-link me?

11.9.19 6:20pm - I have hoped that the "Imitate RCT" approach will give way to the "interrogate nature"
approach, in which RCT is just one way of interrogating nature, as in #Bookofwhy. The difference may seem
academic; no so, when it comes to explanation, see https://ucla.in/2lEOJtO

11.9.19 6:02pm - (Replying to @tanyacash21 and @gnshealthcare) I looked up GNS and found this: "Using Causal
AI to answer  the Patient Response Question Biopharma companies have long struggled to identify the best
responders for specific drugs." What's your method? Is it a secret? #Bookofwhy @smueller

11.9.19 5:30pm - (Replying to @olafhartig) Olaf, on the other hand, German streets are paved with 7,000
"Stolpersteine" (commemoratives stones), as I've learned from: https://tabletmag.com/jewish-news-and-
politics/293667/the-perils-of-forgetting, while American politicians' minds are paved with ignorance.

11.9.19 1:00pm - And I know reporters who are kind to kindness.

11.9.19 11:48pm - I was 2 years old on Kristallnacht, November 9-10 1938, It was a quiet night in Tel Aviv. My
parents understood what's coming, but kept it from us, so we can grow up free. We did. And by the time we
understood what kristallnacht meant, Egyptian war plains came attacking Tel Aviv.

11.9.19 2:29am - (Replying to @renatrigiorese) I'm not familiar with Khrennikov, and what I know about de
Finetti comes from others. i.e., Diaconis & Skirms "Ten Great Ideas About Chance"declares de Finetti "the father
of subjective Bayesianism"but I'm not convinced he could manage the "inevitable regret" problem. #Bookofwhy

11.8.19 3:37am - Here is another "must" in teaching philosophy of statistics: What logic gives us the authority to
say something about an individual by observing other individuals, however similar? The question is raised here:
https://ucla.in/2N6x36Q and I wonder what philosophers say #Bookofwhy

11.8.19 3:10am - (Replying to @klts0 @EpiEllie and @HarvardChanSPH) Enticing title: "Un-conflating causal
effects and intervention effects: saving the counterfactual baby while tossing the well-defined intervention
bathwater." I hope Sharon un-conflates Rung-2 and Rung-3 of the Ladder. Else, re-conflation will triumph.
#Bookofwhy

11.8.19 2:54am - (Replying to @IdoDaniel @GalitPeleg and @TheOnion) I am studying the faces of the delegates
standing there, and I'm thinking: How do they manage to control their smiles? Super-humans! As the Mishna
said: "Who is strong (Eizehu Gibor), he who can conquer his smile" (almost).

11.8.19 1:42am - (1/2) @bzaharatos , your Tweet reminded me how anxious I am to teach a course in
Philosophy of Statistics (if they let me). Because there are so few philosophers who understand that the
foundations of statistics lie in causality, not in statistics. I would start with Chapter 6,
11.8.19 1:42am - (2/2) because through paradoxes we see foundations. I would then go to the chapter on Rev
Bayes, spiced with "why I am only half Bayesian?" https://ucla.in/2nZN7IH and a few more goodies that I will try
tweeting tomorrow, after some search. #Bookofwhy

11.7.19 11:57pm - (Replying to @tanyacash21) "On a bookshelf in a bar"? Can't think of a greater honor to a bar-
avoiding author! (Nothing against the culture, but can't stand the noise -- do they have silent bars someplace?).
But what do I see on Dylan's note: "grain of salt". Was he serious?

11.7.19 6:20pm - (Replying to @Yu_Ke__) Sorry, the expression "n order effect" is not known to me. Can we
perhaps translate it to the language of SCM?

11.7.19 5:37am - (Replying to @Humblefool_14) I dont get it. Do you mean weird effects due to unanticipated
eventualities?

11.7.19 4:13am - (Replying to @hirahira2835) No problem, write to my assistant kaoru@cs.ucla.edu. BTW, what
is "orange bon"?

11.7.19 4:11am - (Replying to @Humblefool_14) HMM. This section could stand expansion, I agree. But see if the
discussion of Inverse Probability Weighting in Primer answers your questions https://ucla.in/2KYYviP Or,
perhaps here: https://ucla.in/2N9nSCV where mass movement is described in vivid colors? #Bookofwhy

11.7.19 4:03am - (Replying to @Humblefool_14) What is "second order effect"? Example please.

11.7.19 1:40am - I am trying to remain unemotional, but seeing these flags blend in solidarity bring tears to my
eyes. Tears of common experience, common purpose, and of pride. Pride in a people that is able to give back
what the world has given them - hope.

11.7.19 12:56am - (Replying to @Navatso_IR @RashidaTlaib and @RepAndyLevin) Of course not! It means only
that their leaders cannot pose as seekers of a "peaceful solution", craving for an "honest dialogue". They should
start with honesty: "We are bent on destroying you, but let's talk about the occupation".

11.6.19 11:53pm - (Replying to @Navatso_IR @RashidaTlaib and @RepAndyLevin) You're right. I've assumed that
other readers are as tuned as I am to the fine nuances in the conflict. To clarify, for Palestinians, "peaceful
solution" is "no Israel". For Israelis, it is "Coexistence: two states for two peoples, equally legitimate and equally
indigenous."

11.6.19 10:30pm - (Replying to @RashidaTlaib and @RepAndyLevin) Those who insist on "an honest dialogue
about the occupation" while refusing "an honest dialogue about co-existence" are incapable of leading an honest
dialogue on any subject.

11.6.19 6:50pm - (Replying to @EinatWilf) I wish I could understand the neural architecture of a person who
chose to head an organization committed to the perpetuation of human misery.

11.6.19 6:38pm - (Replying to @DKedmey) DAGophobia = Egophilia + Patriotism + Paranoia

11.6.19 4:16am - Thanks all for the pointers posted here to Turing-Bayes. As you can tell from #Bookofwhy, my
mind only works with toy examples. I wish I could see a toy Enigma machine with just two wheels, showing the
tradeoff that Turing faced between using Bayes rule and exhaustive search.

11.6.19 3:10am - I just watched Sharon McGayne fascinating talk on Bayes https://youtube.com/watch?
v=8oD6eBkjF9o and I noticed that the scientific essence of Bayes rule is still not clear to the general public.
Does anyone know of a simple description of how Turing used Bayes theorem to break the Enigma code?
11.6.19 1:30am - (Replying to @omaclaren @stephensenn and 4 others) I would hate to impose my own
interpretation of "before". I will take whatever statisticians mean by it in their attempt to improve SE w/o
causal vocabulary. My point is that even temporal information is insufficient. #Bookofwhy

11.6.19 1:10am - (Replying to @mendel_random and @stephensenn) "Causal thinking" was there all the time. In
fact this thread revolves around the discovery that causal thinking drives even in purely predictive tasks. What
#Bookofwhy claims is that there wasn't much beyond "thinking", because the "thinkers" lacked notation. Today
we have it.

11.5.19 11:27pm - (Replying to @stephensenn @autoregress and 3 others) Great!!! #Bookofwhy commends David
Cox (1958) for breaking statistical taboos and using a causal expression "affected by". Now, how about
covariates measured BEFORE randomization. Are they safe, in purely predictive context? What do the textbooks
say? Do you agree with them?

11.5.19 11:13pm - (Replying to @RonKenett and @stephensenn) I promise to keep a qualifier when its absence
can be confounding. So far, your reference to aliasing was the first complaint I got since 1993. But our more
pressing issue concerns purely predictive tasks which, presumably, should require no causal thinking. They do!!
#Bookofwhy

11.5.19 11:00pm - (Replying to @stephensenn @autoregress and 3 others) I thought you will give us a statistical
criterion to decide which variables should safely be controlled for, to reduce SE. Hundreds of statisticians on
this education channel will sigh in relief when you agree that no such criterion exists. #Bookofwhy

11.5.19 10:49pm - (Replying to @RonKenett and @stephensenn) If statisticians have been using the word
"confounding" in other meanings, different from Eq. (1), I will qualify my statement. But sticking to Eq. (1), I
maintain: Confounding is not a statistical notion. And please note, I define things explicitly and unambiguously.
#Bookofwhy

11.5.19 10:28pm - (Replying to @stephensenn @autoregress and 3 others) We are all concentrating on SE,
nothing is missing. So far, all assertions about reducing SE (by variable control) had exceptions. To see it,
please express your favorite (statistical) criterion for reducing SE, and I will produce a counter example for your
enjoyment. #Bookofwhy

11.5.19 10:00pm - (Replying to @RonKenett and @stephensenn) In 1998, I wrote: "Why there is no statistical
test for confounding, why many think there is, and why they are almost right" https://ucla.in/2MrKAkk. It was
later incorporated into Chapter 6 of Causality. "Confounding" is defined as violation of Eq.(1), nothing more.
#Bookofwhy

11.5.19 7:15pm - (Replying to @autoregress @AdanZBecerra1 and 3 others) In a regression context, at least,
there are no "treatments" nor "effects", only "predictors" and "predicted". Can you restate your theorem in these
terms.? #Bookofwhy

11.5.19 5:52pm - (1/3) This thread might come as a shock to regression analysts and perhaps to all statistically-
trained followers of this education channel. One of the least disputed dictum in Statistic textbooks says that, in
purely predictive tasks, to reduce standard errors, one should adjust
11.5.19 5:52pm - (2/3) for variables that are associated with the predicted outcome. Recent conversations on
this thread drove me to refute this dictum and state that, even in purely predictive tasks, the language of
statistics (eg. "associated with") is inadequate for deciding which variables
11 5 19 5:52pm (3/3) should or should not be adjusted for Causal models are needed for the task which
11.5.19 5:52pm - (3/3) should or should not be adjusted for. Causal models are needed for the task, which
cannot be captured in the language of statistics. Counter-dictum examples abound. This means that to do
statistics properly one should (occasionally) snap out of it. #Bookofwhy

11.5.19 2:44pm - (Replying to @stephensenn @AdanZBecerra1 and 2 others) Adjusting for variables that are
associated with outcomes, because Stat textbooks say you should, CAN be a fundamental mistake. In general,
using statistical vocabulary to decide what to adjust for should be a fail in Stat 1. It hasn't been, but it should
be. #Bookofwhy

11.5.19 1:18pm - (Replying to @AdanZBecerra1 @fledglingStat and 2 others) These are good heuristics but
insufficient. Adjusting for vars that are associated w outcome but not exposure may actually increase standard
errors AND change causal effect. The rule is: Do not use statistical vocabulary when speaking "effect", or even
"outcome", use #Bookofwhy

11.5.19 1:04pm - (Replying to @stephensenn @AdanZBecerra1 and 2 others) This is what R. Fisher said, when
asked about mediation, and fumbled with ANOVA. See https://ucla.in/2N7wOIN. "Identifying a mediator" is
easy, using our mind DAG, but how to handle it in assessing mediation is not easy, as Fisher's fumbling shows us.
#Bookofwhy

11.5.19 6:58am - (Replying to @fledglingStat @robertstats and @stephensenn) I do not believe any trick will
enable you to distinguish good from bad "explainer" or "accounter for" of "partitioner" using associations
vocabulary. But I'll keep my mind open. #Bookofwhy

11.5.19 6:39am - (Replying to @fledglingStat @robertstats and @stephensenn) It does indeed. I was talking
about the confounding step. Speaking about "explain outcome variability" is interesting, because "explain",
again, is a causal, not statistical notion. I am curious how statisticians defined it in terms of "associations"
#Bookofwhy

11.5.19 6:27am - (Replying to @stephensenn) I have assumed we are both in the context of estimating causal
effects, where covariates enter into the denominator of PS because they are presumed to reduce bias
(explained in https://ucla.in/2NnfGPQ#page=18). When chosen to reduce variance, statistical considerations
set in. #Bookofwhy

11.5.19 5:53am - (Replying to @stephensenn) Choosing covariates by association is fundamentally wrong,


regardless of what associations are considered. Covariates should be chosen to reduce confounding, and
confounding is a causal, not statistical concept. #Bookofwhy

11.5.19 4:25am - (Replying to @yudapearl @firstmn59 and @Cahliveira) The tyranny has little to do with DAGs
(though those have been tabooed by leadership since 1995) It has more to do with the paradigm imposed by
"experimentalists" which forbids their faithfuls from seeking TPT (Transparency, Power, Testability).
#Bookofwhy

11.5.19 4:11am - (Replying to @firstmn59 and @Cahliveira) The tyranny comes not from econs., but from their
leadership. Try sending an email to the editor of any "top" journal and ask whether he/she can refer you to ANY
paper (or review, or survey) describing methods of mediation analysis. Please share their answers with us.
#Bookofwhy

11.5.19 2:40am - I thought it was a dream: "demanding Iraqi Jews to come back "home"'. I rushed to show it to
my wife, who was expelled from Iraq in 1950, to Israel. Her answer: These people will never understand what
"home" is; we never felt "at home" in Iraq, though we stayed there 2,536 yrs.
home is; we never felt at home in Iraq, though we stayed there 2,536 yrs.

11.5.19 1:42am - Great idea!! I am honored to be one of the first to speak at this forum, and I hope it becomes
an educational tool to the entire data-science community, with lectures from variety of disciplines teaching
what can be done with TPT (Transparency, Power and Testability). #Bookofwhy

11.5.19 1:30am - (Replying to @ArielElyseGold) "deaths of people"? Unheard of!! Champions of humanity and
protectors of human rights would never dream "calling for the deaths of people." If we only psych ourselves into
a self-hating Stockholm Syndrome, we'll see things as they really are, not as we wish them to be.

11.4.19 11:34pm - (Replying to @yudapearl and @kareem_carr) Apropos Psychometric Society, I met two past
presidents. The first, the late Rod McDonald, wrote to me (2009): "The best way to discuss moderation or
mediation is to set aside the entire literature on these topics and start from scratch". The 2nd did not follow
Rod's wisdom.

11.4.19 10:50pm - (Replying to @kareem_carr) Students tend to underestimate the clout they have. If I were
President of the Psychometric Society and a student wrote to me how he saw the name of my organization
tarnished in a serious Tweeter conversation, I would jump 10 ft and become an expert mediation analyst.
#Bookofwhy

11.4.19 4:25pm - Action should overcome depression. How about writing to the leadership of psychometric
society (exists?) whether they are aware of the reputation of their literature. I bet they are waiting for an
excuse to reinvigorate the editorial staff of their leading journals with new blood

11.4.19 2:33pm - This is an amazing finding: 97% studies still in the stone age! Is Psychometrics under same
tyranny as Econometrics? Here's a great PhD topic (ech. Kuhn):"The psychology of scientific revolutions". Is it
institutionally imposed or intellectually chosen? Any historians? #Bookofwhy

11.4.19 6:38am - (Replying to @glarange72) I never practiced "meditation" but, comes to think about it, it does
take some meditation to explain what happened to economic education in the past few decades. Bemoaned
here: https://ucla.in/2mhxKdO, but echoed nowhere enough to shake "top" econ. journals. #Bookofwhy

11.4.19 6:17am - Why I said "especially economists!"? Because there are many enlightened economists among
our readers, educated on "mostly harmless", who cannot wait for "top" econometric journals to open their gates
to modern causal analysis. And economics is all about mediation #Bookofwhy

11.4.19 5:03am - Readers (esp. economists!) interested in a gentle introduction to modern mediation analysis
will find this post useful: https://blog.methodsconsultants.com/posts/how-to-do-mediation-scientifically/ It
contrasts modern analysis with the Baron & Kelley tradition, gives examples, and explains the meaning of
Natural Effects. #Bookofwhy

11.4.19 4:41am - Sharing a space-photo of Israel by lady astronaut (and Howard professor) Jessica Meir, who
remembered her father's first country after escaping from Iraq. We can clearly see the Red Sea, the Sea of
Galilee, and the Suez Canal. My biblical home town is there too, under the clouds.

11.4.19 2:51am - (Replying to @dmi3k) Agree, the article would have been better written with one or two
DAGs, to show the competing theories, and what each tells us about predictions and interventions.But it is still
an interesting article and a nice summary of this line of psychological studies. #Bookofwhy
11.4.19 2:29am - (Replying to @dmi3k) Causality is confusing? God forbid! It's the simplest scientific concept we
have. It was made confusing by statisticians who thought that anything that is not in the data must be either
"confusing" or "super-natural". I hope they think differently today (do they?) #Bookofwhy

11.4.19 2:04am - (Replying to @DataDonors @jan_ni and @welchering) My German speaking neighbors, in the
town where I grew up use to say: "stimme voll und ganz zu" which, to me, sounded like: "Dont you dare say it
again!" I remember them fondly. #Bookofwhy

11.3.19 4:58am - (Replying to @RDMetcalfe @jhaushofer and @PHuenermund) I might agree with you, once we
explicate precisely what "all else equal" means. But I cannot agree on "because you better understand".
"Understanding" was yesterday's excuse for inaction; today we can formalize and quantify whatever needs
understanding. #Bookofwhy

11.3.19 4:28am - (Replying to @jhaushofer and @PHuenermund) In this debate: https://ucla.in/2JfdQcv I


stressed: " Observational studies have the virtue of observing large number of people in their own natural
habitat instead of an artificial experimental setting, marred by selection bias" #Bookofwhy

11.2.19 8:59pm - I used to think that bootstrapping is a statistical estimation technique, of little use in causal
inference. This paper, however, tells about causal bootstrapping
https://arxiv.org/ftp/arxiv/papers/1910/1910.09648.pdf It needs to be listened to. #Bookofwhy

11.2.19 7:56pm - It is well known (is it??) that covariate-adjusted and propensity-score estimators are
asymptotically equivalent (shown eg. here https://ucla.in/2NnfGPQ#page=18 ). This recent paper
https://arxiv.org/pdf/1910.11397.pdf combines the two to improve small-sample performance. #Bookofwhy

11.2.19 7:15pm - A new article crossing my desk provides a procedure, new in SAS/STAT® 15.1, for analyzing
graphical causal models.
https://pdfs.semanticscholar.org/28e0/6aa69ed106798a73c33985d721826b467000.pdf. I hope readers find it
useful #Bookofwhy

11.2.19 3:48pm - True, my article https://tinyurl.com/y4l2emnf deals only with the moral/historical
significance of the Balfour Declaration. For the demographic/political forces behind the scene I recommend
Kramer's eye-opening article

11.2.19 3:06pm - (Replying to @rkarmani) I do not know of any, sorry. But my ignorance does not say a thing
about existence. #Bookofwhy

11.2.19 6:04am - (Replying to @JustinSandefur) And I thought (naively) that I convinced Deaton here:
https://ucla.in/2JfdQcv that some of his critics were outdated. It's hard to convince an economist, but his
paper is fun to read, especially for its survey of economists' thoughts #Bookofwhy

11.2.19 5:46am - (Replying to @Abel_TorresM @dmonett and 2 others) As I am sure you know, temporal
information is not sufficient for eliminating confounding, as the debate about smoking--->cancer reminds us. We
need scientific knowledge about who is listening to whom, eg a DAG. #Bookofwhy

11.2.19 12:18am - Your history professor understood the music of history. I wouldn't be alive if it weren't for the
Balfour Declaration (1917), which gave my grandfather the push to leave boiling Europe , rebuild a Biblical town
Balfour Declaration (1917), which gave my grandfather the push to leave boiling Europe , rebuild a Biblical town
(1924) and then create a miracle called Israel (1948). Kudos!Balfour!

11.1.19 11:30pm - (Replying to @Abel_TorresM @dmonett and 2 others) I do not think it is correct without
explicating what we mean by "based on correlation". If it means "based ONLY on correlation" I would say: No. If
it means "based on correlation and causal assumptions" I would say: Sure, this is what #Bookofwhy is all about.

11.1.19 10:32pm - My,My! It's November 2nd! And I almost forgot to invite you to celebrate with me the Balfour
Declaration: The first international recognition of a people's right to a homeland (1917). I even wrote an article
on its universal significance: https://tinyurl.com/y4l2emnf. One of my best!

11.1.19 5:43am - (Replying to @GuillermoBurr) I did not realize that this is what the "evidence-based"
movement was after all these years. Why did'nt they tell us? They should have called their aspiration "forces-
based" movement, because "evidence" usually means "an observed piece of data".#Bookofwhy

11.1.19 5:03am - We know that institutions can turn creepy, its in their nature. What I cannot understand is
how the delegates, fellow human beings, can face their families, back home, when their children ask: "And you
Papa were part of this Zoo?"

10.31.19 9:00pm - Discovered an interview I gave in Oct 2018, https://tinyurl.com/y6q3sfmh with a punchline
saying: "Knowledge lies not in dry facts but in the forces BEHIND the facts." Sounds obvious? Not to folks who
think big data means big knowledge. #Bookofwhy

10.31.19 12:24am - (Replying to @ClaudeAGarcia) I was not aware of this study, thanks. My paper asks what
does it take for the "illusion of understanding" to be less of an illusion and more grounded in science, data and
logic to earn the title "rational understanding" #Bookofwhy

10.30.19 8:42pm - Sharing the final version of a chapter for The Handbook of Rationality
https://ucla.in/2Iz9myt. It now seems strange to me that philosophers have spent so much effort on rationality
of preferences and so little on rationality of envisioned consequences. #Bookofwhy

10.30.19 5:10pm - To readers in Brazil, and to followers everywhere, a bit of fresh air and good will from Haifa,
where people can still enjoy a fun game of soccer, spiteful of dark clouds and hopeful of (multiple) better
worlds.

10.30.19 2:30pm - (Replying to @David_desJ and @anthonysamir) My argument is based on specific evidence, as
well as long experience in spotting Zionophobes and reading their modus operandi. The "never fails" is not a
premise but a fact among Omar's constituency, in case you have watched them in action.

10.30.19 2:12pm - (Replying to @David_desJ and @anthonysamir) Beg to disagree. "Islamic" is too heterogenous
an entity to serve as a focus of obsession, varying wildly from Erdogan, to Ayatola, to Sisi and Saudia.
Zionophobia is simple, focused, emotionally charged, and never fails to arouse the worst in people.

10.30.19 4:17am - (Replying to @DrMikeH49 and @ErakatSaeb) And as long as no school teacher in Ramallah
mentions this "commitment".

10.30.19 4:05am - What a happy day! When I was a boy, 6-7, I used to visit my aunt who lived in King Cyrus
street (Melech Koresh) in Tel-Aviv. When I asked my teachers who King Cyrus was, they asked my to open my
Bible book of Ezra it was all there!!! Happy King Cyrus day!
Bible, book of Ezra, it was all there!!! Happy King Cyrus day!

10.30.19 3:44am - (Replying to @anthonysamir and @David_desJ) Not "selective attention" but "selective
obsession". She cannot offend Sultan Erdogan as long as he supports Hamas.

10.30.19 2:18am - I actually called PS "ingenious". The only thing I "demolished" was the myth that it has
anything to do with (asymptotic) "bias reduction". It is a powerful estimator, once you make sure that the
covariates satisfy the back-door criterion. See https://ucla.in/2NbS14j #Bookofwhy

10.30.19 1:34am - (1/ ) (Replying to @TatThangVo1 @LauraBBalzer and 9 others) I commend you for taking the
lead and explaining to statisticians how they should think and do meta-analysis in this century. You should
follow it with a glossary of terms to help them and us communicate across jargon barriers. I hope they resonate
favorably to your heresy.I'm
10.30.19 1:49am - (2/2) (Replying to @yudapearl @TatThangVo1 and 10 others) wondering though whether you
think traditional meta-analysts (say S Normand) would feel more comfortable thinking "ignorability" or thinking
"graph separation"? Traditionalists shun both, to their detriment, but the latter is more human. Awaiting to see
feedback. #Bookofwhy
10.30.19 3:10am - (3/ ) (Replying to @yudapearl @TatThangVo1 and 10 others) Apropos, here is Wiki's
definition: Meta-analysis is the statistical procedure for combining data from multiple studies. When the
treatment effect (or effect size) is consistent from one study to the next, meta-analysis can be used to identify
this common effect.
10.30.19 3:19am - (4/ ) (Replying to @yudapearl @TatThangVo1 and 10 others) When the effect varies from one
study to the next, meta-analysis may be used to identify the reason for the variation." Someone should correct
it. (1) The reason CANNOT be identified by ANY statistical means. (2) Today we can do more than just "identify
reasons". #Bookofwhy

10.30.19 1:23am - (Replying to @thebyrdlab) The trouble with the "ignorability" assumption is not that those
who make it do not know how to test it (or even if it can be tested) but that it is so far removed from what they
know about the problem, that no one can argue for its plausibility or implausibility.#Bookofwhy

10.29.19 9:14pm - This is a wonderful paper that should be read by every ML researcher concerned with
explainability. I dont know how it escape my attention. I would shorten it more, skip the "according to so and
so.." and present ML folks with a computer-minded taxonomy of explanation #Bookofwhy

10.29.19 1:47pm - Those who deem "ignorability" crucial are escaping the light of day. Which explains why they
deem it "crucial". #Bookofwhy

10.29.19 5:54am - (Replying to @EpidByDesign @LauraBBalzer and 11 others) Your abstract says: "we show that
even perfect internal validity does not ensure that a causal effect will be unbiased in a specific target
population." Is this surprising? Why would IV have anything to do with EV? #Bookofwhy

10.29.19 4:39am - (Replying to @FriedrichHayek) I have that hunch too, but I need help in translating their
ideas from verbose philosophical writing to computer age: What information do we need, and how to represent
it, so as to generate rational explanations for actions intended as well as actions already taken. #Bookofwhy

10.29.19 1:11am - I have a strong hunch that ML discussions on explainability could benefit from the centuries
of "philosophy of action", for example https://cambridge.org/core/journals/royal-institute-of-philosophy-

supplements/article/representation-of-action/64B00B2225D55A1A1C763DE1772F2EB0. Likewise, philosophers


should benefit from our new ability to represent "actions" formally in a machine. #Bookofwhy
should benefit from our new ability to represent actions formally in a machine. #Bookofwhy

10.29.19 12:56am - (Replying to @andrewthesmart @jainfamilyinst and 19 others) To be more specific, once we
assess the causal connections among variables in a given population, we can also assess "the degree of
discrimination" exhibited by that population. The latter is DEFINED in terms of (not "defined by") the former.
#Bookofwhy

10.29.19 12:46am - (Replying to @matsouaka @LauraBBalzer and 9 others) Trying to advise readers who have
spent many hours studying classical transportability (eg http://ucla.in/2Jc1kdD). Would they be able to solve
new problems after spending many more hours on Issa et al ? Or is it the same set of problems? Perhaps weaker
assumptions? #Bookofwhy

10.28.19 6:16pm - (1/ ) My screen got blessed by an intriguing Open Access volume
https://brill.com/view/title/38676 titled: Free Will, Causality, and Neuroscience. Chapter 3 reads: When Do
Robots Have Free Will? A question we address in #Bookofwhy. Unforturtunately,
10.28.19 6:16pm - (2/2) I got spoiled by Twitter and lost the ability to go over pages +pages of introductions.
Can anyone extract from it the meat: When would we say that a robot passed the Turing test for Free Will?
#Bookofwhy

10.28.19 5:50pm - (Replying to @jwhogan42 @LauraBBalzer and 9 others) Naive question: Anyone knows what
these two papers add to the work on Data Fusion described here: https://ucla.in/2Jc1kdD. The abstracts read
the same, and the texts do not offer comparisons. Any help? #Bookofwhy

10.28.19 3:52pm - (Replying to @andrewthesmart @jainfamilyinst and 19 others) DAGs postulate individual-level
relations. But data only allows us to estimate population-level quantities. BTW, "discrimination" is not a
variable, but a property of behavior, the degree of which can be estimated from data. #Bookofwhy

10.28.19 3:36pm - (Replying to @LauraBBalzer @eliasbareinboim and 8 others) What's the occasion? Is NIH
funding a new initiative? (They should!). My favorite is this one: https://ucla.in/2ocoWqq, primarily because it
raises so many questions that have not been answered, all on a tiny example: X-->Z-->Y. #Bookofwhy

10.27.19 8:33pm - (Replying to @MiriamElman @TheAENetwork and @UMassAmherst) He needs support from
more than one organization. Most importantly, UMass students and faculty should express their commitment to
inclusion and a hate-free campus. Other Chancellors have condemned hateful meetings before, yet no one has
dared do it to the smoke-covered BDS

10.27.19 3:20pm - A moment of silence for the eleven Tree of Life Victims. A personal connection: Joyce
Fienberg was the wife of Professor Steve Fienberg of CMU, a pioneer in categorical data analysis, a renaissance
man and a true friend. #Bookofwhy

10.27.19 5:29am - (Replying to @s_jagabathula @fuiud and @PHuenermund) I've seen this paper before, but
thanks for reposting - I should read it again. Ad-placement is indeed an industry that pays many ML salaries and
is loaded with not-trivial causal and counterfactual considerations, see eg., https://ucla.in/34mq6Uh. Needs
rereading. #Bookofwhy

10.27.19 12:44am - Something wonderful has happened to me on Twitter. I fell in love with my own slogan:
"Transparency, Power and Testability" (TPT, defined below). I am proposing it as a litmus test, to distinguish
genuine CI from its many claimants. What do our TPT followers think? #Bookofwhy

10 26 19 8 44 (1/ ) R d k ki "Wh t' i l b t th Ch ll f UM ?" E i i hi


10.26.19 8:44pm - (1/ ) Readers keep asking: "What's so special about the Chancellor of UMass?" Examime is his
original statement: https://umass.edu/newsoffice/article/statement-university-massachusetts-amherst . As you
can see, he does not stop at the mantra: "Academic boycotts are antithetical to academic freedom" but
continues with a moral indictment:
10.26.19 8:44pm - (2/ ) "the BDS position in general fails to acknowledge the humanity on the Israeli side of the
conflict" and "is the antithesis of our commitment to inclusion". Finally, a Chancellor who speaks like a
Chancellor, and fearlessly tells the campus where BDS stands on the moral scale

10.26.19 7:55pm - (1/ ) I am retweeting, because the question: "Compare the PO vs. DAG approaches on
applied, 'real-life' data" comes up once a week. Here it is: Take any applied paper labeled "PO" and re-label it
"DAG" after adding to it three tiny ingredients: (1) Show how the modeling assumptions
10.26.19 7:55pm - (2/ ) emerge organically from our scientific understanding of the problem, (2) Show if there
is anything we can do if it doesn't thus emerge, and (3) test whether those assumptions are compatible with the
available data. Once added, Bingo! We have an "applied DAG" approach,
10.26.19 7:55pm - (3/3) demonstrated on "real life" data, that can be compared to the "applied PO approach" or
"applied quasi-experimental approach" on all counts, especially on: (1) transparency, (2) Power, and (3)
Testability. #Bookofwhy

10.26.19 7:15pm - (Replying to @SamuelW09371956 @fuiud and 2 others) Why not? Epidemiologists are so much
better than me in conducting experiments that I chose not to compete. Economists are better too, but their
mind has been locked into the "credibility" movement, and only now their students are saying: "Wow! No one
told us" #Bookofwhy

10.26.19 5:37pm - (Replying to @fuiud and @PHuenermund) Most of applied work in epidemiology since 2000
uses DAGs. So much so, that 2018 saw a couple of papers protesting "the tyranny of DAGs in epi". But speaking

of high school, @smueller has been teaching HS using Primer. Kids understand that toy problems are"applied
work"#Bookofwhy

10.26.19 5:20pm - (Replying to @fuiud and @PHuenermund) When I said "people who never used DAGs" I meant:
"people who never used DAGs, even on toy problems, because toy problems force them to deal with issues that
"applied works" allow them to hide. #Bookofwhy

10.26.19 5:54pm - (Replying to @fuiud @SamuelW09371956 and 2 others) When I was in high school I fell asleep
reading other people's experimental work, all I wanted to know was how I can apply the method learned to my
own problems, or to problems I am likely to encounter in the future. #Bookofwhy

10.26.19 4:45pm - (Replying to @fuiud and @PHuenermund) Your description of "DAG approach" comes from
people who never used DAGs. Correct description: Take any applied PO+IV study, and add a tiny question: How
do we know, given our problem description, that Z is a good IV? Once you answer it, you got an "applied DAG"
study. #Bookofwhy

10.26.19 4:30pm - (Replying to @fuiud and @PHuenermund) If it is seeing PO vs DAG, then the answer is even
simpler: Take any applied paper labeled "PO" and re-label it "DAG" after adding to it one tiny ingredient:
showing how the assumption W||X,Y(0),Y(1) emerges from students understanding of the problem. Bingo! PO
vs. DAG #Bookofwhy

10.26.19 4:16pm - (Replying to @fuiud and @PHuenermund) If it is seeing PO vs DAG, then the answer is even
simpler: Take any applied paper labeled "PO" and re-label it "DAG" after adding to it one tiny ingredient:
showing how the assumption W||X,Y(0),Y(1) emerges from students understanding of the problem. Bingo! PO
vs. DAG #Bookofwhy
10.26.19 4:03pm - (Replying to @fuiud and @PHuenermund) If it is seeing PO vs DAG, then the answer is even
simpler: Take any applied paper labeled "PO" and re-label it "DAG" after adding to it one tiny ingredient:
showing how the assumption W||X,Y(0),Y(1) emerges from students understanding of the problem. Bingo! PO
vs. DAG #Bookofwhy

10.26.19 6:41am - (Replying to @RogerFrigola and @DOAJplus) No, I do not know. In what context can it be
useful?

10.26.19 1:58am - (Replying to @fuiud) Did you look into the example I suggested, p.67 of:
https://twitter.com/yudapearl/status/1188004711107969024 I have found it to be excellent for teaching,
because it demonstrates both backdoor and frontdoor and it starts with data (Table 3.1), not a distribution. Try
it. #Bookofwhy

10.26.19 1:27am - (1/ ) (Replying to @gerdienja @analisereal and 2 others) Very interesting paper. I never
heard of Guala theory, stating: "if one has (a) some controlled initial conditions in the laboratory and (b) some
observed experimental result on the one hand and given (c) some observed properties of the target system and
(d) some observational
10.26.19 1:34am - (2/ ) (eplying to @yudapearl @gerdienja and 3 others) data on the other, then (by analogy) c
stands in the same (causal) relation to d as a stands to b (Guala 1998)." Did you check if this rule works on the 3
canonical examples in Fig. 1 of https://ucla.in/2N7S0K9 ?? Off hand it seems it does not. Are you in the mood to
improve
10.26.19 1:47am - (3/3) (Replying to @yudapearl @gerdienja and 3 other) Guala's theory so that it covers what
we have learned about extrapolation in the past decade? A summary of what we can do today is given in my
comments on Deaton & Cartwright https://ucla.in/2JfdQcv which you cite. I think Guala would appreciate it.
#Bookofwhy Cor Fig1->Fig3

10.25.19 7:50pm - Anders, I failed to understand your post. You start by describing the problem nicely but,
then, instead of specifying what information you have or don't have, you rush to tell readers what's wrong with
Bareinboim and Pearl. Completing the specification would help #Bookofwhy #Bookofwhy

10.25.19 7:47pm - (Replying to @AndersHuitfeldt and @analisereala) I was surprised to find out that parametric
assumptions were NOT crucial for so many problems. So I am curious to see one problem, fully specified, where

it IS crucial. #Bookofwhy

10.25.19 5:02pm - (Replying to @AndersHuitfeldt and @analisereal) I am honored to be designated as the one
who changes the way "almost everyone" is talking. If we know Pr(Y=y|do(x)) for every x and y, surely we can
compute ratios, differences, logarithms and arctangents of various combinations of x and y. #Bookofwhy

10.25.19 3:38pm - (Replying to @AndersHuitfeldt and @analisereal) Equating "effect sizes" with "parametric"
analysis is new to me. Surely, "size" stands for a number. But that does not mean that we should commit to a
specific functional form in analyzing questions of identifiability or heterogeneity. #Bookofwhy

10.25.19 3:31pm - (Replying to @AndersHuitfeldt and @analisereal) There are 2 reasons for doing
"nonparametric" analysis: 1. Lacking knowledge of param. form. 2. Generality. If I can write an effect
nonparametrically, eg. ACE=P(y|x), then this should hold for ANY parametric form in which I choose to estimate
P(y|x), depending on data.#Bookofwhy

10.25.19 3:20pm - (Replying to @AndersHuitfeldt and @analisereal) I am not interested in winning friends as
much as I am interested in doing things correctly for the set of problems that I find challenging Please tell us
much as I am interested in doing things correctly for the set of problems that I find challenging. Please tell us
what definition "makes more sense", why hide it? #Bookofwhy

10.25.19 3:15pm - (Replying to @AndersHuitfeldt and @analisereal) I believe my def. of "heterogeneity" is the
same as yours. Here it is: "Treatment effect varies across populations". Why mystify it? Traditional Meta-analysis
deals with "distributional variations across populations" which is a different species. #Bookofwhy

10.25.19 1:26pm - I agree with @analisereal observation that a complete solution to external validity problems
is available for the nonparametric case. Moreover, the many "meta-analytical" papers that I have read, are
simply not dealing with the problem of heterogeneity. #Bookofwhy

10.25.19 1:16pm - (Replying to @fuiud @PHuenermund and @Jabaluck) Every applied paper is using DAG for
identification, some explicate portions of it, some explicate conclusions based on it (eg ignorability) and most
try to keep the DAG secret, informal. A DAG is simply the way scientific knowledge is stored in the mind, no
escape. #Bookofwhy

10.25.19 5:07am - (Replying to @PHuenermund) But what do you make of the fact that so many people "liked"
to believe that ***convolution solves causality**. Are they really unaware of the "curve fitting barrier" or just
expressing a wish that what they understand would solve what they don't ? Puzzled. #Bookofwhy

10.25.19 3:19am - I think I heard you from outside, and I have answered it inside at great length. No, DAGs do
not replace that crucial step, they do much more. They give you a language to demystify that mental step
called "research design" by bringing it to where knowledge resides. #Bookofwhy

10.25.19 1:28am - Dear colleagues at UMass, Amherst, Please convey to your Chancellor, Kumble Subbaswamy,
my deepest thanks for demonstrating exemplary leadership, and condemning the BDS movement on hard moral
grounds, the only effective way of exposing racism on campus.

10.25.19 12:30am - For readers asking for hints in solving the puzzle in this slide: https://ucla.in/2N9w2Z8, the
setting assumes that Z2 and W2 are unmeasurable, while other variables are. We need to identify the causal
effect of Z3 on Y (which is confounded) and we ask how to do it using an IV?

10.25.19 12:13am - (Replying to @boredyannlecun and @ylecun) My twitter receives many: "XYZ solves
causality", "We have been doing causality since 1988" etc. I cannot respond to such statements without seeing
how XYZ breaks through the "curve-fitting barrier" of a simple toy problem, such as the many examples
described in #Bookofwhy.

10.24.19 10:15pm - From theory to algorithms. I call on colleagues, members of the National Academy of
Science, National Academy of Engineering and National Academy of Medicine: Let's jolt these 3 organizations to
take action on a matter that is central to their charters and vital to our community?

10.24.19 9:28pm - One promising idea: The creation of a non-profit organization that, leveraging modern
technology, would compete with giants like Elsevier by promoting free dissemination of scientific research, as
opposed to impeding access to such information. NSF,NIH and NAS should initiate it.

10.24.19 6:00pm - An opinion piece that should worry us all: https://the-scientist.com/news-opinion/opinion--


boycotting-elsevier-is-not-enough-66617 An overhaul is needed to make science publishing fair and open. This, I
believe, should be top priority for NSF, NIH, NAS, even Darpa, ie., all Gov's agencies in charge of scientific
research. #Bookofwhy

10.24.19 5:23am - (Replying to @DaveBrady72) Not sure what kind of "post-treatment control" you find hard to
teach. Does it appear in any of the models discussed in "crash course in good and bad controls"? linked here:
https://ucla.in/2ZcRpRq

10.24.19 3:41am - I had great fun dancing with students over the hurdles of causal inference from A to Z. The
slide I have enjoyed most was this: https://ucla.in/2N9w2Z8, since most students were trained in "mostly
harmless", and had to be updated with practical tools. Solution anyone? #Bookofwhy

10.24.19 4:23am - @Susan_Athey Sylvia turned out a brilliant problem solver. She also taught me that the word
"inference" has different meaning to different people. This is what I tried to do throughout the lecture, to
define 'inference". Enjoyed it thoroughly. #Bookofwhy

10.23.19 12:08am - (Replying to @stephensenn and @kaz_yos) Not necessarily. "What if" asks for "effects of
causes", whereas "Why" asks for "causes of effects". See how beautifully the two are handled in counterfactual
logic: https://ucla.in/2L9JHNu #Bookofwhy

10.22.19 6:02am - (Replying to @jazchaz) Thank you Raelle, and may the Muse of History acknowledge the
music of Charlie and Jasper. I will be singing with them. Today is Shmini Atseret, the day Danny was Bar-Mitzva-
ed at the Western Wall, in 1976, and joined the soul of our people.

10.22.19 4:35am - Following my talk at the Business School, I will be heading to Bing Concert Hall for the Daniel
Pearl World Music Days concert at 7pm, https://live.stanford.edu/calendar/october-2019/daniel-pearl-world-
music-days-concert, and sing for world sanity.

10.22.19 4:31am - Colleagues at Stanford and neighborhood may be interested in a talk I will be giving
tomorrow, Oct. 23, 4:30pm at the Grad School of Business, "Class of 1968" bldg, Room 106 Topic: Econometrics
and Causal Inference

10.21.19 8:45pm - Beyond love and hate, the innocent and curious asks: "Do they cross the 'curve fitting'
barrier?" #Bookofwhy

10.21.19 5:45pm - (Replying to @ShMMor @thomaschattwill and 5 others) Batya should not have left, and her
article should have informed the public that Forward Magazine will be joining others in making Zionophobia the
ugliest word in town, because it IS and because the harassment she experienced will continue till it IS. See

10.21.19 5:09pm - (Replying to @alanmpardo) Sounds like an inspirational battle cry: "Let the reading
commence!". I hope it leads to an easy victory. After all, the enemies are exhausted, and history is on your
side! Charge! #Bookofwhy Batya should not have left, and her article should have informed the public that
Forward Magazine will be joining others in making Zionophobia the ugliest word in town, because it IS and
because the harassment she experienced will continue till it IS. See

10.21.19 4:13am - (Replying to @JaredJacobsen2) Thanks for re-tweeting this beautiful explanation of the
distinction between rung-2 vs. rung-3 causation. This question comes back again and again, primarily because
PO-researchers can't get it (cultural blinders), and they still rule the planet. Or do they? #Bookofwhy
10.21.19 3:50am - (Replying to @patrick_s_smart @KateRaworth) . Curious to know what Doughnt Econ is? If you
can summarize it in one Tweet, I promise to do one for you on #Bookofwhy

10.20.19 11:50pm - Todah Rabah Maestro Zubin Mehta. Every October, since 2002, you have dedicated one of
your concerts to the Daniel Pearl World Music Days. Today, as you lead the Israeli Philharmonic Orchestra for
the last time, we salute your vision, artistry and humanity. To Life!

10.19.19 12:01pm - While some media outlets are interpreting the recent Nobel announcement as a rebuke of
those who challenge RCT hegemony, this is not the dominant view among economists. This article takes a more
balanced view of RCT economics https://thefederal.com/opinion/2019/10/16/the-experimentalist-
economists/. #Bookofwhy

10.18.19 10:54pm - (1/ ) Regression analysts, and this means tens of thousands of smart statisticians, are
recovering from a century of internal anguish, in which they had to think causes and do regression. They now
rejoice the power of writing and speaking the same language -- causal graphs, even
10.18.19 10:54pm - (2/2) when analyzing pure prediction tasks. I would encourage the authors, but make sure

the word "causal" is not added to the title just for riding the causal hype. We need to secure the reputation of
"causal inference" as a new way of viewing scientific problems. #Bookofwhy

10.18.19 2:02am - My fingers are frozen, would I be able to Retweet the Untweetable?

10.17.19 10:07pm - We all lament the herd-like mentality that social media is breeding. But how can we
account for the mentality of sovereign states, members of the United Nations, an organization ordained to be
the moral guardian of the human race??

10.17.19 5:58pm - (Replying to @gfrison) Great "seeing" and "doing" experience. One comment: the term "Causal
Bayesian Network" was defined (Causality, ch. 1) by interventions, not counterfactuals. For counterfactuals
(Rung-3) we need functions eg. y=f(x,z,eps), named Structural Causal Models (SCM). #Bookofwhy

10.17.19 12:17pm - (Replying to @rwolffoot @_julesh_ and 3 others) You are right. The firing-squad example in
ch. 1 of #Bookofwhy is deterministic and computing whether "The prisoner would still be dead had soldier A not
fired" requires non-standard logic.

10.16.19 11:22pm - (Replying to @tvladeck @f2harrell and 14 others) Taking a Bayesian viewpoint (which I do
not recommend), the unique problem of CI is to compare DAGs that have same likelihoods yet produce opposite
answers. #bookofwhy

10.16.19 11:12pm - (Replying to @kareem_carr and @SpenceKjell) Why? Do you know anyone who has not read
it yet? #Bookofwhy

10.16.19 11:10pm - (Replying to @f2harrell @tvladeck and 14 others) @f2harrell , Sticking to my "input-output"
religion, can you tell us in "input-output" terms how this paper proposes to answer causal questions about a
population? #Bookofwhy

10.16.19 10:59pm - (Replying to @FJnyc and @IzaTabaro) Mehdi Hasan and other eloquent racists will continue
their hate-song with impunity until our leaders and writers start shaming them with the one ugly label they
cannot dare to disown: "Zionophobe".
10.16.19 10:43pm - (Replying to @Humblefool_14) I do not see anything wrong with this article, every sections I
start sounds reasonable. But I believe that in order to formulate "general principles of explicability" we need
first to solve a toy explanation, e.g., "what caused the fire?" https://ucla.in/2lEOJtO #Bookofwhy

10.16.19 10:27pm - (Replying to @austinvhuang @zacharylipton and @ipam_ucla) I see it for the first time, and I
can't resonate to it because I can't see the input-output relationship. That is, what kind of questions we will be
answer with this new system, and what do we need to specify to get it running. Does anyone see it? #Bookofwhy

10.16.19 5:30am - (Replying to @f2harrell @stephensenn and 13 others) Fair enough. All graph-based CI tasks
take such uncertainty into account, since none relies on the assumption that "treatment improves" things.
Improvements, if any, are inferred from {data+qualitative assumptions}, not assumed. #Bookofwhy

10.16.19 2:53am - Media hype aside, this paper by Duflo etal https://nber.org/papers/t0333.pdf is a beautiful
exposition of the philosophy of experimental economics in 2007, when economists were isolated from new
developments in causal research. Of primary interest is Section 8, External Validity.#Bookofwhy

10.16.19 2:21am - Strange mediatake: In backdrop of Economics Nobel announcement, (an incriminating) letter
panning RCT surfaces https://business-standard.com/article/current-affairs/in-backdrop-of-economics-nobel-
announcement-letter-panning-rct-surfaces-119101500595_1.html Business Standard Among the signatories:
"Judea Pearl of Columbia University" Pleading innocent on all counts https://ucla.in/2JfdQcv #Bookofwhy

10.15.19 8:32pm - (Replying to @AdanZBecerra1 @boback and 15 others) For a smoker, refraining is a giant
treatment

10.15.19 8:26pm - (Replying to @boback @Lester_Domes and 14 others) I believe the 1964 decision to proclaim
smoking a major cause of lung cancer is an example; it was supported by plausibility arguments and
observational studies. #Bookofwhy

10.15.19 8:17pm - (Replying to @boback @Lester_Domes and 14 others) So far, I've learned that EBM is
whatever physicians do, plus preference to published RCT's studies. Anything I've missed in my attempt to
understand why so many people swear by EBM ?#Bookofwhy

10.15.19 8:10pm - (Replying to @boback @dailyzad and 14 others) That's a nasty thing to say.

10.15.19 5:52pm - (Replying to @dailyzad @boback and 14 others) Twitter is great for transmitting (1) substance
or (2) poetry. Confusing when used for transmitting broad "sounds like" or "seems like" #Bookofwhy

10.15.19 4:56pm - (Replying to @Lester_Domes @stephensenn and 13 others) So, what does "quantification" tell
us that is not told by causal calculus? Is it statistical properties of specific estimators on finite samples? or
identification opportunities under specific parametrization of the model? #Bookofwhy

10.15.19 4:48pm - Reading an interesting paper on algorithmic fairness: https://arxiv.org/pdf/1910.04109.pdf.


"Unfairness" = The presence of undesirable or impermissble path-specific effects of sensitive attributes on
outcomes. #Bookofwhy

10.15.19 2:53pm - (Replying to @f2harrell @stephensenn and 13 others) "general quantification weakness"
10.15.19 2:53pm (Replying to @f2harrell @stephensenn and 13 others) general quantification weakness
seems too general and cryptic to me. Specific weaknesses will be addressed with specific attention or, better
yet, with specific solutions #Bookofwhy.

10.15.19 2:42pm - (Replying to @Lester_Domes @RonKenett and 14 others) "And the sailors said, come and let
us cast lots, to find out who is to blame for this ordeal." (Jonah 1:7)

10.15.19 2:15pm - (Replying to @goodmanmetrics @JadePinkSameera and 13 others) I have to take your
question seriously because our Twitter conversations are loaded with complaints such as "Isn't it just traditional
statistics" or "Isn't it textbook economics?" and the complainers are prominent researchers who need guidance
and respect. #Bookofwhy

10.15.19 9:35am - (Replying to @goodmanmetrics @JadePinkSameera and 13 others) This is easy. The whole
#Bookofwhy is dedicated to this question and shows in 10 chapters what questions we can answer through
"causal inference" that we could not answer by traditional statistics. I'll be glad to answer any specific question
about the book, as I have on Twitter.

10.15.19 9:07am - (Replying to @goodmanmetrics @stephensenn and 13 others) You deserve a Nobel for taking
my question seriously and, if I may continue, does the EBM enterprise come with principles of distinguishing

valid from anecdotal evidence? If yes, are those principles a subset or a superset of those analyzed in the
"causal inference" literature?

10.15.19 7:08am - (Replying to @JadePinkSameera @ShannonBrownlee and 13 others) I would give a Nobel Prize
in Education (if it existed) to any person who can explain to ordinary folks (eg me) what "Evidence Based
Medicine" is, and how it differs from traditional medicine. #Bookofwhy

10.15.19 5:21am - (Replying to @AndersHuitfeldt and @PHuenermund) We are discussing a level of


understanding that allows researchers to judge assumptions HARSHLY. Do you honestly believe that any mortal
can judge the plausibility of "as if nature conducted a RCT...." in even simple situations?Say those discussed
here: https://ucla.in/2ZcRpRq

10.15.19 1:30am - (Replying to @PHuenermund) I love how the "well defined" intervention disambiguates the
do-operator.

10.15.19 1:13am - (Replying to @PHuenermund) Beg to differ. Here is an example of an explicit assumption that
no one can judge harshly because no one understands: "Treatment is conditionally independent of the potential
outcome, given covariate C." See how Imbens & Rubin (2015) struggle to explain what it means. #Bookofwhy

10.15.19 12:02am - (Replying to @analisereal and @edwardhkennedy) Gee, I can't hear what I said! Did I
understand the question? Did I answer it adequately? I wish we had a professional transcriber in the audience,
trained on Israeli accents and their variety -- machine learning 101. #Bookofwhy

10.14.19 7:51am - (Replying to @JadePinkSameera) Likewise. I crossed oceans in Muslim-Jewish dialogues and
ate many Baklawas chanting our common Abrahamic tradition. My moment of truth came around 2009 when I
realized that the idea of two ligitimate nations has not begun penetrating the mind-walls of my Palestinian
partners.

10.14.19 7:23am - The Sept 2019 issue of the Journal of Causal Inference is now out and, as promised, I am
linking you to the Table of Content https://tinyurl com/yy4a5746 My paper: "Sufficient Causes: On Oxygen
linking you to the Table of Content https://tinyurl.com/yy4a5746. My paper: Sufficient Causes: On Oxygen,
Matches and Fires" can be accessed here: https://ucla.in/2lEOJtO. #Bookofwhy

10.14.19 7:09am - Sharing a confession of an Israeli woman who shares my views about religion, people-hood,
history and the prospect of peace.

10.14.19 2:03am - (Replying to @aselbst @mikarv and 19 others) That is what they argued against LaPlace "You
will never know all the boundary conditions, give up" And that is why we invented probabilities, to summarize
the unknowns w/o explicating them. #Bookofwhy

10.14.19 1:51am - (Replying to @MehmerPA @learnfromerror and 31 others) Jayne is forgiven for saying: "It's
just our good old product rule", because he took P(A|B) as primitive. Historians will note however that there
was no "|" symbol in Bayes days and writing the ratio P(A&B)/P(B) meant FORMALIZING the expression "given
that we know B". #Bookofwhy

10.14.19 1:35am - Anti-Zionist bigots can't see the evil in their barking. It is not entirely their fault, for no one
asks them to take a good look at the moral mirror. They will stop when Forward Magazine joins others in making
Zionophobia the ugliest word in town. see https://tinyurl.com/h6l3mrd

10.14.19 1:31am - (Replying to @EWilf) Anti-Zionist bigots can't see the evil in their barking. It is not entirely
their fault, for no one asks them to take a good look at the moral mirror. They will stop when Forward Magazine
joins others in making Zionophobia the ugliest word in town. see

10.14.19 1:10am - (Replying to @yudapearl @KyleCranmer and 6 others) I will try to make it to the reception

Monday, and will pray to the Gods of meta-physics on Wednesday.

10.13.19 10:44pm - (Replying to @MehmerPA @learnfromerror and 31 others) Curious, who is the author of the
paragraph you posted? The phrase: "it is really nothing but the product rule of probability theory" betrays
his/her understanding of the rule. #Bookofwhy.

10.13.19 4:00pm - (Replying to @mikarv @aselbst and 19 others) Witch-doctors got structurally disempowered
by modern medicine. Not a great loss. It is not legibility by computers that we should strive to achieve, but
legibility by social scientists and social activists. Legibility by computers comes naturally once we understand
ourselves.

10.13.19 3:45pm - (Replying to @KyleCranmer @ipam_ucla and 5 others) My, My, thanks for posting. I just
discovered that I am scheduled to talk Wednesday pm. Is it an oversight? Or my forgetfulness?

10.13.19 3:17pm - (Replying to @aselbst @andrewthesmart and 19 others) If we wish to maintain hopes of
changing an undesirable societal phenomenon we must create a scientific language to communicate about it,
not leave it at the mercy of emotional obfuscation. #Bookofwhy

10.13.19 1:52pm - (Replying to @andrewthesmart @jainfamilyinst and 19 others) The language of causal
diagrams does not stop you from "immersion in contexts" or from representing "dynamically changing
constructs". It only stops you from leaving things hazy in your own mind or mis-communicating them to others.
#Bookofwhy
10.13.19 3:31am - (Replying to @yudapearl @jainfamilyinst and 20 others) definition matches their conception
of discrimination. The only requirement is that the definition be formal rather than hazy. In return, the diagram
offers ways of applying the definition to data and estimating the degree of discrimination actually observed.
#Bookofwhy

10.13.19 3:22am - (Replying to @jainfamilyinst @zhitzig and 19 others) Issues connected with Race and
Discrimination invariably evoke pleas for re-definitions, without offering alternatives. Causal diagrams do not
dictate one definition of discrimination. They offer a general representation of human behavior, and invite
experts to propose whatever

10.13.19 12:01am - (Replying to @RonKenett) You may be right about the ignorance of our health-care
colleagues at CUNY, but I try to refrain from using charges of anti-Semitism. Why give bigots an excuse to
debate anti-Zionism vs. anti-Semitism when the former is morally more reprehensible and physically more
dangerous.

10.12.19 7:58pm - Speaking of the history of Bayes Rule, my grandchildren found this cartoon, and complained
that my beard is outdated https://youtube.com/watch?v=BcvLAw-JRss #Bookofwhy

10.12.19 4:38pm - Thanks for sharing, @PHuenermund . I've always said econometric will one day reclaim its
glorious past. Why? Because it was hatched in structural models, free of RCT imitations. Nowadays, while still
under hypnotic spell of pseudo-experiments, restoration is inevitable.#Bookofwhy

10.12.19 4:09pm - And 50 years later, in 2019, a fake feminist named Linda Sarsour said: Zionism and Feminism
don't go together. See https://tinyurl.com/y4rjozub

10.12.19 2:26pm - (Replying to @david_colquhoun @PeterMonnerjahn and 29 others) Agree on induction; beg to
differ on randomization, as we say in #Bookofwhy:

10.12.19 2:12pm - (Replying to @JAdP) The idea of being responsible for someone's career change sends shivers
of responsibility in my spine. I hope it was a positive change.

10.12.19 6:30am - (Replying to @cperek @harari_yuval and @sapinker) Yes, thanks for correcting.

10.12.19 6:28am - (Replying to @scottsantens @ElmSahd and 2 others) And I've left supporting Bernie Sanders
when he chose James Zogby as Middle East "advisor". He is now flirting with Linda Sarsour and other Orwellian
Zionophobes.

10.12.19 5:44am - (Replying to @yudapearl @JWSBayes and 32 others) The quote above was actually written by
Edward C Molina, in his 1940 introduction to "Facsimiles of Two Papers by Bayes". Glenn Shafer's "On Bayes two
arguments" (1982) https://projecteuclid.org/download/pdf_1/euclid.aos/1176345974 is by far the best, and
deserves a more elaborate summary (to appear) #Bookofwhy

10.12.19 5:34am - Sorry for the mis-link. The correct link is https://youtube.com/watch?v=qHSzeijQ95I Enjoy
the cosmos. #Bookofwhy

10.12.19 5:22am - (Replying to @JWSBayes @learnfromerror and 31 others) "The discussions and criticisms of
Bayes' postulate have monopolized so many pages of the history of our subject that little, if anything, has been
published on the most important feature of the essay " (Edwards Demmings 1963) The best is Glenn Shafer's
published on the most important feature of the essay. (Edwards Demmings, 1963) The best is Glenn Shafer s
"On Bayes..."#Bookofwhy

10.12.19 4:55am - (Replying to @juli_schuess) Agree. Woodward never grasped the idea that SCM was designed

to capture human knowledge, not one task or another. He himself uses SCM (DAG) as a basis for defining
"interventions" and then poses "intervention" as a primitive. #Bookofwhy

10.12.19 4:12am - Sharing an illuminating interview with Yuval Harari @harari_yuval and Steven Pinker
@sapinker which would lift your soul as it did mine: https://youtube.com/watch?v=VNIRUigPxA8. I use "soul
lifting" in the AI sense: Take a long-term view of you place in the cosmos, & reset priorities.#Bookofwhy

10.12.19 3:23am - (Replying to @learnfromerror @f2harrell and 30 others) I am not surprised that folks are still
disagreeing on what "being a Bayesian" means. But my question was more innocent: What did Bayes do in 1761
that caused this commotion? After all, he just equated two ways of writing P(A&B), which should have been
equated before. #Bookofwhy

10.10.19 10:23pm - (Replying to @AndrewPGrieve @stephensenn and 29 others) Great question! First, unlike
Bayes Rule, I cannot derive Pythagoras Theorem in one line. So, the question: "Whats the big deal?" does not
apply. Second, Pythagoras Theorem has not been a topic of controversy by philosophers and practitioners alike.
#Bookofwhy

10.10.19 10:09pm - (Replying to @f2harrell @stephensenn and 30 others) Student: What's the big deal? I can
derive it in one line! Professor: Bayes is a way of thinking... Student: Thinking what? And why was this thinking
not done before 1763? Nothing in probability theory prohibits it. I just derived it in one line!! Professor: ???
#Bookofwhy

10.10.19 6:07am - (Replying to @f2harrell @stephensenn and 30 others)

10.10.19 6:07am - (Replying to @stephensenn @omaclaren and 29 others) Sure it is related. So, how do I explain
to students why someone (ie., Bayes) became famous by proving a trivial identity. #Bookofwhy

10.10.19 5:16am - Thank you Raelle for remembering Danny's birthday. Indeed, he would be 56 years old today,
traveling the world with his violin and laptop, spreading truth and understanding, refusing to believe that man
is a predator of another man. Thanks to all who are joining us. HALLELUJAH!

10.10.19 4:26am - (Replying to @stephensenn @omaclaren and 2 others) It may be an excellent thing to do, but
I am concerned only with whether doing so represents "what we know". I am yet to meet someone who can
explain why one prior (on parameters of priors etc. etc...) is scientifically more plausible than another such
prior. #Bookofwhy

10.10.19 3:58am - (Replying to @omaclaren @stephensenn and 29 others) Good article by Hajek, thanks! The
relevant point is: "So while we are free to stipulate that `P(A|B)' is merely shorthand for this ratio, we are not
free to stipulate that `the conditional probability of A, given B' should be identified with this ratio."(p. 101).
#Bookofwhy

10.10.19 3:42am - Interesting observation. Except the 1948 abandonment of Israel was imposed by the U.N.
while the 2019 abandonment of the Kurds is Trump's doing. I always tell Chechoslovakian colleagues we will
never forget what their rifles did for us, targets of the 1948 "momentous massacre".
10.10.19 2:57am - (Replying to @stephensenn @smueller and 28 others) How would Dicing with Death change
the argument that: P(A|B) is just a short hand notation for the ratio P(A,B)/P(B), hence Bayes Rule
P(A|B)=P(B|A)P(A)/P(B) is a trivial identity. We need to explain how one becomes famous for deriving a trivial
identity. #Bookofwhy

10.10.19 2:06am - (Replying to @omaclaren and @Prof_Livengood) Well put. If we "allow probabilities over
whatever", like many Bayesians, then everything is "probabilistic", even Astrology, just assign Pr. to "unicorns
have blue eyes" and other conclusions you desire to derive. Same as spraying priors on parameters (of priors...)
#Bookofwhy.

10.10.19 1:54am - If you examine how "statisticians in the Neyman-Rubin-Holland tradition talk about causation
by way of counterfactuals," you will see that they ignore principle (i) "It is silly to ignore what we know." Their
assumptions reflect what they want, not what they know.#Bookofwhy

10.10.19 1:32am - (Replying to @katchwreck) This is the nature of basic research. But I agree with you that
theoretical understanding of limitations should be part of every research endeavor.

10.10.19 1:21am - (Replying to @leskocar @andrearmolino and 2 others) Disagree here. Stratification and IPW
are two different estimation methods, both requiring a valid choice of covariates from some graphical model,
see https://ucla.in/2KYYviP Section 3.6. #Bookofwhy.

10.10.19 1:10am - (Replying to @leskocar @andrearmolino and 2 others) Agree with @leskocar . The answer
depends on what the regression model is used for, a question that tends to surprise most regression modelers. I
once suggested that we should take the "model" out of "regression models" because they do not model anything.
Do they? #Bookofwhy

10.10.19 12:57am - (Replying to @fledglingStat) Yes there is a rigorous sense of deciding what control is good
and what is bad, see: http://causality.cs.ucla.edu/blog/index.php/2019/08/14/a-crash-course-in-good-and-
bad-control/ As to testing, the answer is NO, see "Why there is no statistical test for confounding..."(Causality
chapter 6) https://ftp.cs.ucla.edu/pub/stat_ser/R256.pdf. #Bookofwhy

10.9.19 11:56pm - (Replying to @learnfromerror @stephensenn and 28 others) In which case P(A|B) is just a
short hand notation for the ratio P(A,B)/P(B), and everything done in the name of Bayes can also be done in
standard probability calculus, absent the "|" symbol. Is that a good summary of your position? #Bookofwhy

10.9.19 11:24pm - Had an invigorating Yom Kippur navigating a new sense of agency. Sang Mi-Haish in memory
of our late son: "... keep you tongue from speaking falsehood..." and spoke on a panel on how we can turn
Zionopobia into the morally ugliest word in town, along: https://tinyurl.com/y66pnnsm

10.9.19 10:28pm - Dana Mackenzie and I are glad that one of the people who helped develop deep learning
understands this limitation. We look forward to writing a second edition of #Bookofwhy where chapter ten will
be replaced with success stories about how AI will overcome its causal blindness.

10.9.19 2:48am - (1/ ) Apologizing to readers for taking a Yom Kippur leave from Twitter. I do not fast. I do not
observe religious rituals. Yet I, like many secular Israelis, observe Yom Kippur "religiously". I wish I could do so
in Israel, where everything comes to a complete halt and "thinking" has
10.9.19 2:48am - (2/ ) a different meaning on that day. I bet our mind is driven by a different algorithm,
relentlessly probing the essence of one's agency. #Bookofwhy
10.8.19 3:53am - An extremely valuable article on the influence of Wikipedia on science. Thanks. A possible
bias: I often use Wikipedia to find citations to articles that I do not want to read. Sometimes to pacify a
reviewer and sometimes to tell readers: "it's there, don't bother" #Bookofwhy

10.8.19 1:40am - Speaking of the spread of "fake waves" into our culture, there is no better example than the
one coming this week from an actual "Minister of Culture": https://youtube.com/watch?v=gxXhI8sC4Og Note the
straight face of an educated person (PhD) as he fakes history to fit an agenda.

10.7.19 7:45pm - And we were taught that mathematics is protected from the "fake wave". It was, until a
colleague sent me this article: https://unherd.com/2019/10/will-maths-succumb-to-the-woke-wave/ Now I
only believe Bible stories, b/c I know the authors and the publisher.

10.7.19 4:49am - Wow! @jazchaz , I did not realize a hashtag #DanielPearlMusicDays exists! Now the whole
world will sing "Mi Haish? Keep your tongue from evil and your lips from speaking falsehood (fake news?). Shun
evil and do good. Seek peace and chase after it."(Psalm). Counterfactual-Hoping?

10.7.19 3:55am - (1/ ) Thanks @jazchaz for posting this photo of my son Danny, showing both his love for music
and his determination to pursue truth and understanding. In two days we will be celebrating Yom Kippur (Day of
Atonement), and I will be singing in his memory: "Who is the person who
10.7.19 3:55am - (2/ ) loves life?... Seek peace and chase after it" (Psalm 34: 13-15). It has become a logo for
Danny's life and legacy. I am not as good as Hava Albertstein: https://youtube.com/watch?v=9NetLhLrqQU, but
I won't miss an opportunity singing it duet with a friend or stranger:

10.6.19 4:14pm - (Replying to @depistemology) I can see the connection now, thanks. Yes, "adjacent possible" is
a good metaphor for the generation of counterfactuals (another is Lewis' "closest possible world"), except that,

cognitively, we perform these local perturbations on a mental model of reality. #Bookofwhy

10.6.19 3:55pm - I see that Elias Bareinboim is speaking at Stanford again, https://stanford.io/336XeOh, so, if
you are around Palo Alto Monday 1pm and want to taste new wine from the causal brewery, it should
strengthen your spine and improve peripheral vision #Bookofwhy

10.6.19 2:17pm - (Replying to @ganesh__s) Well put. I would only add that to appreciate what it does and does
not do we need to simulate big data algorithms on toy models of reality, and test if we get the answers we
expect. #Bookofwhy

10.6.19 2:06pm - (Replying to @depistemology) I'm guilty of missing Kauffman's theory. Would appreciate a
pointer or, better yet, a Twitter-friendly description of principles and "input-output". #Bookofwhy

10.5.19 11:59pm - Ed Leamer's profound and entertaining 1983 classics


https://sites.duke.edu/niou/files/2011/06/Leamer_take-the-con-outsmall.pdf is worth reading for the n-th
time. My favorite slogan: "The mapping is the message", that is, the mapping from assumptions to inferences
does not depend on assumptions. #Bookofwhy

10.5.19 7:13pm - Cheeting with bad controls is easy, as we have seen


here:http://causality.cs.ucla.edu/blog/index.php/2019/08/14/a-crash-course-in-good-and-bad-control/ But it
is getting harder and harder in the age of graphical models, as we can see here:
https://tau.ac.il/~rani/cheating.pdf. #Bookofwhy
10.5.19 2:02pm - (Replying to @DrMoritzWvB @dompagano and 10 others) `People prefer to believe what they
prefer to be true.' True, timely and still, this should not stop us from seeking truth and standing for what we
believe to be true. #Bookofwhy

10.4.19 1:34pm - (Replying to @piamancini and @agentofuser) Glad you found causation interesting. Correlation
is indeed so boring; strange how it kept us mesmerized for over a century. #Bookofwhy

10.4.19 6:18am - @thosjleeper , the high number of "likes" to your post encourages me to retweet it, because I
realize that this crash course : http://causality.cs.ucla.edu/blog/index.php/2019/08/14/a-crash-course-in-
good-and-bad-control/ is really fundamental for anyone using regression and wishing to "see" unbiased effects in
one's own eyes. #Bookofwhy

10.4.19 2:53am - (1/2) As some of you know, October is when we celebrate http://danielpearlmusicdays.org -a
global concert dedicated to "Harmony for Humanity". If you happened to be at Stanford (Oct. 23)
https://tinyurl.com/yypn8lad, Taipei (Oct.28) https://tinyurl.com/y54gcctn, Santa Monica
https://tinyurl.com/y3cjlbs7
10.4.19 2:53am - (2/2) or La Paz https://tinyurl.com/y2l4774o, you are invited to join the celebration and add
an epsilon to world's sanity. There will be many more concerts but, unfortunately, due to malicious hacking, we
can't upload them this year on our website.

10.4.19 2:18am - Sharing a recent paper on mediation analysis, with good introduction to path-specific effects:
https://polmeth.mit.edu/sites/default/files/documents/Zhou-Yamamoto_paths_0.pdf #Bookofwhy

10.4.19 12:17am - (Replying to @firstmn59 @mushfiq_econ and 2 others) Recall that the entire field called
"econometrics" was authored by folks who did not speak DAGs, so the opportunities are enormous. Your topic

should depend on whether your adviser is enlightened or old guard, and if your interests are methodological or
domain specific #Bookofwhy

10.3.19 11:42pm - (Replying to @yudapearl @stephensenn and 29 others) Twitter is a great arena for bouncing
multiple perspectives. I would love to learn from more folks how they explain why Bayes Rule is not a trivial
identity, and what kind of conclusions it allows us to assert that we could not assert from straight probability
calculus.#Bookofwhy

10.3.19 4:59pm - (Replying to @jlanastas @thosjleeper and @tmorris_mrc) We can see it in Carlos and Andrew's
"A Crash Course in Good and Bad Control" https://ucla.in/2ZcRpRq - Models 3 and 5, for example, treatment (X)
is selected by unobservables (U) yet treatment effect is identifiable by OLS. Similar for the napkin problem, etc
etc #Bookofwhy

10.3.19 4:37pm - (Replying to @jlanastas @thosjleeper and @tmorris_mrc) "Selection on unobservables" is not a
hindrance to identification. The phrase was probably coined by folks who did not speak DAGs (check me!)
because DAGs do tell us what is and what isn't hindering identification. I heard economists are updating their
lexicon. Kodos.#Bookofwhy

10.3.19 4:44am - (Replying to @stephensenn @smueller and 28 others) I would be delighted to announce that
I've found a statistician who acknowledges the epistemological essence of Bayes Rule. What does Dicing With
Death say about why Bayes Rule is not a trivial identity? Does it agree with #Bookofwhy page 102?

10 3 19 3:10am (Replying to @smueller @david colquhoun and 29 others) How can a trivial identity inspire so
10.3.19 3:10am - (Replying to @smueller @david_colquhoun and 29 others) How can a trivial identity inspire so
much reverence and controversy? I am yet to find a statistician who agrees that the essence of Bayes Rule is
epistemological, i.e, it formalizes "given that we know". Most think epistemology belongs in psychology, not
statistics. #Bookofwhy

10.2.19 9:44pm - I think it's worth retweeting.

10.2.19 6:55pm - (Replying to @mushfiq_econ @Jabaluck and @amitabhchandra2) Even the most tool-neutral
philosophers cannot but notice that DAGs are second language in epidemiology and taboo in economics. Some
attribute it to differences in research questions, some to types of data. Others, me included, to social &
academic leadership. #Bookofwhy

10.2.19 2:52pm - (Replying to @learnfromerror @twainus and 24 others) @twainus , I would like to join this
request, I tried but failed to understand your theories.

10.2.19 2:54am - (1/2) (Replying to @RonKenett @david_colquhoun and 29 others) There are so many labels
around, that we must start talking substance (eg input-output), away from labels. What in havens is the purpose
of "data integration" if one does not specify the needed output under diverse populations? As to the "domain
adaptation" it has been around
10.2.19 3:05am - (2/2) (Replying to @yudapearl @RonKenett and 30 others) for at least 15-20 years and,
according to my ML trusted colleagues, it has been given up as "undoable". Why? Because if we do not specify
the reasons for data disparities, we can't fuse together heterogenious data properly. Stat. cant fuse Apples and
Oranges.#Bookofwhy

10.1.19 10:48pm - (1/2) Thanks @jlanasta for the link to my new paper "Sufficient Causes: On Oxygen Matches
and Fires" which will appear in the next issue of JCI. See: https://degruyter.com/view/j/jci.ahead-of-print/jci-
2019-0026/jci-2019-0026.xml . I will post here the Table of Content of the entire issue, as soon as it is
available. (Cont. )

10.1.19 10:48pm - (2/2) It is important especially to students entrapped in the PO tradition who cannot make
the transition from rung 2 to rung 3 because, for them, Causal Effects emanate from counterfactuals, and
counterfactuals require a "Treatment". Not so. #Bookofwhy

10.1.19 7:23pm - Totally different? I doubt it. Let's see. What type of descriptions do these "statistical methods"
require about the disparate populations which they aim to "integrate"? As a part-time robot, I can't read Wiley
books, but I can understand "input-output" #Bookofwhy

10.1.19 5:49pm - (Replying to @Lester_Domes @twainus and 11 others) Input-output specification is the only
way to disambiguate the confusion of multi-meaning concepts like "information" "fusion" "integration". It works
in the Ladder of Causation, and it should work elsewhere. Think robots to understand humans. #Bookofwhy

10.1.19 5:37pm - (Replying to @stevesphd @eliasbareinboim and 2 others) DAGs are there to help you precisely
at this stage of your "real life" problem, by allowing you to explicate what you DO and DON'T know about that
"life". Once you do that, you can view the DAG as a TOY, and turn to enjoy "real-life" again. #Bookofwhy

10.1.19 5:13pm - (Replying to @RonKenett @david_colquhoun and 29 others) As a former engineer (Technion
1960) I've learned the power of input-output thinking. So, What is the input needed to do "data integration"
compared with "data fusion" (ala B&P) compared with "meta-analysis"??? Taxonomies based on input-output
distinctions are useful #Bookofwhy

10 1 19 9 12 (R l i R K d id l h d 29 h )I i l i b
10.1.19 @9:12am - (Replying to @RonKenett @david_colquhoun and 29 others) I mention meta-analysis because
Jeske and Xie mention it in their introduction to this special issue. (BTW, my definition of meta analysis:
Averaging apples and oranges to get properties of bananas) Moreover, what is "data integration"? Formally
speaking? #Bookofwhy

10.1.19 @7:34am - (Replying to @RonKenett @david_colquhoun and 29 others) To determine whether these
methodologies are "much wider" or "much narrower" I would propose a litmus test: Solve the toy problems
described in https://ucla.in/2Jc1kdD or https://ucla.in/2N7S0K9 . Meta Analysis is certainly helpless in dealing
with causal disparities #Bookofwhy

10.1.19 @7:20am - (Replying to @JadePinkSameera @stephensenn and 29 others) Heretic? Me? In my circle of
friends I am known as "defender of commonsense"

10.1.19 @7:00am - (Replying to @JadePinkSameera @stephensenn and 29 others) Watch the speakers list and,
unless it contains one or two heretics, I doubt it would be interesting.

10.1.19 @6:55am - (Replying to @david_colquhoun @twainus and 28 others) Who is the offender? Where is the
offense? Willing to defend innocence. #Bookofwhy

10.1.19 @6:50am - (Replying to @DavidSFink @amitabhchandra2 and @AdanZBecerra1) Well put. I am not sure
about (b) but (a) certainly slows down conversations with economists. See my attempt to get them interested in
answers to questions they themselves asked: https://ucla.in/2mhxKdO, #Bookofwhy

10.1.19 @6:34am - (Replying to @GilmanStephenE @DavidSFink and 2 others) Not right! You have thus identified
the causal effect of X on Y contingent on the assumptions that (1) A and B are the only confounders and (2) the
system is linear. Contingent effects are weaker than absolute effects, but they can be VERY useful. #Bookofwhy

10.1.19 @5:58am - (Replying to @david_colquhoun @twainus and 28 others) Yes, its Nancy Cartwright, and she
is also mentioned here: "Challenging the Hegemony of Randomized Controlled Trials" (Deaton and Cartwright
2018.) And in my comments: https://ucla.in/2JfdQcv #Bookofwhy

10.1.19 @5:48am - (Replying to @amitabhchandra2 and @AdanZBecerra1) A DAG is a carrier of knowledge,


hence I find it indispensable in thinking about MTE, or other properties of populations.Not alone, of course, but
in combination with counterfactual logic. It is easier to educate those who misuse DAGs than those who misuse
other texts #Bookofwhy

10.1.19 @5:29am - (Replying to @david_colquhoun @twainus and 28 others) In the language of causal logic, your
contention is a "theorem" - no causes in, no causes out (Cartwright). The wisdom comes in answering: {Some
causes in - which causes out?}. Once we solve this, we can go to: {Some causal assumptions in - which causes
out?}. Not easy #Bookofwhy

10.1.19 @4:45am - (Replying to @david_colquhoun @twainus and 28 others) Thanks for linking me to this
interesting conversation. I see two points over which I normally disagree with statisticians: Bayes rule and
causality. The wisdom of causal logic lies not in assertions from data that don't warrant it, but from data that
DO warrant it. #Bookofwhy

10.1.19 @2:45am - (Replying to @DrColinWPLewis and @d_spiegel) I love the quote "If the headline is
interesting, it probably ain't true". @d_spiegel was also a pioneer in developing graphical models for expert
systems (80's) so, if he says things are hard, he knows what he is saying, and he is ripe for joining the era of
y ( ) y g y g p j g
causation #Bookofwhy

9.30.19 @11:04pm - (Replying to @Jabaluck and @amitabhchandra2) No one is safe from bad assumptions - we
are all humans. But making a bad assumption that is explicitly stated and can be communicated and repaired is
better than concealing one's assumptions under cryptic names. I havn't mentioned "economists"-I know the
sensitivities.#Bookofwhy

9.30.19 @9:31pm - I am retweeting this answer, because the question comes up in almost every interview that I
give: "And what if we insist on using our Gold Standard, RCT, and none other!" Same question comes in ML
context: "And what if you are a model-blind RL expert?" #Bookofwhy @amitabhchandra2

9.30.19 @5:42pm - (Replying to @AdanZBecerra1) DAGs are needed even if one vows to ban observational
studies. Just to store and make sense of the results of several CRT's require some organizational structure (ie
DAGs), not to mention fusing data across populations or correcting for selection bias. #Bookofwhy

9.30.19 @3:19pm - (Replying to @societyforepi @EpiEllie and 5 others) Can someone summarize: What does
social epi bring to the table? The papers I could read do not tell you what. New kind of problems? New kind of
tools? New discoveries? Curious.

9.30.19 @2:21pm - (Replying to @hcp4715 @psforscher and 3 others) I hope Primer did not reverse your faith in
rung 2 and rung 3. As they say in the Talmud (or should have said) "commonsense is irreversible"

9.30.19 @4:12am - (Replying to @Kkushaj and @stanfordnlp) Plenty. See this thread:

9.29.19 @11:45pm - Now, that top quality educational material is available to statistics instructors, what do you
think is the main hindrance keeping them from teaching causality in statistics? notation? mindset? If I was such
instructor, I would jump at the opportunity. Do they? #Bookofwhy

9.29.19 @9:04pm - My sister in Tel Aviv reminds me that 2019 (5,779) also marks the 110 anniversary of the city
of Tel-Aviv, where I happened to get a first peek of the cosmos. Who could imagine this miracle:

9.29.19 @3:41pm - The new year, 5,779-th from creation, is just an hour away and, as an AI student, I can't stop
thinking: What an ingenious programming trick it is (for a robot): "Step back, observe the cosmos, take a tally,
re-set priorities" and make believe it makes a difference. Shana Tovah

9.29.19 @7:05am - Congratulations! Julian, on co-winning the ASA "causality in statistical education Award", and
sharing this interesting thread on how you entered CI from neighboring disciplines. I did not realize they
selected winners in 2019. I hope you help others make the entry. #Bookofwhy

9.29.19 @4:50am - Sharing an insightful article by @vardi https://zimbra.cs.ucla.edu/service/home/~/?


auth=co&loc=en_US&id=746755&part=3, touching on both: the early history of Neural Networks and the
"Usefulness of Useless Knowledge" https://zimbra.cs.ucla.edu/service/home/~/?
auth=co&loc=en_US&id=746755&part=4. On page 111 of #Bookofwhy I describe how NN's influenced my research
in the 1970's.

9.28.19 @11:01pm - (1/2) Welcoming the 2019 Jewish New Year (tomorrow), I wish all readers and followers,
Jews and non-Jews, a Happy New Year, a year of peace and prosperity, progress and understanding, in science
and in the global human condition. As a piece of history, I am sharing this top view
9 28 19 @11 01 (2/2) f th t f th W t W ll i J l
9.28.19 @11:01pm - (2/2) of the stones of the Western Wall, in Jerusalem:
https://twitter.com/HananyaNaftali/status/1178112193881104385 which reminds me how I saw those stones
for the first time, a 3 years old, and how my mother said: "This is why I came to this country." It was 1939,
when her family was stranded in Poland, by the Nazis.

9.28.19 @12:57am - (Replying to @dan_marinazzo @manlius84 and 12 others) Whether causal inference is tricky
cannot be seen in http://stat.columbia.edu/~gelman/arm/chap9.pdfseen because this text is paralyzed by
"ignorability" blinders and, oddly, skips modern developments that make causal inference fun and transparent.
For contrast, see #Bookofwhy or http://bayes.cs.ucla.edu/PRIMER/

9.27.19 @8:36pm - This edited version is an improvement, but my favorite video is still the one I gave in Paris:
https://youtube.com/watch?v=bcRl7sXR1hE (Microsoft, 2012) perhaps because the ideas were still new to me,
fresh from the brewery, or because I was 7 years younger. #Bookofwhy

9.27.19 @2:11pm - Who is doing the work, and who can spoil the work of others is shown visually in Carlos and
Andrew's "A Crash Course in Good and Bad Control" https://ucla.in/2ZcRpRq -A road map guide for the
perplexed traveler, novice and seasoned. Enjoy! #Bookofwhy

9.27.19 @2:02pm - (Replying to @ProfMattFox) Your thought is valid. Adjustment for a parent (surprisingly) does
not "partially adjust" for the child. This is shown in https://ucla.in/2LcpmHz, section 3.2 and explained in
Appendix I. #Bookofwhy

9.27.19 @5:04am - An improved video of my talk at the Why-19 symposium is now available here:
https://youtube.com/watch?v=nWaM6XmQEmU It provides captions of the questions asked during the Q&A
sessions (press the CC button), as well as copies of the slides used. Welcome to the CI-framework. @Bookofwhy

9.26.19 @7:01am - I assume you are as embarrassed by Dr. Mahthir Mohamad as many Malaysians are, and you
are offering an excuse for his mentality. The excuse, however, is more embarrassing than the offense, for it

denies one people the right to self determination. It begs for a major revision.
https://twitter.com/rula_awad/status/1177178637579296769

9.26.19 @5:34am - An interesting paper on causal mediation analysis from Political Science perspective: Tracing
Causal Paths from Experimental and Observational Data X Zhou, T Yamamoto - 2019
https://polmeth.mit.edu/sites/default/files/documents/Zhou-Yamamoto_paths_0.pdf #Bookofwhy

9.26.19 @3:33am - Yesterday we saw a creepy performance by Mahthir Mohamad, PM of Malaysia, who told the
world: "I am proud to be called an anti-Semite", and made the whole world wonder: Why would a Prime Minister
of a respected country expose the level of education he received in that country?

9.26.19 @3:07am - Thank you, @juli_schuess for creating this thread. It is both an honor to the winners and an
invitation to statistics educators to enrich the perspectives of their students with new tools and new
understanding. The class material is all there: Books, slides and notes. #Bookofwhy

9.26.19 @1:08am - Your personal invitation to a Causal dance. A video of my Stanford Why-19 talk is now
available here: https://youtu.be/95aMVpVrkSU Apologizing for the faint audio during Q&A, but at least you can
see the slides regardless of where the speaker stands. Enjoy the dance of WHY #Bookofwhy

9.25.19 @3:53pm - (Replying to @NeuroStats @learnfromerror and @PHuenermund) My "strange birds" comment
p ( py g ) y g
referred to interest in "causal inference", not in "graphical models". Still, I remain open minded, curious to
examine 2020 statistics textbooks, plus next ASA presidential address on the frontiers of stat research.
#Bookofwhy

9.25.19 @12:25pm - (Replying to @learnfromerror and @PHuenermund) There are a few converts, e.g.,
Lauritzen , Dawid, Didelez. But they are considered strange birds in mainstream. #Bookofwhy

9.25.19 @3:49am - This new book: Handbook of Graphical Models


https://stat.ethz.ch/~maathuis/papers/Handbook.pdf captures well how the field has grown since its inception
in the 1980's. Written by top statisticians, it can serve as a good introduction to causal modeling for
traditionally-trained statisticians. #Bookofwhy

9.25.19 @3:03am - And I would challenge any of my peers to convince us that the racist words that Malaysia PM
Mahathir Mohamad is about to spew at Columbia today
https://twitter.com/yudapearl/status/1176379230482886657 are just a matter of "political sensitivity", not a
human right issue, and of no concern to true "scientists".

9.25.19 @2:37am - (Replying to @patrick_s_smart @ForecasterEnten and 3 others) HMM, now we know what
political scientists do in their spare time. Causality cannot explain this link, but it can help reconcile all these
interesting opinions on what makes one factor a better predictor than another. I can see how they can benefit
from a causal graph #Bookofwhy

9.25.19 @2:16am - (Replying to @gomezramirez_ac) Scientists are very much like entertainers. They rely so
heavily on peers approval that one can rarely tell what they truly stand for. Some even use the cover of "no-
politics" to justify their over-sensitivity to peer approval. We need more thoughtful straight-shooters
#Bookofwhy

9.24.19 @5:52pm - (Replying to @gomezramirez_ac and @f2harrell) I do not understand what you deem to be at
odds with scientific ethos. My scientific ethos was authored by people like John McArthy, who threatened to
cancel an AI conference in Tbilisi (1975), unless soviet dissidents get permission to talk.

9.24.19 @5:34pm - (Replying to @exactsake and @f2harrell) Of course I support the ban, this is what the
Massachusetts initiative is all about.

9.24.19 @5:31pm - (Replying to @f2harrell) It's the price one has to pay for trying to be 1. helpful, 2.
thoughtful, 3. truthful and 4. principled. When scholars are being told to shut up, that's when politicians take
charge, and scholars come complaining: "How? On my watch?" Yes, on your watch! #Bookofwhy

9.24.19 @1:45pm - (Replying to @f2harrell) Students at Columbia and other followers expect my support on this
Twitter handle, and they consider their plight to be a moral imperative, not "political views". I concur, and I
can't let them down.

9.23.19 @11:13pm - I join Columbia University students call on President Bollinger to issue a denunciation of Dr.
Mahathir Mohamad, saying: "I can't stop you from speaking, but it's my duty to tell you that you are not
welcome in this University. Your values are not ours."

9.23.19 @10:50pm - (Replying to @malpaso) Having been brought up in Mogadishu, Omar surely met a few
victims of FGC, and many Minnesota voters have hoped that she would raise hell in Congress against the
i d h d ii lik h i M h M i i k l d Wh h '
practice and spearhead petitions like the one in Massachusetts. My point is to make people wonder: Why hasn't
she?

9.23.19 @8:49pm - (Replying to @cmirzayi) Having been brought up in Mogadishu, Omar surely met a few
victims of FGC, and many Minnesota voters have hoped that she would raise hell in Congress against the
practice and spearhead petitions like the one in Massachusetts. Has she?

9.23.19 @5:06pm - My! My! We are talking Massachusetts 2019, not Mogadishu 1600. I bet Congresswoman Ilhan
Omar signed this petition; she was born there, in Mogadishu.

9.23.19 @4:56pm - (Replying to @omaclaren and @HenningStrandin) Good question! Formal Causal Inference is
still in its Embryonic formative stage, with tiny influence in academia, industry and other power houses, so it is
still protected from the dangers of dogmatism and out-datedness that mature disciplines should watch for.
#Bookofwhy

9.23.19 @2:55pm - (Replying to @HenningStrandin) This immediately brings to mind: statistics, economics and
machine learning but, given the professional-personal sensitivities involved, it is risky business to bring anything
to mind. So lets just quote Venn (1834-1923), as if it does not pertain to modern sciences. #Bookofwhy

9.23.19 @2:44pm - (Replying to @gescher) One of my favorite quotes. #Bookofwhy

9.23.19 @5:04am - Sharing an answer posted on Quora to the question: "Why is it that machine learning systems

are black-boxes." https://quora.com/Why-is-it-that-machine-learning-systems-are-black-boxes-even-to-the-


researchers-that-build-them/answer/Judea-Pearl-1?__filter__=&__nsrc__=2&__snid3__=5287085749 It touches
on the two notions of "explainability". #Bookofwhy

9.22.19 @11:53pm - (Replying to @Sisyfuzz) No. I am not familiar with this docuseries. Bill Gates was
instrumental in pushing Bayesian Networks in Microscope, but I have not heard his current opinion on the causal

revolution. Any pointer? #Bookofwhy

9.22.19 @10:26pm - (Replying to @EllieAsksWhy and @wtgowers) But then, how do you handle a new structure,
that does not fit into the one you learned VERY well? If you go through at least one structure-smashing
experience in your life, you develop the muscles to view structures as tools, not sanctuaries. #Bookofwhy

9.22.19 @10:02pm - (Replying to @EWilf and @UNWatch) May the bells of history inspire you as you stand up for
Human Rights and commonsense. And don't forget to use the key word, Zionophobia, which tells HRC delegates:
Recall, your children will one day be reading your words from the record, would you be able to bear the shame?

9.22.19 @4:20pm - (Replying to @mpbennett @Josh_Bersin and 2 others) I still do not see there the distinction
between explaining the program vs. explaining the world. Consider: Q. "Why the increase in crimes?" Ans.
"Because we hired more policemen" #Bookofwhy

9.22.19 @3:58pm - (Replying to @osazuwa) Good point. The difference between SCM and PO however is not that
great, since both are about "world outcomes" not "program outcomes". The latter, is wedded to "treatments",
the latter to "undoing of past events", not necessarily "treatments" of "deliberate actions" #Bookofwhy

9.22.19 @6:13am - (Replying to @wtgowers) We need indeed the more standard paths, but I would not worry
9. . 9 @6: 3a ( eply g to @wtgowe s) We eed deed t e o e sta da d pat s, but would ot wo y
about depleting their rank; the incentives to remain within the bubble far exceed those given to the cross-field
explorers. #Bookofwhy

9.22.19 @5:27am - (Replying to @mendel_random) Indeed, Haldane's quote crossed my desk from your inspiring
tribute to Jerome Cornfield, concluding: "Perhaps one of the advantages that Cornfield had was his lack of any
sustained formal training in either epidemiology or biostatistics". Lesson: To impact stat. - study history.

9.21.19 @10:40pm - (Replying to @AvigailFerdman) Hitcharatnu?

9.21.19 @10:01pm - A brilliant quote crossed my desk which I could not let go un-shared (from Haldane JBS. ,
Science and life , 1969): "I consider it desirable that a man's or a woman's major research work should be on a
subject in which he or she has not taken a degree." #Bookofwhy

9.21.19 @7:54pm - (Replying to @PHuenermund) I was waiting to hear: "... they take you to a special room &
tell you how to get an academic appointment in another econ. department but ...". At great physical risk...

9.21.19 @6:18pm - (Replying to @EWilf) From my explorations, the idea of two-states has not begun to scratch
the surface of their heads. I searched for it in what Palestinian intellectuals write on Israel's independence day

9.21.19 @7:13am - This philosophical paper, http://ceur-ws.org/Vol-2445/paper_7.pdf, titled "Instantiating


Sapience", describes #Bookofwhy as "An engineering approach to AI." I must confess that I cannot understand
the many non-engineering alternatives described in the paper, perhaps one of our readers could.

9.21.19 @5:51am - Beware of an important distinction between "counterfactual explanations" in ML, eg.
https://arxiv.org/pdf/1909.06342.pdf, and in science. The former explains when "system output" would change.
The latter explains when "world outcome" would change. #Bookofwhy

9.21.19 @4:36am - (Replying to @quantumciaran) Oh, I looked at the wrong paper. It is


https://arxiv.org/pdf/1804.04119.pdf that deals with the instrumental inequality, thanks. I was not aware of
these extensions and ramifications in quantum systems. #Bookofwhy

9.21.19 @4:27am - (Replying to @quantumciaran) Fascinating paper. But it is not clear to me which of your
results generalizes the instrumental inequality of https://ucla.in/2pFa9pc ?

9.20.19 @3:41pm - (Replying to @EWilf) Speaking of Zionism and "displacement," I always find it illuminating to
read a page or two in the history of the Yishuv : https://tinyurl.com/yxufyohe and listen again and again to the
clash between the "Me, Me, Me!" and the "We, We, We!"

9.20.19 @4:32pm - I dont dig this theater. It seems like a waste of money to fund a billion $ Apollo program
without knowing what the moon looks likes, or how far it is. I would rather pass another mini-Touring test (as in
#Bookofwhy), or build a robot with sparkling free-will.

9.20.19 @2:40am - (Replying to @smueller) I think the culture that produces Israeli researchers was created by
the Zionist pioneers of 1917-1947 who came there without their professors - a pre-requisite to creative thinking.
(Hard to do in America.) #Bookofwhy
9.20.19 @2:04am - (Replying to @quantumciaran) Thanks for the link. Indeed, the causal discovery perspective
makes sense here: What structure can explain the quantum correlations predicted (and empirically confirmed)
by Bell. They also touch on the smart & under-utilized IV-inequalities https://ucla.in/2pFa9pc #Bookofwhy

9.20.19 @1:37am - It was a pleasure reading your review of #Bookofwhy, partly because you confessed the
pleasure of reading it, and partly because you described it from a fresh perspective, uncharged with
preconceived molds that insiders usually carry. I posted it on http://bayes.cs.ucla.edu/WHY/

9.20.19 @1:37am - It was a pleasure reading your review of #Bookofwhy, partly because you confessed the
pleasure of reading it, and partly because you described it from a fresh perspective, uncharged with
preconceived molds that insiders usually carry. I posted it on http://bayes.cs.ucla.edu/WHY/

9.19.19 @8:31pm - As an American researcher, I am worried about US standing in the Spectator Index. As a
former Israeli, I am surprised at the low number: 8250. In my home town, everyone, even the butcher and the
grocery man were accomplished researchers -- they always knew all the answers.

9.19.19 @7:23pm - The literature on Bell Inequality in quantum physics has been growing fast in recently years,
but remained fairly cryptic to ousiders. This paper turns things around: https://arxiv.org/pdf/1909.05434.pdf.
If you know causal models, you can understand Bell Inequality. #Bookofwhy

9.19.19 @3:39pm - (Replying to @ZahraBilloo) @ZahraBilloo You just dont get it. It is Zionophobia that is
growing and dangerous, as is Islamophobia and so much more. Spell it: Zionophobia - The obsessive animosity
against a homeland for the Jewish people. Please spell it. @WomensMarch @ZionessMovement

9.19.19 @8:22am - Glad you asked, the definition is simple: Zionophobia -- An obssessive animosity toward the
idea of a homeland for the Jewish people. I have defined it in more details here: https://ucla.in/2SxjgWH

9.19.19 @7:32am - Greatest news since I woke up! A Zionophobic bigot identified and voted out. But the
@WomensMarch will remain a fake movement until they add a member of @ZionessMovement to their national

board. After all, don't they speak for ALL women? Have they forgotten feminism's pioneers?

9.19.19 @6:58am - (Replying to @rickwahs and @melb4886) They might, but they don't, because mathematics
overcomes terror through its usefulness. Every high-school kid notices it when terrified in the first class of
algebra. Try it on your favorite research question! #Bookofwhy

9.19.19 @6:49am - (Replying to @rickwahs and @melb4886) Indeed I remember this book and, now, going over
the chapters, I can't help but thinking how clearer many of them be had the authors used the mathematical
language of Beebee Hitchcock and Price. Day and night. Some of the chapters are still in the dark age.
#Bookofwhy

9.19.19 @5:48am - (Replying to @rickwahs and @melb4886) Please teach us what is incremental about the
dramatic shift of philosophers' writings on causation, from informal to mathematical, and what "other people",
beside those recognized and revered in #Bookofwhy we should credit for this shift. Eager!

9.19.19 @5:35am - This picture convinces me that democracy has a life of its own, detached from egos and
personalities, and that, against all odds, history may be on the side of this tiny democracy. Lets hope.
9.19.19 @5:00am - (Replying to @rickwahs and @melb4886) Good question. Could the answer be that many
people are interested in learning new things rather than judging things or people from a distance? Or could it be
that most people, after learning it, understand that causal inference is not a incremental passing hype?
#Bookofwhy

9.19.19 @4:41am - (Replying to @AdanZBecerra1 and @AlexBroadbent) This "end of the day" "practical setting"
aspiration, and how it is informed by the do-operator is shown here: https://ucla.in/2LY8M2X. #Bookofwhy

9.19.19 @4:36am - (Replying to @p_realism and @davidpapineau) I met @davidpapineau in a conference late
1980's and I remember that his lecture was different from those of his fellow philosophers. Can you remind me
of his main contribution? #Bookofwhy

9.19.19 @12:52am - (Replying to @AlexBroadbent and @AdanZBecerra1) Vanderweele does not go all the way,
as I explain here: https://ucla.in/2EpxcNU

9.19.19 @12:45am - (Replying to @EsserHartmut @litgenstein and @tomdrabowicz) I am still eager to know what
principles, methods or tools this literature has spawned that we, lay persons would be deprived of, if not read.
What kind of causal problems can they solve or formulate that we have neglected to address? #Bookofwhy

9.18.19 @11:20pm - (Replying to @EWilf) For years I was wondering if foreign journalists really don't see the
formula, or just play blind, to keep the conversation going. I now know; its the former. Why? b/c to see it you
need to read a chapter or two in the history of the conflict and dig the red lines of the 2 sides.

9.18.19 @10:36pm - (Replying to @PHuenermund @eliasbareinboim and 2 others) I am interested in those who
are NOT using DAG methods. Can they do any meaningful fusion beyond taking weighted averages of two or
more sources? #Bookofwhy

9.18.19 @10:27pm - (Replying to @EsserHartmut @litgenstein and @tomdrabowicz) I must have missed this
literature on Explanative Sociology". Do you know if any of the tools presented in #Bookofwhy owes its existence
to this literature? I hope its not too late to make up for the oversight, in case I failed here
https://ucla.in/2Nbq7W2

9.18.19 @10:13pm - (Replying to @FabItMart and @causalinf) The self-referencing is in your mind. I am pointing
to a dramatic shift in philosophers writings, from informal to mathematical (crediting Wright), which was
surprising to me. By fearing self-referencing, you risk missing the benefits of the shift. Enjoy them first.
#Bookofbook

9.18.19 @8:38pm - (Replying to @AdanZBecerra1) I can only answer such questions if told the principles by
which a package does what it says it does. For a quick answer, if it does not use do-calculus, I doubt very much
a package can solve the simple examples presented here https://ucla.in/2Jc1kdD Please check. #Bookofwhy

9.18.19 @8:27pm - (Replying to @FabItMart and @causalinf) This is precisely what #Bookofwhy does. It starts
with the Garden of Eden, and goes to great pain of explaining the difference between "having interest in
causation" and "mathematizing causation". What are you (and others) objecting to? Eden? The Garden? or the
mathematization?

9.18.19 @7:28pm - For anyone who is interested, Carlos is going to demo the "Causal Fusion" software
https://forms.gle/q1sLgtqkfZ4Nph1p9 tomorrow, during his talk at the SoCal 2019 Methods Conference, UC
Riverside You can see the papers and the agenda at: https://docs google com/document/d/10
Riverside. You can see the papers and the agenda at: https://docs.google.com/document/d/10-
ZO1ElXeABnl_y21gfHxfC2rwAFItsd7n6Lp845w54/edit #Bookofwhy

9.18.19 @7:13pm - (Replying to @FabItMart and @causalinf) It goes back even to the Garden of Eden. See
#Bookofwhy chapter 1. Who does not have interest in a form of thinking that governs everything we know and
do in the world.

9.18.19 @7:08pm - (Replying to @Prof_Livengood @learnfromerror and 6 others) Great bibliography. Refreshing
and proving my point. #Bookofwhy.

9.18.19 @4:47pm - (Replying to @yudapearl @learnfromerror and 6 others) Its hard to swallow. As I discussed on
Amstat News http://bayes.cs.ucla.edu/WHY/AugAmstat_Judea-Pearl.pdf. Someone must be over-generalizing,
it's truly hard to swallow. #Bookofwhy.

9.18.19 @3:36pm - (Replying to @litgenstein @EsserHartmut and @tomdrabowicz) "Interest of philosophers" goes
back to Democritus and earlier. Even statistics emerged from "interest in causality". Who doesn't have such
interest?" I should have said "interest of philosophers in modern causal inference". Thanks. #Bookofwhy

9.18.19 @3:28pm - (Replying to @learnfromerror @JadePinkSameera and 5 others) No I do not overgeneralize.


When I wrote "show no interest" I did not mean "have no interest", I meant "show"!! In writings, in meetings,
etc. Can we count the number of times the word "causal" appeared in your recent summer symposium? Why
deny "show no interest"?#Bookofwhy

9.18.19 @3:14pm - (Replying to @litgenstein @EsserHartmut and @tomdrabowicz) Certain ideas are surely new
in that book, others are old. I am holding Hume (1739) and Mill (1843) in my hands(literally) and, for the life of

me, I can't see how they could resolve Simpson's paradox or other puzzles. Could any pre-1990 philosopher? Why
deny progress? #Bookofwhy

9.18.19 @2:54pm - (Replying to @JadePinkSameera @StatModeling and 5 others) I am of the impression that
such forum would be boring, because the differences, commonalities and miscommunications have been
rehashed to exhaustion in the literature and Twitter. Note also that, sadly, "philosophers of statistics" show no
interest in causal models. #Bookofwhy

9.18.19 @7:45am - (Replying to @yotambarnoy) I am aspiring to the same ending, hoping that acting as equal
will end their separate and hostile existence.

9.18.19 @4:32am - (Replying to @EsserHartmut @ingorohlfing and @tomdrabowicz) In social science, following
Blalock Duncan et al, the ideas of testing and identifications are not new. But if you were a student of Suppes
or Lewis, these ideas are new, b/c you cannot identify the effect of something (eg action) that you cannot
represent. #Bookofwhy

9.18.19 @4:17am - While we are at it, here is another book on experimental philosophy, also dounlaudable,
http://40.114.28.106/pdf/04-08-2019_medium/pdf/9780191028861.pdf#page=302 with a chapter (11) on
"causal reasoning". #Bookofwhy

9.17.19 @11:58pm - An unexpected opportunity emerging from the Israeli election.


9.17.19 @11:32pm - Philosophers of science are beginning to take interest in causal inference. This new book
(downloadable) http://40.114.28.106/pdf/04-08-2019_medium/pdf/9780191063916.pdf#page=130 shows how
traditional problems in philosophy come to light through the new lens #Bookofwhy

9.17.19 @3:08pm - Sharing a nostalgic photo. Some place in my army days (1953-1956) I remember driving a
truck, collecting ballots from remote army units, to be counted. A most heroic endeavor on the roads of those
days. Our mission: Not one soldier left uncounted. It worked!

9.17.19 @2:36pm - (Replying to @jprwg @strangecosmos and 2 others) Knowledge (or "information") is what
constrains your answers and drives them from "maybe" to yes/no or probable. I am reluctant to use
"information" because people tend to confuse it with Shannon's information - a purely probabilistic notion.
"Mechanism" is too narrow #Bookofwhy

9.17.19 @2:10pm - (Replying to @on_clusters @ABravoBiosca and @IGLglobal) The walls of NBER are taller than
the Himalaya. No access to anyone but club members.

9.17.19 @1:47am - (Replying to @hofbeezy) Not only doesn't it hurt, it is also a beautiful poetry; like chanting
Kiddush or singing Hatikva. Its not the words but the melody.

9.17.19 @1:28am - (Replying to @strangecosmos @GaryMarcus and @ylecun) If by "ML techniques" you mean any
future algorithm, then you are justified in saying "I dont know". But if by "ML techniques" you mean algorithms
based on data only, we can tell even today that the answer is NO. We can't compute 3-dim volume from a 2-dim
shadow. #Bookofwhy

9.17.19 @1:17am - (Replying to @RivasElenaRivas) Let's do it together. Fit a 100-layer NN to data coming from
ice-cream sales and crimes. Interpret the fitted NN as structural causal model and ask it: "Would crime increase
if we ban ice-cream?" What answer would we get? #Bookofwhy

9.17.19 @12:54am - (Replying to @strangecosmos @GaryMarcus and @ylecun) Why speculate on what Sutton
means or meant? Do you @strangecosmos believe that any ML technique can solve any of the toy problems in
#Bookofwhy or Primer, given all the data in the world, and no information beside data?

9.17.19 @12:43am - Today, it is an election day in Israel. Public transportation is free, from anywhere to
everywhere. I wish I could take one of those trains and fulfill my duty (I have dual citizenship). For a deep
understanding of issues and moods see

9.16.19 @10:18pm - (Replying to @womensmarch) @womensmarch has just replaced a fake feminist
https://tinyurl.com/y4rjozub with a confessed Zionophobe. Is there anything more weird, fraudulent and
contradictory than an anti-Zionist feminist?

9.16.19 @10:11pm - (Replying to @SethAMandel)9.16.19 @10:10pm - (Replying to @Marissa_Jae)

9.16.19 @9:57pm - (Replying to @haileybanack @EpiEllie and @AmJEpi) Just in case you wish to inflame the
paradox with more fuel, here is a gallon from my own brewery https://ucla.in/2EpxcNU. In #Bookofwhy we
took the liberty of calling the birth-weight phonomenon a "paradox", though it has been explained. Paradoxes
enjoy cultural immortality.

9.16.19 @9:22pm - Retweeting my comment on a controversial change in the leadership of the Women March
p g y g p
movement. An interesting thread.

9.16.19 @8:40pm - (Replying to @ZionessMovement) I just can't agree with your last sentence, implying that
"anti-Zionism" is less racist than antisemitism. The former is not only bigoted but also eliminationist and borders
on genocidal. Are you implying that anti-Zionism is the lesser of the two evils?

9.16.19 @8:25pm - (Replying to @ZionessMovement) Women March, Inc. has just replaced a fake feminist
https://tinyurl.com/y4rjozub with a confessed Zionophobe. Is there anything more weird, fraudulent and
contradictory than an anti-Zionist feminist?

9.16.19 @6:09pm - (Replying to @strangecosmos @GaryMarcus and @ylecun) There are theoretical impediments
that even "general methods" cannot circumvent. If you search for a common point of two parallel lines, you can
have the best search-and-learning method in the world and you won't find one. ML confronts such impediments
to reaching GAI. #Bookofway

9.16.19 @4:22am - (Replying to @ewerlopes) Yes. Causality is heavier, with all the proofs, and histories, and
arguments with philosophers and economists and statisticians. Primer is game-like: "Look Ma, I can do today
what I couldn't yesterday, and it makes so much sense!" You dont want to miss it. #Bookofwhy

9.15.19 @6:38pm - So, perhaps I was too naive in assuming that my colleagues in the stat dpt have been doing
non-parametric estimation for the past two centuries. But now, that they are awaken to the importance of
causal estimands, do they need outside help? #Bookofwhy

9.15.19 @6:04pm - Agree. But why do we have to "think hard" if the task of estimation has been the sole target
of super smart statisticians for over a century, backed by an empire that produced thousands of PhD's. Are we
"smarter" then they? or are CI estimands new to them? #Bookofwhy

9.15.19 @4:32pm - (Replying to @LauraBBalzer and @GilmanStephenE) By all means, assuming someone younger
does the writing, and someone older does the diplomacy. #Bookofwhy

9.15.19 @2:43pm - (Replying to @GarridoWainer) No official document, just a long discussion on Twitter,
followed by a survey whether it is ethical for an author to publish the review he/she received. I believe
majority voted YES. #Bookofwhy

9.15.19 @2:35pm - After #Bookofwhy, the next entry into causal inference is the PRIMER
http://bayes.cs.ucla.edu/PRIMER/. It was praised already by so many readers, so I won't add, except to note
that Wiley is coming up with a clean version next month. In the meantime, the corrected chapters are
accessible.

9.15.19 @2:23pm - (Replying to @GilmanStephenE and @LauraBBalzer) By the HUGE gap, do you mean going
from a finite sample to an estimate of the estimand?

9.15.19 @4:36pm - (Replying to @manuelbaltieri) To make the Ladder of Causation more connected to "thinking"
and cognitive functions, I was considering labeling the rungs: 1. Foresight, 2. Control. 3 Understanding. It rings
better with Toulmin, 1961, "Forecast and Understanding" #Bookofwhy

9.14.19 @7:52pm - (Replying to @vardi) I wonder what Melinda Baldwin's opinion would be on our proposal to
make all reviews public, 5 years after decision (anonymously if requested), so as to remind reviewers of the
hi h j d t f hi t #B k f h
higher judgment of history. #Bookofwhy

9.14.19 @7:52pm - (Replying to @vardi) [You Retweeted] I wonder what Melinda Baldwin's opinion would be on
our proposal to make all reviews public, 5 years after decision (anonymously if requested), so as to remind
reviewers of the higher judgment of history. #Bookofwhy

9.14.19 @4:05pm - (Replying to @PHuenermund) I see nothing wrong in self-citing, especially if no one else has
articulated the idea that you deem relevant. The problem I do see is that most self-citings point to irrelevant
publications, bearing little relationship to the discussion in the text. #Bookofwhy

9.14.19 @3:50pm - Some symbolic gifts fill your heart with gratitude, but thinking what this gift would do to the
spirit of the rockets-stricken children of Sderot, stops your heart from beating.

9.14.19 @2:29pm - In ancient Greece, "over-Democracy" led to the invention of formal logic; they had to reign-
in the endless arguments. What invention will Israel's democracy lead to? A crucial election will take place on
Tuesday. Stay tuned.

9.14.19 @10:56am - In this paper on Off Policy Evaluation https://arxiv.org/pdf/1909.03739.pdf I was


particularly interested in Appendix C: Bridging the Gap between Reinforcement Learning and Causal Inference.
#Bookofwhy

9.14.19 @10:43am - Another causal discovery paper


https://static1.squarespace.com/static/59d5ac1780bd5ef9c396eda6/t/5d4736ce8fefae000125bbf7/1564948177059/Zheng.pdf
. This one applied to diabetes data. #Bookofwhy

9.14.19 @10:33am - This Economics Bulletin paper http://accessecon.com/Pubs/EB/2019/Volume39/EB-19-V39-


I3-P192.pdf applies causal discovery to future prices in the Chinese Stock Index 300. It seems someone is about
to get rich soon. #Bookofwhy

9.13.19 @1:26pm - (Replying to @manuelbaltieri) I trust readers of #Bookofwhy are equipped with night-vision
glasses that tell them right away that no definition of causality can be constructed in probabilistic vocabulary,
no matter how sophisticated or skillful.

9.12.19 @7:12pm - (Replying to @katchwreck) By enlarge, the literature on "dimensionality reduction" stands
orthogonal to causality, by virtue of being statistical. There is however a point of contact when we minimize
the number of covariates we need to adjust for. Beautiful algorithms exist for this task. #Bookofwhy

9.12.19 @4:23pm - Every time a reader praises Primer https://ucla.in/2KYYviP I take a minute and read a
paragraph or two, and come back with an urge to reply: "You are so right!". Today I succumb to this urge and
recommend it to all readers, especially free Ch4, https://ucla.in/2G2rWBv #Bookofwhy

9.12.19 @4:46am - (Replying to @harrydq and @TelegraphTech) Many thanks for this pointer. For the life of me,
the last things we contemplated was getting embroiled in the Brexit debate. On the other had, if the WHY rules
the world, why should this debate be excused. #Bookofwhy

9.12.19 @3:50am - (Replying to @TelegraphTech and @harrydq) Too bad I was blocked by the Pay Wall as it
became interesting. #Bookofwhy
9.11.19 @11:07pm - As 9/11 day comes to an end, I am sharing one of the most profound requiems I have ever
heard: "Kaddish para Daniel" https://danielpearlmusicdays.org/kaddish_para_daniel.php Written by Benjamin
Lapidus, it combines Hebrew, Aramaic and Spanish in a rythm that shakes the foundations of our souls. Kaddish
for 9/11

9.11.19 @10:02pm - (Replying to @BarryOSullivan and @UCC) A milestone in the history of AI.

9.11.19 @8:31pm - (Replying to @BarryOSullivan and @UCC) Wow! And my copy of Boole's Laws of Thoughts
(1854) has the signature of James William Warren, AM, MRIA, Sep. 1864. Was he a professor at Cork? We, book
collectors, form a bond of ownership. A weak form of immortality.

9.11.19 @8:08pm - (Replying to @BarryOSullivan) UCC!! Cork Ireland!! Thats where George Boole wrote The
Laws of Thoughts (1854) and Boolean Algebra was born!! Glad you are still marking the pre-Christian calendar. I
do too!! Though mine is in pre-Christian Hebrew, going (logically) from right to left. Cheers, and many springs

9.11.19 @5:27pm - It is still 9/11, and it is LA, my home town, where a Temple of ex-Morroco Jews has been
vandalized. https://jewishjournal.com/news/los_angeles/304267/baba-sale-congregation-vandalized-with-
free-palestine-graffiti/ The populist slogan "Free Palestine" has become a license for every disgruntled group to
spread its "My-Grievance-Above-All" mentality wherever it clicks.

9.11.19 @4:44pm - The first victim of 9/11, Danny Lewin (Z'L), was an internet innovator and a graduate of my
alma mater, the Technion in Haifa, Israel. He was on board AA Flight 11 from Boston to LA, and was murdered
as he struggled to advance toward the cockpit. Watch

9.11.19 @12:43am - An opportunity to learn DAGs from Felix Elwert, one of causal-inference top teachers:
https://statisticalhorizons.com/seminars/public-seminars/directed-acyclic-graphs-for-causal-inference .
#Bookofwhy

9.11.19 @12:14am - It's 12 midnight, September 11, in LA, an hour I can't forget, b/c my son was murdered by
the same people who made this day unforgettable. I salute the people of Israel who remember my fellow
Americans through these columns of lights in the "9/11 Living Memorial", in Jerusalem.

9.10.19 @5:49am - This paper on "counterfactual fairness" has reached my desk:


https://ijcai.org/proceedings/2019/0199.pdf and reinforced my conviction that "fairness" is a counterfactual
notion, and must hence be managed by structural models - the breeding grounds of counterfactuals.
#Bookofwhy

9.9.19 @10:04pm - History enthusiasts will probably find the discovery of this colorful mosaic to be a proof that
man is a history-seeking machine. The place, Tabgha, is were I spent some of my army days, laden with sweet
nostalgic memories of the sea of Galilee, Kinneret in Hebrew.

9.9.19 @5:04pm - I have added a link to Maudlin's review of #Bookofwhy on https://ucla.in/2KYvzau and,
following our lively discussion here, I've added comments to clarify some not-so-obvious points in the book,
especially the difference between Rung Two and Rung Three in the Ladder of Causation

9.9.19 @2:42pm - Our AI-minded readers should find this NYT interview with Gary Marcus and Ernest Davis to be
illuminating: https://nytimes.com/2019/09/06/opinion/ai-explainability.html #Bookofwhy
9.8.19 @4:49am - (Replying to @PHuenermund and @nickchk) An important distinction that makes the
difference between those who can take a causal problem and bring it to a stage where it can be estimated by
random trees and other statistical tools and those who forget the first stage and assume it was prepared by
someone else. #Bookofwhy

9.7.19 @4:46pm - (Replying to @autoregress) Sorry it I am/was going after the wrong economists, but think
about what an enlightened econ. student feels upon hearing from leaders in his/her field: "We are waiting to be
shown the money". Is this the spirit of Haavelmo, Marschack and Arrow? #Bookofwhy

9.7.19 @4:30pm - (Replying to @soupvector and @aregenberg) Thanks for posting, and welcome to this Twitter
Zoo, where you can meet other readers enjoying their ability to do things they always wanted to do, including
convincing economists that they, too, can do things today they always wanted to do. #Bookofwhy

9.7.19 @4:23pm - (Replying to @autoregress) "I like my flashlights demonstrated in the real world!" but I won't
try them myself, not even near my faithful lamppost, and if someone demonstrates them elsewhere, I say: "We
are different, for us .... for us "real world" is what we read in good econometric journals.#Bookofwhy

9.7.19 @4:00pm - (Replying to @autoregress) The "flashlight" works around your IV lamppost as well, see Fig.
5.1 in Causality (2000), you just need to press on the right button. #Bookofwhy

9.7.19 @3:53pm - (Replying to @autoregress) Imagine how many lost wallets and "good experimental studies"
are awaiting their owners in train stations while economists cling to one lamppost, and won't try a flashlight,
even one that performs well in epidemiology and other train stations. #Bookofwhy

9.7.19 @2:47pm - To all my Brazilian readers, colleagues and students - Happy Independence Day.

9.7.19 @2:38pm - (Replying to @HananyaNaftali) Thank you @HananyaNaftali for helping me start the day on a
positive note. ps. I am trying to contact you by email: judea@cs.ucla.edu

9.7.19 @2:15pm - (Replying to @jon_y_huang @AdanZBecerra1 and 7 others) "will never accept" is a pretty bold

statement for a discipline that prides itself on inventing structural approaches. As a student of history, I am
fascinated by what happened to those economists, and whether modern day econ. students will shake them
away from the new lamppost.

9.7.19 @1:50pm - (Replying to @DorotheaBaur) Not only "he doesn't entirely agree with the ladder" he actually
misses the key transition from Rung-2 to Rung-3, as I explain in an earlier thread:

9.7.19 @12:58pm - (Replying to @DorotheaBaur) Thanks for posting, but note that "Pearl seems to think they
are loaded with philosophical significance" is too humble. The #Bookofwhy actually claims that they are
essential in science, & of practical significance in legal, medical and policy decisions. https://ucla.in/2G2rWBv

9.7.19 @3:33am - (Replying to @sapinker) My first reaction to Maudlin's review was:


https://twitter.com/yudapearl/status/1169340697444413445. The profound separation between interventions
and counterfactuals is shown in https://ucla.in/2Qb1h6v (App. I), and by "Mute!" I mean: It would surprise me
to find an idea from pre-1990 philosophy that I missed.

9.6.19 @12:38pm - (Replying to @djinnome) You are right, thanks for noting.
@ p ( py g @ j ) g , g

9.6.19 @12:35pm - (Replying to @_Srijit) Correct. Like "token" vs. "type", individual vs. population, actual vs.
average.

9.6.19 @5:58am - (Replying to @_Srijit) What you and Maudlin are missing is remembering that interventional
studies are non-deterministic estimating averages over populations (or unknown factors.)The statement
"knowing that I am going to win" is not what the study gives you. I discuss it here: https://ucla.in/2Qb1h6v

9.6.19 @5:36am - (Replying to @yudapearl @arturlsc and 6 others) I am surprised that the computational
aspects of DAGs is so underestimated. DAGs permit us to answer questions which otherwise are intractable.
E.g.,"Tell me if the partial correlation R_{XY.Z} is zero", or "Tell me which parameter is estimable by OLS"
#Bookofwhay

9.6.19 @4:08am - (Replying to @arturlsc @Jabaluck and 5 others) You must be kidding. Can you name another
representation scheme, out of the many others, which allows you to see the testable implications of your causal
assumptions ? #Bookofwhy

9.6.19 @3:02am - (Replying to @asweinmann) Appreciating your kind words. After a week of arguing with on-
lookers, it is refreshing to hear from an unspoiled reader, interested in science, not in arguments. #Bookofwhy

9.6.19 @2:27am - Echoing our discussion of interventions and counterfactuals, I have summarized part of it in
this paper https://ucla.in/2Qb1h6v submitted to the next issue of JCI. Comments are welcome, pointing to
omissions, disagreements and improvements; my deadline = Sep 12 #Bookofwhy

9.6.19 @12:58am - (Replying to @Jabaluck @PHuenermund and 4 others) You have very low opinion of
experimentalists, assuming they are incapable of generalizing from a toy problem to patterns of impediments
they see in their substantive works. #Bookofwhy is written for enlightened experimentalists who, once aware,
will dismantle those impediments.

9.6.19 @12:42am - (Replying to @Jabaluck @PHuenermund and 4 others) Every modeling task assumes away
many difficult things, but at least we can represent them explicitly, and reason about ways of overcoming those
"most difficult parts." I have not seen this ability demonstrated in "mostly harmless". #Bookofwhy

9.6.19 @12:33am - (Replying to @Jabaluck @PHuenermund and 4 others) Beauty! Please teach us the vocabulary
of that "own language" and its internal logic, perhaps it is more effective than DAGs, since quasi-exps. think it is
more "reliable", and has resulted in a "credibility revolution". I am truly curious, as computer scientist should
#Bookofwhy

9.6.19 @12:19am - (Replying to @yudapearl @Jabaluck and 5 others) And what's the added wisdom? Are the
impediments to fitting identification strategies in "real-life studies" different than 1. confounding,2. non-
exclusion, 3.selection bias etc. all of which are representable in "make up" examples if one is serious about
handling them.#Bookofwhy

9.6.19 @12:09am - (Replying to @Jabaluck @PHuenermund and 4 others) Please show us one challenge that
cannot be represented in a made-up example and that only reveals itself in "real-life studies" where, instead of
symbols, variables are decorated with: "return to school" "return to prison" "Years in service" What's the big
deal?#Bookofwhy
9.5.19 @10:36pm - (Replying to @PHuenermund @Jabaluck and 4 others) @PHuenermund please remind me
what a tDAG is. Is it a mental representation of knowledge which "quasi-experimentalists" consult when fitting
an identification template, and which they refuse to commit to paper for fear of appearing traditional, and
exposing assumptions?#Bookofwhy

9.5.19 @10:26pm - (Replying to @Jabaluck @PHuenermund and 4 others) Your skepticism reveals how powerful
DAGs are in unveiling assumptions that DAG-averse cultures hide under the rug, ensuring that no one ever
questions the confidence with which exogeneity or exclusions are overstated in quasi-experiment. The hidden
invites no question#Bookofwhy

9.5.19 @10:01pm - (Replying to @Jabaluck @PHuenermund and 4 others) What about "DAG-invented" strategies
like "conditional IV" or "Instrumental set" or "bow-free" https://ucla.in/2NbSLqj or even Fig. 5.10 ?? Some are
"DAG-invented" and some are "DAG-synthesized". The jury is not stupid, it's just kept ignorant by tribal zealots.
#Bookofwhy

9.5.19 @8:48pm - (Replying to @Jabaluck @PHuenermund and 4 others) I surely believe that DAG as a good
working hypothesis about the process (with Friedman's critics of secondary importance) and, more importantly,
as a model for a huge number of real-life cases in which the mediator is physically "shielded" from the
confounder. #bookofwhy

9.5.19 @8:35pm - (Replying to @Jabaluck @PHuenermund and 4 others) And to clarify, I do it too. If you look at
the text below Fig. 5.10, I search the DAG for patterns of identification and, when not found, I repair the ones
that are repairable. Except, I do it on a drawn DAG, not on my mental DAG, which I believe to be less reliable.
#Bookofwhy

9.5.19 @8:21pm - (Replying to @Jabaluck @PHuenermund and 4 others) This is how Phillip and Sewall Wright
"stumbled upon" IV's, now a celebrated "research design". They started with 4-5 variables, causal relations
among them, Sewall's trick of converting them to covariance constraints and faint hope of inverting them back,
no "design". #Bookofwhy

9.5.19 @8:07pm - (Replying to @RashidaTlaibz) Strangely! I do not see any Jewish Stars among the marchers!
What happened to my old comrades? Sad! @EWilf

9.5.19 @7:31pm - (Replying to @y2silence) Interesting! Does Rosenbaum use the words "cause of crying"? Rubin
proclaimed such questions to be "more of a cocktail conversation topic than a scientific inquiry", thus purging
the word "why" from the vocabulary of his disciples. I thought his verdict still reigns. #Bookofwhy.

9.5.19 @7:14pm - (Replying to @_Srijit) Causal Bayesian Networks are called "Causal" because, unlike ordinary
Bayesian Networks, which are purely associational, they provide answer to all interventional questions: What if
we raise taxes? ie., all policy related questions - just what economists should rejoice #Bookofwhy

9.5.19 @7:02pm - The coin-game (below) exemplifies this kind of contradiction. Win-on-correct-guess is one
model of the world, random winning (ignoring your guess) is another. Both are compatible with experiments,
yet the second says:"No, you would't have lost had you acted differently"#Bookofwhy

9.5.19 @5:43pm - I owe readers an explication of what I mean by: "Intervention studies CANNOT ANSWER
counterfactual questions". "Cannot answer" means that two different world models, both compatible with the
studies, can generate two contradictory answers to the same question. #Bookofwhy
9.5.19 @2:19pm - Glad you re-posted this explanation of the difference between rung-2 (intervention) and rung-
3 (counterfactuals) in the Ladder of Causation. Many researchers still find it hard to swallow (eg Maudlin)
especially RL folks, for whom the world is just "interventions"#Bookofwhy

9.5.19 @5:59am - For a simple example consider a game where we win upon guessing the outcome of a fair coin
and lose otherwise. The action "guess head" has no effect on winning, neither has "guess tail". Yet, upon
winning, we can assert: "Had we acted differently we would have lost".#Bookofwhy

9.5.19 @5:47am - To elaborate, Causal Bayesian Networks [Causality Ch. 3] enable us to compute the effects of
all possible actions, compound actions and actions conditioned on observed covariates and, still, none can
answer the couterfactual:"What if we have done things differently? #Bookofwhy

9.4.19 @10:16pm - (Replying to @Jabaluck @PHuenermund and 4 others) "Real-world" again? Take any one of
those revered "quasi experiment" successes in which exogeneity and exclusion were contested. Done. But I
thought the discussion revolved around what comes first, the DAG or the ident. strategy. So now we start with
an oracle, no DAG.#Bookofwhy

9.4.19 @9:42pm - (Replying to @yudapearl @PHuenermund and 5 others) "Starting with nothing" means no DAG,
no SEM, no IV, just an oracle that can tells you, for every variable that comes to your mind, what the "sources
of variations" are for that variable. #Bookofwhy

9.4.19 @9:20pm - (Replying to @PHuenermund @Jabaluck and 4 others) I tried to stay out of this discussion
because it went over my head. I think @Jabaluck methodology could become clearer if he tells us how one
should handle the problem in Causality Fig. 5.10 (below), step by step, starting with nothing but the desire to
estimate beta.#Bookofwhy

9.4.19 @1:05pm - I just read Maudlin review. It's largely sensible, save for two mistakes: (1) ".. it is not possible
to think causally but not counterfactually." Causal Bayesian Networks demonstrate that it is possible. (2) I was
intimately familiar with the Tetrad project of the 1980's. Mute!

9.4.19 @5:53am - (Replying to @stephenpollard) This precious lady is so saintly innocent, that I begin to believe

she really does not understand why people would call her a racist for supporting a racist movement. Many BDS
supporters can't stomach it: Me? A racist? See https://ucla.in/2SxjgWH

9.4.19 @3:00am - (Replying to @PHuenermund @pierre_azoulay and 4 others) What is the canonical example(s)
economists used to demonstrate "selection on observables"? What are students told to do in such cases?
#Bookofwhy

9.4.19 @2:54am - (Replying to @saurabh_jha21) Causal inference goes beyong ml/dl models, so I do not think
you can implement the former with the latter.

9.4.19 @2:52am - (Replying to @FelixThoemmes and @y2silence) I heard about the Venn diagrams used in the
context of "variance explained". Can you explicate what intuition they support?

9.4.19 @1:26am - (Replying to @y2silence) I interpret Horst's surprise (upon finding a suppressor in the data,
1941) as evidence that regression analysts expect correlations to behave like separation in graphs, that is, if a
node Y is separated from X and from Z it must also be separated from the pair (X Z) #Bookofwhy
node Y is separated from X and from Z, it must also be separated from the pair (X,Z). #Bookofwhy

9.3.19 @11:11pm - Should have used it as a trailer for #Bookofwhy. Don't knock it, the script writer was a
thoughtful philosopher.

9.3.19 @2:42pm - (Replying to @y2silence) And the winner is Yongnam Kim @y2silence !!! The example I had in
mind had S and X interchanged, which works just as well, and tells us that either S or X need to be a collider for
a suppressor to play tricks on us. Now, back to Horst (1941): Why was he surprised? #Bookofwhy

9.3.19 @1:40pm - (Replying to @FelixThoemmes) Keen observation. But the example I have in mind has no
cancellation, and is familiar to every UG student of probability or statistics. #Bookofwhy

9.3.19 @1:36pm - (Replying to @djinnome) The independencies stated are not shown in the graph. eg X and S
are shown dependent. So they need to rest on some compelling process.

9.2.19 @11:45pm - Hate to keep you in suspense. Yes! A super-suppressor does exist! Its a variable S,
uncorrelated with X and Y, that, if added to the regression, turns X from a useless to a perfect predictor of Y.
Can readers guess who S is? His name will tell us what suppression is. #Bookofwhy

9.2.19 @2:09am - (Replying to @matt_vowels) On the other hand Horst (1941) was interested in prediction,
something I did not realize. Evidently, prediction-minded people also have intuition. It comes, I surmise, from
causal assumptions that sneak secretly into intuition about predictions. Deserves some thought. #Bookofwhy

9.2.19 @12:39am - For readers who wrote to me last month about anti-Semitism and anti-Zionism, this latest
article by Gil Troy is the best I've ever read: https://jewishjournal.com/cover_story/303681/the-rise-of-anti-
semitism-and-what-to-do-about-it/ And Weiss's book that Troy reviews:
https://penguinrandomhouse.com/books/616727/how-to-fight-anti-semitism-by-bari-weiss/ is an insightful,
eye-opening microscope of our generation.

9.1.19 @11:46pm - Social Scientist: Look what I found! A suppressor! Statistician: Big deal, it shows in regression
analysis. Computer Scientist: Why were you surprised? Statistician: This is a question for psychologists. Social
Scientist: No, its a question for all scientists Why was I surprised?

9.1.19 @11:35pm - To understand "suppressors", it is instructive to examine a "super-suppressor": A variable S


that is uncorrelated with the regressor X and with outcome Y, yet, when added to the regression equation,
turns X from useless to perfect predictor of Y. Puzzle: Does S exist?#Bookofwhy

9.1.19 @1:41pm - (Replying to @NunezKant) Agree. We sometimes forget what science is all about, and it is
amazing that 12 years after, I still need "words" and "caps and gowns" to be reminded. Thanks for re-posting.
#Bookofwhy

9.1.19 @4:59am - An unprecedented development at the UN. First time this world body addresses the inner core
of the Arab-Israeli conflict, which was discussed earlier on this Tweeter. Kudos to the Brazilian delegate.

9.1.19 @3:25am - Today, September 1st, marks the 10-year anniversary of the publication of Causality (2009,
2nd ed.) https://ucla.in/2BDDTLK I am proud to see that the book has stimulated 14,850 citations on Google
Scholar and that, oddly, I am still agreeing with everything it says. #Bookofwhy
9.1.19 @2:40am - For readers who are tired of listening to my Israeli accent, here is my incredible co-author,
Dana Mackenzie, introducing #Bookofwhy in plain English and lucid eloquence.

8.31.19 @5:05pm - (Replying to @nbarrowman) "Contributing cause" stands between "sufficient" and "necessary"
cause. Note this interesting Wash. Post. definition of "Analysis": Interpretation of the news based on evidence,
including data, as well as anticipating how events might unfold based on past events. All predictive!

8.31.19 @12:19pm - (Replying to @yudapearl @ang_hermann and 3 others) Specifically, the probabilities that
annotate arrows emanating from "action nodes" in a decision tree are P(y|do(x)), not P(y|x), as classical
textbooks might suggest. The former need DAGs to be estimated. #Bookofwhy

8.31.19 @11:55am - (Replying to @ang_hermann @furtadobb and 2 others) DAGs ARE used in the theory of
decision. The reason we estimate P(y|do(x)) is to "Find x that maximizes E[U(y)|do(x)]" where U(y) is the utility
of outcome y. For use in decision trees, see
https://ucla.in/2NnfGPQ#page=50
https://ucla.in/2NnfGPQ#page=52
https://ucla.in/2NnfGPQ#page=54
#Bookofwhy

8.31.19 @11:43am - (Replying to @TimFooler) No, I haven't given any such thought, worth looking into, while
keeping in mind what we want to know and what we do know, ie, input--> output.

8.31.19 @1:45am - (1/3) This is an excellent paper, that every regression analyst should read. Primarily, to
appreciate how problems that have lingered in decades of confusion can be untangled today using CI tools.
What I learned from it was that the "suppressor surprise" is surprising even when
8.31.19 @1:45am - (2/3) cast in a purely predictive context: "How can adding a second lousy predictor make the
first a better predictor?" Evidently, what people expect from predictors clashes with the logic of regression
slopes. The explanation I offered here https://ucla.in/2N8mBMg (Section 3)

8.31.19 @1:45am - (3/3) shows how the phenomenon comes about, but the reason for the clash is still puzzling:
What exactly do people expect from predictors, and why? #Bookofwhy

8.30.19 @12:45am - (Replying to @analisereal @reflecmec and 4 others) One should add here that the second
kind of intervention is identifiable whenever ETT is (ie, effect of treatment on the treated), as demonstrated in
Primer (p. 109-111) primer-ch4: https://ucla.in/2G2rWBv, #Bookofwhy

8.30.19 @3:58am - (Replying to @ClaudeAGarcia) And they dare tell us they have hard time recruiting subjects
for randomized treatment ..... Hoping you have a great summer.

8.30.19 @1:24am - (Replying to @Moshe_Hoffman) Very interesting thread, touching on a long debated concept
"the actual cause" [Causality ch. 10]. An important distinction may illuminate your analysis: "necessary vs.
sufficient" causes. A recent post https://ucla.in/2Qb1h6v demonstrates it in the Oxygen-Match story
#Bookofwhy

8.29.19 @11:28pm - (Replying to @boredyannlecun) According to Von Neumann we should all research
computers. Thermostat controls are an xxx trillion $$$ industry. Counter that fact! #Bookofwhy

8.29.19 @9:45pm - (Replying to @analisereal @thosjleeper and 3 others) Great, you just proved that "linear
models" imply homogeneous effects, but not the other way around. Linear combinations of nonlinear functions
l t ff t h it [A i th t b " ff t " diff E[Y|d ( 1)] E[Y|d ( 2)]
also guarantee effect-homogeneity. [Assuming that by "effects" we mean differences eg E[Y|do(x1)]-E[Y|do(x2)]
@Bookofwhy

8.29.19 @8:02pm - (Replying to @thosjleeper @DanielNevo and 3 others) This has been my consistent usage,
Yes. Although I would not be surprised if someone discovers a "nonlinear model" exhibiting the "effect
homogeneity" property. #Bookofwhy

8.29.19 @7:56pm - (Replying to @ildiazm and @mgaldino) Agree. Parametric regression models comes in two
varieties: 1. Carriers of statistical assumptions, eg. "E[Y|x] is linear in x" and 2. Statements of estimation
strategies, eg. "Find the best linear estimate of Y, given x, regardless of the actual shape of E[Y|x]".
#Bookofwhy

8.29.19 @1:57pm - (Replying to @DanielNevo @ildiazm and 2 others) I believe @analisereal summarized the
"linear model" issues fairly well here: https://twitter.com/analisereal/status/1167158287868891136 A
distinction between "linear in parameters" and "linear in variables" is highly warranted. #Bookofwhy

8.29.19 @5:16am - (Replying to @bwundervald @nickdaleburns and 3 others) Good try, but the nonlinear
function x3=x1*x2 makes the model nonlinear. It is not a matter of convention; it is substantive. In linear
models, all causal effects are the same for all units (ie, all values of the error terms) We cant change this
property by renaming. #Bookofwhy

8.29.19 @2:59am - (Replying to @nickdaleburns @RduvalH and 2 others) A "linear function of the separate
predictors" is nice and explicit. But the words "linear model" may lead to some confusion, for the reasons I
mentioned.

8.29.19 @2:48am - (Replying to @davekarpf) OK, now that you are a national celebrity, can you explain to us
mortals what were you bedbugs a metaphor for? What did you mean to say in that comparison? Who are your
heroes and your bedbugs?

8.29.19 @2:07am - (Replying to @thosjleeper and @mgaldino) Interesting. I was not aware of this confusion. Any
reference to a Social Science book using this nomenclature? (preferably by authors who know the difference
between regression eqs. and structural eqs.)

8.29.19 @1:50am - (Replying to @thosjleeper and @mgaldino) I am eager to learn, which community is it that
labels a regression equation with product terms a "linear model"? Who are the careless authors who would do
so? Why would they do it? #Bookofwhy

8.29.19 @1:41am - (Replying to @causalinf and @Undercoverhist) As tribute to the great fun we had today, I
would like to dedicate a day each month to meet with potential authors of economics textbooks. Authors better
be: 1) Tenured, ie free of peer pressures 2) Aspiring to brighten up the dark sides of econometric education
#Bookofwhy

8.29.19 @12:27am - (Replying to @mgaldino) Sorry if any confusion, but I naively assumed that a regression
equation containing a product term would not be classified as a "linear model". Two reasons: 1. it is not linear.
(2)It is not a "model" of reality (ie a carrier of assumptions), but a tool of estimation. #Bookofwhy

8.28.19 @5:18am - (Replying to @emc2G @DanzigMD and 3 others) And you buy into this racist propaganda?
Would you use the "separate but equal" metaphor on any other 2-states, say US & Canada? or US & Mexico? This
comparison to racial segregation was manufactured by enemies of coexistence, and people in your sphere of
comparison to racial segregation was manufactured by enemies of coexistence, and people in your sphere of
information buy it?? SAD

8.28.19 @5:05am - (Replying to @emc2G @DanzigMD and 3 others) You make it sounds like "subjugation" is an
Israeli pastime recreation, rather than a predicament forced upon them by neighbors who openly declare their
intentions. Don't you read what Palestinians tell their children?

8.28.19 @2:55am - (Replying to @tdietterich) I would be very interested in your collection, when done. I think
scientific creativity and theory formation are structurally similar to improvisational problem solving, as both
invoke "modular template breaking" like the Lion Man in #Bookofwhy Chapter 1.

8.28.19 @1:24am - (Replying to @DanzigMD @emc2G and 3 others) I, for one, never understood what people are
trying to achieve by saying the conflict is "COMPLICATED". An excuse from solving it? A diversion from addressing
its core? To me, it's baby simple: A clash between two legitimate national movements, one says WE, the other
says ME.

8.27.19 @9:45pm - (Replying to @emc2G @DanzigMD and 3 others) The comparison may sound inaccurate if you
do not consider Zionophobia a form of racism, with genocidal intentions. I do, for the reasoned explained here:
https://ucla.in/2SxjgWH

8.27.19 @8:06pm - (Replying to @emc2G @DanzigMD and 3 others) @DanzigMD compared David Duke racism to
the racist rhetoric and activities of Rashida Tlaib. Ashrawi and Miftah are more sophisticated, they know how to
apologize once the damage is done. I have not heard Tlaib apologize for her Zionophobic outbursts. She can't!
betray her base.

8.27.19 @2:57pm - (Replying to @emc2G @DanzigMD and 3 others) You would never get a racist to admit to
what he/she is. The litmus test: Charge them with Zionophobia and see how proud they sing.

8.27.19 @5:40am - (1/2) A bunch of new papers have reached my screen which seem related to discussions we
have had here on tweeter.
[PDF] Transcriptomic Causal Networks identified patterns of differential gene regulation in human brain from
Schizophrenia cases versus controls [Three more...]
8.27.19 @5:40am - (2/2) [PDF] Data Management for Causal Algorithmic Fairness
Counterfactual Reasoning for Process Optimization Using Structural Causal Models
[PDF] Reinforcement Learning is not a Causal problem #Bookofwhy

8.26.19 @11:33pm - @RepSchneider has stood up to @IfNotNow with "shades of gray", but I prefer Joe Biden's
answer to the same bullies (paraphrased): "Occupation! Occupation! I have not met a single Palestinian leader
who is willing to accept Israel's right to exist". Its black and white! No gray!

8.26.19 @9:51pm - (Replying to @f2harrell @EpiEllie and 6 others) This article was written 2008 and, yet, I see
no sign of causal definition of "diagnosis". I wonder if probabilistic notions of diagnosis are still ruling the field?
Or have they been replaced by Diagnosis = Best Explanation, as in https://ucla.in/2G2rWBv ??#Bookofwhy

8.25.19 @10:47pm

8.25.19 @10:47pm - I had the privilege of knowing Danny Cohen in the 1990-1980's and of watching his brilliant
mind at work. Unlike us, back-seat academics, he was an adventurer in real life - a pilot, a fighter, a system
builder and a tough skeptic of AI. He will be missed. #Bookofwhy
8.25.19 @10:18pm - (Replying to @ayusharms) In more advanced studies one need indeed to accommodate
cycles, (See for example Causality p.215 https://ucla.in/2BDDTLK ). DAGs however allow us to leverage the full
power of do-calculus, which has not been matched yet in cyclic systems. #Bookofwhy

8.24.19 @5:23pm - (1/ ) (Replying to @Chris_Auld) 1/I guess what you take as "formally similar" I take as vastly
dissimilar. In one case (X-rest.) I can immediately write down the OLS estimand of EVERY parameter and in the
other (cov-rest.) it is still an open question whether some parameters are identified, awaiting a decision
8.24.19 @5:38pm - (1/ ) (Replying to @yudapearl and @Chris_Auld) of whether those parameters can be solved
uniquely from the covariance matrix [!!! decision] The mystery may be dissolved if you can just walk my
students by the hand in Fig.5.10, & starting with the 3 eqs., show them why adding W-->Z spoils beta and W<--Z
does not. #Bookofwhy

8.24.19 @5:01pm - (Replying to @Chris_Auld) Where discussion has gone? I am trying to extract the set of
principles that leads econ. students towards the solution of Fig. 5.10. So far, a failure. Along the way, you said
"cov. restrictions are just extension of exclusion restrictions" which blew me off 2 miles. #Bookofwhy

8.24.19 @4:39pm - (Replying to @Chris_Auld) My glitch! I meant the effect of Z on X is not identified. The same
as delta in Fig. 5.10. Same as beta, if arrow W-->Z is added. In contrast, with exclusion restrictions alone (& all
eps's uncorrelated), all parameters are OLS identified; not 2SLS, but straight OLS. #Bookofwhy

8.24.19 @12:39pm - (Replying to @Chris_Auld) Here is something it cannot do:


Z----->X---->Y
with Eps(Z) correlated with Eps(X), and Eps(Z) uncorrelated with Eps(Y) By "cannot do" I mean "it cannot
identify effect of Z on Y." Moreover, it cannot tell us in general which system of eqs. is identifiable.
#Bookofwhy

8.24.19 @6:52am - (Replying to @Chris_Auld) I dont see why covariance restrictions are straight forward
extension of exclusion restrictions. The latters permit all effect to be identified by OLS, the former are still an
open problem. Why is Z exogenous? Dont we need to examine the Z eq.? eg. what if we add W-->Z #Bookofwhy

8.24.19 @1:48am - European readers may have interest in this workshop on causality by @RonKenett . I hope he
will share the slide with us, so we can learn what "fishbone diagrams" can do for CI, and how business
applications can benefit from the causal revolution. #Bookofwhy

8.23.19 @6:33pm - (Replying to @Chris_Auld) I am not disputing the necessity or advantage of doing algebra vs.
other methods. I am just asking: What are econ. students taught to do in cases like Fig. 5.10 ? No traps to my
question. Just trying to learn something I could not get from the econ. literature. Help! #Bookofwhy

8.23.19 @3:39pm - (Replying to @Chris_Auld) Agree. But how is this "determination" done? By symbolic algebra?
(super exponential) or, by step by step reasoning, as in Causality p. 153? #Bookofwhy

8.23.19 @3:28pm - (Replying to @analisereal) So, for my understanding, given model 5 below, DR will instruct
us to adjust for both U and Z, correct?

8.23.19 @3:13pm - (Replying to @oacarah @AndersHuitfeldt and 4 others) The content probably add clarity, but
the abstract speaks of protecting "consistency", not "precision" which, to me, adds to the confusion. #Bookofwhy

8.23.19 @10:04am - (Replying to @Jsevillamol) Thank you for making confounders simple (is it an oxymoron?) in
@ ( py g @ ) y g p ( y )
this post. I would suggest another warning, against proxies of mediators, as demonstrated here:
http://ucla.in/2ZcRpRq #Bookofwhy

8.23.19 @8:26am - (Replying to @AndersHuitfeldt @AdanZBecerra1 and 3 others) Thanks for making this
distinction clear. Curious, why has "double robustness" developed in the context of causal inference tasks and
not in classical statistics? Or has it? #Bookofwhy

8.23.19 @2:31am - (Replying to @jon_y_huang @AdanZBecerra1 and 2 others) What is the simplest model to
demonstrate this preference? Doesn't it violate the heuristic advocated in: "A Crash Course in Good and Bad
Control" https://ucla.in/2ZcRpRq #Bookofwhy

8.23.19 @1:42am - (Replying to @AdanZBecerra1 @jon_y_huang and 2 others) I understand that "double
robustness" provides protection against misspecification in your model. Thus, naturally, your model (ie DAG)
needs to be consulted before deciding if protection is needed and, if so, what protection would be adequate.
#Bookofwhy

8.23.19 @12:41am - Watch this video of my friend @DanzigMD about Congresswomen @Ilhan & @RashidaTlaib &
how they were caught partnering w Miftah, a Zionophobic NGO that accused us of using "the blood of Christians
in the Jewish Passover."

8.22.19 @11:47pm - (Replying to @mehdirhasan) One thing you would never get a Zionophobe to accept: that
anyone else has a right to dignity and self determination. They win debates on anti-semitism, but will never
debate the ugliness of Zionophobia.

8.22.19 @10:46pm - I believe my guiding mantra would not be inappropriate here: "Only by taking models
seriously we learn when they are not needed". #Bookofwhy

8.22.19 @10:23pm - (Replying to @josephpapptheat @ADL and 2 others) I believe by "our trauma" they mean the
Zionophobic bigotry of Rep. Rashida Tlaib and the way some Democrats embrace this bigotry.

8.22.19 @2:47pm - (Replying to @FJnyc @mehdirhasan and 2 others) The new Orientalism: Mehdi Hasan is
defining Jewish identity. And CNN pretends he knows what he is talking about.

8.22.19 @12:54pm - (Replying to @Chris_Auld) I am unfamiliar with the evaluation method you mention. How
would your students tackle Fig.5.10 ?? What would be the first step? Input --> output? #Bookofwhy

8.22.19 @12:14am - (Replying to @JohannesTextor @PHuenermund and 2 others) Interesting paper on ranking
efficiency of adjustment sets in linear models. I presume this coincides with the ranking in
https://ucla.in/2PORDX2 for non-parametric models. A skeptical economist may still ask: What about 2SLS? My
answer: Where do econs. see 2SLS? #Bookofwhy

8.21.19 @10:41pm - (Replying to @n_iccolo and @FriedrichHayek) This is the crucial first step. Next, all you
need is to read Primer https://ucla.in/2KYYviP and, believe me, you will be way ahead of most ML folks.
#Bookofwhy

8.21.19 @9:46pm - (Replying to @yudapearl and @Chris_Auld) A question to economists and other folks curious
about the role of "reduced form equation" (RFE). Q. If I were to ask 100 econ students what the RFE's are in this
model: https://twitter com/yudapearl/status/1163370384479019008 would I get one answer? three answers or
model: https://twitter.com/yudapearl/status/1163370384479019008 would I get one answer? three answers, or
ten answers? Can we see one? #Bookofwhy

8.21.19 @9:29pm - (Replying to @FriedrichHayek) This is indeed an awesomely important work, that requires at
least a few days/weeks to digest, with the aim of sorting out what primitive causal templates infants possess,
and how advanced causal structures evolve from those templates. #Bookofwhy

8.21.19 @7:07pm - (Replying to @PHuenermund @djvanness and @Chris_Auld) True, DAGs are not expected to
talk about efficiency, nor does any model of the world, they nevertheless do, to the maximum extent that
efficiency considerations are dictated by the world. Are the alternatives more informative about efficiency?
@Bookofwhy

8.21.19 @6:58pm - (Replying to @analisereal and @Chris_Auld) My earlier tweet seems to have gotten lost from
this thread. Here is is https://twitter.com/yudapearl/status/1163945199258574849 And I wish a seasoned
economist would tell us how his/her students are taught to solve this problem, step by step. #Bookofwhy

8.21.19 @6:18pm - (Replying to @tytung2020) The two laws are described concisely on page 168 here
https://ucla.in/2mhxKdO

8.21.19 @6:47am - Ten years ago I wrote this survey paper https://ucla.in/2N9f28c, which provides a
panoramic view of the various approaches to causal inference. Its aim was to unify, rather than differentiate. I
believe it was successful in showing how they all emerge from two laws. #Bookofwhy

8.21.19 @4:18am - (1/ ) It was not the mean-spirited tone of "uninformed" that triggered my reaction, but the
realization that so many well-intentioned people can psych themselves into believing that slogans are
knowledge. One of the reasons that I often fear left tyranny as much as right tyranny is
8.21.19 @4:18am - (2/ ) that the former strives to base its claims on "universal knowledge". "Everyone knows
that there was no Temple in Jerusalem" said Arafat to Clinton. "Everyone knows that Israel is 'apartheid' state"
chant BDS cronies, even intellectuals. I flip by the sound of those chants.

8.21.19 @1:39am - I haven't seen this intriguing thought experiment before. It is amazing how much truth can
be unveiled through a mechanical transposition of just two words Jewish<---> Muslim. It highlights how crucial
counterfactuals are to scientific thought. #Bookofwhy

8.21.19 @12:41am - (Replying to @lewbel) I thought non-parametric models include non-additive and non-

separable errors. If true, then exog. ensures the OLS identifiability of RFE [Assuming, of course, we agree on
what exogeneity and identifiability mean]. Oh God, when will we have a consensual glossary??? #Bookofwhy

8.20.19 @11:50pm - Thanks for noting the connection between Causality and the search for truth. I have
explained here https://twitter.com/yudapearl/status/1163702856702054400 why I decided to continue both on
@yudapearl . Someone has to counter the sirens of deceit that continue to dishonor the floor of the US
Congress.

8.20.19 @11:28pm - (Replying to @DSPonFPGA and @SoniaCuff) Always a thrill to read ancient papers, like

reading Greek mythology, Gee, what we used to believe in those days! #Bookofwhy

8.20.19 @3:45pm - (Replying to @Chris_Auld) Do we have a list of requirements that would guide my students
towards an econ. solution of this problem? Something in the form: 1. examine the equations, 2. check if there
p g q
exists...such that... 3. next check... #Bookofwhy

8.20.19 @3:38pm - (Replying to @Chris_Auld) This is a very interesting observation, deserving a mention in 3rd
edition. The DAG only says: "use either beta1 or beta, both are consistent". The efficiency comes from noticing
that beta1 invites 2sls est., which is not in the DAG, but in the mind of the modeller. #Bookofwhy

8.20.19 @5:23am - (Replying to @Chris_Auld) According to the definition in Phil Haile slides, Z is not
exogeneous, as it is correlated with W. Can you outline just the conceptual steps of how LII is achieved.
#Bookofwhy

8.20.19 @5:04am - Retweeting Elias talk on "Causal Data Science." If I were in Boston, I would not miss it. But
given the circumstances, we will wait for the utube video. #Bookofwhy

8.20.19 @3:20am - (Replying to @CalumDavey) Appreciate your applaud but, for me, this is not "politics." It is a
matter of being true to myself and my identity. It is paying back to a community that has invested dearly in my
education and is now unable to fend for itself under this new barrage of populist slogans.

8.20.19 @2:39am - (Replying to @melvinwevers) Would "singularities" be more acceptable? But please do not ask
me to change "morally deformed" - this is the lens through which people must judge the consequence of their
words. This is my first book in traditional Chinese. Just the thought that students in remote areas of China are
learning to speak cause and effect sends shivers in my spine. I hope the govt tolerates this revolution.
#Bookofwhy

8.20.19 @2:29am - (Replying to @Chris_Auld) RF isn't useful here because X is not exog. by RF definition. As to
precision, DAGs are not very helpful here, with two exceptions: 1. the partial order defined in
https://ucla.in/2PORDX2. 2. Ratios of correlation coefficients can be estimated by 2SLS. #Bookofwhy
@PHuenermund

8.20.19 @2:13am - (Replying to @Chris_Auld) In Fig.5.10 (Causality p.153) Z is not exogenous, and X is not a
valid IV in the traditional econ sense (exclusion is violated). So, I am not sure it can be solved using traditional
econ methods, and would be eager to learn otherwise. #Bookofwhy

8.20.19 @1:08am - (Replying to @melvinwevers) I would replace "contaminate" with "stain". But "weed" connotes
"undesirable" and "unintended" odd balls. I don't believe my fellow Democrats contemplated this kind of
embarrassment, and most of them would rather see it disappear, if it didn't appear as Trump's victory.

8.20.19 @12:54am - (Replying to @thehuntinghouse) Thanks for granting my people right to a homeland,
something Rashida can never do. But your use of the word "apartheid" proves how easy it is for decent people to
fall victim to deceitful propaganda. Did you really fall for it?

8.20.19 @12:23am - (Replying to @pythiccoder) Thanks for the song. Phil Ochs was my favorite singer in the late
1960's, when I was part of the counterculture revolution. Yes, the fear of getting Blacklisted by Zionophobic
big-mouths is what prevents my fellow Democrats from calling out Rashida's racism. Someone has to do it.

8.19.19 @11:42pm - (1/ ) Hating to insult or disappoint any of my followers, I was seriously considering your
suggestion to create a new twitter handle. But one word you said made me change my mind: "uninformed". I
have been reading, writing and researching the Middle East for the past 83 years. I was
8.19.19 @11:42pm - (2/ ) there when Azam Pasha declared (Oct. 11, 1947) "a war of extermination and
momentous massacre" on a nation of refugees of which I was a son. And I was here at UCLA (2014) when BDS's
Omar Barghouti re-denied my people right to self determination: https://ucla.in/2SxjgWH. A new
8.19.19 @11:42pm - (3/ ) twitter handle will give people of your persuasion the illusion that it is impossible for
an "informed" person to disagree with their bubble of self righteousness and that "informed" people must be
blind to the genocidal aims of BDS and its spokeswomen Rashida and Ilhan.
8.19.19 @11:42pm - (4/ ) I can't do it. I feel an obligation to truth and to history to let followers in your bubble
know that "well informed" people exist who view them as gullible instruments in the service of a racist
movement called BDS. Many of my colleagues feel same, but keep silent. I cant.

8.19.19 @8:22pm - (Replying to @StevePittelli) I would like very much to learn from a native English speaker
what "time in history" these words harken, and why they seem insulting for some people. I chose them as
carefully and as informedly as I could, given what I know about these two ladies stand for. Lifelong learning.

8.19.19 @7:42pm - (Replying to @Jacobb_Douglas @PHuenermund and 2 others) Can you elaborate your "how
far" question? Please use several tweets and be pedantic about the references. By 3.2, do you mean Section
7.3.2? I cannot parse: "If Y=y and U=u, then X's PO=x." Please help

8.19.19 @6:49am - (Replying to @JonAMichaels) I honestly thought I was charitable, given that these two ladies
have been dehumanizing a whole nation on a daily basis. Do you think their hatred is less than "sickly"? Is it
controllable? Is it less dangerous if treated as "healthy"? Truly perplexed.

8.19.19 @4:08am - (Replying to @jmtroos) I tried to be charitable.

8.19.19 @3:32am - (Replying to @mom2phd) I am seriously worried about it. However, given that it can also be
used for identifying patients "most in need" of a given treatment, I hope the net benefit to society will be
positive. BTW, was the technique you mentioned based on counterfactual bounds? #Bookofwhy

8.19.19 @3:18am - (Replying to @BenWinegard @dabblingfrancis and @clairlemon) Interesting!. Do you have the
source? Was I right about the correlation+plausibility combination?

8.19.19 @3:07am - The task of identifying individuals who are "susceptible to persuasion" (or "gullible"), has an
enormous range of applications. Ang's slides tell you how it can be done using counterfactual logic. The
technical paper, with proofs, is here: https://ucla.in/2EgRzvl #bookofwhy

8.19.19 @1:28am - (1/ ) (Replying to @yudapearl and @Chris_Auld) 1/I did some further reading in Mann and
Wald (MW) 1943, and I am fairly convinced now that their motivation was to facilitate identification of
structural parameter and, not knowing S. Wright's, nor any other method, they identified the RF and tried to
solve for the parameters.
8.19.19 @1:41am - (2/2) (Replying to @yudapearl and @Chris_Auld) This brings us to the question of whether
today, that we know many other methods, should RF's be as revered as they were in the past? See what we can
do today with other methods, and where RF will fail. Causality page 153. #Bookofwhy

8.19.19 @1:16am - Another new arrival, for those who do not speak Portuguese:

8.19.19 @1:01am - New arrival: My first Portuguese translation in paperback #Bookofwhy = O Livro Do Porque

8.19.19 @12:53am - Omar and Tlaib are already a "national scandal". Two morally deformed weeds in my party
that have contaminated the US Congress with their sickly hatred of a certain country, and will continue to
embarrass American democracy till someone (their voters?) say: Enough!
y ( ) y g

8.19.19 @12:31am - (Replying to @dabblingfrancis and @clairlemon) In the case of smoking, to be precise, it
was a combination of correlations and "plausibility judgement", which is a type of causal assumptions.
#Bookofwhy Chapter 5 https://ucla.in/2KYvzau

8.18.19 @11:23pm - (Replying to @stephensenn @AlexJohnLondon and 3 others) Perhaps you can also tell us
what functions were assumed for the arrows of Fig. 6.9(b), before you ran the simulation. How you made sure
that the ellipses would be aligned as in Fig. 6.9(a) and, most importantly, what did you expect to learn from
the simulation.#Bookofwhy

8.18.19 @8:16pm - A bit of history. The first BDS-style campaign started April 1, 1933, when the Nazi's
boycotted Jewish businesses. It followed 3 years later, 1936, by the Palestinians, in their efforts to prevent
European Jews from escaping - the most inhumane immigration policy in human history

8.18.19 @4:58pm - (Replying to @AlexJohnLondon @el_hult and 3 others) Don't miss Lord's Paradox, in its
unmolested version. Just two plans, A and B, just two statisticians, just an innocent story, no ghosts, no red
herrings. #Bookofwhy.

8.18.19 @3:11pm - (Replying to @stephensenn) The quoted passage has suffered the wrath of several
misinterpretations, some suggesting a dining Hall serving several diets, and other complications. Fig. 6.9(b)
disambiguate the data generation process. Were your two figures generated by this process? #Bookofwhy

8.18.19 @2:44pm - (Replying to @stephensenn @RonKenett and 2 others) If I have not provided explanation it
must be that I do not understand those figures or, more specifically, how the data were generated, and
whether both were generated in accordance with Fig. 6.9(b) -- the data generating model of Lord's Paradox.
#Bookofwhy

8.18.19 @1:53pm - (Replying to @RonKenett @stephensenn and 2 others) Disagree! None of the versions
separates Diet from Hall-of-Residence from Dining-Hall. Why complicate things? WHY? As summarized here:
https://ucla.in/2YZjVFL Lord Paradox is simple, and the decision between the two analysts is a simple exercise
in causal analysis #Bookofwhy

8.18.19 @2:50am - (Replying to @broudsov and @rkarmani) Much of what was debated for centuries should be
re-debated in our century, because AI brings to the table the first operationalization of ideas that were debated
in the abstract. #Bookofwhy

8.18.19 @1:05am - (Replying to @broudsov and @rkarmani) I feel uncomfortable mixing reality "learning the
cause" with fantasy: "rationalization". If you learn it, then it is there. The word "rationalization" weakens the
necessity of such an understanding, as if there an alternative way of explaining observed regularities.
#Bookofwhy

8.17.19 @10:26pm - (Replying to @heymanitshayden @stephensenn and 2 others) The Primer delves into the
mathematics, and has many illustrating examples, thus empowering you to DO causal inference, as opposed to
TALK ABOUT it. #Bookofwhy

8.17.19 @9:56pm - (Replying to @rkarman) My take: "Statistics is a torturous workaround until we learn the
cause, then it begins to make sense."#Bookofwhy
8.17.19 @9:49pm - (Replying to @stephensenn @MadelynTheRose and @NP_Jewell) To read Primer
https://ucla.in/2KYYviP is an irreversible decision. Like #Bookofwhy, it is going to be painful and, like
#Bookofwhy, it is going to be transformative. Let me know when I can tweet our comrades in the trenches:
"Stephen Senn has joined the revolution."

8.17.19 @4:44am - (Replying to @EngineerDiet @optempirics and @FatWhiteFamily) The words "mathematically
convenient" are misleading when we compare designing digital circuits using Boolean algebra vs. equations of
electrons and holes in semiconductors. The difference is between the doable and the undoable, and can only be
appreciated by doing #Bookofwhy

8.17.19 @2:06am - I retweet this reply because I get many such inquiries: "Everything you can do with DAGs you
can do with ...." The analogy is "Everything you can do with computers you can do with the theory of
semiconductors, yet Boolean algebra helps, and so does programming language"#Bookofwhy

8.16.19 @6:26pm - (Replying to @optempirics and @FatWhiteFamily) I confirm your suspicion that everything
you can represent in a DAG you can also represent in math equations. The former is an abstraction of the latter.
Now, to appreciate the former, take any 3 variables in the latter and check if X is independent of Y given Z.
#Bookofwhy

8.16.19 @12:55pm - (Replying to @mattshomepage) A quick scan warns me of something fishy here: all terms
are probabilistic, I see no causal model, no causal assumptions, hence - no causal conclusions. Unless I am
missing the key, ie. the input model, and then the question arises, why hide it?. #Bookofwhy

8.15.19 @2:57pm - (Replying to @Chris_Auld) Thanks so much Chris. I was about to post a query for the origin of
the name. Now we can examine what Mann and Wald had in mind before the concept got distorted.
#Bookofwhy.

8.15.19 @3:04am - (Replying to @stephensenn and @RonKenett) What's wrong with assuming " the diet being
varied between Halls." or as #Bookofwhy says it: "the students eat in one of two dining halls with different
diets". Each hall serves its own diet. What's so "WRONG" in assuming it, and moving forward to the paradox.

8.15.19 @2:34am - This wonderful road map on "good and bad controls" http://ucla.in/2ZcRpRq reminds me of
a paper I wrote with S. Greenland on "Adjustments and their Consequences" https://ucla.in/2N8vDbX. Here the
same issues are discussed in epi-vocabulary -- good for our dictionary #Bookofwhy

8.15.19 @2:18am - Replying to @stephensenn and @RonKenett) I am incapable of such offense. First, because
terms such as "varying treatment within centers" is not in my vocabulary and, second, because I don't see such
variations in Lord's story, nor in the "clean" data generation process of fig. 6.9. #Bookofwhy

8.15.19 @12:03am - (Replying to @stephensenn and @RonKenett) The adjustment equation is this: P(Y|do(Diet))
= SUM W_I P(Y|Diet,WI) P(WI) taken from https://ucla.in/2YZjVFL, and telling us precisely how things are
estimated. No weaknesses, no "two cases", no complications -- straight causal analysis and a paradox dissolved.
#Bookofwhy

8.14.19 @11:47pm - (Replying to @jdramirezc) I am familiar with the examples cited. But here we are trying to
understand the logic behind them. i.e., what information is ADDED when an economist says "This is a REDUCED
FORM EQ." Or, What would we miss if we haven't heard him say it? Anything useful? #Bookofwhy
8.14.19 @11:33pm - (Replying to @jdramirezc) The supply-demand example was also in Haile's slides, but he
insisted on all arguments being exogeneous. Relaxing this, turns RF into a regression equation - impossible. Our
question: What would science miss if, suddenly, economists forget such a concept ever existed? #Bookofwhy

8.14.19 @10:41pm - (Replying to @jdramirezc) Your interpretation of Reduced Form is much broader than
anything we heard here from economists, and would fit almost anything. Low and Maghir, likewise, are not
providing a definition, they just talk around it. One day, we outsiders will get it too, I am sure. #Bookofwhy

8.14.19 @9:05pm - Continuing our efforts to improve communication with economists, Carlos and Andrew have
posted: "A Crash Course in Good and Bad Control" https://ucla.in/2ZcRpRq -A road map guide for the perplexed
traveler, novice and seasoned. Enjoy! #Bookofwhy

8.14.19 @8:03pm - (Replying to @jdramirezc) Thanks for the reference. We are exploring how economists think,
so every such source is valuable. Oh, almost forgot, what does REF mean to you? #Bookofwhy

8.14.19 @6:04pm - (Replying to @Chris_Auld) Great!! So, if RF is useful "because they can be solved on a
computer", so can many other equations, especially those that are OLS identifiable. Agree? And you and I (not
sure about Imbens) can easily tell who those equations are. #Bookofwhy

8.14.19 @5:59pm - (Replying to @Chris_Auld) I do not view RF as arbitrary. On the contrary, it is very well
defined, once you have a structural model. I am still exploring though how economists think about it, and why
they chose (historically) to give those sort of equation a special name "REDUCED FORM". #Bookofwhy

8.14.19 @5:54pm - (Replying to @jdramirezc) Not really "simpler". Imagine a system with 100 exogenous
variables, in which one structural equation has only three variables. The latter is "simple", "simpler than any
reduced form one can write, yet we are fobidden from calling it "reduced form". #Bookofwhy

8.14.19 @5:05pm - (Replying to @stephensenn and @RonKenett) i) 'controlling for the initial weight' is achieved

through the adjustment equation in my tweet. ii) fig. 6.9b would look the same if diets had been independently
varied at the level of student. 6.9b assumes some students deviate (randomly) from the dictated diet.
#Bookofwhy

8.14.19 @4:44pm - (Replying to @Chris_Auld) Are we in agreement then that identification considerations
played a role in econometricians deciding which "solution" deserve the RF name and which do not? Else, why are
we insisting on "exogeneity" of ALL arguments? #Bookofwhy

8.14.19 @3:57pm - (Replying to @Chris_Auld) This is my point. We have a "solution" to the system of equations
and, yet, we DO NOT call this solution "reduced form". Why, because all observed variables in the RFE must be
exogenous. I have no opinion of my own, just trying to understand how economists think about RFE.

8.14.19 @2:29pm - (Replying to @Chris_Auld) There are many solutions to the system of equations which were
not baptized with a NAME; why RF? E.g., consider the front door: X-->Z--->Y, with X<-U->Y . Y=f(Z,U) is a
solution, and so is Y=g(X,U). Behold, no name and no OLS. Further, why do u say "sometimes"? #Bookofwhy

8.14.19 @6:23am - (Replying to @stephensenn and @RonKenett) To reiterate. #Bookofwhy aims only to resolve
the paradox under the assumptions stated in Fig. 6.9 (b). Here all assumptions are stated, because it is a data-
generating process. Still, the paradox persists in our minds, then resolved., mission accomplished.
8.14.19 @5:38am - (Replying to @JadePinkSameera) Great! Please tell us if you find Lord's paradox to be
paradoxical and if #Bookofwhy left anything undone in the way it explains away the paradox.

8.14.19 @5:28am - Foreign Affairs Magazine just published an excellent review of "Possible Minds" where the
future of AI is discussed from 25 angles. (link: https://www.foreignaffairs.com/reviews/review-essay/2019-06-
11/ready-robots) foreignaffairs.com/reviews/review-essay/2019-06-11/ready-robots . My chapter can be
viewed here (link: https://ucla.in/2wj4pox) ucla.in/2wj4pox proposing model-based ML to overcome Deep-
Learning limitations.

8.14.19 @5:21am - (Replying to @stephensenn and @RonKenett) #Bookofwhy may fail to do many things under
the sun, but one thing it does not fail to do -- resolve a clash between the two intuitions, a clash that has

baffled many analysts before, and still baffles them today, even under the "clean" data generation conditions of
fig. 6.9(b)

8.14.19 @3:43am - (Replying to @RonKenett and @stephensenn) Different tasks do not make differences in
conclusions. Accounting for "blocking structure" and "study design" and perhaps "quantum uncertainties" are
unrelated the task of resolving a clash between two intuitions that persist EVEN in the "assumed causality links"
#Bookofwhy

8.14.19 @2:05am - (Replying to @yudapearl @Chris_Auld and 7 others) Continuing our explorations of "Reduced
Form Equations" (RFE) and what they mean to economists, I have tweeted this thread: (link:
https://twitter.com/yudapearl/status/1161562085567815682)
twitter.com/yudapearl/status/1161562085567815682 . I hope RFE experts approve of the way I explain it to my
students. #Bookofwhy

8.14.19 @1:55am - (1/ ) Continuing our exploration of "Reduced Form Equations" (RFE) and what they mean to
economists, let me address some hard questions that CI analysts frequently ask. Q1: Isn't a RFE just a regression
equation? A1. Absolutely Not! A RFE carries causal information, a regression
8.14.19 @1:55am - (2/ ) equation does not. Q2: Isn't a RFE just a structural equation? A1. No! Although a RFE
carries causal information (much like a structural equation) the RFE may not appear as such in the structural
model; it is derived from many such equations though functional composition.
8.1 .19 @1:55am - (3/ ) (The output-instrument in the IV setting is a typical example). Q3: One may derive
many equations from a structural model; what makes a RFE so special to deserve its own name. A3: It is
exceptional because it comes with a license of identification by OLS. This is not usually
8.1 .19 @1:55am - (4/ ) the case for other derivable equations, say those relating two endogenous variables.
Q4: But some of those other equations ARE identified by OLS; why haven't they been baptized with a name?

A4:Because traditionally, economists did not have an easy way of telling which equation
8.14.19 @1:55am - (5/5) enjoys identification by OLS and which does not. Now they do, so it is quite likely that
next generation econ. texts will introduce new names. For example, every structural equation identifiable by
OLS should be so recognized. #Bookofwhy @PHuenermund @causalinf @analisereal

8.13.19 @11:02pm - (Replying to @EpiEllie) Intrigued by your "hotter take". Are descriptive analysts using the
term "adjustment" in lieu of "stratification"? How would you explain to them what the difference is? #Bookofwhy

8.13.19 @10:45pm - (Replying to @RonKenett and @stephensenn) Why assume that there are "differences"
between us? I am trying to explain a clash between two strongly held intuitions (no building systems, no
automation) and Senn is trying to do something else (no intuitions no causal assumptions). I do not see
"differences" #Bookofwhy
8.13.19 @9:33pm - (Replying to @Prof_Livengood and @OUPPhilosophy) John to painter: "Paint the wall either
green or purple". Painter to John: "Your wife will get pretty angry." John was imperative, Painter was
indicative. Both used a disjunctive do-operator, as interpreted here: (link: https://ucla.in/2N9nSCV)
ucla.in/2N9nSCV #Bookofwhy

8.13.19 @5:05pm - (Replying to @Prof_Livengood and @OUPPhilosophy) I will be happy to have a long
conversation and to learn what the distinction is between "imperative" vs. 'indicative" in the context of the do-
operator. Do they make different claims? #Bookofwhy

8.13.19 @3:53pm - My latest offering on "Lord's Paradox and the Power of Causal Thinking" with a few goodies
for the curious and empowered (link: https://ucla.in/2YZjVFL) ucla.in/2YZjVFL #Bookofwhy

8.13.19 @8:48am - (Replying to @NikosTzagarakis and @intoolab) Glad another person beside me sees how
important vocabulary is to thinking straight.

8.13.19 @8:39am - (Replying to @BruceTedesco) There is something addictive in the word "Baysianism" (as in
the word "induction"), perhaps because it seems to capture so many cognitive functions that blinds out from
seeing what it can't. #Bookofwhy

8.13.19 @2:31am - (Replying to @imleslahdin and @spyrosmakrid) I commented on this paper a few days ago,
and so did other ML folks.

8.13.19 @2:05am - Very interesting post. I wonder what it would take for a philosopher to read (link:
https://ucla.in/2nZN7IH) ucla.in/2nZN7IH and change the title of Section 2 to read: "What a Great Many
Phenomena Bayesian Decision Theory CANNOT Model" #Bookofwhy @OUPPhilosophy

8.12.19 @4:06pm - (Replying to @AndrewPGrieve) When I see a quote I dont remember, I begin to realize how
old I must be. 105 !

8.12.19 @1:39pm - (Replying to @yudapearl @Chris_Auld and 7 others) Do experts on "reduced-form" approve of
the licenses I am attributing to Y=f(X,Z,U) ? If so, we can continue with what f says about the world, or perhaps
with more licenses. #Bookofwhy

8.11.19 @9:46pm - (Replying to @analisereal @autoregress and 7 others) I bet there is some added value to
saying "reduced forms", perhaps a pedagogical value, but, since I've never used it, its up to those who do to

explain: "what is added?" Perhaps it is just a verbal assurance that someone already taken care of exogeneity.
Hard to guess #Bookofwhy

8.11.19 @9:07pm - (Replying to @Chris_Auld @autoregress and 6 others) Sorry, I thought you said that (quoting):
"equation 3.46 in Causality is a ratio of two reduced forms", so I looked at 3.46 and found there a ratio of two
cond. expectations. I will try to be less pedantic. (Unless you would you like me to be more pedantic?)
#Bookofwhy

8.11.19 @8:55pm - (Replying to @yudapearl @Chris_Auld and 7 others) 2/ infer: E[Y|do(X)] = E[Y|X], E[Y|do(Z)]
= E[Y|Z], and E[Y|do(X),do(Z)] = E[Y|X,Z], regardless of the form of f. License #2. Y=f(X,Z,U) is RFE then so is
W=g(X,Z,U) for some g, where W is ANY variable not in {X,Z}. Hence, E[W|do(X)] = E[W|X]. Will continue if
allowed #Bookofwhy
allowed. #Bookofwhy

8.11.19 @8:39pm - (Replying to @yudapearl @Chris_Auld and 7 others) 1/ What is a Reduced Form Equation
(RFE) - a layman interpretation. Introduction: an RFE is a causal, not statistical statement. It makes causal
claims about the world, and provides licenses for causal inference. License #1- If Y=f(X,Z,U) is RFE, then thou
have the license to

8.11.19 @8:13pm - (Replying to @Chris_Auld @autoregress and 6 others) Sorry, but Eq. 3.46 in Causality reads: b
= E[Y|z]/E[X|z] It is a ratio of two conditional expectations, not "two reduced forms". The former are
descriptive, the latters are causal -- like water and oil. IOW, E[Y|x] conveys no causal information, RF's do!
Can't equate.#Bookofwhy

8.11.19 @7:19pm - (1/ ) (Replying to @Chris_Auld @autoregress and 6 others) I am the last one to claim that a
concept is not "useful" its usefulness however needs to be clear, else it turns into "abusefulness." I am happy to
finally hear from someone what it means to that someone, not to another person, or to some textbook, and I

will try to refine it,


8.11.19 @7:33pm - (Replying to @yudapearl @Chris_Auld and 7 others) so as to explain it to my students, for
whom "meaning" comes either as a claim about the world, or as a license to conclude something about the
world. I will try translating your "meaning" of RF to their language, starting with the licences. Pls check for
promiscuity #Bookofwhy.

8.11.19 @6:01pm - (Replying to @Chris_Auld @autoregress and 6 others) So, @Chris_Auld perhaps you can tell us
in plain language what the sentence "Y=f(X,Z,U) is a reduced form equation" means to YOU, so I can tell my
students how to communicate with economists. Let's forget MHE and ancient history. What does the label RF
add? #Bookofwhy

8.11.19 @5:45pm - (Replying to @steventberry @autoregress and 5 others) So, if we agree that economists
today are not clear on what is meant by RF equation, perhaps we should give them a few hints on how to
explain RF when communicating with DAG-minded folks. Right now, they dont really know what the explanation
should entail. Ready? #Bookofwhy

8.11.19 @3:52pm - (Replying to @steventberry @MariaGlymour and 5 others) We shouldn't argue about ancient
history. I am trying to inform my students and other DAG-minded researchers how to communicate with
economists, i.e., when an economist tells you "This is a reduced form equation" what does he/she mean? What
does the equation claim etc.#Bookofwhy

8.11.19 @3:05pm - (Replying to @omaclaren @fuzzydunlop123 and 7 others) When my grandson insisted on
understanding what makes a water molecule accelerate when it goes through a constriction, I said: its

neighbors from behind are pushing it stronger than the neighbors in front, they are closer, hence more pressing.
#Bookofwhy

8.11.19 @2:33pm - (Replying to @lewbel) I am not aware of the notions of "incompleteness" or "incoherence". Is
there a simple example for the uninitiated?

8.11.19 @2:12pm - (Replying to @onnlucky @dchackethal and 2 others) I am convinced that, if Popper was alive,
he would buy the Ladder of Causation from #Bookofwhy, after realizing that "induction" is too general a term to
be operationalized.
8.11.19 @3:36am - (1/3) Econ. readers asked if they can get hold of those magical night-vision goggles that tell
us which causal effects in an econ. model are identifiable by OLS (and how). The answer is embarrassingly
simple: Consider Model 2 in p. 163 of (link: https://ucla.in/2mhxKdO) ucla.in/2mhxKdO
8.11.19 @3:36am - (2/3) Take any two variables, say Z_3 and Y. If you can find a set S of observed variables
(e.g., S={W_1,W_2}), non-descendants of Z_3, that block all back doors paths from Z_3 to Y, you are done; the
coefficient a in the regression Y = a*Z_3 + b1*W_1 +b2*W_2 gives you the right
8.11.19 @3:36am - (3/3) answer. A similar goggle works for a single structural parameter (See page 84 of Primer
(link: https://ucla.in/2KYYviP) ucla.in/2KYYviP). Duck soup. This re-begs the question whether the restricted
notion of "reduced form" is still needed in 2019. @EconBookClub @lewbel #Bookofwhy.

8.11.19 @2:25am - (Replying to @mnjp) Glad you took the firing squad seriously. Some people think it is just a
"toy problem" and never learn what counterfactuals mean. #Bookofwhy

8.11.19 @2:20am - (Replying to @Jabaluck @fuzzydunlop123 and 6 others) Adding my interpretation of Haile's
interpretation of "reduced form". I hope I am right, or at least faithful:

8.10.19 @6:09pm - (Replying to @omaclaren @fuzzydunlop123 and 7 others) You are confirming my observation
that people tend to cite the name of the conservation law as an explanation of observed phenomena. Yet naive
me always wanted to know: why would a fluid particle accelerate when getting into the narrow part of the
tube. Now I know why. #Bookofwhy

8.10.19 @5:25pm - (Replying to @omaclaren @fuzzydunlop123 and 7 others) How would Dowe "explain" Venturi's
effect (pressure measured in a fluid flowing in a pipe of varying cross section).?

8.10.19 @5:07pm - (Replying to @fuzzydunlop123 @omaclaren and 7 others) Conservation laws are non-causal
emergent properties of dynamic systems. The latters are causal, because each particle is "responding" to the
forces in its neighborhood. Interestingly, people tend to accept the name of the conservation law as an
explanation of events #Bookofwhy

8.10.19 @4:39pm - (1/ ) (Replying to @MariaGlymour @Jabaluck and 5 others) 1/ The translation of Phil's
definition to DAG language is simple: Y=f(X,Z,U) is "reduced form" iff X, Z and U are ALL the root nodes that are
ancestors of Y, both observed and unobserved. The question remains why would an analyst write down such an
equation when all its claims
8.10.19 @4:44pm - (2/ ) (Replying to @yudapearl @MariaGlymour and 6 others) are obvious from the graph? For
example, the claim that the causal effect of X (and Z) on Y can be estimated by OLS. The answer I believe has
to do with the fact that the notion of "reduced form" emerged in a period when economists where desperate for
conditions to justify
8.10.19 @4:51pm - (3/3) (Replying to @yudapearl @MariaGlymour and 6 others) identification by OLS. Today
they are fortunate to have night-vision goggles that tell them immediately which effects are estimable by OLS,
and which regressors to include/exclude (link: https://ucla.in/2mhxKdO) ucla.in/2mhxKdO This was not always
the case in the history of economics. #Bookofwhy

8.10.19 @6:48am - (Replying to @mathtick) I dont recall commenting on it. Will add to my "to do" pile.

8.10.19 @6:44am - (Replying to @tarinziyaee) No changes except for the many errata shown here: (link: (link:
http://ucla.in/2KYYviP) ucla.in/2KYYviP) (link: http://ucla.in/2KYYviP) ucla.in/2KYYviP and marked in red.
The new edition will incorporate them in a clean text.

8.10.19 @2:13am - (Replying to @RonKenett and @ShalitUri) I do not recommend going in Hastie's direction (no
pun) of starting with black-box and then asking: when would some statistical estimate have a "causal
p ) g g
interpretation". #Bookofwhy goes the other way: start with what you know and ask if you can estimate what you
want to know.

8.10.19 @2:07am - Correcting a link in a previous post. The new errata for the Causal Inference Primer book can
be accessed here: (link: https://ucla.in/2YWojRK) ucla.in/2YWojRK (marked in red). #Bookofwhy

8.9.19 @10:57pm - (Replying to @fuzzydunlop123 @lewbel and 6 others) Here we are, trying to out-guess each
other what Haile meant by "reduced form", and MHE, and so and so... Something is terribly wrong in 2019
science if we can't find a red-blooded economist willing to take a stand and say: TO ME, "reduced form"
means... How about it? #Bookofwhy

8.9.19 @10:37pm - (Replying to @fuzzydunlop123 @IvanWerning and 6 others) I agree here, "model" cannot
replace "structural model," because statisticians think that the assumption of normality is a "model" and some
MHE students think that a "regression equation" is a "model". The word "structural" tells us what assumptions it
carries. #Bookofwhy

8.9.19 @10:18pm - (Replying to @lewbel @steventberry and 5 others) I have no doubt that "reduced form" exists
or can be concocted. But the question is: What useful information does it provide to @lewbel (not MHE). Put
differently, what would science miss if, suddenly, all economists get amnesia and forget such a concept ever
existed? #Bookofwhy

8.9.19 @9:29pm - And while we are speaking of Causal Inference - Primer, (link: https://ucla.in/2KYYviP)
ucla.in/2KYYviP, good news comes from Wiley: They are preparing a revised edition to hit the shelves soon. It
will include new errata collected and composed by Scott Meuller. See (link: http://ucla.in/2KYYviP)
ucla.in/2KYYviP #Bookofwhy

8.9.19 @8:19pm - (Replying to @thanhnguyentang) You can start with #Bookofwhy for fun,history and
philosophy, but if you want to delve straight to the mathematics, and still have fun, there is no better
introduction than Primer (link: https://ucla.in/2KYYviP) ucla.in/2KYYviP.

8.9.19 @8:01pm - (Replying to @IvanWerning @Jabaluck and 5 others) @IvanWening, the more I read your
Tweets the more I agree with you (r u sure u r an economist?). The rearrangements you are permitting are
permissible as long as they preserve information. Algebraic rearrangements are NOT information preserving. Nor
are "Reduced forms" #Bookofwhy

8.9.19 @7:41pm - (Replying to @lewbel @steventberry and 5 others) "Reduced form" vs. "Confused form": Can
anyone define what it means to him/her (not to Phil or MHE) and why it should not be purged from econometric
discourse w/o loss of information? #Bookofwhy

8.9.19 @7:33pm - (Replying to @lewbel @steventberry and 5 others) I must be missing something important. If
the "reduced form" is descriptive, how can it impose untestable assumptions? More basic: Does the "reduce
form" preserve the structural assumptions carried by the structural model of which it is a "reduced form"?
#Bookofwhy

8.9.19 @7:16pm - (Replying to @IvanWerning @Jabaluck and 5 others) Totally agree that "reduced form" is a
source of much confusion. It gives the impression that it is merely a syntactic transformation of "form" with no
loss of information. But if it is "descriptive", then we loose all the causal information that SEM provides. Shun!
#Bookofwhy
8.9.19 @4:55pm - Machine Learning enthusiasts will be interested in George Lawton's new post titled "Causal
Deep Learning Teaches AI to ask why" (link: https://tinyurl.com/y6x7479o) tinyurl.com/y6x7479o I am not
familiar with all the actors mentioned in the story, but I am glad ML is moving beyond curve fitting #Bookofwhy

8.9.19 @6:19am - (Replying to @Jacobb_Douglas @maximananyev and 6 others) The draft lottery, the price of
beans in China, the court decision (see @Bookofwhy Chapter 1), etc. its a variable that only sick/creative
stretch of imagination would deem it relevant to the context of interest.

8.9.19 @4:39am - (Replying to @sarahmrose @SHamiltonian and 2 others) As I said in my confession (link:
https://twitter.com/yudapearl/status/1159066293120598017)
https://twitter.com/yudapearl/status/1159066293120598017: I spend time on such debates knowing that there
are dozens of silent and bright students out there, listening to the conversation, and gathering ammunition for
future defense of commonsense. #Bookofwhy

8.9.19 @1:14am - (Replying to @fuzzydunlop123 @maximananyev and 6 others) One should add though that this
is not done with the intention to "fool." When you don't SEE your assumptions you tend to believe that your
"experiment" takes care of them. The Talmud says: If our eye were given the power to see our assumptions, we
won't get out of bed #Bookofwhy

8.9.19 @1:04am - (Replying to @paulgp @maximananyev and 4 others) But, in addition to saving those answers
from oblivion, the DAG also permits you to COMBINE them, and thus answer questions that DAG-averse folks
CAN'T. E.g.,"Is the treatment ignorable?" or "Are there any testable implications". That's why these folks rarely
ask them. #Bookofwhy

8.9.19 @12:10am - (Replying to @maximananyev @Jabaluck and 5 others) A DAG is none other but a collection
of answers to the question: "What is the source of variation in variable X?" recorded as an arrow into X. What is
often called "The DAG Approach" is consulting those answers, instead of re-asking when they're needed in
estimation. #Bookofwhy

8.8.19 @11:44pm - A question on Quora read: ML is becoming too competitive! Should a person wishing to
become a ML researcher give up? As an aspiring ML researcher, I had to tell them the truth:

8.8.19 @7:54pm - (Replying to @DonskerClass @Jabaluck and 5 others) This is one way of handling potential
misspecifications, (good paper) another is sensitivity analysis (if done correctly) which passes the burden on to
a Plausibility judgment on the likelihood that certain edges can attain a certain strength. #Bookofwhy

8.8.19 @6:54pm - (Replying to @causalinf @steventberry and 3 others) Lewbel's taxonomy notwithstanding,
"Design" cannot replace "identification". "Design" connotes an option an analyst may or may not choose,

"identification" is a property of your model; a causal query is either identifiable or not (given the model) no
matter what you do.#Bookofwhy

8.8.19 @4:41pm - (Replying to @causalinf @steventberry and 2 others) @causalinf, Are you really going to
replace "identification" with "design"???. Clarity with escapism?? "Identification" is perhaps the clearest notion
developed by economists --"Design" is the foggiest . @Undercoverhist, as historian, watch the last days of
clarity.#Bookofwhy

8 8 19 @4:17pm - (Replying to @fuzzydunlop123 @Jabaluck and 5 others) If someone succeeded in eliminating


8.8.19 @4:17pm - (Replying to @fuzzydunlop123 @Jabaluck and 5 others) If someone succeeded in eliminating
the word "design" from the literature (perhaps leave it in RCT contexts) clarity will shine brightly on causality-
land. A catch-all for "I wish to be more specific". I hope it's avoided in #Bookofwhy. Just checked, it's used
mostly harmlessly.

8.8.19 @2:38pm - (Replying to @Jabaluck @Jacobb_Douglas and 2 others) Please do not put words in my mouth.
Conditional ignorabitlity assumptions are inscrutable because they are far removed from what we know, as
proven by (1) PO textbooks and (2) PO folks refusing the test: Given a story (no DAG), tell me if "Y_x is ind. of X
given Z" #Bookofwhy

8.8.19 @2:24pm - (Replying to @fuzzydunlop123 @Jabaluck and 6 others) I can't tell if the Yale folks stand
behind the slides, nor can I tell whether the slides reflect outdated terminology or advocated terminology. Its
healthy to see them going through some sort of soul-searching catharsis. Let's summarize what we learned by
email judea@cs.ucla.edu

8.8.19 @1:55pm - (Replying to @fuzzydunlop123 @Jabaluck and 5 others) @fuzzydunlop123, I cant follow the
barrage of tweets in the wake of Phil Haile slides, but you make consistent sense. Bringing up (link:
https://ftp.cs.ucla.edu/pub/stat_ser/r395.pdf) ftp.cs.ucla.edu/pub/stat_ser/r395.pdf) (link:

http://ftp.cs.ucla.edu/pub/stat_ser/r) ftp.cs.ucla.edu/pub/stat_ser/r was timely -- we haven't seen a drastic


improvement in Econ. texts yet. Is it coming? #Bookofwhy

8.8.19 @11:42am - (Replying to @steventberry @Jabaluck and 4 others) Very interesting and helpful. Thanks for
posting. Is this set of slides representative of the current thinking at Yale-Economics ? Or are they still
undergoing internal debate regarding the precise meanings of terms? #Bookofwhy

8.8.19 @6:05am - (Replying to @georgemsavva @stephensenn and @statsepi) But if Hall is perfectly correlated
with Diet, why won't the effect of Diet on Gain not be the same as the effect of Hall on Gain. Once we decide
what effect we need, the diagram delivers it for you, since we do not have unobserved confounders in the
model. #Bookofwhy

8.8.19 @5:37am - (Replying to @_amirrahnama) I agree with your general statement above, though I have not
read the paper by @tmiller_unimelb . I would be delighted to find out that causality and explanation are well
understood in AI. #Bookofwhy

8.8.19 @3:01am - (Replying to @georgemsavva @stephensenn and @statsepi @georgemsavva, you hit it on the
nail, thanks for making it explicit, and for stressing that #Bookofwhy deals with Lord's paradox, not with
experimental design. Moreover, if we suspect Hall and Diet have separate effects on Weight, another diagram
would resolve it just as well.

8.8.19 @2:39am - For those concerned with issues of Free Press and East-West relations, a Panel on Aug 15 in LA
would be most illuminating. See (link: http://danielpearl.org/freepresstoday) danielpearl.org/freepresstoday
RSVP required.

8.8.19 @2:17am - (Replying to @stephensenn and @statsepi) You have to trust my honesty if I say: I don't have
the slightest idea how this is related to Lord's dilemma with TWO HALLS, each serving ONE diet, and students
are SHOWN what diet is served in each. Please start here, and tell us what factor is neglected in #Bookofwhy

8.8.19 @2:08am - (Replying to @statsepi and @stephensenn) Great news indeed. Humble advice: Try to focus on
why YOU (not Nedler) are surprised by Lord's dilemma, and see for yourself whether #Bookofwhy does not pacify
y ( ) p y , y y p y
your surprise to the point where no such story can ever surprise you again.

8.8.19 @2:01am - (Replying to @stephensenn and @RonKenett) "reasonable under common circumstances " is
insufficient. Resolution comes from understanding why two analyses, which were reasonable, even compelling
under "common circumstances" suddenly cease to be reasonable under the current circumstances. Watch it
explicated in #Bookofwhy

8.8.19 @1:50am - (Replying to @zaffama) You just gave a clear example. In nature, the sun's motion determines
its angle, generating sun rays that are reflected from the atmosphere, sensed by the rooster, and make him
crow. The sun's motion also determine the sun's angle next hour, which we call "Sunrise". #Bookofwhy

8.8.19 @1:35am - (Replying to @zaffama) My point was to show that the direction of causal effect can be
opposite to the times of which events are observed or reported, but (God forbids!) not opposite to time's in the
data-generating process. #Bookofwhy

8.8.19 @1:28am - (Replying to @yudapearl @RonKenett and @stephensenn) To be consistent with my request, I
believe I did an honest job in confessing my "surprises" in my 3rd comment on Senn's post. Here: (link:
https://errorstatistics.com/2019/08/02/s-senn-red-herrings-and-the-art-of-cause-fishing-lords-paradox-
revisited-guest-post/comment-page-1/#comment-184346) errorstatistics.com/2019/08/02/s-senn-red-herrings-
and-the-art-of-cause-fishing-lords-paradox-revisited-guest-post/comment-page-1/#comment-184346 . And note
my conclusion: Yes, causal analysis does dissolve the clash of intuitions in Lord's paradox. #Bookofwhy

8.8.19 @1:11am - (Replying to @RonKenett and @stephensenn) Another comment: I find it hard to discuss this
topic without knowing where my discussant stands on the issue of "surprise", or whether he/she is at all
surprised by Lord's story. So, I beg you, and future discussants to start with an "analysis of surprise" #Bookofwhy

8.8.19 @12:58am - (Replying to @RonKenett and @stephensenn) Never mind engineering challenges and
robotics. The #Bookofwhy chapter on Lord's paradox is an exercise in scientific psychology: Why are scientists
surprised and vexed by the story? Can causal logic dissolve this surprise? The rest is irrelevant to our discussion.
#Bookofwhy

8.8.19 @12:45am - (Replying to @DanielNevo @Jacobb_Douglas and @PHuenermund) The differences between
the (in)dependencies of Y(a) and Y(not-a) do not show in the structure of the SWIG, this information comes
from outside the graphical model. #Bookofwhy

8.8.19 @12:38am - (Replying to @PHuenermund @MariaGlymour and @Jacobb_Douglas) OK, Yielding to Maria
and Paul. But please try to make the context clear to an engineer (like me), who can only think in terms of (1)
what we know, (2) what we want to know and (3)what data we have available. Input-Output. #Bookofwhy

8.7.19 @11:18pm - (Replying to @DanielNevo @Jacobb_Douglas and @PHuenermund) Agree on "alleviated". But
you do not need SWIGs for that task, an ordinary DAG can tell you all about d-separation and back-door tells you

all about ignorability and, if you really need to, it shows it to you explicitly, see Causality p.343, (link:
https://ucla.in/2NnfGPQ#page=11) ucla.in/2NnfGPQ#page=11, #Bookofwhy

8.7.19 @7:06pm - (Replying to @Jacobb_Douglas and @PHuenermund) Best place to ask is by email. A "typical
case" for PO is "we assume, as usual, that D is conditionally ignorable given X". A "typical case" in science is:
"Given a story on 3-4 variables, lets see if measuring X would help us get the effect of D on Y . #Bookofwhy
8.7.19 @4:53pm - (Replying to @roieki and @SamHarrisOrg) Eize Perek?

8.7.19 @2:01pm - (1/2) (Replying to @PHuenermund and @Jacobb_Douglas) That is why the answer to the
question "Can it be done in PO ?" is plain NO. If we say YES, the impression is created that it is just a matter of a
few more computations. It is hard for people to appreciate the difference between tractable and intractable.
And that is why PO
8.7.19 @2:07pm - (2/2) (Replying to @yudapearl @PHuenermund and @Jacobb_Douglas) PO folks dread toy
problems, where the intractability shows immediately. It is like "solving" a polynomial assuming that someone
else already computed its roots. #Bookofwhy

8.7.19 @4:38am - (1/4) I would like to welcome the 500 new followers who have joined us on Tweeter since
@SamHarrisOrg posted our conversation on his podcast. Welcome to the wonderful land of WHY and, please, be
aware of what you are getting yourself into. Our main theme is the #Bookofwhy and
8.7.19 @4:38am - (2/4) the way it attempts to democratize the science of cause and effect and apply it in
artificial intelligence, philosophy, and the health and social sciences, see (link: https://ucla.in/2KYvzau)
ucla.in/2KYvzau. We alert each other to new advances in causal reasoning and new methods of answering
causal
8.7.19 @4:38am - (3/4) questions when all we have are data, assumptions and the logic of causation. We also
debate detractors and nitpickers who mistrust fire descending from Mt. Olympus for use by ordinary mortals. I
spend time on such debates knowing that for every nitpicker there are dozens of
8.7.19 @4:38am - (4/4) silent and bright students out there, listening to the conversation, and gathering
ammunition for future defense of commonsense.
Overall, I hope you find this forum entertaining, challenging, and idea driven.

8.6.19 @10:49pm - Thank you @SamHarrisOrg for having me on your podcast and for our lovely discussion on
cause and effect, counterfactuals, free will, and the future of AI. I believe it was your podcast that caused the
# of followers on this Tweeter to cross the 20K mark. They'r welcome!#Bookofwhy

8.6.19 @5:13am - (Replying to @stephensenn) The answer is simple: The #Bookofwhy produces a consistent
answer to problems defined in the #Bookofwhy, not to problems defined elsewhere, involving several dining
halls, or dining halls shifting their diets, or other variations. One thing at a time. The idea of "surgery
estimators" is ingenious, it would not occur to me that you can get extra mileage on top of "pruned estimators".
However, is Figure 2(b) the simplest example to show this extra mileage?? Would c-equivalence help here? (link:
https://ucla.in/2NbJUVq) ucla.in/2NbJUVq #Bookofwhy

8.5.19 @11:18pm - (Replying to @jaketapper and @RashidaTlaib) @Jaketapper is 100% right, Congresswoman
@RashidaTlaib lies. She knows that Palestinians want one more thing, in addition to posing as human right
advocates -- dismantling their neighbors. If I am wrong, let us say so publically. She can't!

8.5.19 @7:14am - (Replying to @doinkboy @GRich_Cinci and 2 others) This is what causal inference is all about:
"interpret it causal, given a set of assumptions (i.e., a causal model)." Except that the "interpretation" is no
longer whimsical, as it used to be, it must obey the logic of causation. And this is what makes it "causal"
#Bookofwhy Their numbers is still rising, I hope one of you continues --an amazing phenomenon.

8.5.19 @4:16am - I have just posted my 3rd comment on Lord Paradox (link:
https://errorstatistics.com/2019/08/02/s-senn-red-herrings-and-the-art-of-cause-fishing-lords-paradox-
revisited-guest-post/comment-page-1/#comment-184346) errorstatistics.com/2019/08/02/s-s.... Here, I
returned to my original goal of empowering readers with an understanding of the origin of the paradox and
what we can learn from it. Red herrings have been taking too much of our time. #Bookofwhy
8.5.19 @4:16am - It is important to add that the huge literature on Simpson's and Lord's paradoxes attests to a
century of scientific frustration with a simple causal problem, deeply entrenched in out intuition, yet helplessly
y p p , py ,y p y
begging for a formal language to get resolved. I can count dozens of
8.5.19 @4:16am - red herrings thrown at this stubborn problem, to deflect attention from its obvious resolution.
Why? Because the latter requires the acceptance of a new language - the hardest transition for adults. I am
grateful for the opportunity to talk to thousands of young followers here
8.5.19 @4:16am - on Twitter and convey to them my honest conviction: Chapter 6 of #Bookofwhy (paradoxes
galore) is a condense summary of a century of confusion, and a powerful recipe for deliverance from that
confusion. Additional references are: (link: https://ucla.in/2JeJs1Q) ucla.in/2JeJs1Q (Lord's paradox),
8.5.19 @4:16am - ucla.in/2Jfl2VS (Simpson's paradox) (link: https://ucla.in/2NTbnrS) ucla.in/2NTbnrS (Sure-
thing Principle) And Chapter 6 of Causality (link: https://ucla.in/2BDDTLK) ucla.in/2BDDTLK , where I
document dozens of red-herrings 1900-2000.

8.5.19 @1:16am - Replying to @mendel_random and @oacarah) The point is: "discovered using a causal model,
later depicted as diagrams". Red herrings do not stop us from discussing an issue. OK. We get your point: DAGs
dont handle cycles and calculus does't handle non-differentiable functions. Can we get on with the discussion?
#Bookofwhy

8.4.19 @7:22pm - (Replying to @mendel_random) I dont think it was worked out in folks' heads too well, they
had to use some mathematics, and the mathematics they used came as "equations" which, at the time looked
like algebraic but, in time, came to be recognized as non-algebraic, asymmetric, or "structural", #Bookofwhy

8.4.19 @3:13pm - (Replying to @stephensenn) Oh, @stephensenn, you never told us what's surprising you in
Lord's story. In other words, must we go to finite sample before you can describe to us the reason for your
surprise? #Bookofwhy

8.4.19 @3:09pm - (Replying to @stephensenn) Not wrong. Just irrelevant to Lord's paradox where we assume
(using Wainer's model) that students are allowed to choose their own Hall and that each Hall serves ONE diet. If
I am surprised with this simple story, I'll try to resolve it HERE, before complicating it. #Bookofwhy

8.4.19 @2:25pm - (Replying to @stephensenn) We differ here. The notion of "probability distribution function" is
not meaningless is statistics. Most stat texts start with distributions, then go to "samples from distributions".
Lord introduced his paradox at that level, and managed to surprise us; why quit? #Bookofwhy

8.4.19 @2:09pm - Now, speaking specifically about Lord's paradox, the paradox was introduced to us in
"asymptotic" terms (ie, using distributions, not samples) and we were surprised. Is it likely that we can resolve
our surprise by going to finite samples? or to "block design"? #Bookofwhy

8.4.19 @1:54pm - That is why I am begging folks: "Please, do not tell me 'I am not entirely satisfied' before you
tell me why you are surprised (by the paradox) ". I am proud that #Bookofwhy addresses this question (of
"surprise") head on, before offering "a resolution".

8.4.19 @1:36pm - (Replying to @mendel_random) No big deal. Replace "using Wright's DAGs" with "using Wright's
equations, later depicted as DAGs" and the rest of the Tweet follows, especially "using what you know". The
Tweeter discussion was about "using a model", which some folks shun. #Bookofwhy

8.4.19 @1:09pm - Any discussion of "paradoxes" is really an exercise in psychology. Yet we, quantitative
analysts, are trying to avoid psychology at all cost. We can't. We must explicate why two strong intuitions seem
to clash, and the conditions under which our intuitions fail. See #Bookofwhy

8.4.19 @6:06am - (Replying to @f2harrell @stephensenn and @Lester_Domes) I will expand, in a day or two. But
( py g p ) p y
it would help if you reconstruct Lord's paradox in your own way and pin point: What was paradoxical in the
story? What was surprising there that deserved the word "paradox"? #Bookofwhy

8.3.19 @7:06pm - (Replying to @EpiEllie) I tried to look into it, but I am missing your research question, ie, the
query.

8.3.19 @6:43pm - Good news for missing-data analysts. Karthika Mohan @karthica is joining the Editorial Board
of Journal of Causal Inference jttps://www.degruyter.com./view/j/jci It's a welcoming invitation for articles
on modern ways of recovering what we thought to be missing. #Bookofwhy

8.3.19 @4:42pm - I've just posted a comment (link: https://errorstatistics.com/) errorstatistics.com on S. Senn's
n-th attempt to deconstruct Lord's paradox. It ends: "I hope we can now enjoy the power of causal analysis to
resolve a paradox that generations of statisticians have found intriguing, if not vexing." #Bookofwhy

8.2.19 @9:25pm - Coming from my fellow statisticians it reminds me of King Solomon's saying: "Let a stranger
appraise your work, not your mouth"(Proverbs 27:2). And the Mishna saying: "The baker does not judge his own
bread" (Tosefta Yom Tov 3:7). I hope my bread tastes well to others. #Bookofwhy

8.2.19 @9:25pm - Many thanks! Carlos Cinelli came back from JSM-2019 and brought me a gift that I would
cherish dearly:

8.2.19 @8:03pm - (Replying to @jasonintrator @glogauer_jakob and @TheIHRA) A group HAS a self. What makes
an individual have self is memory, telling him: your experience yesterday informs your decision today. Same
with groups, except here the pertinent experience spans centuries, and will easily get lost unless members
attain self-determination

8.2.19 @7:26pm - (Replying to @jos_b_mahoney @jasonintrator and @glogauer_jakob) Unlike Jason, I would feel
very very lonely, to know that I do not have Canaanite role models who spoke my language, who left my kids
beautiful legends and who encoded their experience in a culture baked over 3,000 years. Woefully lonesome!

8.2.19 @5:15pm - (Replying to @jos_b_mahoney @jasonintrator and @glogauer_jakob) Charging people with
anti-semitism is stone-age. Edward Said is guilty of a worse offense: Zionophobia, denying Jews the right to
define their collective identity after writing volumes on "Orientalism"-the right of Arabs to define themselves.
I'm sensitive to logical consistency.

8.2.19 @5:05pm - (Replying to @jasonintrator and @glogauer_jakob) The HE in that prayer (Birkat Hamazon) is
God himself, called Ha'Rachaman (the merciful one). True, it was not about "now", but it reflects the
continuous aspiration of the Jewish people to eventually regain sovereignty in their historical birth place, i.e.,
Zionism

8.2.19 @4:54pm - (Replying to @jasonintrator and @glogauer_jakob) Universal liberalism is central indeed to
Jewish life. But how: "Thou shall not oppress the stranger, because YOU know what it means, YOU were once a
stranger..." Namely, if you forget YOUR collective memory, gone is your universal liberalism. Can't have one
without the other.

8.2.19 @4:44pm - (Replying to @jasonintrator and @glogauer_jakob) Their rejectionism started in 1920's, there
was no loss of homes or political rights. Not even FEAR of losing homes or rights, as is documented in the Arab
newspaper Carmel See "Early Zionists and Arabs"
8.2.19 @4:36pm - (Replying to @jasonintrator and @glogauer_jakob) Herzl, Jews in the diaspora have been
divided about the role of Israel as a practical solution to the immediate problems Jews faced then. That role
was never debated in my (and your?) grandfather house, where they prayed 3 times a day: "He will walk us in
sovereignty to our land"

8.2.19 @4:27pm - (Replying to @jasonintrator and @glogauer_jakob) Your friends from JVP are dangerous
indeed, not because they wish us death, but because they recklessly blind themselves to our death, ostensibly
in pursuit "universal liberalism." I call them "Jews of Discomfort" here

8.2.19 @4:00pm - (Replying to @jasonintrator and @glogauer_jakob) Yes, I work on this. What is it that I
highlighted "over" other facts? Isn't Palestinian rejectionism a fact that cannot be "over highlighted" when it
comes to prospects for peace, lifting the occupation, and almost every issue youngsters care about. Do you talk
to them about it?

8.2.19 @3:49pm - (Replying to @jasonintrator and @glogauer_jakob) I would never accuse JVP of being
motivated by antisemitism; they are motivated by more dangerous forces which I have outlined here:"Our New
Maranos"

8.2.19 @2:54pm - (Replying to @jasonintrator and @glogauer_jakob) What is it about the way I represent
Palestinians that you think is (1) not factual and (2) not effective with younger American Jews? I hope you are
not suggesting young people are not moved by facts? These youngsters were: (link:
https://jewishjournal.com/culture/lifestyle/first_person/300661/inspiration-and-a-rallying-cry-for-graduates/)
https://t.co/Ko3RFa6AUw?amp=1 What about those you meet?

8.2.19 @1:45pm - (Replying to @glogauer_jakob) Truth has its secret way of prevailing, despite BDS's key
slogan: "if you repeat a lie long enough, people will fall for it."

8.2.19 @1:32pm - (Replying to @AdanZBecerra1) Agree. And my reason: If we introduce regression before DAGs,
students are likely to get trapped in the "regressional confusion" of the 20th century, unable to distinguish
structural from regression equations. (link: https://ucla.in/2NnfGPQ#page=38) ucla.in/2NnfGPQ#page=38
#Bookofwhy

8.2.19 @5:29am - (Replying to @GivingTools and @causalinf) In my efforts to make causal diagrams palatable to
economists I am trying build as much as possible on identification strategies devised by empirical economists.
Can you point us to an author or two who came close to outlining diagramatic ideas informally? Thanks.
#Bookofwhy

8.1.19 @9:13pm - (1/ ) I am recommending this paper to every data analyst educated by traditional textbooks,
which start with regression equations, add and delete regressors, estimate and compare coefficients before and
after deletion, and then ask which coefficient has "causal interpretation"
8.1.19 @9:13pm - (2/ ) I was shocked to realize that the majority of data analysts today are products of this
culture, trapped in endless confusion, with little chance to snap out of it, since journal editors, reviewers and
hiring committees are trapped in the same culture. The new PO framework does
8.1.19 @9:13pm - (3/ ) offer a theoretical escape route from this culture, through the assumption of
"conditional ignorability" but, since it is congnitively formidable, practicing analysts must rely on regression
arguments. Keele et al examine a causal model (Fig.2) and ask: suppose we regress Y
8.1.19 @9:13pm - (4/ ) on all observed variables; which of the coefficients has any causal interpretation. I have
alerted economists to such questions here (link: https://ucla.in/2mhxKdO) ucla.in/2mhxKdO (3.2.7) so, we can
assume they have mastered the techniques by now. Students of #Bookofwhy seeking a gentle way to
8.1.19 @9:13pm - (5/5) approach their mentors or professors or their peers, this is a great channel to motivate
them.

8.1.19 @8:35pm - (Replying to @VenkatNagaswamy) I was over-intoxicated by scientific poetry

8.1.19 @8:34pm - (Replying to @yudapearl @Jabaluck and @eliasbareinboim) But, lets deal with what they CAN
DO. Let's ask their students to examine Fig. 2 of Keel etal (link: https://doi.org/10.1017/psrm.2019.31)
doi.org/10.1017/psrm.2) (link: http://doi.org/10.1017/psrm.2) doi.org/10.1017/psrm.2 and decide which
regression coefficient coincides with a structural parameter. I challenged them here (link:
https://ucla.in/2mhxKdO) ucla.in/2mhxKdO #Bookofwhy

8.1.19 @8:12pm - (Replying to @Jabaluck and @eliasbareinboim) I happened to go over MHE last night. Read:
"we follow convention and refer to the difference between the included coefficients in a long regression and a
short regression as being determined by the OVB formula (p.44). You say: "They think you should write down a
model and then
8.1.19 @8:23pm - (Replying to @yudapearl @Jabaluck and @eliasbareinboim) figure out an estimation strategy
..." I have not seen "a model" written down of what one believe about the world. Even Eq. (3.2.8), which is
supposed to be a structural equation, is written after CIA is assumed (conditional ignorability) which we know is
cognitively impossible.

8.1.19 @7:58pm - Too bad they cut off my song, just before the tenor!! Not too late. If any reader of
#Bookofwhy has ANY unresolved question on DAGs, please check if that question is not answered here (link:
https://ucla.in/2Pnj22p) ucla.in/2Pnj22p. If still unresolved, I am here on Tweeter, with an army of resolvers.

8.1.19 @7:48pm - Extremely inviting. "With a book of verse, and thou, beside me, singing in the wilderness, and
wilderness is paradise anow." #Bookofwhy

7.31.19 @4:53pm - (1/2) As I was re-reading "Wagged by the DAG" (link: http://ucla.in/2Pnj22p)
ucla.in/2Pnj22p I've found a few collectibles that are relevant to our other discussions: (1) "It would be a
disaster if models were allowed to produce information unintended by the modeler." (2) Tools ....
7.31.19 @4:53pm - (2/2) (2) "Tools that are indispensable in solving simple problems are unlikely to become
dispensable when problems become more complex." (3) "[In some unnamed cultures,] selecting covariates for
confounding control is still a black magic," #Bookofwhy

7.31.19 @2:25pm - (Replying to @EpiEllie and @BUSPHEpi) No one is really anti-DAG. But many propose
extensions, enrichment and complementing methodologies. I was wondering if Dr. Krieger still believes the
extensions she proposed in "Wagged by the DAG" are worth pursuing. (link: https://ucla.in/2Pnj22p)
ucla.in/2Pnj22p #Bookofwhy

7.31.19 @1:54pm - (Replying to @EpiEllie and @BUSPHEpi) Curious. Does Nancy Kreiger still upholds the views
she expressed in: "The Tale Waggs the DAG?" (link: https://ucla.in/2Pnj22p) ucla.in/2Pnj22p #Bookofwhy

7.31.19 @4:00am - Why I call it "upside-down culture"? Because the logical way to start is with what you KNOW,
eg. structural equations, then use regression to estimate what you WISH TO KNOW. Confusion erupts when
people think regression equations represent what you know. #Bookofwhy

7.31.19 @3:45am - (Replying to @TariqTaha123 @CNNSotu and @jaketapper) Straight from the book of slogans
of BDS - the racist movement that shows no respect for truth or other people's identity. See (link:
https://ucla.in/2SxjgWH) ucla.in/2SxjgWH
https://ucla.in/2SxjgWH) ucla.in/2SxjgWH

7.31.19 @3:23am - (Replying to @willjharrison and @svenohl) Causation precedes manipulation. In #Bookofwhy
we describe manipulation as a way of interrogating nature to reveal the causal forces that tie variables

together.

7.31.19 @1:43am - (Replying to @isli_amar) Glad #Bookofwhy reached Algeria, probably from UK, swimming. I
hope you start teaching it before young minds get frozen into "regression". The temporal constraints paper
almost escaped my memory - so much has happened. Enjoy!

7.31.19 @1:30am - Link broken? Please try this one: (link: https://www.cambridge.org/core/journals/political-
science-research-and-methods/article/causal-interpretation-of-estimated-associations-in-regression-
models/4488EC8925CF8F623CDE655E01268F6F) https://t.co/b41GKvI4rL?amp=1

7.31.19 @1:11am - In contrast, here is a new paper from political scienists (link:
https://zimbra.cs.ucla.edu/service/home/~/?auth=co&loc=en_US&id=728638&part=2)
https://t.co/SFJV1O4RiO?amp=1 who address the upside-down culture of starting with "regression" and
"controlling for" (eg Angrist's MHE) and then asking: Do our findings have any "causal interpretation?"
#Bookofwhy

7.31.19 @12:53am - (Replying to @JamesLNuzzo and @NicoleBarbaro) Cofield's findings can be partially
excused, for this was published 2010, only 10 years after epidemiologists started using graphs. But how come
students in quantitative behavioral science do not rebel? #Bookofwhy

7.31.19 @12:17am - Very interesting paper. Published in 2019; 25 years after causal language has been
mathematized and separated from statistical language! The author seems unaware of the causal revolution. No
wonder! Has Psychometrika ever published an article on modern causal modeling?#Bookofwhy

7.30.19 @11:27pm - (Replying to @TariqTaha123 @CNNSotu and @jaketapper) My father was one of those Jews
who came peacefully to his historical homeland. He was not silenced at all, I was there. He actually offered our
Arab neighbors peaceful co-existence. Do you know what answer he got? I hate to embarrass you in public but,
beware, I was there!

7.30.19 @9:44pm - (Replying to @JennieBrand1) Got it, thanks. Was I right in summarizing FE as: "An assumption
of equality of two or more structural parameters which, in certain DAGs, leads to identification that otherwise
will not be achieved." Q. The X's in your figs have no parents, what does it mean? #Bookofwhy

7.30.19 @9:04pm - (Replying to @JennieBrand1) Thanks Jennie, can you send an active link? Evidently, Oxford
took UCLA off their university list, and they won't give me access to your paper.

7.30.19 @9:02pm - (Replying to @TariqTaha123 @CNNSotu and @jaketapper) I an going to show your Tweet
when people ask me: What kind of neighbors Israel has? what kind of mentality drives them? and why it is so

hard to reason with them? Dont they see that the two nations are equally indigenous to the land?My answer: see
above "Me, Me, Me!"

7.30.19 @1:40am - (Replying to @maximananyev @Jabaluck and 3 others) Apropos:The more explicit the
assumption, the more criticism it invites, for it triggers a richer space of alternative scenarios in which the
assumption may fail. Researchers prefer therefore to declare threats in public and make assumptions in private.
assumption may fail. Researchers prefer therefore to declare threats in public and make assumptions in private.
(link: https://ucla.in/2N7S0K9) ucla.in/2N7S0K9

7.29.19 @10:41pm - (Replying to @Psylocke42356 @ZachWritesStuff and 2 others) This is precisely what I tell
them in almost every piece I write. Can you reciprocate? Can you repeat after me (and ask Rashida to join):
"Two states for two peoples, equally legitimate and equally indigenous". No ifs, no buts, just say "equally
indigenous". Can you?

7.29.19 @10:10pm - (Replying to @Psylocke42356 @ZachWritesStuff and 2 others) Nations have a right to
freedom and existence to the extent that they confirm such rights to their neighbors.

7.29.19 @9:58pm - One of the best quotes of the century. I hope it changes at least one heart.

7.29.19 @9:46pm - (Replying to @Jabaluck @fuzzydunlop123 and 2 others) That is why it is so important to
write "The history of bad-control in pre-optics econometrics", to see precisely if the fumbling came from
doubting the validity of models, or from inability to handle even a simple and valid model. Anyone writing? I'll
help (anonymous) #Bookofwhy

7.29.19 @7:17pm - (Replying to @Jabaluck @fuzzydunlop123 and 2 others) And you insist you could have run a
similar conversation with Angrist using the language of potential outcomes, where even "bad controls" are
subject to embarrassment. BTW, is anyone writing the history of "bad controls" in pre-Telescopic econometric
literature? #Bookofwhy

7.29.19 @3:13pm - (Replying to @fuzzydunlop123 @Jabaluck and 2 others) What am I watching? I thought DAGs
are good for pedagogical purposes only. Now I see discussions on whether to condition on a variable or
not...Dont tell me DAGs are good for discussions too, or for evaluating ID strategies, etc. etc. Beware of harsh
consequences. #Bookofwhy

7.29.19 @2:31pm - Good news for causality research! Congratulations to Elias and co-authors for best paper
award at the UAI-2019 conference, straight from Tel-Aviv marina, where I got my wind-surfing diploma in 1980
(framed in my office). They told me creative surfing is the secret to success.

7.29.19 @7:40am - (Replying to @ZachWritesStuff @CNNSotu and @jaketapper) What you teach your child is
what you intend to do. Not one teacher, not two, but EVERY teacher. This is what you, as a journalist should
decry and labor to change.

7.29.19 @7:31am - (Replying to @ZachWritesStuff @CNNSotu and @jaketapper) No one says Israel should not
exist? You hav'nt heard what Palestinian teachers say repeatedly, nor what Omar Barghouti said at UCLA in
2014. See (link: http://bayes.cs.ucla.edu/DPF/anti-zionism-on-campus-jp-ch16.pdf) https://t.co/RRFwcCRKAi?
amp=1 .

7.29.19 @7:15am - (Replying to @glarange72 @CNNSotu and @jaketapper) The asymmetries on the ground are
grim consequences of asymmetries in intention, with one side dreaming "We, We, We" and the other
threatening "Me, Me, Me". Quite stark.

7.29.19 @6:19am - (Replying to @y2silence @stephensenn and 14 others) What is the simplest, canonical
example of MLM showing that if you ignore the hierarchy you will not answer your research question properly. I
am asking because the MLM papers I've chanced to read start and end with no explicit research questions.
#Bookofwhy
7.29.19 @5:59am - (Replying to @ZachWritesStuff @CNNSotu and @jaketapper) States, indeed, just "exist", they
have no rights. But when you tell your child: "Our neighbor has no right to exist" you are telling him: "We are
not going to honor any peace agreement, forever". And you telling Israelis, you have compelling reasons to
control those territories.

7.29.19 @4:50am - (Replying to @Vic1Nobody @ZachWritesStuff and 2 others) I am not a right-winger and I do
not think it is fair to mention right-wing habits to shut down debates on core issues: Can Palestinians claim
rights which they deny their neighbors. Current affairs are surface manifestations of this core issue.

7.29.19 @4:35am - (Replying to @shanbhardwaj @CNNSotu and @jaketapper) How? By tuning in, day and night,
to what Palestinian leaders, clergy, and educators are saying (to their constituents, in Arabic) as I do, and my
friends in the Israeli peace camp do, hoping to detect a seed of acceptance. Thus far - Nada.

7.29.19 @4:25am - (Replying to @glarange72 @CNNSotu and @jaketapper) Hypothetical questions have practical
implication. In our case the implication is that one cannot demand rights to Palestinians that they deny their
neighbors.

7.29.19 @2:51am - (Replying to @glarange72 @CNNSotu and @jaketapper) My position is that Jake Tapper
should have asked her: "What should Israelis do if Palestinian leaders tell them what they do (I hope you heard
them the past 75 years)???"

7.29.19 @2:40am - (Replying to @yudapearl @CNNSotu and @jaketapper) Poor Rashida, she just lost 90% of her
voting base by saying: "Of course, but..." when Jake Tapper's asked her: "Does Israel have the right to exist?" It's
tough to be a Zionophobe on CNN, torn between viewers norms of justice and voters push for elimination.

7.29.19 @1:40am - (Replying to @CNNSotu and @jaketapper) The punchline was missing: "What if the
overwhelming majority of the Palestinians, leaders, educators and clergy, deny the right of Israel to exist, and
openly declare that, occupation or not, they will to continue their arm struggle against Israel, in any borders?"

7.29.19 @1:32am - (Replying to @ZachWritesStuff @CNNSotu and @jaketapper) Zachary, Serious Person,
associates a people's right to self-determination with "right-wing talking points". A serious person indeed.

7.29.19 @12:19am - (Replying to @AdanZBecerra1 @Lester_Domes and 14 others) I fail to see the relevance of
MLM in this paper. I see constancy of effects across time, fine, but I do not see clusters or Set-Subset
relationships in the examples. My blindness?

7.29.19 @12:03am - (Replying to @stephensenn @Lester_Domes and 13 others) I see no "strong undeclared
assumption" in #Bookofwhy, perhaps because I read assumptions from DAGs and, looking at Fig.6.9, I see no
assumption left out. Nada! The point is: {Data + DAG} determine which analyst was correct. No need to recruit
Nelder to make this simple point.

7.28.19 @9:57pm - (Replying to @Lester_Domes @melb4886 and 13 others) Glad we are in agreement. I could
never understand why certain folks, who are reluctant to learn causal inference, always excuse themselves with
"we can do it in MLM". What is it in MLM that gives people the illusion that it can answer causal questions?
#Bookofwhy

7.28.19 @9:09pm - (1/ ) Great paper. First time I understand what "fixed effect" is. I used to confuse it with
h it b t Fi 1 t ll it i ti f lit f t t t l t hi h i t i
homogeneity, but Fig. 1 tells us it is an assumption of equality of two structural parameters which, in certain
DAGs, lead to identification that otherwise will not be achieved. So simple,
7.28.19 @9:09pm - (2/ ) "Why didn't they tell us?" I should complain, but I won't, because it was probably all
there, in the papers and the lengthy motivations, and the indexed regression equations that I was too lazy to
digest. Glad it is over, one figure did it. #Bookofwhy

7.28.19 @4:21pm - You are right. In case other readers of #Bookofwhy got stuck on Eq. 7.2 page 227, replacing
Z with U would make it less mysterious in the context of the Z=Tar story. I hope we can make the change before
the paperback edition hits the shelves. Thanks @the_aiju

7.28.19 @4:00pm - (1/2) I love your Advisor, and I think his "hide it" was an honest expression of the prevailing
culture. I am sure that if he and colleagues start using DAGs in hiding, under the cover of "only for education,"
they will eventually use them everywhere, from seeing assumptions, to
7.28.19 @4:00pm - (2/2) to testing assumptions, to validating ID strategies, to discovering new ID strategies.
Recall, IV was discovered using Wright's DAGs. Why? Because "using DAGs" simply means "using what you know". I
presume the Church lifted its ban on telescopes "only for education"#Boodofwhy

7.28.19 @2:06pm - (Replying to @nyarlathotepesq) d-separation is valid for linear CYCLIC models as well. So it
is easy to identify bad controls in simultaneous eq-ns. #Bookofwhy

7.28.19 @5:41am - FYI, I could not resist answering a new Quora question: "How does the Rubin causal model
differ from graph-theoretic approaches like Pearl's do-calculus?" (link: https://www.quora.com/How-does-the-
Rubin-causal-model-differ-from-graph-theoretic-approaches-like-Pearl-s-do-calculus/answer/Judea-Pearl-1)
https://t.co/Of6jbZhyql?amp=1 There is nothing new there that our readers do not know; it is just compiled
succintly #Bookofwhy

7.28.19 @5:34am - (Replying to @thosjleeper @PHuenermund and 4 others) No surprise that this is how IV is
(still) taught in econ. They are talking about "structural equations" and call it "regression" and the variables
"regressors". It is OK, as long as: (1) Econ. students can endure the confusion, and (2) a Glossary is in the
making.#Bookofwhy

7.28.19 @3:51am - (1/ ) (Replying to @PHuenermund @dlmillimet and 3 others) In looking over other tweets in
this thread I got the feeling that many of the questions are answered in the #causalinference literature, but the
jargons are almost incompatible. I hope you plan on writing a glossary one day. For example, 2sls is an
estimation method, hence it
7.28.19 @4:00am - (2/ ) (Replying to @yudapearl @PHuenermund and 4 others) has nothing to do with
interpretation or with coefficients. Also, in the IV model the IV variable is the only "exogenous" variable.
Finally, in a regression model "exogeneity" is not defined. The glossary seems like an endless job, but where
would we be w/o it?#Bookofwhy

7.28.19 @2:20am - (Replying to @PHuenermund @dlmillimet and 3 others) @PHenermund I ventured to read
your post and met your question: "How can we be sure that what we're estimating for the compliers is
representative for the whole population?" Why don't Balke's bounds provide a general answer to your question?
eg: (link: https://ucla.in/2pQ0Gvr) ucla.in/2pQ0Gvr #Bookofwhy

7.28.19 @12:08am - (Replying to @stephensenn @AdanZBecerra1 and 12 others) Is there room for single-level
reasoning anywhere in science? If not, I'll scrap 99% of my science books. If yes, I would first cast Lord's and
Simpson's dilemmas in single-level context, see if I can solve them, then proceed to multilevel if needed, but
only if needed.#Bookofwhy
7.27.19 @8:30pm - (Replying to @AdanZBecerra1 @stephensenn and 12 others) I am not questioning the need to
do multilevel modelling when needed, be it in causal or predictive tasks. I am questioning the wisdom of forcing
multilevel modeling on single-level causal questions (eg Lord's and Simpson's examples) w/o the tools of causal
modeling. #Bookofwhy

7.27.19 @5:08pm - (Replying to @FJnyc @intelligence2 and 11 others) Mehdi Hasan will continue to stage these
Kangaroo debates until someone charges him with "Zionophobic bigotry" in front of an audience, which would
corner him to compare his perpetual denial of Jewish identity as people with his perpetual whining of
"Islamophobic bigotry."

7.27.19 @3:49pm - (Replying to @AdanZBecerra1 @stephensenn and 12 others) @AdanZBecerra1 I do not think
one should delve into multilevel models before acquiring the tools of handling single-level problems, like the
one posed by Lord, or Simpson, or #Bookofwhy. Multilevel modeling, if imposed on every single-level problem
can become counterproductive.

7.27.19 @3:11pm - (Replying to @matt_vowels) We are now hearing a 9th myth "A causal model is a special case
of a predictive model". This one is particularly misleading, for it tells stat. students: "Dont bother to change
your thinking, what we have been doing this past century is sufficient". No it ain't. #Bookofwhy

7.27.19 @3:02pm - (Replying to @marcelogelati) I can't decipher what they are trying to say, but if the're saying
you can't infer causality from observationial studies alone, w/o a model, they are right. If they deny the power
and ubiquity of qualitative models -- they are wrong. #Bookofwhy

7.27.19 @2:23pm - (Replying to @intelligence2 and @EWilf) Thanks for posting. I would never debate under such
title, which pre-fixes its verdict. Show me one titled: "Is Anti-Zionism RACISM on its own merit" and I will show
you a YES verdict earned. That is why I always use "Zionophobia", not anti-Semitism, e.g.

7.27.19 @4:37am - (Replying to @laurencepearl) It used to be Perl (on my grandfather visa, as he arrived at the
Holy Land in 1924, from Poland). Legend says Jews bought their surnames centuries ago, from their European
masters, based on their professions and the money they could raise. Seems Jewelry was a good profession.

7.26.19 @7:10pm - (Replying to @_asubbaswamy) Now I see it. But why did it take me 24 hours? Because I was
missing one word: "E[T|do(a), c] is stable and IDENTIFIABLE". Naive me thought you are recommending an
experiment with do(a). I suggest you stress this point explicitly in future papers.#Bookofwhy

7.26.19 @3:49pm - (Replying to @laurencepearl) Don't understand!! ??? You must be kidding! How do you
estimate causal effects? Or, how do you find stable predictors? More importantly, how do your colleagues do

that? You'r kidding!!! #Bookofwhy

7.26.19 @3:41pm - (Replying to @_asubbaswamy) Thanks for the refinement! Agree. S goes into X. A question to
you: Why didn't you use a simpler example than Fig. 2(b) to show advantage over "pruned estimators" ?? It took
me an hour to believe that such examples exist. #Bookofwhy

7.26.19 @3:06pm - Just occurred to me: Is't the front-door a simple example of getting an unbiased and stable
estimator of Y, given observations on X and Z (with U varying)? [ X--->Z--->Y with confounder U, X<-U->Y. ]
E[Y|x,z] is unstable, whereas the front-door formula is. #Bookofwhy
7.26.19 @2:43pm - (1/ ) Saying that "a causal model is a special case of a predictive model" is like saying "sailing
is a special case of swimming, since it is conditional of something floating". In general, saying that task-A is a
special case of task-B depends on what the sayer is trying to get:
7.26.19 @2:43pm - (2/2) an excuse for not doing A, or a license for doing A using methods used in B. The latter
would be justified, if it was possible. Unfortunately, causal models require information (+methods) not
available in traditional prediction modeling. So why say "special case"? #Bookofwhy

7.26.19 @2:19pm - You are right, the word "Zionophobia" is gaining traction, including the Justice Department
Summit, see (link: https://www.jns.org/opinion/the-truth-about-anti-israel-incitement/)
https://t.co/05mmrOzHIc?amp=1. It is becoming "the ugliest word in town", at least among people of
conscience. It's the most effective defense weapon I know.

7.26.19 @6:00am - (Replying to @suchisaria @KordingLab and 4 others) The idea of "surgery estimators" is
ingenious, it would not occur to me that you can get extra mileage on top of "pruned estimators". However, is
Figure 2(b) the simplest example to show this extra mileage?? Would c-equivalence help here? (link:
https://ucla.in/2NbJUVq) ucla.in/2NbJUVq #Bookofwhy

7.26.19 @5:44am - (Replying to @nyarlathotepesq) You are right, the bet is not a proof, that is why I suggested
an Appendix with some proofs. But the two polar poles is not a good analogy; going from linear to
nonparametric amounts to monotonically removing constraints. Are you considering writing it? #Bookofwhy

7.26.19 @3:10am - (1/2) Great question! Three answers. 1) In practice, we are really seeking "safe control", not
"bad control" and nonparametric analysis gives it to us. 2) If something is bad in both nonparametric and linear
analyses, you can bet it is bad in between. 3)
7.26.19 @3:10am - (2/2) (3) All the fumbling and stumbling I have seen in the econometric literature occur
already in linear systems, where the proof of "badness" is easy, see (link: https://ucla.in/2LcpmHz)
ucla.in/2LcpmHz. Plus, you can discuss the question in the Appendix. #Bookofwhy

7.26.19 @2:07am - (1/2) BAD-CONTROL. If I were a young economist, seeking visibility and impact in my field, I
would sit down and write an article titled: "Bad-Control - A lingering challenge and its resolution." I got this
thought upon reading (link: https://www.mostlyharmlesseconometrics.com/2009/05/comments-on-bad-
control/) noting that Angrist (2017)
7.26.19 @2:07am - (2/2) still calls this elementary econ. exercise: a "difficult problem". Such article will be a
highly appreciated eye-opener to many of your peers, but it must be written diplomatically -- lot's of
professional honor involved. #Bookofwhy

7.26.19 @12:34am - (Replying to @neurosutras and @JonAMichaels) Causal thinking does not mean purging
predictive and associative relations; it means using such relationships properly whenever they emerge from a
causal model of the world. #Bookofwhy

7.25.19 @6:38am - (Replying to @KordingLab @fhuszar and 4 others) Everything Econs did is correct, but what
they did is of limited scope. I once asked: how many economists can do X, or Y etc. and I lost all my econ
friends. See Causality p.216 fn10 - I will not repeat this mistake - do not force me, please. #Bookofwhy

7.25.19 @6:23am - (Replying to @brunofmr and @suchisaria) "Stability" under environmental change should not
be confused with "stable distribution" which is a purely probabilistic notion, and has nothing to do with
environmental changes. I hope no confusion results. #Bookofwhy

7.25.19 @5:03am - (Replying to @suchisaria @KordingLab and 4 others) Will do.


( py g g )

7.25.19 @5:02am - Not many people realize that the strength of a DAG comes from building ALL its logic on ONE
primitive question: "Why does a variable vary?" All the rest is mechanically derived, demanding no further
judgment. Poetically, I crowned it: "going where knowledge resides." #Bookofwhy

7.25.19 @4:44am - (Replying to @KordingLab @fhuszar and 4 others) I dont see it this way, since I hardly
touched on "estimation" (ie, going from finite sample to distribution); I call the stat dept once I get an
estimand. Econometrics has been doing some limited identification correctly. How limited? See: (link:
https://ucla.in/2mhxKdO) ucla.in/2mhxKdO, #Bookofwhay

7.25.19 @1:26am - (1/ ) I just read (link: https://arxiv.org/pdf/1905.11374.pdf) and I agree with @suchisaria .
Anyone concerned with stability (or invariance) should start with this paper to get a definition of we are looking
for and why. Arjovsky etal paper should be read with this perspective in mind, as an attempt to
7.25.19 @1:26am - (2/ ) secure this sort of stability w/o having a model, but having a collection of varying
datasets instead. The former informs us when the latter's attempts will succeed or fail. It is appropriate here to
repeat my old slogan: "It is only by taking models seriously that we learn
7.25.19 @1:26am - (3/ ) when they are not needed". I wish I could quote from Aristotle but, somehow, the
Greeks did not argue with their Babylonian rivals, the curve-fitters. They just went ahead and measured the
radius of the earth AS IF their model was correct, and the earth was round. #Bookofwhy

7.24.19 @4:03pm - (Replying to @fhuszar @yudapearl and 3 others) It makes sense to me to start from what is it
that we know must hold true. Discovering which invariances you want to guarantee based on your data is OK but
the data are only a sample and it's important to declare which invariances are desired and why.

7.25.19 @12:44am - (Replying to @RichmanRonald) This paper (link: http://web.stanford.edu/~hastie/Papers)


.is an interesting window into statistics education of 2019. Contrary to causal logic, statistics students start
with data visualization routines and then ask: when do these have causal interpretation. I hope #Bookofwhy will
change this order.

7.25.19 @12:19am - (Replying to @KordingLab @fhuszar and 4 othersz) There has always been an understand
between ci and ml -- I do identification and you do estimation. Nothing has changed with all the talk about
"causal ML", except perhaps ML folks internalizing the limits of the Ladder of Causation #Bookofwhy

7.24.19 @6:20pm - (Replying to @Jabaluck @PHuenermund and @fuzzydunlop123) I happened to pass by this
thread. There is no testable implication for IV from observational studies unless treatment is discrete. See (link:
https://ucla.in/2pFa9pc) ucla.in/2pFa9pc #Bookofwhy

7.24.19 @6:03pm - (Replying to @suchisaria @tdietterich and 2 others) Yes, these papers are in scientific
language. Not because they are using DAGs, but because they provide theoretical guarantees under meaningful
assumptions. I've vowed to read them and comments. #Bookofwhy

7.24.19 @4:22pm - (Replying to @RichmanRonald) Thanks for the pointer. I will try reading it tonight. It is
always a valuable learning experience to see how statisticians think about causal problems.

7.24.19 @3:40pm - (Replying to @fhuszar @tdietterich and 3 others) How about adding it as appendix to your

blog? "From ppp to SSS - a declarative summary"

7 24 19 @2:35pm - (Replying to @yudapearl @fhuszar and 4 others) Most importantly having read Arjovsky etal
7.24.19 @2:35pm - (Replying to @yudapearl @fhuszar and 4 others) Most importantly, having read Arjovsky etal
paper, do we understand what ppp and SSS are? Or, at least what the claim is? Such translation will help
evaluate the claim under the light of existing theories. #Bookofwhy

7.24.19 @2:25pm - (Replying to @yudapearl @fhuszar and 4 others) "If a pattern ppp is seen in the data then
something (SSS) must hold true in the world". The language of "my algorithm tries" of "my algorithm optimizes"
etc reminds me of AI in the 1970's "MYCIN tries" "ELIZA optimizes" which was later replaced with declarative
writing style.

7.24.19 @2:09pm - (Replying to @fhuszar @tdietterich and 3 others) Agree, but at this point we are trying, not
to convince ML folks but to understand what their new method provides, and under what assumptions. To
understand what a ML paper offers we need someone to translate the paper into a declarative language, that
goes:

7.24.19 @5:23am - For those interested in the history of UAI, I wrote a personal memorandum of those days
(link: http://ucla.in/2Qg0Rfs) ucla.in/2Qg0Rfs #Bookofwhy

7.24.19 @4:56am - (Replying to @yudapearl @fhuszar and 4 others) we would say: "If we observe x,y pairs from
multiple E's, and we find Y||E|X,W and NOT-X||E|Y,W, then something must hold true in the world." Now, let's
continue from here: What is it that must hold true? Can you get it from the paper? #Bookofwhy

7.24.19 @4:50am - (Replying to @fhuszar @tdietterich and 3 others) Your summary is ALMOST Causal language,
but not quite. In Causal language we do not invoke man-made algorithms (IRM) to describe environmental
properties such as invariants; we speak declaratively. For example, instead of saying: ""IRM observes" x,y pairs
from multiple E's,"

7.24.19 @4:34am - (Replying to @sobu_18) No, No, the word Zionophobia has a unique magic to it, unmatched
by antisemitism. I've never met an anti-Semite who admitted to being one, and I've never met a Zionophobe
who denied being one. Quite a difference for a mentality that denies a homeland to a people.

7.23.19 @11:38pm - (Replying to @suchisaria @tdietterich and 2 others) Great! I was't aware of these two
papers. Now we know that searches for "stability" and "invariants" are on firm scientific grounds: When we
suspect a certain relation is "stable" we can check and see that it is truly stable and what conditions will make
it unstable.#Bookofwhy

7.23.19 @8:59pm - Sorry to have missed it, hoping a video was taken. I always regret not having a video from
the first UAI, 34 years ago, at UCLA-1985, when probabilities first infiltrated the forests of AI. I see causality
following a similar route. #Bookofwhy @eliasbareinboim

7.23.19 @7:02pm - (Replying to @tdietterich @zittrain and 2 others) I understand that IRM leverages the
availability of multiple data sets, usually absent from CI. So let's represent the multiple data sets in the
language of CI and see what must hold in the world for IRM to find the invariants it is searching for. Anyone
done it? #Bookofwhy

7.23.19 @3:54pm - (Replying to @sobu_18) Careful reading of (link: https://ucla.in/2SxjgWH) ucla.in/2SxjgWH


reveals that #BDS is not charged with antisemitism, but with Zionophobia: The irrational animosity toward a

homeland for the Jewish people. Populist slogans like "apartheid" tend to tarnish the credibility of their
chanters.
7.23.19 @3:08pm - (Replying to @tdietterich @zittrain and 2 others) @tdietterich, Since you are familiar with
both CI and Invariant Risk Mininization (IRM), can you (or anyone else) explain in input-output terms how IRM
extracts from non-experimental data information that CI thought must be obtained from either a model or
intervention. #Bookofwhy

7.23.19 @3:59am - (Replying to @pcastr) If you take seriously my experience with BDS activists (link:
https://ucla.in/2SxjgWH) ucla.in/2SxjgWH you will see that "boycott" is the last thing on their agenda. Their

aim is to silence pro-coexistence voices. Is it really tenuous to see that absent such voices Israel's demise is
almost certain?

7.23.19 @3:12am - I, likewise, just want to voice my conviction that support of BDS may have serious
unintended consequences, a genocidal demise of Israel is one, as I describe here: (link:
https://ucla.in/2SxjgWH) ucla.in/2SxjgWH. From there to the demise of the Jewish people takes another leap
of logical deduction.

7.23.19 @12:34am - If you know a BDS promoter who bought into its rhetoric without checking the destructive
aims of its leadership, you have found someone who has not studied BDS as thoroughly as I have. See (link:
https://ucla.in/2SxjgWH) ucla.in/2SxjgWH

7.22.19 @8:25am - (Replying to @azuur @acastroaraujo @Jabaluck) I beg to differ. It discards whatever does not
have a MODEL, which need not be a DAG. Since every CI study must rely on a model, not having one means
keeping one in your head and giving the impression that you operate model-free. Both are ill-advised.
#Bookofwhy

7.22.19 @8:16am - (Replying to @azuur @acastroaraujo @Jabaluck) Completamente de acuerdo!!

7.22.19 @7:47am - (Replying to @bzaharatos @learnfromerror) I am also going to miss JSM this year (Carlos
Cinelli will represent me at the Fellow Reception). As to Philosophy+Statistics seminar, glad someone will be
there who can translate back to us, down in the trenches, speaking cause and effect. #Bookofwhy

7.22.19 @6:16am - (1/ ) (Replying to @RonKenett @learnfromerror) 1/ I am not sure that we are talking about
the same notions of "generalizability". For me, this word means taking experimental results from one population
and applying it to another, potentially different. I am not sure @learnfromerror means the same thing; I would
be
7.22.19 @6:22am - (2/ ) (Replying to @yudapearl @RonKenett @learnfromerror) curious to know if she does.
Why I am not sure? Because the way the two populations may differ may be non-statistical, namely they may
have the same joint distribution functions on all variables, and differ ONLY in the causal forces holding the
variables together #Bookofwhy
7.22.19 @6:30am - (3/ ) (Replying to @yudapearl @RonKenett @learnfromerror) Under such circumstances
statistical methods cannot deliver remedy unless they are guided by a causal model of those underlying forces.
Such models are absent from standard statistical writings, they are introduced back-door in the PO literature,
as in https://ucla.in/2L6yTzE

7.22.19 @1:20am - (1/ ) (Replying to @PHuenermund) Thanks Paul for starting this important discussion on
@EconBookClub of Imbens paper. Since the paper is full of objectionable macro and mini-statements, I believe
it is wise to focus on its core, summarizes precisely in the paragraph you posted. "Little is said about what
7.22.19 @1:30am - (2/ ) (Replying to @yudapearl @PHuenermund) comes before the identification question and
h t ft th id tifi ti ti " Th fi t fl t i d t di f h t t t l i
what comes after the identification question" . The first reflects misunderstanding of what structural economics
is about, while the second decries a "before vs. after" distinction that should be welcome with joy. The rest of
Imben 7.22.19 @1:39am - (3/ ) (Replying to @yudapearl @PHuenermund) paper follows as a corollary of these
two basic misunderstandings. Once we illuminate these two, it should be easy to clear the rest, especially the
identification stage itself, which he admits to be virtually absent from his tool set. (if # of variables exceeds 3)
#Bookofwhy

7.21.19 @11:47pm - An old saying goes: "When mathematicians notice an interesting problem it becomes
SCIENCE". Today, Notices of The American Mathematical Society took notice of #Bookofwhy, thanks to Lisa
Goldberg : https://www.ams.org/journals/notices/201907/rnoti-p1093.pdf ... I hope #causalinference gets
enriched with new insights.

7.21.19 @11:18pm - This is funny! Cornel West is "standing in moral solidarity with four sisters?" Last I wrote
about him and his moral deformity, it was he who needed moral solidarity
https://jewishjournal.com/opinion/judea_pearl/170625/open-letter-cornel-west/ ...

7.21.19 @10:45pm - (Replying to @bzaharatos) In what way?

7.21.19 @9:34pm - (Replying to @oacarah) On the mark! Which proves the algebra is for the birds.

7.21.19 @6:02pm - Your daughter, Isadora, is adorable. If she could only get that partial differential equation
right, we could have had a perfect afternoon -- next time!

7.21.19 @4:49pm - Look Ma! I'm a statistician! They tell me that at the upcoming JSM meeting in Denver I'll be
ordained as an ASA Fellow. I assume it means that, starting July 30, statisticians can treat #Bookofwhy as a
homegrown tomatoe, and each of its toy problem as 10,000 "real-life" examples.

7.20.19 @3:01pm - (Replying to @AdanZBecerra1 @PHuenermund) Do you know what blog or forum do
thoughtful statisticians use to communicate ideas about the philosophy of statistics, including causal inference
??

7.20.19 @4:41am - (Replying to @AdanZBecerra1 @PHuenermund) Another remarkable observation: none of the
discussants had any clue on how to handle conditioning on post-treatment variables. The advice they got was:
"Control for as many pre-treatment variables as you can." Is this the best place to learn how statisticians think?
#Bookofwhy

7.19.19 @11:37pm - (Replying to @yudapearl @AdanZBecerra1 @PHuenermund) @AdanZBecerra1, I remember


you from Gelman's blog. You reminded the discussants of DAGs and got bullied by one of the big guys who told
you to go solve a "real life" problem. I was tempted to come to your rescue but those guys knew so much about
"real-life"- got scared #Bookofwhy

7.19.19 @8:11pm - (Replying to @PHuenermund) Totally agree. Discussions should start with the question: Why
are DAGs "pedagogically" more "transparent". Is it just "taste"? Or is it something universal in our minds that
makes DAGs good "displayers of assumptions"? The rest follows from this cognitive phenomenon.#Bookofwhy

7.19.19 @7:37pm - (Replying to @PHuenermund) Interesting. Guido may have changed his mind - I did not. I
truly believe that if any of my economist colleagues actually roles sleeves and solves a couple of toy problems,
instead of talking ABOUT them, he/she will never go back to talking ABOUT them. #Bookofwhy
7.19.19 @12:28pm - Each toy problem is 10,000 "real life" examples, in which you cannot hide behind messy
data, unobserved confounders, or other excuses and, moreover, you can check your method against ground
truth. Those who shun toy problems do have something to hide, watch them. #Bookofwhy

7.19.19 @2:23am - (Replying to @aqsaqal) Sounds like a beautiful example of missing-data. And I bet it is
recoverable, since the missingness of Weight is not caused by Weight itself. See a graphical approach
https://ucla.in/2LdEjZW (and stay away from imputation) #Bookofwhy @Carthica

7.19.19 @1:39am - (Replying to @thosjleeper @matloff) My My! Good point! It is better that they leash out on

me than to let truth reduced to "gossip". Thanks, #Bookofwhy

7.18.19 @11:32pm - (Replying to @yudapearl @thosjleeper @matloff) Why is quoting unnamed statisticians a
bad thing? Would naming them make less hostile? Recall, they are still alive, deciding on promotions, and
perhaps repenting for what they once thought. Why embarrass them? #Bookofwhy

7.18.19 @11:27pm - (Replying to @thosjleeper @matloff) Comments on your thread. The statement P(Y|do(x))
is causal but not counterfactual, because it implies no contradiction, and it can be estimated from RCT.
P(Y|see(x)), on the other hand, is statistical, not causal, because we can evaluate it without experiments.
#Bookofwhy

7.18.19 @11:12pm - (Replying to @anirudhacharya1 @oacarah) How is this: https://ucla.in/2O2z4St+ ?? Any


luck? #Bookofwhy

7.18.19 @10:56pm - (Replying to @klausmiller @PHuenermund and 3 others) Judea Pearl Retweeted Judea Pearl
As I Tweeted here: https://twitter.com/yudapearl/status/1151970273148596224 ... this is the greatest thing
that happened to DAGs: Sunrays are the best de-confounders. #Bookofwhy @quantadan @analisereal

7.18.19 @10:39pm - I like this title: "When in Doubt, DAG it Out. " @oacarah new commentary on "Analyzing
Selection Bias for Credible Causal Inference" https://zimbra.cs.ucla.edu/service/home/~/?
auth=co&loc=en_US&id=723335&part=2 ... It confirms my mantra: "DAG goes where knowledge resides," which
PO folks are about to internalize. #Bookofwhy

7.18.19 @8:11pm - (Replying to @vauhinivara @WSJ) Delighted to read your thread and to see that the WSJ
Internship set up in memory of my son Daniel was instrumental in lifting you forward. Pitching story, so it
seems, is much like pitching scientific articles - enlightened editors are scarce. Good luck #Bookofwhy
@craigmatsuda

7.18.19 @2:41pm - Thanks for pointing to Imben's new paper on PO and CI. We finally have a window into the
thinking of leading PO researchers. It will give econ. students a chance to compare alternatives and ask: Do we
really want to think that way? #Bookofwhy @StatsPapers @causalinf @PHuenermund

7.18.19 @12:16am - (1/2) (Replying to @omaclaren @djvanness and 2 others) 1/2 We have this world; it is
called classical mathematical modeling. What this world was missing (when I last fiddled in it) was the EXTREME
case called DAGs. Namely, the miracle of how much can be accomplished with so few and weak assumptions. I
might go back to classics, once
7.18.19 @12:22am - (2/2) (Replying to @yudapearl @omaclaren and 3 others) we fully understand the full
potentials of this new miracle. But it seems to go on and surprise us with new capabilities: external validity,
missing data, sensitivity analysis, fairness,...Its pouring! And you want us to quit? #Bookofwhy
7.17.19 @10:53pm - (Replying to @omaclaren @djvanness and 2 others) You have this strength in SCM if you
want to express something you are really sure about. But if you are only sure about "who listens to whom" use
DAGs. Perhaps you are after a calculus that works in between these two extremes? Happy sailing; watch the two
extremes #Bookofwhy

7.17.19 @10:43pm - (Replying to @omaclaren @djvanness and 2 others) Easy! You write Y = OR(X,Z) or
Y=AND(X,Z). This is how you express it if you think you can. If you can't commit to one function or another you
just write Y=f(X,Z) and you let data decide. If your data must be taken in isolation, it is mathematical
impossibility.#Bookofwhy

7.17.19 @10:35pm - (Replying to @omaclaren @djvanness and 2 others) You are repeating what I said, so we are
settled. Recall the DAGS were devised to minimize the amount of information (ie assumptions) demanded from
researchers, and limit it to those relations only that reside in the scientist's comfort zone --variable listening
#Bookofwhy

7.17.19 @10:20pm - (Replying to @LMCheongTS) RL can handle some interventional problems, like predicting
effects of actions tried before. As to counterfactuals, RL can potentially produce bounds on counterfactuals,
the same as what Rungs 1 & 2 produce. A miracle shown here: https://ucla.in/2G2rWBv . #Bookofwhy

7.17.19 @10:11pm - (Replying to @omaclaren @djvanness and 2 others) Yes, `interaction' a concept that is
formalised in SCMs . Compute P(Y|do(x),z)) and check if it depends on z. Note that this quantity is computable
from full SCM specification. Data is needed only to supplement missing specifications, say P(u), or functional
forms #Bookofwhy

7.17.19 @10:04pm - (Replying to @omaclaren @djvanness and 2 others) To state directly that interaction does

not exist SCM must resort to some parametric specification, say "linear family". Or, indirectly, SCM can tell us:
"Check for yourself. I have given you part of me, a DAG, which implies a recipe for interrogating the data; use
it. #Bookofwhy

7.17.19 @9:56pm - (Replying to @ntabari8221 @fulhack) The #Bookofwhy argues that too much time is spent on
"confounders," What we should be looking for are the "de-conounders".

7.17.19 @9:52pm - (Replying to @djvanness @ConiByera and 2 others) Yes, it is "legal" with or without the U.
And, remarkably, one cannot tell who is the modifier and who is the "cause". This distinction still demands a
"penalty" definition: What would I lose if I incorrectly switch the two. #Bookofwhy

7.17.19 @9:45pm - (Replying to @djvanness @ConiByera and 2 others) Not really. The topology of the DAG (with
or w/o U) tells us how to get the Z-specific causal effects. Whether or not those effects depend on Z would be
revealed from the data, once we estimate P(Y|do(x), z) correctly. #Bookofwhy

7.17.19 @8:33pm - (Replying to @djvanness @ConiByera and 2 others) The last warning is given to you by the
DAG, coupled with a method of quantifying the penalty. So, what is missing? #Bookofwhy

7.17.19 @8:30pm - (Replying to @djvanness @ConiByera and 2 others) I am still missing what that "additional
extra-statistical structural information" is that you need "to get the policy estimates right." What kind of answer
you expect the system to deliver to you if you had this extra information encoded in the DAG #Bookofwhy
7.17.19 @8:25pm - (Replying to @djvanness @ConiByera and 2 others) If your target quantity is P(Y|do(x),z) for
different z's, you can identify it by controlling for C. The DAG with U gives you the same information, ie
"interaction is possible". You probably want an early warning saying: "dont bother, no interaction here". Right?
#Bookofwhy

7.17.19 @8:16pm - (Replying to @djvanness @ConiByera and 2 others) If all you wish is P(Y|do(X=1)) then you do
not need to account for Z. But you really wish to know more. You must start by telling the inference-engine
what it is. Perhaps you seek the difference P(Y|do(X=1), Z=0) - P(Y|do(X=1), Z=1)? Or ratio? or something else?
#Bookofwhy

7.17.19 @8:06pm - (Replying to @djvanness @ConiByera and 2 others) I will try to answer slowly each one of
your tweets. With DAGs, no one is left confused. The DAG you drew just says: I don't rule out interactions
between Z and X. And he ( Mr. DAG) is now waiting for you to specify you research question. What is it?
#Bookofwhy

7.17.19 @7:51pm - (Replying to @omaclaren @djvanness and 2 others) SCM assumes the existence of functions
Y=f(x,z,w...), with the help of which we define other concepts such as causal effects, explanation etc. Whether
or not we can identify those concepts from data is a separate issue. Those who wonder whether DAG has a
concept C need first to
7.17.19 @7:59pm - (2 )(Replying to @yudapearl @omaclaren and 3 others) define what they mean by C. They
can use mathematics, or examples, or the method of "penalty", which goes: "If I know C, I can escape from the
penalty that I would suffer not knowing C." So far I have not seen a penalty definition of the concept sought by
Dave. #Bookofwhy

7.17.19 @7:43pm - (Replying to @mgaldino) I care less about those "some economists" than I care about our
"current and living economists". Do they tell their students: "We rely on models"? Are they proud or ashamed of
using models? Are they reading models or hiding them? @Bookofwhy

7.17.19 @6:39pm - Every time I read Quora I wonder: Do empirical economists in Angrist etal school consider
themselves "model-based" or "model-free". Some say: "unlike statisticians, we use models". Others say: "unlike
traditional econs, we are not tied to models". I wish someone would clarify.

7.17.19 @6:17pm - (Replying to @djvanness @ConiByera and 2 others) I tried to follow this thread, and still can't
see what "logical interaction" you wish to represent. DAG in itself just says "there can be an interaction". SCM
says more: it tells us the function Y=f(X,Z,U,W...). But I sense that you want more, what is it? and what for?
#Bookofwhy

7.17.19 @5:12am - Apropos "Zionophobic Thuggery", I am delighted to learn that the words are gaining traction
all the way to the Department of Justice. They were mentioned 5-6 times in this Summit meeting of DOJ
https://www.c-span.org/video/?462607-1/justice-department-anti-semitism-summit , and they are making my
students vow: No more of this racism.

7.17.19 @4:50am - And I need to check my blog, to make sure I ever wrote it. Who knows how many useful
nuggets are buried in that blog, written before Twitter, when I had time to write long answers, and when
people expected in-depth answers to questions that textbooks tended to ignore. #Bookofwhy

7.17.19 @3:41am - Funny, 1/4 of a century later, the introduction to "Graphical Models, Causality, and
Intervention" (1993) https://ucla.in/2pJqtW3 sounds like it was taken straight from #Bookofwhy
7.17.19 @2:33am - The first paper using do() notation was "A probabilistic calculus of actions" (1994)
https://ucla.in/2pFa0Ca , though the idea was used earlier (1993), using set(X=x) notation, in
https://ucla.in/2pJqtW3 #Bookofwhy

7.16.19 @12:39am - (Replying to @djvanness @y2silence and 2 others) Good example where most people agree
that Z is a modifier of the effect of X on Y, and not that X is a modifier of the effect of Z on Y. So, back to
penalty analysis. What will go wrong if I mislabel them and I mistake X to be a modifier of the effect of Z on Y?
#Bookofwhy

7.16.19 @12:26am - (Replying to @VladZamfir @lostinio and 3 others) Curious, what is your favorite
epistemology of uncertainty? Plus, what is a good way to assess the merit of such epistemology? #Bookofwhy

7.15.19 @9:56pm - (Replying to @VladZamfir @oliverbeige and 2 others) Two remarks. (1) There are hordes of
identities one can extract out of probability theory. Why did Bayes' receive such prominence? (2) I am still
searching for a definition of Bayesianism, not for what's wrong with its practitioners. #Bookofwhy

7.15.19 @8:41pm - (Replying to @djvanness) The way I usually get to the "essence" of a distinction is to ask
myself: How would I be penalized if I did not know the distinction. In our case: What would I do wrong if I were
to mistake an "effect modifier" for a "cause"?? From the penalty comes the "essence." #Bookofwhy

7.15.19 @6:20pm - (Replying to @djvanness) Many feel that DAGs and other models do not capture the "essence"
of "effect modifier", as opposed to just "cause". But few have taken time to explicate, even semi-formally what
that "essence" is, and how they would benefit if a model does capture it. Care to try? #Bookofwhy

7.15.19 @3:28pm - (Replying to @VladZamfir @oliverbeige and 2 others) The frequentist interpretation of Bayes
rule is a one-line high-school algebra trick. This is not what made Bayes famous, controversial and revered.

#Bookofwhy (p. 102-3) proposes an explanation of what made Bayes rule justifiably revered.

7.15.19 @2:48pm - (Replying to @ipyadev @_MiguelHernan @JohannesTextor) Thanks for posting this paper, by
Clarice, which skipped my awareness. The key sentence is: "Any 2 direct causes of D are effect modifiers for
each other on at least 2 scales, which can make a reasonable person question the utility of the concept. " I'll try
to clarify #Bookofwhy

7.15.19 @5:46am - (1/ ) (Replying to @oliverbeige @VladZamfir and 2 others) For some people, Bayesianism is
spraying priors on parameters and waiting for the posteriors to peak. For others, especially those who read
Bayes (1763), Bayesianism means 1) leveraging subjective knowledge and 2) processing evidence by Bayes' Rule.
I was in the second camp
7.15.19 @5:55am - (2/ ) (Replying to @yudapearl @oliverbeige and 3 others) When I coined the name "Bayesian
networks" (1985), but I am camp-friendly and would not mind switching if you can tell me what I am switching
to. In the meantime I also wrote "Why I am only half Bayesian" https://ucla.in/2nZN7IH which will affect my
next camp #Bookofwhy

7.15.19 @12:42am - (Replying to @brettjgall @BrendanNyhan @NateSilver538) This is an under-understatement.


Causal inference is actually a mathematical impossibility without a model, more accurately, without a "causal
model", namely one that cannot be articulated in the language of statistics. #Bookofwhy

7.14.19 @9:51pm - (Replying to @ZionessMovement @CoryBooker) Thank you @ZionessMovement for joining my
Twitter and for informing me of how @CoryBooker and VC Biden responded to IfNotNow squeeks. I neve say the
Israeli Palestinian conflict is "COMPLEX". It is baby simple: One side says "we, we, we" and the other "me, me,
y p y , , , ,
me". Baby simple.

7.14.19 @7:54pm - (Replying to @ipyadev @_MiguelHernan @JohannesTextor) In DAGs all variables are
presumed to be "effect modifiers" by the nonparametric nature of the assumptions. You probably meant to ask:
How do we mark variables that are NOT effect modifiers. My question: for what purpose? ie What would those
marks enable you to do? #Bookofwhy

7.14.19 @1:55am - (Replying to @ERMANigeria) In Quora we are witnessing another phenomenon, not
parametric addiction. Angrist for example, https://www.quora.com/What-are-the-differences-between-

econometrics-statistics-and-machine-learning ... talks as if economics has a unique & exclusive ownership on all
CI questions, and everyone else is doing data-fitting. It's a remarkable phenomenon. #Bookofwhy

7.14.19 @12:19am - (Replying to @ERMANigeria) Please share. I have not seen that. Perhaps because, immersed
in non-parametric modeling, I failed to notice what my colleagues have been doing. #Bookofwhy

7.14.19 @12:17am - (Replying to @henning_lars) I agree. But how do they intend to "aggregate data across a
variety of sources," about which "good old statistical methodology" is totally helpless. [And I include Meta
Analysis in this state of helplessness]. I am afraid they will fall irreversibly for same dead-ends #Bookofwhy

7.14.19 @12:03am - Sharing. I have noticed that all CI-related questions on quora are answered by
econometricians who, as we have noticed on this Twitter, are not too familiar with SCM (yet). I therefore
posted a new answer: https://www.quora.com/What-are-some-of-the-important-things-to-know-about-causal-
inferences/answer/Judea-Pearl-1 , in the hope of stimulating awareness #Bookofwhy

7.13.19 @6:43pm - (Replying to @akelleh) Beautiful introduction to Inverse Probability Weighing of do-
adjustment. I hope you have more users than argumentative skeptics. #Bookofwhy

7.13.19 @6:25pm - (Replying to @stuartbuck1 @asacarny and 3 others) I do not see any condescending
comments around. I see a request to accept the premise that "graphs go where the knowledge resides". If
accepted, the rest is corollaries. If not, lets discuss what relationships the mind may find easier to discern than
"Y listens to X" #Bookofwhy

7.13.19 @5:30pm - Thanks for posting these quotes from Karlin and Pearson. Interestingly, today we hear a
slightly different song: "Yes, the process generating the data is important," - end of song. No! It's the beginning:
How do we represent that "process"? How do we operationalize it? #Bookofwhy

7.13.19 @4:39pm - (1/ ) Interesting News: Google sister-company Verily is teaming with big pharma on clinical
trials https://www.cnbc.com/2019/05/20/alphabet-verily-doing-clinical-trials-with-novartis-sanofi-pfizer.html?
utm_content=95923950&utm_medium=social&utm_source=twitter&hss_channel=tw-3377714967 ... Quotes: 1)
Will try to find ways to modernize their clinical trials and speed up the time it takes to bring a new drug to
market. 2) aggregate data across a
7.13.19 @4:39pm - (2/ ) variety of sources, 3) Clinical trials have historically been expensive processes that rely

on outdated technologies. Does anyone know the technical leaders of this initiative? e.g., Are the versed in the
less outdated literature on "aggregating data"? #Bookofwhy

7.13.19 @3:05pm - (Replying to @dkipb) It runs deeper, agree. It's a product of screwed up education in all
fields. Today's urgency is to get over those tribal narratives and get things right. And to get things right we need
y g y g g g g g g g
accept that business is not as usual -- those distinctions must be operationalized #Bookofwhy

7.13.19 @5:56am - (Replying to @jack_bowdenjack @RobinGomila) Sure, but this does not make the difference
a "causal effect". What makes it a "causal effect" is the model and what the model says about the ignorability
condition, which is here taken to hold apriori, thus assuming away the causal component of the problem.
#Bookofwhy

7.13.19 @4:13am - (Replying to @NoahHaber @economeager and 2 others) @NoahHaber, I overlooked your
thoughtful thread on the comparison of epi to econ. Allow me to reiterate another element: Epi use DAGs
because "DAGs go to where knowledge resides". Econ. do not use DAGs because it is culturally prudent to do
things in your head. #Bookofwhy

7.13.19 @2:54am - I've just posted an answer on quora https://www.quora.com/What-are-probabilistic-


graphical-models-and-why-are-they-useful/answer/Judea-Pearl-1 ... which ends with a catchy phrase of what
#Bookofwhy is trying to do: It tries to "dispel the myth that AI is about to be colonized by statistics. " -- Sharing

7.13.19 @2:35am - I still can't figure out what makes statistics leadership so eager to tell their constituency:
Don't panic, there is no confusion, our ancestors have been here before, they touched, they recognized and
pointed out, they discussed and distinguished, business is as usual #Bookofwhy

7.13.19 @2:09am - (1/ ) (Replying to @RobinGomila) I think the sentence: "I draw on econometric theory and
established statistical findings to demonstrate that linear regression (OLS) is generally the best strategy to
estimate causal effects on binary outcomes" will make some economists rethink if "Almost Harmless" is truly
7.13.19 @2:20am - (2/ ) (Replying to @yudapearl @RobinGomila) as harmless as some claim. It seems to me that
you are dealing with strategies to estimate conditional expectations, not causal effects. This potential
confusion is prevalent in the new culture schooled by structure-less econometrics.

7.12.19 @2:57pm - This "Mediation Formula" always restores my confidence in the power of very simple
mathematics to produce very meaningful results. #Bookofwhy

7.12.19 @2:38pm - (1/ ) (Replying to @asacarny @paulgp and 2 others) I am not an NBER member, but
ocassionally producing results that are of interest to economists. The exclusive policy of NBER prevents its
members from learning these results. This in itself may not be sufficient reason to change policies, but the
reputation of econometrics
7.12.19 @2:44pm - (2/ ) (Replying to @yudapearl @asacarny and 3 others) as an insular and outdated field does
hurt its members, and increasingly so in the age of data-science. I therefore welcome your proposal to open
another channel of communication between economics and the rest of the scientific community. #Bookofwhy

7.12.19 @3:31am - (Replying to @yudapearl @AndroSabashvili @_MiguelHernan) Gee, I think I know why they
discarded explanations; it goes back to the RCT's roots and the philosophy of "well defined interventions". You
cannot define "undoing past events" unless you have manipulations; plain "events" cannot explain things, only
manipulations can. #Bookofwhy

7.12.19 @3:21am - (Replying to @AndroSabashvili) I am not sure "explanation" is part of @_MiguelHernan


taxonomy, because explanations demand retrospective counterfactuals, not "predictive counterfactuals". Thus,
scientists quest for explanation remains outside data science. I wish I knew why they changed the ladder.
#Bookofwhy

7.12.19 @3:02am - (Replying to @AndroSabashvili) You are very right. I wrote to @miguelHernan: 5.7.19
7.12.19 @3:02am (Replying to @AndroSabashvili) You are very right. I wrote to @miguelHernan: 5.7.19
@1:47pm - I am questioning the benefit of separating "description" from "prediction", skipping "diagnosis" and
lumping together "intervening" and "retrospecting" under one opaque category "causal inference". No-Ans.
#Bookofwhy

7.11.19 @10:21pm - (1/ ) (Replying to @PavlosMsaouel) 1/ Glad I found another fan. Shmueli's paper had many
(1,200) citations which is encouraging. Statisticians hit the roof when I tell readers how causality was neglected
in 20-century statistics. Glad they have accepted Shmuei's story. It seems that David Hand wrote his new paper
7.11.19 @10:40pm - (2/ ) (Replying to @yudapearl @PavlosMsaouel) to refute Shmueli's conclusions but, in my
opinion, he actually reinforces them with additional evidence provided by the refs to Box, Breiman (2001) and
Cox (1990). I hope #Bookofwhy sparks renewed discussion on the nature of modeling.

7.11.19 @8:43pm - (1/3) While struggling to find a ladder or demarcation lines between rungs, Hand's paper
called my attention to a comprehensive article by Shmueli (2010) "To Explain or to Predict?", rich with
references and quotes, which states: "Although not explicitly stated in the methodology
7.11.19 @8:43pm - (2/2) literature, applied statisticians instinctively sense that predicting and explaining are
different." https://projecteuclid.org/download/pdfview_1/euclid.ss/1294167961 ... Going through the quotes
and references in Hand's paper, I believe they confirm Shmueli's verdict: "the statistical literature lacks a
thorough
7.11.19 @8:43pm - (3/3) discussion of the many differences that arise in the process of modeling for an
explanatory versus a predictive goal." One may also note that, adding to a "thorough" discussion, the Ladder of
Causation provides "formal distinctions" between those differences. #Bookofwhy

7.11.19 @4:49pm - (Replying to @namalhotra @matt_blackwell and 18 others) Whence do you think comes the
urge to condition, despite "Didn't we know this already". Is there some textbook guideline that makes good
people forget what they know? #Bookofwhy

7.11.19 @3:53pm - (Replying to @Scientific_Bird @elizpingree @BridgetPhetasy) You will not regret it. PRIMER is
ideal, especially to Inquisitive-minded Birds.

7.11.19 @3:47pm - (Replying to @cutearguments @rlmcelreath) I was not aware of "Virtual History" when
writing #Bookofwhy. Thanks. Gelman, on the other hand, is wrong; Rubin denies any meaning to "undoing past
events". That's why he replaced "counterfactuals" with "potential outcomes".

7.11.19 @3:35pm - (Replying to @namalhotra @matt_blackwell and 18 others) This is quite shocking: "Overall,
we find that 46.7% of the experimental studies published in APSR, AJPS, and JOP from 2012 to 2014 engaged in
posttreatment conditioning (35 of 75 studies). " Something to keep in mind in scrutiniaing "real-life"
experimental studies. #Bookofwhy

7.11.19 @6:48am - (Replying to @cdsamii @ingorohlfing and 14 others) What are the cultural roots of path
models in political science? Duncan and LISREL? or some other source? When and how did PO enter into this
field, as we can see in the works of King and Imai?

7.11.19 @3:56am - (Replying to @nasim_rahaman @MPI_IS) Gee, I recognize the books behind me, but not the
Balcony. I hope it is not a "cross world" illusion. #Bookofwhy

7.11.19 @3:51am - For the physicists among us, especially those fascinated by quantum mechanics and Bell
Inequality, DAGs appear to enlightened the conversation when we allow communication among observers, see
https://arxiv.org/pdf/1607.08182.pdf
7.11.19 @1:34am - (Replying to @ingorohlfing @thosjleeper and 13 others) This is a good paper, thanks for
posting. DAG's are used here not merely to convey assumptions, but also to visualize violations of assumptions
and how these correspond to PO expressions. I can't imagine any paper on racial discrimination skipping DAGs in
those roles #Bookofwhy

7.11.19 @1:11am - (Replying to @yskout @Jabaluck and 11 others) Robustness comes from making assumptions
in the language of stored knowledge. And logical equivalence is not "modeling equivalence. Ask an economist to
tell you if HIS latest model has testable implications. Can he do it? Sure! But compare his effort to students of
DAGs#Bookofwhy

7.11.19 @12:49am - (Replying to @thosjleeper @PHuenermund and 12 others) I am an optimist. I imagine 18K
out of our 19.1K followers to be silent observers today and vicious rebels tomorrow; they are not dumb. Yes,
they have to worry about tenure, publications and other suppressants of academia, but they are no dumb. They
see the potentials. #Bookofwhy

7.11.19 @12:31am - (Replying to @thosjleeper @PHuenermund and 12 others) I believe Imai and his students are
recent converts and, in social science, we have Morgan, Winship and Elwert. #Bookofwhy

7.11.19 @12:21am - (Replying to @JDHaltigan @WiringTheBrain @PeterJungX) Agree. Mediation analysis is one
of those areas where regression culture is deeply entrenched, and you can still find PhD theses written in stone-
age style. I noted it in pp.324-5 of #Bookofwhy, https://ucla.in/2KYvzau

7.11.19 @12:12am - (Replying to @WiringTheBrain @PeterJungX) I am not sure to what extent your concern is
valid today, 30 years after the Causal Revolution. I know there are pockets of regression analysts who refuse to
elevate themselves from statistical thinking, but most of those in my circle are aware of ANCOVA's dangers
#Bookofwhy

7.10.19 @11:05pm - (1/4) Happy Anniversary! About a year ago, I started Twitting and posted this: 6.27.18 - Hi
everybody, the intense discussion over The Book of Why drove me to add my two cents. I will not be able to
comment on every tweet, but I will try to squeak where it makes a difference....
7.10.19 @11:05pm - (2/4) A year later, I can hardly believe the 2,300 Tweets behind me, 19.1K followers, and a
pleasant sense of comradeship with the many inquisitive minds that have been helping me demystify the
science of cause and effect. I have benefited immensely from seeing how causality is
7.10.19 @11:05pm - (3/4) bouncing from your angle and how it should be presented to improve on past faults.
To celebrate our anniversary, I am providing a link to a search-able file with all our conversations (my voice)
sorted chronologically http://web.cs.ucla.edu/~kaoru/jp-tweets ... Today I re-read selected sections
7.10.19 @11:05pm - (4/4) and, believe me, there are quite a few non-trivial nuggets out there that should be
retrieved, re-weaponized and reused. Finally, access to individual chapters of PRIMER is now available by
clicking here https://ucla.in/2KYYviP Cheers, #Bookofwhy.

7.10.19 @6:38pm - (Replying to @MatthewMOConnel @GoldsteinBrooke @TuckerCarlson) We are shifting from


defense to attack and, with your help, will make Zionophobia the ugliest word in town. Use it!

7.10.19 @5:21pm - (1/ ) (Replying to @fuzzydunlop123 @autoregress and 11 others) Economics? "Open to new
methods"? The last open-minded economist I met was Hal White (1950-2012). When he passed away, his former
students could no longer publish in top journals, and had to revert back to old methods (my interpretation).
Now observe how hard Heckman and Pinto
7.10.19 @5:26pm - (2/ ) (Replying to @yudapearl @fuzzydunlop123 and 12 others) labor to preserve old methods
https://ucla.in/2L8OCyl ; pages after pages of derivations just to refrain from using d-separation. "Open to new
methods"? Let's see how long it will take to Editors of top Econ Journal to invite ONE review paper on what
graphical models can offer
7.10.19 @5:41pm - (3/ ) (Replying to @yudapearl @fuzzydunlop123 and 12 others) econometric researchers.
This is 2019, almost 40 years after graphical models came into being and impacted almost every data-driven
discipline, except economics. It's hard to believe, I agree, and surely our silent rebellious students will not call
it "open minded" #Bookofwhy

7.10.19 @5:14pm - (Replying to @autoregress @fuzzydunlop123 and 11 others) Having been burned before by
ungrounded theory is no excuse for refusing an eye-glass that helps you navigate your OWN theory.

7.10.19 @4:36pm - (1/ ) (Replying to @Jabaluck @yskout and 11 others) I agree, but would phrase it a bit
differently. Any economist NOT familiar with DAGs would rejoice knowing that his/her most intimate chunks of
economic knowledge can now be expressed in a scientifically prudent language, uncontaminated by parametric
or statistical baggage
7.10.19 @4:51pm - (2/ ) (Replying to @yudapearl @Jabaluck and 12 others) ready to be submitted for

mathematical or algorithmic analysis, in which each step is meaningful, that is, scrutinizable by the knowledge-
providing economist, and which readily delivers answers to questions that otherwise take hours to answer. How
about this phrase? #Bookofwhy

7.10.19 @3:36pm - (1/ ) (Replying to @Jabaluck @yskout and 11 others) 1/ Allow me to use your favorite: "You
are missing my point". The key point that I am trying to make is that there is such a thing as "human
representation of knowledge" and it has a cognitive library of primitive relationships from which more complex
relationships are composed.
7.10.19 @3:44pm - (2/ ) (Replying to @yudapearl @Jabaluck and 12 others) The compound relation "prehospital
behavior that might independently effect outcomes", is composed of many primitives relationships, each of the
type "X directly affect Y." DAGs tap those primitive relationshiop directly, hence reliably. Compound relations
require mental
7.10.19 @3:52pm - (3/ ) (Replying to @yudapearl @Jabaluck and 12 others) construction effort and are more
vulnerable therefore to human error. This is the key to the clarity of DAGs, not "familiarity". "DAGs go to where
knowledge resides" I once said. Any discussion of DAGs should start with this key observation; the rest are
corollaries #Bookofwhy

7.10.19 @2:54pm - (1/2) (Replying to @Jabaluck @yskout and 11 others) 1/2 @Jabaluck Your resistance to DAG
betrays your cultural upbringing (Rubin? Angrist? Imbens?) and refutes your own words: "We control for
differences in prehospital behavior that might independently effect outcomes." Anyone who can judge if a
difference "INDEPENDENTLY AFFECT
7.10.19 @3:04pm - (2/2) (Replying to @yudapearl @Jabaluck and 12 others) THINGS" can surely judge if "one
variable directly affects another". The latter type of judgments is all that is required for constructing a DAG,
hence it could not be as "incredibly challenging" as you describe it. This is universally true for any
representation of knowledge
7.10.19 @3:11pm - (2/4) (Replying to @yudapearl @Jabaluck and 12 others) 2/4 and it does NOT depend "on
whether people are familiar with the terminology." To semi-prove my point, notice that your recipe "We control
for differences in prehospital behavior that might independently effect outcomes" is deficient, for it misses
controls for variables that 7.10.19 @3:30pm - (4/4) (Replying to @yudapearl @Jabaluck and 12 others) 4/4
affect treatment, not outcome. So, it is hardly the case that "The language economists use seems efficient and
unambiguous here." The language is ridden with ambiguities, which calls into question the credibility of key
judgments issued by DAG-avoiding economists. #Bookofwhy

7.10.19 @2:13pm - (1/2) (Replying to @PeterJungX @WiringTheBrain) Oh, how I wish to see responses of
statisticians to this question! How they interpret the words "works", "control for" "confounders" etc. How their
answers vary depending on what rival camp they declare allegiance to, and more. Can we get a summary? My
answer: Yes, if you are
7.10.19 @2:26pm - (2/2) (Replying to @yudapearl @PeterJungX @WiringTheBrain) versed in causal inference (CI)
and its eye glasses graphical models No if you are a mainstream statistician believing that "confounding" is a
and its eye-glasses - graphical models. No, if you are a mainstream statistician, believing that confounding is a
statistical notion. For a more detailed answer, see "confounding bias" pp. 53-60 of PRIMER
https://ucla.in/2KYYviP #Bookofwhy

7.10.19 @5:12am - (Replying to @EpiSconroy @EpiEllie) Stat theory include probability theory, regression
analysis, hypothesis testing, confidence intervals, etc. all are theories of the joint distribution functions that
govern the observed data. #Bookofwhy

7.10.19 @3:58am - (1/ ) (Replying to @Jabaluck @autoregress and 10 others) I did not say I'll never convince
you. I said: the reason I am spending time on tweeter is not in hope of convincing you or Angrist. I do it to
empower the curious yet silent rebels among econ. students, what seems like an easier task. You (not sure
about Angrist) will be
7.10.19 @4:15am - (2/ ) (Replying to @yudapearl @Jabaluck and 11 others) convinced (I hope) when one of
these students asks in class: "Why can't we test the exclusion restriction by checking if E(Y|x, z) depends on z?"
which will make you sorry for not teaching d-separation. You asked: what I have learned from economists, a
question bothered me a
7.10.19 @4:36am - (3/ ) (Replying to @yudapearl @Jabaluck and 11 others) a lot, because I was hoping to tell
my students: Economists have developed methods of solving problem A, R and T. But, aside from LATE, I have
hard time giving A, R and T content that I can easily describe and comprehend. When I ask colleagues they send
me to fancy articles
7.10.19 @5:03am - (4/ ) (Replying to @yudapearl @Jabaluck and 11 others) without telling me what nuggets of
wisdom I can expect to find there if I try really hard. I made this plea on Twitter with not much success. But I
have not given up; I know those nuggets exist and are waiting to be excavated. Perhaps by tomorrow's rebels
#Bookofwhy

7.9.19 @10:56pm - (Replying to @yudapearl @EpiSconroy @EpiEllie) a carrier of scientific assumptions about the
world outside the data (say populations, or individuals reactions to exposures) and should be used to exemplify
ideas about scientific methods, hypotheses, evidence, predictions, abduction, inductions, etc etc..

7.9.19 @10:51pm - (Replying to @EpiSconroy @EpiEllie) The theories we learn in psych and stat are different
indeed from the epi conception of a "theory". The first, because it was mainly verbal, the second because it was

about the data, not the process generating the data. A DAG is a good embodiment of what we mean by a
"theory" ,ie

7.9.19 @10:01pm - (1/ ) (Replying to @eliasbareinboim) As a strong advocate of "reality first, algorithm second"
I should note that the level at which we model reality is sometime chosen to enable an algorithm. For example,
an inference engine may issue the output: Sorry, your model does not allow for the identification of query Q
7.9.19 @10:05pm - (2/ ) (Replying to @yudapearl @eliasbareinboim) However, if you can only think of a variable
Z that lies on the arrow X---->Y and satisfy additional properties, I would be able to identify Q using Algorithm-
1. Likewise, if you can only think of W that...using Algorithm-2. Thus, the model is refined as we go
along.#Bookofwhy

7.9.19 @3:41pm - (1/2) (Replying to @pablogerbas @Jabaluck and 9 others) I think you should get this published
someplace, for the sake of people really interested in this applied research, so that they can see the layout of
the problem clearly, and discuss substantive issues if any. Alternatively, you can publish it as an educational
device to
7.9.19 @3:56pm - (2/2) (Replying to @yudapearl @pablogerbas and 10 others) enlighten X-econs with the way CI
folks think about a problem, w/o dismissive calls for "real life problems". Or, perhaps a joint paper by all
tweeting discussants. It will turn into a classics, perhaps even an underground "bubble-burster." I'll support it
fully. #Bookofwhy
7.9.19 @3:30pm - (Replying to @autoregress @PHuenermund and 10 others) Before Galileo pointed his telescope
towards the moon, he tried it on a tree, 2 km away, and saw nothing new, just the familiar old tree, but smiling
in freshness. This story is fiction, and I am no Galileo, but DAGs are the eyeglasses of CI, no devils behind them.
#Bookofway

7.9.19 @1:17pm - (1/2) (Replying to @autoregress @PHuenermund and 10 others) Sorry, I did not say
"incapable". I curiously asked "what % of students can solve it?" Why? Because one can read "Harmless
Economics" and "Mastering Metrics" 10 times over and find no clue on how to solve it. Moreover, I am sure that
there are many secret rebels among those
7.9.19 @1:30pm - (2/2) (Replying to @yudapearl @autoregress and 11 others) students and readers, you may be
one of them, who could not sit still seeing CI advancing to new heights and continue to act as if these advances
have no bearing on economic problems. So, I am not surprised that % is rising and will continue to rise. I am
tweeting here
7.9.19 @1:46pm - (3/3) (Replying to @yudapearl @autoregress and 11 others) not to convince Angrist or Jason,
but to empower their curious and capable students to see through the X-Eco-bubble. And, yes, I am confident,

very confident, that the bubble will burst as soon as one of them tells the others: Hey, look at this microscope!
#Bookofwhy

7.9.19 @1:03pm - (Replying to @Jabaluck @PHuenermund and 9 others) You are again distorting my words,
perhaps for realizing their truth. I never uttered the word "inferior". I said "suspect, for lack of ground truth". A
doctor who insists on operating on patients w/o studying anatomy may be a great surgeon, but still highly
suspect. #Bookofwhy

7.9.19 @6:30am - (Replying to @stephensenn @omaclaren @eliasbareinboim) Let's forget then the iid
assumption and estimate E[Y|X=x] by OLS regression. Do I need to know about block design? I just randomized
people to treatment and control by a fair coin. Where will I go wrong? #Bookofwhy

7.9.19 @6:10am - (Replying to @stephensenn @omaclaren @eliasbareinboim) Consider: (1) I randomize a


treatment X and record data in the form of x,y pairs. (2) I forget that these pairs came from RCT and imagine
that they are iid samples from some distribution P(x,y) (3) I estimate E(Y|X=x) under this illusion. Will the
illusion hurt me? #Bookofwhy

7.9.19 @5:08am - A new explanation-seeking paper that hit my screen: https://mobiuk.org/2019/abstract/S4-


P2_Hasthanasombat_TowardsExtractingExplanations.pdf

7.9.19 @4:27am - (1/2) (Replying to @omaclaren @eliasbareinboim @stephensenn) Yes, now that we find one
reader thinking that the complete version makes a difference, we will try to include it in the upcoming
paperback version. I personally think that it was his "reduction of data" mantra that defined 20th Cent.
statistics agenda. Recall that Fisher
7.9.19 @4:37am - (2/3) (Replying to @yudapearl @omaclaren and 2 others) did not have notation for "causal
effects". Even in the context of RCT, his concern was the reduction of data obtained from a randomized trial,
not to assure correctness or unbiasedness. When he tried that (in mediation context), he blundered (ref =
Rubin), #Bookofwhy. He
7.9.19 @4:44am - (3/3) (Replying to @yudapearl @omaclaren and 2 others) definitely did not have the concept
of "causal assumption", which is essential for every task of modern "causal inference." Going from RCT-data to
"causal effect" is indeed a matter of data reduction, the causal part is already prepared for you in the design.
#Bookofwhy

7 9 19 @1:54am - Thanks you @PHuenermund for reading a "real life" study and showing us that it is made up
7.9.19 @1:54am Thanks you @PHuenermund for reading a real life study, and showing us that it is made up
of the same biological tissues as "toy problems", only more of them. #Bookofwhy

7.9.19 @12:20am - (1/ ) (Replying to @Jabaluck @fuzzydunlop123 and 7 others) Please do not distort my words.
You make it very unpleasant to interact with you when you do so. In our exchange I did not mention
"economists" but X-econs, namely, model-avoiding economists of the quasi-experimental school. Nor have I
mentioned DAGs, I spoke "models", which
7.9.19 @12:29am - (2/ ) (Replying to @yudapearl @Jabaluck and 8 others) includes structural economics.
Finally, I never said: "economist do things wrong". I said there is no way of knowing if X-econs do things wrong,
in the absence of ground truth, but it sounds very funny when a whole field prides itself on solving huge "real
life" problems, but
7.9.19 @1:07am - (3/ ) (Replying to @yudapearl @Jabaluck and 8 others) ONLY huge "real life" problems, not
"toy problems" which have ground truth and where everyone can see if you solved it correctly or not. Such
problems are avoided like a plague, labeled "toy", "made-up" and worse, but never discussed in good company. I
said "it sounds funny"
7.9.19 @1:14am - (4/ ) (Replying to @yudapearl @Jabaluck and 8 others) not in the hope of convincing you to
take a step back and see how funny it is, but in the hope of confirming the feelings of hundreds, perhaps
thousands of econ. students who I know are listening silently to this twitter exchange, perhaps after reading
#Bookofwhy in hiding,
7.9.19 @1:22am - (5/ ) (Replying to @yudapearl @Jabaluck and 8 others) and asking themselves 3 times a day:
Isn't funny that my professor can solve "real life" problems of such magnitude and importance, and he/she cant
tell which scenario (scenario! not DAG) contains a legitimate IV? I tweet here to tell this silent student: You are
not alone,
7.9.19 @1:38am - (6/ ) (Replying to @yudapearl @Jabaluck and 8 others) it is mighty funny, but it is going to
change in a few years, and you can prepare yourself for the day when your field will change from "funny" to
"well informed". Just make sure you spend the 2-3 hours it takes to acquire the art of causal modeling. I'll be
there for you.

7.8.19 @11:50pm - (Replying to @Jabaluck @fuzzydunlop123 and 7 others) Show me one simple case where the
exclusion requirement need not be justified. Now lets go to any "real life" problem and examine HOW it was
justified, if at all. I cant find a model (not a DAG, a model) in the papers you ask me to read, so how can one
tell if the reported
7.9.19 @12:03am - (2/ ) (Replying to @yudapearl @Jabaluck and 8 others) results are not biased substantially by
violations of exclusion? You know that there is objective validity test to the results reported, and you are asking
me to go through the numbers and show that they can do better with DAGs. The 4 scenarios we discussed tell us
more about
7.9.19 @12:11am - (3/ ) (Replying to @yudapearl @Jabaluck and 8 others) problems with exclusion that 100-
page "real life" study where those problems are not modeled (forget DAGs, by ANY model). If you think I could
learn something from a "real life" article, please tell me what principle I can gain from it that I may not know
already. #Bookofwhy

7.8.19 @11:09pm - (Replying to @Jabaluck @autoregress and 6 others) Retweeting: Sorry, these are not "toy
problems"; they haunt each & every IV exercise, albeit suppressed by practitioners to the mercy of intuition, so
as to escape "struggle". Each one represents 10,000 "real world" problems in which "exclusion" is/was justified.
#Bookofwhy

7.8.19 @9:28pm - (Replying to @Jabaluck @autoregress and 6 others) Wrong. I answered it head on: "econs use
too many in lieu of other approaches". I even proposed an explanation for this imbalance: X-econs avoid models
because they do not know how to handle them mathematically. Care to estimate of % of Angrist's students who
know how?#Bookofwhy

7.8.19 @9:11pm - (Replying to @Jabaluck @autoregress and 6 others) Sorry @Jabaluck, these are not "toy
problems", because they haunt each & every IV exercise, albeit suppressed by practitioners to the mercy of
intuition, so as to escape "struggle". Each one represents 10,000 "real world" problems in which "exclusion"
is/was justified. #Bookofwhy

7.8.19 @8:59pm - (Replying to @omaclaren @eliasbareinboim @stephensenn) I wish someone would plot the
frequency of the words "cause" or "causal" vs. time, from 1900 to 2019 and tell us, straight face, "No! There was
no causal revolution in 1990s". (There were 13 such papers in JSM 2003) Why are u resisting the idea of a
paradigm shift? #Bookofwhy

7.8.19 @8:29pm - (Replying to @omaclaren @eliasbareinboim @stephensenn) Statistics as a field is defined in


two arenas. (1) What its leaders say and do. (2) What its textbook say and do. For (1) I read the presidential
addresses in the past 20 years. For (2) I look at the index. #Bookofwhy

7.8.19 @8:24pm - (Replying to @omaclaren @eliasbareinboim @stephensenn) I dont think #Bookofwhy states
that 1)Statistics doesn't deal with causality. It quotes Fisher's definition: "Statistics is summarizing data," and it
decries the vacuum in stat textbook, and it places Fisher's DoE precisely where it belongs in history of CI. What
would you add?

7.8.19 @4:59pm - (1/ ) @autoregress I am not an extremist, so I am not worried about econs missing out on
complex DAGs. I am concerned about econs missing out on their simple IV models like the ones discussed here:
http://causality.cs.ucla.edu/blog/index.php/2019/06/01/graphical-models-and-instrumental-variables/ ... 4
scenarios, 4 variables and we havn't found an X-eco who
7.8.19 @5:28pm - (2/ ) (Replying to @yudapearl @autoregress and 6 others) 2/ was able to tell us which of the 4
scenarios has a legitimate IV. Bad sample? Perhaps. So let me ask: How many of Angrist's @metrics52 students
can do it? And if this does not jolt X-econs to do some honest soul-searching, what will? Their credibility is at
stakes. #Bookofwhy

7.8.19 @1:37pm - (Replying to @Jabaluck @autoregress and 5 others) @Jabaluck I believe you misunderstood
me. I have not used the words "back-door" in any of my tweets (to u). I used "sources of variation" which was
your term. My main point is: X-Eco refrain from using a model for two reasons: (1) The mistrust the assumptions
behind the model and
7.8.19 @2:00pm - (2/ ) (Replying to @yudapearl @Jabaluck and 6 others) (2) They mistrust models period, even
when they trust a set of assumptions. I am only concerned about reason (2). If I am wrong, I will convert back to
X-Eco. How can you prove me wrong? Show me one X-Eco paper that explicates and combines assumptions
formally #Bookofwhy

7.8.19 @1:32am - (Replying to @vijayant_k @Canadian_JACD and 2 others) We are concerned with the effects of
BP not with its causes. If it has an effect on cardiac stress we should be allowed to include the arrow BP---
>Cardiac Stress in our model, and not to ask for supreme court permission.

7.8.19 @12:57am - (1/ ) (Replying to @mc_hankins @stephensenn @eliasbareinboim) 1/ Since we started


working on external validity, transportability and data fusion in 2010 (eg, https://ucla.in/2N7S0K9 ) we have
been hearing the whisper: "This sounds like meta-analysis." Yes we are yet to find a meta-analyst expert who
can to tell us how to handle the simplest
7.8.19 @1:07am - (2/ ) (Replying to @yudapearl @mc_hankins and 2 others) example. (and I went to the very
top). Why? Because we can show that the answer depends on the causal relationships between X1,X2,Y and
other factors in the problem. Meta-analysis is a statistical pooling method that is oblivious to those
relationships. Therefore, I take the
7.8.19 @1:17am - (3/ ) (Replying to @yudapearl @mc_hankins and 2 others) liberty to assume that meta-
analytic efforts are orthogonal to the kind of problems we are trying to solve. However, I will immediately
change my mind if I find a Meta-Analyst who solves the examples of https://ucla.in/2N7S0K9 or Fig. 1 of
http://bit.ly/2Xu3uMP #Bookofwhy
7.7.19 @11:02pm - (Replying to @yudapearl @Jabaluck and 6 others) 4/ explicitly what I understand about each
of those sources separately, and try to combine those understandings mathematically to decide if Z is a good IV.
I was tempted to do (2) but, then, I realized to my horror that I am re-committing DAG-heresy. What shall I do?
#Bookofwhy

7.7.19 @10:36pm - (1/ ) (Replying to @Jabaluck @MariaGlymour and 5 others) @Jabaluck your arguments are so
persuasive that I decided to quit DAGs and convert to X-Eco (short for Experimental Economist). The moment of
conversion was truly enlightening and, as if by divine revelation, I began to "understand the sources of
variations" of variables,
7.7.19 @10:44pm - (2/ ) (Replying to @yudapearl @Jabaluck and 6 others) something DAG folks never
understand. Fired by enlightenment, I chose Z as my potential IV, and I understood the "source of variations" of
many other variables, X,Y,Z,W,S, T, V, all related to Z, to X and to Y. I truly understood those "sources of
variations, and I felt elated
7.7.19 @10:51pm - (3/ ) (Replying to @yudapearl @Jabaluck and 6 others) Now I had to decide whether Z, which
looked like a promising IV before, is still a good IV given what I understood about all the other background
variables. I had two options: (1) Decide yes or no based on my understanding of all those sources of variations
or (2) articulate

7.7.19 @10:22pm - (1/2) Sailors and passengers on this voyage of CI research should be interested in these new
results that just reached my screen, https://bit.ly/2Xu3uMP , by @eliasbareinboim & team. Suppose we
randomize treatments X1 and X2 in two separate studies, can we estimate the causal
7.7.19 @10:22pm - (2/2) effect of their conjunction (X1=x1, X2=x2) ?? I see dozens of pharmaceutical companies
rushing to join our voyage. #Bookofwhy

7.7.19 @9:39pm - (Replying to @eliasbareinboim @EpiEllie and 2 others) I am willing to adopt the new
construction method if it helps science. But we are still in need of a rule on when to include an arrow BP--->Y
and when not to, assuming that we sear never to ask for the effect of BP on anything. What should we think
about to decide? #Bookofwhy

7.7.19 @8:56pm - (1/ ) This is a brilliant idea, to use the causal hierarchy in "critical thinking" classes: Given an
English expression, classify it into rung-1, 2 or 3. Some of the top researchers of our time should take it. I
remember when our kids were young we got them a game called "propaganda"
7.7.19 @8:56pm - (2/ ) Each card had a false argument and players had to classify it into one of several types of

falsehood. Great!! As to a simpler/gentler version for high schoolers? The #bookofwhy is all we have, but it
should be fun to do one, with a creative cartoonist. Great!!

7.7.19 @5:59pm - (Replying to @yudapearl @raymondshpeley @f2harrell) certain aspect of the distribution, eg
P(label|features), which is no more of model than say the estimated regression coefficient R, fitted to a cloud
of samples. Would you call R a model? Does it carry any assumptions? Only if you havn't seen any of the data.
#Bookofwhy

7.7.19 @5:51pm - (Replying to @raymondshpeley @f2harrell) Mathematically, models are carriers of


assumptions. Stat models carry assumptions about the distribution, eg normality, binomial,.. and causal models
carry assumptions about causal relationships, eg. who listens to whom. A learned neural net provides efficient
representation to a

7.7.19 @4:50pm - Answer to What are the differences between the "experimentalists" and "structuralists"
approaches to econometrics? by Judea Pearl

7 7 19 @2 57 (R l i t @ ij t k @C di JACD d 2 th )A di t ll t I t lk t BP
7.7.19 @2:57pm - (Replying to @vijayant_k @Canadian_JACD and 2 others) According to all experts I talk to, BP
is not just an indicator of some risk factor, but it actually causing bad things, eg. heart stress. We do not want
to go to extremes and claim that we are only seeing the display on the measuring device, not the BP itself.
#Bookofwhy

7.7.19 @12:44pm - (Replying to @f2harrell @raymondshpeley) A point which I should have added (based on
@f2harrell tweet) is that whereas statistics can give us an (estimated) distribution, DL gives us one aspect of
the distribution, eg P(label|features). DL folks may argue that, in principle, they can learn the P as a function
X-->(0,1)

7.7.19 @12:33pm - (Replying to @MariaGlymour @PHuenermund and 5 others) Who sold us "empirically
validated" ointment? When I say a "valid instrument" I mean one that satisfies the the IV requirement according
to the model that was proposed. Plus, that model may have testable implications.

7.7.19 @5:48am - (1/2) (Replying to @PHuenermund @Jabaluck and 5 others) I guess what the experimentalists
are arguing is as follows: If you are committed NOT to put down any model on paper, and to work purely by
intuition, then it is easier to start by asking yourself: "Is there an exogenous variable around that is somehow
related to X and Y"? as
7.7.19 @5:58am - (1/3) (Replying to @yudapearl @PHuenermund and 6 others) as opposed to asking yourself: "Is
there a way to control ALL confounders of X and Y?". Next comes the handling of "somehow related". If you are
still committed to model-blind thinking, you will have to intuit handwaving about the exclusion restriction,
which experimentalists
7.7.19 @6:07am - (3/4) (Replying to @yudapearl @PHuenermund and 6 others) are willing to stomach given the
ease of having found SOME relevant exogenous variable in the mental forest. What they forget is that there is a
middle ground: Find your favorite exogenous variable, model what you know about the relationships with X,Y
and other factors, then
7.7.19 @6:19am - (4/5) (Replying to @yudapearl @PHuenermund and 6 others) use your modeling tools to
decide if you have a valid IV, if not, can you repair it, can you find another, can you test your assumptions etc
etc, all the nice thing that models give you, rather than remaining in the intuitive world. The trade off will be
settled only when
7.7.19 @6:28am - (5/5) (Replying to @yudapearl @PHuenermund and 6 others) experimentalists agree to learn
how to model problems with 5-10 variables, and see for themselves what they have missed by resisting it.
#Bookofwhy

7.7.19 @4:51am - (Replying to @pentagoniac) Evidently I violated some cosmic rule of Quora protocol. No idea
what was violated nor who vowed for my innocence.

7.7.19 @4:16am - Victory! My appeal was accepted. My answer on "econometrics, statistics, and machine
learning" was reinstated on Quora, and my honor restored. I bet they were impressed by the clean life I live.
#Bookofwhy

7.6.19 @7:51pm - (Replying to @mattshomepage) It is a mystery which I tried to answer in chap. 10 of


#Bookofwhy. Millions of years of playful experiments with bows and arrows are encapsulated in our culture, our
language, our elderly wisdom, our books etc. The practical question is: how to represent it and how to exploit
it.

7.6.19 @7:28pm - (Replying to @y2silence @PWGTennant and 3 others) No one knows what they say. All we hear
is: If A is not randomized (or randomizable) then "it all depends".

7.6.19 @4:49pm - (Replying to @imleslahdin) Answer deleted!! Thanks for telling me. I guess the vice-squad just
found out about me. I appealed. Lets see how they handle "appealers".
7.6.19 @4:18pm - (1/2) (Replying to @AngeloDalli @EpiEllie and 2 others) An arrow BP--->Y has two
interpretations: (1) Y listens to and responds to changes in BP. (2) Manipulating BP changes Y. (1) implies (2),
but some may argue that (2) is all we can observe, hence it is "scientific" while (1) is "meta-physics". Now, since
BP is not manipulable,
7.6.19 @4:26pm - (2/2) (Replying to @yudapearl @AngeloDalli and 3 others) adding an arrow BP--->Y to your
model makes you suspect of membership in the "meta-physics" camp, which carries harsh consequences. This is
why the BP issue is foundational, and that is why we see such hesitations and "it all depends" from the
manipulationist camp. #Bookofwhy

7.6.19 @4:04pm - (Replying to @f2harrell) Do you think statistics offers more than "taking us from samples to
properties of distribution functions" ?? If you do, then I confess to underplaying the role of #statistics. But did I?
#Bookofwhy

7.6.19 @2:16pm - (Replying to @neuro_data @danilobzdok) Progress in plain geometry was very very advanced
before someone said: lets add a 3rd dimension. If we want to remind people that the world is 3-dimensional it
is helpful, I think, to label 3D-geometry "advanced".

7.6.19 @2:07pm - (Replying to @yudapearl @JaapAbbring) Here is a brilliant idea. As co-Editor of the Journal of
Causal Inference I will invite you to tell our readers how they can benefit from tools developed in Econ. and you
will reciprocate by convincing a mainstream Econ. journal to do likewise. Do we have a deal? #Bookofwhy

7.6.19 @1:59pm - (Replying to @yudapearl @JaapAbbring) involves graphical models, do-calculus,


counterfactual logic etc. which do not rule out tools produced in econ. to deal with part (1). I see wide
awakening on both sides to examine and evaluate each other tools. But it is only a Twitter-awakening, not in
journals or NBER. Wait!!

7.6.19 @1:49pm - (Replying to @JaapAbbring) The words "relatively simple" are Imben's words, with which he
justifies why he can avoid graphical models. See http://causality.cs.ucla.edu/blog/index.php/2014/10/27/are-
economists-smarter-than-epidemiologists-comments-on-imbenss-recent-paper/ ... SCM has two parts: (1) a
model of reality and (2) tools for dealing with (1). The first is identical to structural economics. The second

7.6.19 @1:29pm - (Replying to @danilobzdok @f2harrell) Neural network is a man-made artifact, crafted to
capture some aspects of the data, say the relationship between symptoms & disease. By a "model" we mean a
picture of reality, from which the data emerged. e.g. X--->Y and X<---Y are two different models that may yield
same NN.

7.6.19 @1:21pm - (Replying to @f2harrell) I agree. Most neural networks do not involve models of the data-
generation process. Agree. I was afraid you have found points of disagreement.

7.6.19 @2:36am - (Replying to @robertwplatt @Canadian_JACD and 3 others) Would adding or not adding an
arrow from BP change whatever you are interested in doing? If it does, Eureka! We just discovered the scientific
meaning of an arrow going out of nonmanipulable variable. Now we ask: How would you decide whether to add
or not to add? #Bookofwhy

7.6.19 @1:49am - (1/2) (Replying to @JaapAbbring) I do not recall describing econometric models as "relatively
simple," especially not structural econometrics, which can be viewed as a subset of SCM, lacking graphical tools
and do-calculus. I may have used the phrase "relatively simple" to describe the kind of problems
7.6.19 @1:58am - (2/2) (Replying to @yudapearl @JaapAbbring) that can be handled by IV methods Angrist
style, where no model is laid down for analysis and scrutiny, or PO methods Imbens-Rubin style, where
style, w e e o odel s la d dow o a alys s a d sc ut y, o O et ods be s ub style, w e e
ignorability assumptions beg to be discerned. Structural models + graphical tools is a winning combination
#Bookofwhy

7.5.19 @11:49pm - (Replying to @davidmanheim @kareem_carr @EpiEllie) True, basic econometrics was meant
to do CI, but look at the tools they are still using in the 21st century, not even realizing what tools are missing
from the Econ-Echo-Chamber. As to DNN's etc. they were "advanced" five years ago; times they are achangin.
#Bookofwhy @causalinf

7.5.19 @10:32pm - (Replying to @dkipb @EpiEllie and 3 others) I have not received comments on "Q as a limit"
or on https://ftp.cs.ucla.edu/pub/stat_ser/r486-reprint.pdf ... (which is a beautiful paper, no bias), so we do
not know if it helped resolve the hesitations. But "first pass" is a cop out; this is a foundational issue: Listening
vs. manipulating. #Bookofwhy

7.5.19 @10:00pm - (Replying to @PWGTennant @EpiEllie and 2 others) God forbid. What I can or cannot obtain
has nothing to do with the meaning of an arrow coming out of BP, which simply says: someone is listening and

responding to BP. I have not heard Ellie&Miguel say "yes", perhaps b/c the notion of "Well-defined" is not well
defined? #Bookofwhy

7.5.19 @9:51pm - (Replying to @PWGTennant @EpiEllie and 2 others) Disagree. I believe the inconsistencies and
hesitations of the "Not-Well-Defined" school have created more confusion than usefulness. All in the name of
mimicking RCT's. #Bookofwhy

7.5.19 @9:44pm - (Replying to @PWGTennant @EpiEllie and 2 others) Identification should not mar questions of
meaning. And would stay away for obesity because it is also ill-measured. So its better to focus on BP, which is
well-measured and still debatable (at least in some circles). #Bookofwhy

7.5.19 @9:40pm - (Replying to @PWGTennant @EpiEllie and 2 others) I wish I could, like you, agree with both
sides. But I still do not know what one side says about an arrow out of BP. When is it legit?

7.5.19 @9:33pm - (Replying to @melb4886 @EpiEllie and 2 others) By saying "because alcohol raises BP" you just
admitted the presence of an arrow out of BP. Did you get a license from the guardians of Well-Definedness.?
They are still debating if/when such an arrow violates scientific right and wrong. #Bookofwhy

7.5.19 @9:28pm - (Replying to @vijayant_k @Canadian_JACD and 2 others) But if BP is merely an indicator for
some other risk, we should not see an arrow going out of BP. But in all DAGs presented here we did see this
arrow. Can you check with the experts? Still, the fact that we ask this question means that drawing such arrow
make a difference.

7.5.19 @9:10pm - (Replying to @vijayant_k) Last I heard from ML folks was they are all working on "combining CI
and ML". I heard if even from people who have no idea what CI means (names withheld). So why would anyone
object to be working on "advance ML"? #Bookofwhy

7.5.19 @9:03pm - (Replying to @s_monterohdz) "Advance" does not rule out "super-advance" in the future. I
believe I did define the scope of SCM in my writing in "real phenomena", Each DAG you draw represent millions
of "real phenomena" that fit the structure. #Bookofwhy

7.5.19 @8:56pm - (Replying to @ADAlthousePhD @EpiEllie) Thanks, I'll try to reply, instead of retweet.
7.5.19 @1:26pm - (Replying to @kareem_carr) Good point. I have been trying hard to have statisticians embrace
CI and call it "advanced statistics" https://ucla.in/2v72QK5 watch where they are. As to econometrics,
intervention was the goal of even "elementary econometrics,", only the tools that are lacking. #Bookofwhy

7.5.19 @1:10pm - (Replying to @f2harrell) Let's hear ONE disagreement. We settled the previous ones, we can
do same with this ONE.

7.5.19 @7:21am - Answer to What are the differences between econometrics, statistics, and machine learning?
by Judea Pearl

7.5.19 @2:17am - (Replying to @EpiEllie) We are almost getting there. Quoting: "If.....then we can ask causal
questions involving THAT INTERVENTION". But you did not say: "...causal questions involving Blood Pressure".
Was it on purpose or by oversight? Moreover, would you then allow an arrow from BD? @Bookofwhy

7.5.19 @1:03am - (1/2) (Replying to @autoregress @EpiEllie) To interpret what @MariaGlymour wrote I would
add: g-methods become valid "methods" only after you have a DAG to help you decide what variables to
condition on, as in the back-door condition. Additionally, having or not having an IV is also a task dedidable by a
DAG. Further,
7.5.19 @1:16am - (2/2) (Replying to @yudapearl @autoregress and 2 others) deciding whether a valid "g-
method" exists (namely if identification can be done by regression" takes one glance over a DAG, it takes
forever using ignorability assumptions as those used by Rubin, Imbens and Angrist. Lastly, I would not dismiss
do-calculus #Bookofwhy

7.4.19 @11:07pm - (Replying to @EpiEllie) So, the answer is YES. We can include BP in the DAG, with an arrow
going out of BP, since we do not require that BP be directly manipulable. It is enough that there are Well-
defined interventions someplace in the model. Right? Can I tell this to my students? #Bookofwhy

7.4.19 @10:49pm - (Replying to @yudapearl @Canadian_JACD and 3 others) So, by all means, please post more
and more DAGs with arrows emanating from BP. The more you post the less credible would their claim become,
of the "new philosophy", that they are doing "experimental science" whereas you are doing "metaphysics".
#Bookofwhy

7.4.19 @10:34pm - (Replying to @Canadian_JACD @EpiEllie @BL4PublicHealth) The "new philosophy" is the new
school of "no causation w/o manipulation" that is brewing at Harvard (perhaps other places?) according to which
you cannot talk about X causes Y unless X is manipulable. See @_MiguelHernan papers and
https://ucla.in/2LY8M2X for full view #Bookofwhy

7.4.19 @9:30pm - No No, please, continue. You are giving us straight answers: "blood pressure can be an
important cause of things like the thickening of arteries". Please continue, because the very notion of "BP can
be a cause" is under the danger of extinction, in the new philosophy #Bookofwhy

7.4.19 @9:18pm - I was waiting for this answer "the role of blood pressure in the causal question". It is the kind
of answer @EpiEllie could not give, because it violates the dictates of RCT imitation. BP has no well-defined
"role", only interventions have "roles". See https://ucla.in/2LY8M2X

7.4.19 @9:07pm - (1/2) Not so easy. So the answer is YES. After thinking, weighting, assuming, understanding,
specifying, etc. there might come a moment where you would add BP to the DAG. Now innocent me asks: why?
Who needs this ill-defined entity there? Or, what can we do after adding it that
7.4.19 @9:07pm - (2/2) we couldn't do without, with all the thinking, weighting, assuming, understanding,
specifying, etc. that gave us the license to add it? After all, if our research question is well-defined: "the effect
of some drug", why do we need all this BP nuisance? Who cares? #Bookofwhy

7.4.19 @8:34pm - Not a simple question? OK, I will make it MUCH simpler: Is there any circumstance, after you
do the thinking, thinking and more thinking, that you would add "blood-pressure" to your DAG? #Bookofwhy

7.4.19 @7:44pm - (Replying to @JadePinkSameera) It is a good question. And a very simple one too. And that's
what makes it so hard to answer.

7.4.19 @7:22pm - "Are economists smarter than epidemiologists?" Our recent Twitter posts made me re-read the
blog discussion we had in 2014: http://causality.cs.ucla.edu/blog/index.php/2014/10/27/are-economists-
smarter-than-epidemiologists-comments-on-imbenss-recent-paper/ ... which I still find to be illuminating of the
interplay between cultural and technical forces in scientific progress #Bookofwhy

7.4.19 @7:02pm - So, the simple answer to my innocent question is: NO. @EpiEllie and @causalinferenc and
@_MiguelHernan would NOT include a variable "blood pressure" in a DAG before deciding how to weigh all the
well-defined interventions on diet, drugs, exercise, etc, etc Am I right? #Bookofwhy

7.4.19 @6:29pm - (1/2) (Replying to @kareem_carr @jenniferdoleac and 2 others) @kareem_carr , you have
infinite memory. You just confirmed my last Tweet. Many readers will resist my theory that an innocent
curiosity of just two individuals can account for such profound differences between two disciplines. Many will
seek differences in substance or
7.4.19 @6:39pm - (2/2) (Replying to @yudapearl @kareem_carr and 3 others) philosophy, or type of data etc
etc. No way. It is as simple as your Tweet. Escaping from the echo-chambers of social bubbles is the strongest
force that drives scientific progress. And it is becoming harder and harder in the age of internet and social
media. #Bookofwhy

7.4.19 @6:15pm - (1/2) On the difference between Econ and Epi, I also said there are no substantive differences
whatsoever between the two: http://causality.cs.ucla.edu/blog/index.php/2014/10/27/are-economists-
smarter-than-epidemiologists-comments-on-imbenss-recent-paper/ ... The current differences in practice
emerge from one fluke of history: Robins and Greenland were epidemiologists, not economists. Thus,
7.4.19 @6:15pm - (2/2) they followed curiosity and asked: what can DAG do for us? The rest is history. Why
didn't economists follow their curiosity? Let others answer it because I do not want to spoil it for the many
econs who ask this very question today. @causalinf , @jenniferdoleac #Bookofwhy

7.4.19 @5:31pm - All these complications need consideration, I agree. But let's ask an innocent question that
comes before complications strike: "Can we assume that the variable "blood pressure" may appear in one of your
DAGs and, when it appears, there is a arrow going out of it?"#Bookofwhy

7.4.19 @5:18pm - (Replying to @_gbmari) This self-fulfilling effect should be balanced against the spite-him
effect of employees that are driven by the challenge to "prove him wrong". We agree I hope that both
motivations demand causal models and cannot be accounted for by statistical considerations alone. #Bookofwhy

7.4.19 @4:07pm - Am I right to assume that the variable "blood pressure" may appear in one of your DAGs and,
when it appears, there is a arrow going out of it, into other variables. Can we assume that much? #Bookofwhy

7.4.19 @2:02am - Phelps aside, do you think economists today are equipped to define+manage issues of

"fairness" with their current tool set? #Bookofwhy


fairness with their current tool set? #Bookofwhy

7.4.19 @1:23am - Your paper is an eye opener, glad you posted it. If it was not 1am I would have continued
reading it, for it is written in a CI language and a well-structured style. The fact that it won @SIGMOD2019 best
paper award may signify a new age for fairness folks. #fairness #Bookofwhy

7.4.19 @12:20am - I do not doubt that adequate causal defs of discrimination and fairness can be constructed
by CI folks; they know causal models and counterfactual logic. I questioned the readiness of mainstream ML
folks who are lacking those tools. Same goes for economists. #Bookofwhy

7.3.19 @10:35pm - (1/2) Yes, structural counterfactuals escape the torture chambers of the RCT's imitation-
game. I think epidemiologists too are on their way to escape those chambers; how else can we interpret the
arrows emanating from "blood-pressure" -- a well-measured yet
7.3.19 @10:35pm - (2/2) ill-manipulated variable that certainly has "effects" and often finds itself in DAGs
drawn by epidemiologists? @EpiEllie @_MiguelHernan @Lester_Domes @MariaGlymour A puzzle. #Bookofwhy

7.3.19 @7:12pm - Thanks for re-posting, Brooke, and let our mantra for July be: "Make Zionophobia the ugliest
word in town"

7.3.19 @7:03pm - For the next step, after #Bookofwhy, I will continue to recommend the PRIMER
https://ucla.in/2KYYviP , https://ucla.in/2G2rWBv , as long as it finds no match in clarity, examples and
philosophy. Totally liberated from RCT's and other hangups, and essentially free. #Bookofwhy

7.3.19 @6:54pm - (1/3) Economists' "statistical discrimination," as it turns out, is both (1) the use of statistical
associations to discriminate and (2) an attempt to define discrimination using statistical vocabulary alone.
According to Phelps (1972), you discriminate whether you hire people by
7.3.19 @6:54pm - (2/3) education, race or zip code, as long as you base decisions on PREDICTED performance,
rather than performance itself. So, all learning is "discriminatory" for it is based on past experience which
PREDICTS, yet is not equal, the situation at hand. Conclusion: I was right
7.3.19 @6:54pm - (3/3) to suspect criteria entitled "statistical discrimination" and their ability to capture
notions such as "fairness," in which causal relations play a major role. @JaapAbbring @steventberry #Bookofwhy

7.3.19 @6:24pm - (Replying to @swadhin_pradhan @geomblog) I dont recall giving such tutorial. Do you have a
source?

7.3.19 @5:44pm - The counterfactual framework with which I am familiar (eg https://ucla.in/2G2rWBv ) needs
no interventions to have "effects". The paper you cite represents the "potential outcome" framework, a relic of
an age when causation was forced-married to RCT (or imitations of). #Bookofwhy

7.3.19 @4:20pm - Statistics itself came from causal motivation for, surely, good predictions are essential for
good decision (eg carry an umbrella). The missing ingredient is getting confidence to let these motivations out
of the closet and articulate them mathematically #Bookofwhy https://t.co/C0tkz1jtj9

7.3.19 @3:31pm - I was under the impression that the issue of "manipulativity" was settled by explicating the
non-manipulative aspects of "effects". Examples are: https://ucla.in/2NnfGPQ#page=31 ,
https://ucla.in/2LY8M2X . #Bookofwhy

7.3.19 @3:02pm - Footnote (1) in this paper (https://arxiv.org/abs/1810.08810 ) confirms your point and
explains why causal vocabulary is not at the "Frontiers" of Fairness research; it takes a generation to undo the
statistical thinking that rules ML textbooks, classrooms and research practices.#Bookofwhy

7.3.19 @2:51pm - I cannot see how causal inference could "miss" injustices that statistical inference discovers,
when the former represents reality and the latter a silhouette of reality, as projected in data. Example??
#Bookofwhy

7.3.19 @2:43pm - (Replying to @Adelaee) You made my day, Adelaee. The thought that your teenagers are
reading my words compels me to write more, to make Zionophobia the ugliest word in town.

7.3.19 @1:48am - (Replying to @_onionesque @geomblog) I join you in this optimism. I have even blessed the
pre-scientific hype for channeling good people to study causal and counterfactual models, then turn hype into
science. It is happening, agree, because the tools for defining "fairness" correctly are available.#Bookofwhy

7.3.19 @12:48am - Thanks for correcting me on what economists mean by "statistical discrimination". I was not
critical of the argument, but of the title, which alerted me to a possible new oxymoron, along with
"probabilistic causality", "statistical confounding" "statistical mediation"..#Bookofwhy

7.3.19 @12:37am - (Replying to @NeuroStats) The first paper you cite is indeed the first I have seen on the right
track. And I wish ML-folks will take notice. The second is an IV ACE-estimator, a rung-2 exercise, which cannot
capture the counterfactual nature of "discrimination" (rung-3). #Bookofwhy

7.3.19 @12:19am - (1/3) If by "statistical discrimination" we mean the use of statistical associations in the
decision, then it is perfectly harmonious with the causal definition of "discrimination" and I see nothing wrong
with it. What makes me suspicious are attempts to DEFINE discrimination by
7.3.19 @12:19am - (2/3) statistical criteria. Note that in your example, causal considerations are unavoidable,
for if the observed characteristics causally produce/prevent necessary skills, our notion of discrimination would
change. This sensitivity to causal relations is absent from the fairness
7.3.19 @12:19am - (3/3) literature I have sampled thus far, and which seems to dominate recent discussions in
ML. Note also that we now have the tools to define and manage criterion based on combined statistical+causal
relations. BTW, the link you gave us is blocked. #Bookofwhy

7.2.19 @11:30pm - I've never dismissed an argument in which statistics and causality appear together. I am
suspicious however of arguments that attempt to define inherently causal notions (eg discrimination) in terms
of statistical vocabulary ALONE, void of causal relations. Wouldn't U? #Bookofwhy

7.2.19 @7:44pm - The title "statistical discrimination" worries me, because "discrimination" is a causal, not
statistical notion. The words "from economics" makes me doubly worried, for reasons explained in #Bookofwhy.
But I would love to hear how statisticians define "fairness". Truly curious.

7.2.19 @7:34pm - (Replying to @ahmaurya @alexdamour and 8 others) So how can we explain how FATE-
motivated folks can study, write and speak FATE without applying some of the tools that CI is offering the ML
community? For example I cant find the tools of Attribution Analysis and Causal Mediation invoked, Or am I
missing them? #Bookofwhy

7.2.19 @5:12pm - (Replying to @eschisterman1 @AmJEpi) Congratulations!! Enrique.

7.2.19 @4:00 - (1/2) @marypcbuk @katecrawford @eliasbareinboim @benedictevans All definitions and
examples of "fairness" that I have seen revolve around causal and counterfactual considerations, since they
concern EFFECTs of policies on different segments of society. Thus, I venture to predict

7.2.19 @4:00 - (2/2) today's "fairness" hype will channel good people to study causal and counterfactual models,
and help turn hype into scientific disciplines that define and algorithmitize the kind of fairness we wish AI
systems to exhibit. There is virtue in pre-scientific hype. #Bookofwhy

7.2.19 @3:07pm - Agree. With two points of caution:(1) We need to study carefully the theoretical impediments
to automated generation, to avoid falling for fake-gold, and (2) We need to study carefully what we can and
cannot do with a causal model once we have it. The rest is just math #Bookofwhy

7.2.19 @6:38am - Thank you. I begin to understand. "Bias" is used here in the social sense, like discrimination. I
am relieved. In our corner of the wood "bias" has technical meaning, standing for systematic deviation from
expectation. eg confounding bias or "selection bias". Relieved #Bookofwhy

7.2.19 @6:26am - Interesting indeed, because I normally find people complaining of deep learning being
opaque, here is one complaint: https://ucla.in/2wj4pox . Optical illusions are known to be mysterious.
#Bookofwhy

7.2.19 @6:20am - I am not disputing the article, just expressing surprise at a term that I have not heard before.
Can I conclude then that people who talk "AI Biases" mean the same thing as "limitations and biases of
correlational ML." like those attributed to Rung-1 of the Ladder in #Bookofwhy?

7.2.19 @6:11am - Replying to @marypcbuk I couldn't pass the pay wall, but your myths are ML myths, which I
can understand. I am still to understand what "AI bias" is which is not a myth. Are any of your 9 myths
remediable by any of the 7 tools here: https://ucla.in/2HI2yyx

7.2.19 @5:20am - (1/2) I did not realize that people call these problems "AI-bias". Thanks for bringing it to my
attention and to the attention of my Twitter followers. Honestly, I have been swimming in the AI pool since the
1970's and never heard the term "AI bias" used before, especially not to
7.2.19 @5:20am - (2/2) describe problems that AI folks are about to solve in the near future. But, as they say in
the Talmud: "Never too late to learn". Now, suppose I want to define the term on Twitter. Is every difficulty
encountered by an AI program an "AI-bias"?? Which ones are not? #Bookofwhy

7.2.19 @4:36am - (Replying to @pentagoniac @benedictevans) I would love to see an example of a causal
problem (say Simpson's paradox) that "*can* be corrected for without causal models. #Bookofwhy

7.2.19 @4:33am - (Replying to @benedictevans) But by saying "AI-bias" you give people the impression that such
biases are inherit to AI, namely, permanently irredeemable by any AI program. Do we want to give general
audience this impression?

7.2.19 @2:45am - I read @benedictevans on "AI-bias" and I still do not know what he means by "AI-bias", why not
call it "ML-bias" or "curve-fitting bias" and how those biases can be avoided w/o attending to causal models as
outlined here https://ucla.in/2HI2yyx or here #Bookofwhy? A ML puzzle.

7.2.19 @2:22am - (Replying to @zacharylipton) I dont exactly understand what the time-management problem
is. My students only worked on things that other people said are impossible, so mentoring was part of speaking
to colleagues. I am not finished though, some colleagues still say they can do astronomy without telescopes.

7.2.19 @1:59am - (1/2) Nice and concise summary of #Bookofwhy. If I were forced to make a critical comment,
7.2.19 @1:59am (1/2) Nice and concise summary of #Bookofwhy. If I were forced to make a critical comment,
it would be the way it tries to stimulates readers interest by pointing to existing interest. What if no ML
conference had any session on causation, would that make the ideas of #Bookofwhy
7.2.19 @1:59am - (2/2) less compelling for enlightened ML folks trying to build intelligent systems? #Bookofwhy
was written to change, not to follow habits. I hope it does.

7.2.19 @12:05am - (Replying to @mimblewabe) I considered changing "do" to avoid criticism like "You cant do
this" or "some 'do's change everything". But I decided against it b/c there was no sub. We, Sapiens, do not
distinguish between reality and models of reality. Even Abraham asked "what if there were 50...". #Bookofwhy

7.1.19 @11:56pm - Thanks for adding your smiles to this memorable event.

7.1.19 @11:52pm - For something more mathy, I will continue to recommend the PRIMER
https://ucla.in/2KYYviP until something better shows up in the jewelry store. Note that it is now (essentially)
open-accessed. #Bookofwhy

7.1.19 @11:22pm - (Replying to @causalinf @dlmillimet and 2 others) Astronomy will never be the same.

7.1.19 @11:19pm - (Replying to @IndexLlc @IARPAnews) I saw the abstracts and pre-BAA material, and wrote to
the program manager for more information on the "counterfactual prediction" project. Awaiting his reply.

7.1.19 @11:12pm - DAGs are like optical lenses. First you use them as spectacles, to see things you already
know, then you try them as telescopes and microscopes, to see things you never knew existed. Good luck my
fellow econs; astronomy will never be the same. #Bookofwhy.

7.1.19 @12:22pm - (1/2) Mehdi Hasan's embarrassment will not end here. Once the conversation shifts to
discuss the right of Jews to a homeland, he will be facing a dilemma: (1) To be honest and declare (like
barghouti) that Jews are not a people or (2) To say that Jews, if well behaved, are entitled
7.1.19 @12:22pm - (2/2) to some semblance of sovereignty in the land. Theoretically, option (2) would get him
off the hook but, unfortunately, he can't lose his support base of Arab rejectionists for whom such an admission
amounts to a betrayal of 120 years of bloody wars and uncompromising denial.

7.1.19 @6:22am - The do-operator is not limited to physically "doing". It is an operation on your model of reality
and it informs others about your model and about physical interventions that are feasible. On the interpretation
of do(x), see: https://ucla.in/2LY8M2X #Bookofwhy

7.1.19 @2:40am - (Replying to @LARichwine @NimaCNN @LAPressClub) Thanks @LARichwine for immortalizing
this moment of grace. And thank you @NimaCNN for being part of a most memorable evening, and for honoring
our son Daniel with your courage and integrity.

7.1.19 @2:32am - This quote, which sounds obvious in today's standards, was much debated in the 1980's, when
AI folks labored to find proper formalisms to represent uncertainty in expert systems. This paper
https://ucla.in/2RetYPS uses data insufficiency to expose a weakness in belief functions.

7.1.19 @1:48am - (1/2) We should all live to see Mehdi Hasan's face the first time he is called "Zionophobe" in
front of an audience. On the one hand he is proud of being anti-Zionist, so he cannot deny the charges. On the
other hand he will resist the analogy with Islamophobia which would force
7.1.19 @1:48am - (2/2) to philosophize on what "identity" is, be it religious, national or historical which, again,
will steer the conversation to where we can win hands down: the moral imperatives of Zionism and the racist
deformities of Zionophobia - the ugliest word in town.
deformities of Zionophobia the ugliest word in town.

7.1.19 @1:27am - (Replying to @EWilf @intelligence2) @EWilf , The title should have warned you of what the
organizers tried to achieve: clearance from charges of AS. I stopped participating in such debates unless they
change the title to: "Is Zionophobia racism?" or "Zionophobia on trial".
https://jewishjournal.com/culture/lifestyle/first_person/300661/inspiration-and-a-rallying-cry-for-graduates/
... "AS-- NO MORE"

7.1.19 @12:30am - (Replying to @zacharylipton) My goodness!! I never thought of becoming an icon of longevity,
no way, playfulness yes, but longevity? Not in my worst dreams. Wait a minute! Perhaps playfulness is the
secret to longevity? But isn't the latter simply a means to getting the former in greater quantity?#Bookofwhy

6.30.19 @2:58pm - (Replying to @IndexLlc @IARPAnews) Thanks for posting. I was not aware of this program and
it is hard to tell from the BAA text whether its authors are aware of the fact that counterfactuals have been
tamed, domesticated and algorithmitized, as in here https://ucla.in/2NceWNb . We need to find out.
#Bookofwhy

6.30.19 @2:47pm - (Replying to @FJnyc @questionsin2014 @AshagerAraro) I share your suspicions. Zionophobic
Jews-by-birth deny Jews rights that they grant to any other collective -- the right to define themselves. Does
this make them racists? Painfully so!

6.30.19 @2:35pm - (Replying to @bnaibrithcanada) Thank you, @bnaibrithcanada for retweeting my speech to
your followers. I am confident that, with your help, Zionophobia will become the Ugliest Word in Town. ps. Is
there a Bnai Brith US? I wish they join us, in the trenches.

6.30.19 @2:34am - A transcript of a graduation speech I gave at UCLA last week is now posted on line:
https://jewishjournal.com/culture/lifestyle/first_person/300661/inspiration-and-a-rallying-cry-for-graduates/
... It explains what Jewish students are rallying for, and the kind of changes your campus will hopefully see
next academic year.

6.29.19 @11:09pm - (Replying to @ShalitUri) Gee, I did not know about the Miao etal paper, perhaps because I
was already immersed in #Bookofwhy. God bless Twitter for keeping us updated.

6.29.19 @3:39pm - (1/3) This paper identifies causal effects by proxies, a task shown feasible here
https://ucla.in/2N5GNOK and here https://ucla.in/2N9icIX . Blocking back doors is a sufficient condition for
identification, not necessary; it turns out that, under certain circumstances, a proxy
6.29.19 @3:39pm - (2/3) can replace a blocker Z. Going from ATE to ITE ("individual" treatment effect) is not
really hard (theoretically) if by "individual" we mean "c-specific effects" where c is a set of characteristics
marking the individual, namely X=x. The nice thing about the paper you cited,
6.29.19 @3:39pm - (3/3) is that everything is spelled out in a language CI folks can understand, so that
mysteries can be de-mystified. The authors should be commended. Effect-restoration was a breathtaking
mystery to me in 2010, and it shows in the writing; I called it "far from obvious" #Bookofwhy

6.29.19 @12:42am - (1/2) To illuminate our discussion on PS, I am providing free access to Causality, Section
11.3.5 "Understanding Propensity Scores " https://ucla.in/2NnfGPQ#page=18 . Note, in particular, Rubin's
referring to PS matching "as if they had been randomized," and my closing remarks:
6.29.19 @12:42am - (2/2) "it is not enough to warn people against dangers they cannot recognize; to protect
them from perilous adventures, we must also give them eyeglasses to spot the threats, and a meaningful
language to reason about them." Readers of this Twitter understand it. #Bookofwhy
6.28.19 @12:54am - (Replying to @rlmcelreath) I've found that barriers among disciplines are lowered when we
stress the questions to which CI seeks answers: (1) Effects of pending interventions, (2) Effects of undoing past
events. PS. I couldn't read how the match-oxygen problem is related to attributable fraction #Bookofwhy

6.28.19 @12:16am - (Replying to @nghushe) Composure! Mom, Listen! This is the most complimentary message I
received since my Bar-Mitzva. Three days after a prominent statistician accuses me of "calumny, caricature &
confusion." I checked the dictionary, yes, the word exists; it just was not in my vocabulary. Thanks.

6.27.19 @11:21pm - (Replying to @edwardhkennedy) Fine. But I asked for a published paper that states so
explicitly, to warn readers against assuming that PS has anything to do with asymptotic bias.

6.27.19 @11:17pm - (Replying to @jasonhartford @causalinf) HHMMM! you weren't talk about the real PRIMER
(http://bayes.cs.ucla.edu/PRIMER/ ) ?? This too is a huge return on investment. For $20 and a few simple
examples you get a glimpse at what causal inference can do for us, which is quite a lot. #Bookofwhy

6.27.19 @11:00pm - (1/2) I have no doubt that proving the asymptotic equivalence is immediate. Indeed, you
can find such a proof in Causality ch 11 Eq. 11.10. What I asked however was whether anyone knows of a paper
that actually states it EXPLICITLY, WITHOUT the assumption of non-confoundedness. The
6.27.19 @11:00pm - (2/2) paper you cite mentions Nonconfoundedness 17 times, thus contributing to the
marketing myth that PS matching somehow contributes to bias reduction. See footnote 9, Causality p.349.
#Bookofwhy

6.27.19 @5:14pm - I have not seen the connection, nor shown one. But I am a slow learner. The connection to
complexity fascinated me in my youth: "On the Connection Between the Complexity and Credibility of Inferred
Models," https://ftp.cs.ucla.edu/pub/stat_ser/pearl-1978-connection.pdf ... But I haven't tuned in since.
#Bookofwhy

6.27.19 @3:19pm - Statistics as a discipline that helps us go from samples to distributions should be embraced
and promoted. Statistics as an intellectual blinder that prevents one from seeing beyond distributions should be
abandoned and shunned. #Bookofwhy

6.27.19 @7:11am - (Replying to @ildiazm) I never object to inclusion, especially inclusion of basic building block
like stats. I object to exclusion. Like math teachers who exclude multiplication from arithmetic because they
can always add a number to itself n times, i.e., the "classical way". #Bookofwhy

6.27.19 @2:46am - (Replying to @JDHaltigan) I agree with your take, and our challenge is to make 1>>2 in a
climate where 2 control academia and demand submission from 1. #Bookofwhy

6.27.19 @2:39am - (Replying to @stephensenn) Exactly the way Lord describes what the 2nd statistician's does:
Compare W_F of Diet-1 to that of Diet-2 for students of equal W_I. But Stephen, it is your turn now to teach me
what I do not know about adjustment, I am listening carefully to new ideas, Just listening. #Bookofwhy

6.27.19 @2:21am - Thanks for posting, Jennie, now everyone can see what I mean by "causal inference" and why
it needs a new logic, different from "the classical approach" we have been discussing here. #Bookofwhy

6.27.19 @12:41am - (1/3) I am curious, do you know of any paper that states explicitly the asymptotic
equivalence of PS and the "adjustment formula"?? I have only seen it in Causality, ch 11, not elsewhere, Why?
Perhaps someone is interested in marketing it as a magic wand? I am also curious to know
6.27.19 @12:41am - (2/3) how many readers hear this equivalence for the first time. It says that regardless of
6.27.19 @12:41am (2/3) how many readers hear this equivalence for the first time. It says that regardless of
what covariates you use, as the number of samples increases, the bias of the PS estimator converges to the bias
of the adjustment estimator. If many hear about it for the first time, it will
6.27.19 @12:41am - (3/3) serve as an example of information stifling that CI needs to liberate itself from.
Anyone knows what the new CI textbooks say about PS? I would use it as a litmus test for authors understanding
of modern CI. #causalinference #Bookofwhy @causalinf

6.26.19 @11:44pm - By all means. I hereby advocate, as always, that data-science rests on two equally
important pillars: causal inference and statistical estimation. However, a glaring asymmetry can be seen today
in academia: researchers in the CI pillars are thirsty for new
6.26.19 @11:44pm - (2/3) tools from the estimation pillar but not the other way. By and large, leaders of the
stat pillars have zero interest in advances emerging from the CI pillar. The great majority of them truly believe
in "do it the classical way" and "a causal model is a special case of a
6.26.19 @11:44pm - (3/3) predictive model". Moreover, in academia, the stat pillar dominates its CI partner by
100:1 ratio, and now insists on total dominion in the name of "its just a special case". Thus, to achieve
equilibrium, I think CI needs academic autonomy, at least for a while. #Bookofwhy

6.26.19 @11:08pm - (1/2) Why the anger? Was I wrong in pointing to Gelman's blog as a "stronghold of statistical
thinking"? (He has 21K followers, more than any other statistics-minded blog that I know.) Did I misquote it as
advocating "do it the classical way"? or in stating that a "causal model is
6.26.19 @11:08pm - (2/2) a special case of a predictive model" or as advocating adjustment for all "pre-
treatment differences among groups". I am always willing to learn more about what "adjusting" means. What is
it? Teach us #Bookofwhy

6.26.19 @7:13pm - (Replying to @oacarah @PWGTennant) I think these papers miss the point. Agree? They treat
PS as another method of identification, rather than an efficient estimator of the adjustment formula.
#Bookofwhy

6.26.19 @7:01pm - (1/2) There was a time when statisticians were the guardians of prudence and caution, and
causal folks the unruly children of unprincipled adventures. Today, if you peek at the strongholds of statistics
(eg, Gelman's blog) you will find the opposite. CI folks hold the leash
6.26.19 @7:01pm - (2/2) on scientific principles, while statisticians advocate "do it the classical way", namely,
rush into the mine-field of causation w/o a metal detector, and tell students that curve fitting is "causal
inference". Yes, The Times They Are A-Changin' #Bookofwhy

6.26.19 @3:39pm - To summarize: (1) Is it just more explicit? No it is less explicit. (2) Helps with checking
'balance'? Sometimes, but 'balance' is the wrong criterion for covariates. (3) Encourages better model building?
On the contrary. The advantage lies in mapping high-dim to 0--1 interval.

6.26.19 @3:39pm - In case you skipped Section 11.3.5 in Causality, titled "Understanding Propensity Score"
(p.348), I would highly recommend it, for it clears up some of the myths connected with PS. This one
https://ucla.in/2NbS14j may also be illuminating #Bookofwhy #EpiTwitter #Causalinference

6.26.19 @2:43pm - (Replying to @elizpingree @Jamie_Woodward_) Agree. Aside from its artistic qualities, the
significance of the Lion-man in human development cannot be underestimated. Inspired by Harari's "Sapiens" , I
featured it as the seed of counterfactual reasoning in the #Bookofwhy . Thanks for posting.

6.25.19 @11:57pm - (Replying to @lisabodnar) How powerful and truthful your uplifting words sing. I wish I
could see them each time an academic troll unveils his/her pain. #Bookofwhy
6.25.19 @10:45pm - Thanks for your kind words. As to Aesthetics, our publisher insisted that general-purpose

books should look different from technical books. Who are we to resist? #Bookofwhy

6.25.19 @10:40pm - (Replying to @adamdedwards) Good question. Thus far, I know of only two US philosophy
dpts elevated to the age of causation: CMU and CalTech. Anyone knows of more? #Bookofwhy

6.25.19 @8:14pm - Agree, Epidemiologists are 98% there. All it takes is a final snip of the umbilical cord to
mother-stat with a firm commitment to listen to what causal models tell us. #Bookofwhy

6.25.19 @1:13pm - (Replying to @DanielNevo) Same question can be asked about statistics. But it so happened
that, in order to attract top scholars to the field, statistics insisted on academic autonomy, and dominion over
data analysis, rather than being unappreciated minority in each data-using department #Bookofwhy

6.25.19 @3:06am - (Replying to @vkehayas) We have two additional sources of information: (1) playful
manipulations (often called "interventions" or "experimentation") and (2) hearsay (often called "education").

#Bookofwhy

6.25.19 @3:01am - (Replying to @phi_nate) I twitted my take about two months ago, saying essentially: Can
anyone translate what he is doing to our language so that we can prove that the claims do not violate any of the
impossibility theorems we derived mathematically? Waiting. #Bookofwhy

6.25.19 @2:56am - (Replying to @zaffama) Agree, the culprit is in the interpretation. But notice an interesting
phenomenon: the interpretation controversy did not rise when we accepted the axioms of probability, it has
arisen only when we derived a consequence of those axioms, and interpreted it. What took us so long?

6.25.19 @1:50am - (Replying to @mariotelfig) We are talking about a one-line proof, not one-line statement of
the theorem.

6.25.19 @1:48am - Eventually, I am sure, there will be more Causal Inference PhD programs than statistics PhD
programs, possibly under the title "data science - causal inference" The question is which departments will
launch it first, statistics or computer science?

6.25.19 @1:33am - Name another one-line theorem that has remained controversial for 250 years, and books
like "the theorem that never dies" are written about it, and people are hired and fired in its name, etc etc. My
take: it is not merely a theorem; it is a statement. #Bookofwhy

6.25.19 @12:26am - (1/2) Apropos Bayes. Does anyone thinks Bayes' Theorem is really a theorem? If it is, then it
is the most trivial theorem in the cosmos, with a one-line proof. Can someone, even a Reverend, become
immortal with a one-line proof? If it is more than just a "theorem", whence comes its
6.25.19 @12:26am - (2/2) added value? The #Bookofwhy answers this question (p. 102) from computational
perspective, since I could not find any discussion of it in the statistical literature (@learnfromerror, @f2harrell,
@stephensenn ) I believe it has more to do with psychology than with statistics.

6.24.19 @11:14pm - Agree. Bayesians first reacted to frequentists zeal, who persecuted them for contaminating
statistics with "subjective knowledge". In time, they defined club membership by "priors on parameters",
regardless if those priors conveyed knowledge or habits. I go Bayes 1763 #Bookofwhy
6.24.19 @8:58pm - (Replying to @kareem_carr @EpiEllie) Causal inference has been unified 10 years ago, see
https://ucla.in/2N9f28c . True, the "schools" are still singing different anthems, laden with egos and tribalism,
but you, as a rebellious champion of commonsense should look at the content, not the label. #Bookofwhy

6.24.19 @8:36pm - (1/2) My, My, Carlos, this paper was published 10 years ago and, so untypical of me, I am
still behind every assertion. Remarkably, 10 years have passed, and statisticians are still resisting the distinction
between causal and statistical notions (Section 2.1). Just this week,
6.24.19 @8:36pm - (2/2) Andrew Gelman @StatModeling wrote: "So I think it's a mistake to think of causal and
predictive inference as being two different things." Your posting this review makes me both sad (0 inches - 10
years) and hopeful -- I detected sparks of awakening on Gelman's blog #Bookofwhy

6.24.19 @8:03pm - (Replying to @kareem_carr @EpiEllie) I wish you were right. My brief (83) encounter with
statistics textbooks reveals a slightly different picture -- a complete prohibition on all assumptions,
philosophical, causal or otherwise, except statistical assumptions that are but a tiny part of decision making.
#Bookofwhy

6.24.19 @7:54pm - (Replying to @lisabodnar @EpiEllie) Congratulations, Lisa! Judea

6.24.19 @7:47pm - (Replying to @kareem_carr @EpiEllie) The tough question is: Is this a "well-defined"
intervention?

6.24.19 @3:14pm - I think this discussion would benefit from a glimpse into the Bayesian vs anti_Bayesian
controversy in AI, in the 1980's, about how to model epistemological uncertainties in expert systems. My
recollections are here https://ucla.in/2Qg0Rfs Biased, but honest #Bookofwhy

6.24.19 @4:28am - (1/2) When statisticians learned how to spray priors on parameters (of distribution
functions) they formed an exclusive club called "Bayesian Statistics" and decided that he who won't spray priors
on parameters is "not a Bayesian". My definition if Bayesianism is more broad, it
6.24.19 @4:28am - (2/n) follows Bayes paper of 1763, and it has to do with his interpretation of the phrase
"given that we know X=x" and a license to invoke prior information. When I coined the name "Bayesian Network"
(1995) I justified it on these grounds and I added Bayes' fascination with "cause"
6.24.19 @4:28am - (3/3) as another reason. Today I am only half Bayesian for reasons explained here
https://ucla.in/2nZN7IH , mainly because the bulk of our knowledge is causal, which cannot be captured by
priors over parameters, definitely not parameters of a distribution of observations #Bookofwhy

6.24.19 @2:25am - True, I am very fond of my 1988 book "Probabilistic Reasoning", but I left Baeysian analysis in
favor of causal inference, partly because most of medical reasoning is causal, not probabilistic. Its a whole new
paradigm https://ucla.in/2KYYviP , and much fun #Bookofwhy.

6.24.19 @12:47am - (1/2) On a different occasion we will debate what DAGs cannot do, promise. Right now, the
question is "Can a mechanism-seeking researcher answer causal questions from partial information alone, given
in the form of "who listens to whom". Here are some examples:
6.24.19 @12:47am - (2/2) (1) The prisoner is dead. What if Rifleman-A refrained from shooting? (2) More people
died from inoculation than from smallpox. Should we ban inoculation? (3) Is there a drug that is good for men,
good for women, and bad for a typical person? #Bookofwhy

6.23.19 @11:35pm - (Replying to @yourbirlfriend @EpiEllie) @yourbirlfriend Fascinating! But cellular automata
are driven by local forces and "minimum thread" is a global feature. How can you accomplish it? @yudapearl
6.23.19 @11:25pm - (Replying to @NeuroStats @shravanvasishth and 4 others) For computer scientists, the
helplessness of Bayes analysis in causal reasoning comes glaring already in the notation. BDA invokes only one

conditioning symbol, the vertical bar |X=x). Causal reasoning requires |do(X=x) Bingo! Done! You cant get "do"
from "see" #Bookofwhy

6.23.19 @4:01am - (1/3) This article is worthy of our attention:


https://www.atsjournals.org/doi/full/10.1513/AnnalsATS.201808-564PS ... 47 journal editors are offering
guidelines to authors on ways to report results of causal inference studies. It is refreshing to see 47 editors
reach consensus on a topic that only a decade ago was a sure ticket
6.23.19 @4:01am - (2/3) to discord. I believe the availability of DAGs as a communication language helped the
process. I have trouble with some of the terminology (e.g., "causal association") but, overall, I welcome the
timely rejection of "traditional" approaches of wholesale adjustment for
6.23.19 @4:01am - (3/3) everything one can conveniently measure. See @StatModeling for a lively discussion of
opposing viewpoints, especially my explanation of why blindness to DAGs is an invitation to bias amplification
https://ucla.in/2N8mBMg . #Bookofwhy

6.23.19 @2:35am - (1/3) When you have a chance, please explain to readers on this Twitter how Bayesian Data
Analysis (BDA) can help one think about causality. I have heard it from many statisticians and data analysts but I

have never been able to understand what they find to be helpful and why.
6.23.19 @2:35am - (2/3) Is it the "model selection" part offered by BDA? Or the idea that you are properly
combining prior knowledge with data? In my opinion, BDA is a siren song that lure people away from properly
"thinking" about causation, as I argue here https://ucla.in/2nZN7IH and in many
6.23.19 @2:35am - (3/3) other forums. I am appealing to you because, as an accomplished reader of
#Bookofwhy I think you would be able to pin point to us where precisely BDA enthusiasts see the connection to
Causal Inference, and why I am missing this connection. As always, a toy example is the KEY.

6.22.19 @7:59pm - (Replying to @RMartinChavez) I could not resist clicking on "like" to such a compliment, but
the main thing is it is true: "At last, a science of causation!" And as much as I would try to minimize it the facts
will scream at my face: a science of causation. #Bookofwhy

6.22.19 @5:58pm - (1/n) My remarks about being interested in seeing a Multi-level problem solved, and about
traditionalists fearing toy-problems were in the context of the discussions on Gelman blog. As to #Bookofwhy,
you are asking to be shown the "how" first and the debates second. This is exactly
6.22.19 @5:58pm - (2/n) what the book does. Chapter 1 tells you about the inference engine and what kind of
questions it solves. What you see as a "fight" among competing approaches does not exist, because the history
of debates about causation was not among "approaches" but among "ideologies"
6.22.19 @5:58pm - (3/n) 3/n that do not qualify for the title "approach". An "approach" should be armed with
definition of what problems the approach attempts to solve and theoretical guarantees of reaching adequate
solutions under certain conditions. Thus, the reason readers get the impression that
6.22.19 @5:58pm - (4/n) there is only one "approach", (ie SCM,) is that there is really only one approach armed
with the needed theoretical guarantees, the alternative themes where merely "themes". Currently, I know of
one competing "approach", ie, Rubin's potential outcome , which is logically
6.22.19 @5:58pm - (4/n) equivalent to SCM under the same assumptions but makes it hard for researcher to
represent assumptions. It is described humbly and respectfully on pages 272-280. (I just made a similar point on
Gelman's blog.) Its tough to tell the naked truth to readers who expect
6.22.19 @5:58pm - (5/n) something else. Some even view my claim about a transformative revolution to be
nothing but "hype". I hope now, that you have finished the #Bookofwhy and had a chance to seek alternatives,
you see that I had no choice but risk being called "hype" - "Si Muove" it really moves!`

6.22.19 @6:23am - (Replying to @djvanness) The calculus cannot help you with this choice, since you have not
ifi d h fk l d hi h h h i d N i h l l ki k i l f
specified the state of knowledge upon which such choices are made. Now, since the calculus kicks in only after
you made the choice, it can be considered an "assumption". #Bookofwhy

6.22.19 @5:42am - (Replying to @sakrejda @shravanvasishth and 2 others) You are being unfair by assuming

apriori that what you see as "pettiness" is not genuine attempts to learn from each other. I, for one, would very
much like to hear more about what we can learn from the "broader search" that goes beyond "causal inference".
Care to explicate?

6.22.19 @5:31am - In causal modeling we do not call it a "choice" but an "assumption" (item (2)). If you think it
is not plausible, go ahead and tell the calculus that Hall affects weight independently of Diet. (Perhaps one Hall
has a dancing floor??). But we need a story to proceed #Bookofwhy

6.18.19 @5:35am - (Replying to @gjcampitelli @Lester_Domes and 4 others) I would stick to my SCM religion:
Solutions must start by articulating the questions. What do we mean by "handling"? What are we expected to do
after we differentiate between FE and RE in the model construction? Or, put more bluntly, why would anyone
care? #Bookofwhy

6.22.19 @5:23am - It was great fun talking at your conference and assuming (wrongly) that audience never
heard about causal inference. Thanks for posting this very slide "What is Causal Inference?" which is so timely in
view of conversations we are having on Twitter and on @StatModeling #Bookofwhy

6.22.19 @5:09am - (Replying to @stephensenn) In Fig. 4 "Dining Hall " appears because it is taken from Wainer
etal where it is another name for "Diet". Would you prefer we proceed with Fig. 4 and compute P(weight gain|
do(Dining Hall)) ??? Give me a story. #Bookofwhy

6.22.19 @4:27am - (Replying to @stephensenn) Are you happy with the story of Fig. 6.9(b) ??. If not, please
modify, if yes, we will go to: (3) data available. #Bookofwhy

6.22.19 @4:22am - (Replying to @stephensenn) These questions are vividly answered in Fig. 6.9 b. "Hall" is not
in the graph, which means it is irrelevant to any research question. It shows W_I ----> D, which means "Diet may
attract slim/heavy students", and so on, all assumptions are vividly displayed. #Bookofwhy

6.22.19 @4:04am - (Replying to @stephensenn) We need to analyze each version and see. So far we do not have
a story -- what's the role of the "Hall". Is it just a place where Diet is served or a place that also affects weight
or a place that attracts slim students, or a place that attracts students craving for a given diet?

6.22.19 @3:50am - (Replying to @stephensenn) Causal calculus does not "handle" things; it answers (1) research
questions, given (2) assumptions and (3) data (experimental or observational). So far we got the question: (1)
Find the causal effect of diet on weight gain, or P(gain|do(diet)). Now we need (2) and (3). #Bookofwhy

6.22.19 @2:23am - (1/2) I thought we agreed that in these versions Lord's paradox disappears. If you are still
interested in "handling" these versions, independently of Lord, by all means, but you need to specify them as we
specify any "toy problem", namely, (1) what do you wish to estimate?
6.22.19 @2:23am - (2/2) (2) What (causal) assumptions are you willing to make? and (3) what data is available
to you? Let's start with ONE version, not TWO, to avoid going back and forth. #Bookofwhy

6.22.19 @2:01am - Unfortunately, I had to remove the Babylonian vs Greek analogy from the final version of the
paper https://ucla.in/2HI2yyx ; a reviewer insisted that this would offend ML folks. #Bookofwhy
6.22.19 @1:28am - Why do you assume that I am trying to "advocate' one approach or put down others? Why not
assume that I am genuinely craving for ANY approach that enriches causal inference, and I can't satisfy this
craving (my weakness) without seeing toy examples solved in other approaches.Why?

6.21.19 @10:48pm - I have the feeling that @StatModeling folks did try toy problems and the reason they dread
them so consistently is that they realize they havn't got the tools. I would be very interested in seeing a simple
Multilevel problem well posed and toy-like solved.

6.21.19 @9:55am - I like the term "susceptible to persuasion" (or "gullible"), because Ang Li has just finished a
paper that formalizes this notion and gives it algorithmic teeth https://ucla.in/2EgRzvl #bookofwhy

6.21.19 @12:37am - I just finished an adventurous visit to Gelman's blog


https://statmodeling.stat.columbia.edu/2019/06/18/causal-inference-i-recommend-the-classical-approach-an-
observational-study-is-understood-in-reference-to-a-hypothetical-controlled-experiment/?
unapproved=1064803&moderation-hash=4003e2b37d332ac7a4195c59d997cbd5#comment-1064803 ... and I am
sharing here the last sentence: "if Causal Inference is 'Statistical Inference given causal assumptions' [as some
claim] then Car making is car-painting given an engine and a body."#Bookofwhy

6.20.19 @10:30pm - Yes, Just in case you are at UCLA tomorrow, and have not gotten tired yet from listening to

my Songs of the Revolution, join me tomorrow 3:30 pm where I will be singing one Aria to a slightly new libretto
in front of social scientists and other mostly harmless folks. #Bookofwhy

6.20.19 @1:32am - I was truly honored to speak at the Algemeiner Gala, to be introduced by Sharon Stone and
to warn against "Zionophobic Thuggery," which I hope will be recognized as the ugliest form of hostility on US
campuses.

6.19.19 @4:52pm - @JewishPub. The more I read it, the more strongly I feel about including it in the list of 5
books that an "educated Jew" could not live without.

6.19.19 @3:49pm - (1/2) Many thanks. This is a good place to start chewing the eco. literature. One hurdle is
the definition of "identified" which in eco. usually means identify the functions or their parameters, and in CI
means identify a query Q which depends on the functions. I have just
6.19.19 @3:49pm - (2/2) noticed that the example you brought up is the usual IV setting, as in here
https://ucla.in/2pKdx1U , So it is already in our arsenal of "Powerful tools." Will see if more nuggets can be
excavated. #Bookofwhy

6.19.19 @1:44am - (1/2) Great!! Here are some powerful results that hold for the linear extreme of the
spectrum: https://ucla.in/2L8jUFg . Can you or one of your students pin point to me which of them is
extendable in some way to nonlinear systems? For example, Eq. (20) shows us how to estimate ANY
6.19.19 @1:44am - (2/2) counterfactual expression in ANY linear system. Can this capability be extended in
some way to binary monotonic nonlinear systems? Just a hint, or eq.(#) would do, but please do not send me
without guidance to a vast unchartered literature. Thanks #Bookofwhy

6.19.19 @4:30am - (Replying to @DanielOberski) There is a habit, especially on Twitter, by people who have not
read things with sufficient depth to quote authors names, or provide links to cryptic papers and say: "It is
treated here...". Done. "Without links" means: give me the basic idea, dont send me elsewhere. #Bookofwhy

6.19.19 @4:21am - (Replying to @WhitneyEpi @AdanZBecerra1 and 12 others) One should note that these
6.19.19 @4:21am (Replying to @WhitneyEpi @AdanZBecerra1 and 12 others) One should note that these
""classical" rules from the RCT context were merely "rules of thumb", lacking theoretical underpinning. It turns
out that some post-intervention factors are actually safe, while some pre-intervention factors are unsafe. See
Causality p. 339 #Bookofwhy

6.19.19 @3:14am - (1/3) I am partially familiar with Matzkins works, and I know that many economists refer to
them. However, I have not been able to supplement the arsenal of powerful results now available for linear and
NP extremes of the spectrum, with similar results applicable to mid-spectrum
6.19.19 @3:14am - (2/3) I examplified what I mean by "powerful results" here:
https://twitter.com/yudapearl/status/1141104139553783808 ... Perhaps you or some other expert economist
can fill me in on what is known about the mid-spectrum. Or, what is the one most important identification
opportunity one would miss by not
6.19.19 @3:14am - (3/3) this literature as thoroughly as it deserves. For example, is the counterfactual
estimability of linear models extendable in some way to monotonic models? or Binary models? #Bookofwhy

6.18.19 @10:43pm - I have posted a comment on Gelman's blog, explaining why causal and predictive inference
are not the same thing, and why it is beneficial to solve each task in its own distinct vocabulary. #Bookofwhy

6.18.19 @3:03pm - (1/2) "Powerful results" are those that one remembers. For example, "Every causal effect in
every NPSEM is either identified or quickly proven to be non-identifiable". "Every linear SEM that contains no
bow-arc is identifiable" "Every counterfactual in linear SEM is estimable" etc.
6.18.19 @3:03pm - (2/2) Are there similar results for the spectrum between linear and nonparametric? For
example, what do we know about the loglinear variety? What identification results from linear SEM are
preserved? To be "powerful" they need to be articulable verbally w/o links . #Bookofwhy

6.18.19 @12:30pm - My thoughts? Hilarious!!! Here is one gem: "So I think it's a mistake to think of causal and
predictive inference as being two different things." Here is mine: "So I think it's a mistake to think of any two
things as being two different things - it's all arithmetic" #Bookofwhy

6.18.19 @5:35am - (Replying to @gjcampitelli @Lester_Domes and 4 others) I would stick to my SCM religion:
Solutions must start by articulating the questions. What do we mean by "handling"? What are we expected to do
after we differentiate between FE and RE in the model construction? Or, put more bluntly, why would anyone
care? #Bookofwhy

6.18.19 @4:49am - (Replying to @fr_amodeo) If by "handling" you mean "provide distinct specification" then for
the standard SCM the answer is NO. But why not add a marker "L" to any family that combines linearly, and
leave the others unmarked. The beauty would be to decide if an arbitrary causal effect is fixed #Bookofwhy

6.18.19 @2:12am - (Replying to @juli_schuess @PHuenermund) Paul, why the wonder? This is the classical IV
setting and, as @juli_schuess notes, all parameters are identifiable using either Wright rules
https://ucla.in/2LcpmHz or the method of moments. For generalizations, I would use https://ucla.in/2MqNBRU
#Bookofwhy

6.17.19 @8:50pm - (1/2) Got it. Thanks. And I am retweeing because many CI folks are likely to stumble on this
jargon. Writing y=bx+u gives fix-effect, because P(y|do(x+1))-P(y|do(x)) is independent of u. But writing
y=f(x,u) gives random-effect. So, nonparametric SCM assumes random effects, and the
6.17.19 @8:50pm - (2/2) distinction does not show in the DAG, only in one's declaration "Assume a linear
model". An entirely different question is: "Is there a test that detects the presence of RE in a given population?"
Yes, there is, and beautiful one too: https://ucla.in/2uqDjdp #Bookofwhy

6.17.19 @7:59pm - (Replying to @Corey Yanofsky @Lester Domes and 2 others) A partial meeting of minds was
6. . 9 @ :59p ( eply g to @Co ey_ a o s y @ este _ o es a d ot e s) pa t al eet g o ds was
achieved here, on Twitter, when my mind realized that what Senn objects to is the perfect alignment of the
two ellipses, which would rarely occur in real life. This I consider to be orthogonal to Lord's question: "Who is
right?" #Bookofwhy

6.17.19 @7:52pm - (Replying to @Lester_Domes @Corey_Yanofsky and 2 others) I will try, if only I knew what RE
and FE are. (With examples please). Are the the causal effects estimated in an ideal RCT RE or FE ????

6.17.19 @6:59pm - (Replying to @PWGTennant @lisabodnar @ProfMattFox) Sorry to be missing your dazzling
party. Now I have two reasons to attend SER-2020. Cheers!

6.17.19 @1:15pm - So, according to your perspective, modern SEM maintains the same input-output relation as
SCM. Namely, INPUT = data + qualitative causal assumptions. OUTPUT= causal effect sizes. And no one is
uncomfortable about the word "causal" there, even in published papers. #Bookofwhy

6.17.19 @12:59pm - (Replying to @andres_fandino) There is no such notion as "appropriate". If your


understanding of the problem demands N exogeneous and M endogeneous variables, so be it. Reality comes
first, estimability second. #Bookofwhy

6.17.19 @12:43pm - (1/2) Yours is a very encouraging perspective of modern SEM, namely, SEM = "SCM that
accommodates parametric assumptions." Its badly needed. I am familiar with the linear and nonparametric ends
of the spectrum; are there any powerful results in between? Say binary, or monotonic?.
6.17.19 @12:43pm - (2/2) Also, just out of curiosity, what percentage of your colleague/students can answer
"Tell me if the partial correlation R_{XY.Z} is zero", or "Tell me which parameter is estimable by OLS". I have
not communicated with them for several years. #Bookofwhy

6.17.19 @4:50am - (Replying to @y2silence) Beautifully put. And I wish more people learn to appreciate this
miracle, which is truly unique. In fact, what else can one ask of a modeling methodology? #Bookofwhy

6.17.19 @3:28am - (1/3) In the last conversation I had with Peter Bentler he defined SEM as a "compact and
meaningful representation of a covariance matrix". We know that "meaningful" is a round-about way of saying
"causal". Fine. But since the purpose is declared as fitting a covariance matrix,
6.17.19 @3:28am - (2/3) one should ask: To what use can one put the conclusions of an SEM study (aside from
getting your dissertation published in Journal of SEM), which usually comprises a huge list of path coefficients
with their confidence intervals. I wish someone can explain why not list the
6.17.19 @3:28am - (3/3) estimated covariance coefficients themselves; a much simpler task. #Bookofwhy

6.16.19 @11:58pm - It is hard to discuss SEM since its practitioners are still not sure what it is. I tried to
describe their confusion here https://ucla.in/2Nbq7W2 and sway them to use SCM. Not successfully, and
Psychomentika-type papers are my witnesses. Interesting social phenomenon #Bookofwhy

6.16.19 @11:29pm - (Replying to @ulusdd) True, but I dont know anyone who starts with a cyclic system and
expects DAGs to handle it. DAG techniques are expected to compute properties of the input model, namely
directed-acyclic systems. Besides, d-separatioin holds in linear cyclic systems. #Bookofwhy

6.16.19 @11:22pm - My perspective today: SEM is a community of researchers using SCM who refuse to commit
to the causal reading of their models, argue endlessly about what that reading is, and refuse to benefit from
the comp. power of DAGs. Too harsh? Would welcome opposing evidence #Bookofwhy

6 16 19 @11:09pm (Replying to @omaclaren) I am not familiar with "rate parameters" "ODE models" and
6.16.19 @11:09pm - (Replying to @omaclaren) I am not familiar with rate parameters ODE models and
'nonlinear least square". Can you explicate in terms of the input model, which is a DAG, be it linear or
nonparametric. Of course a DAG cannot compute how much money you have in bank account, or halting
problems etc.

6.16.19 @8:57pm - (1/2) "tools with which to think about models" is only one usage of DAGs. Another usage is
computational; DAGs permit us to answer questions which otherwise are intractable. E.g.,"Tell me if the partial
correlation R_{XY.Z} is zero", or "Tell me which parameter is estimable by OLS"
6.16.19 @8:57pm - (2/2) I am surprised that the computational capacity of DAG is under-appreciated by most
researchers. Am I exaggerating? Is there another computational tool with which the questions above can be
answered? Is there a question you wish to ask which DAGs do NOT answer? #Bookofway?

6.15.19 @6:28pm - (Replying to @robanhk) I love those "intellectuals" and have written a poem in their honor

6.15.19 @6:26pm - (Replying to @robanhk) I was not sure if you were sarcastic or honest. Now that you are
equating "not loyal to Israel" with "denying Jews the right to a homeland" you made it clear. I do not know
anyone who denies a people's right to a homeland who is also considered an organic part of that people.

6.15.19 @4:21pm - (Replying to @robanhk) Yes. I thought that the fringe group of Jewish-born intellectuals who
deny Jews the right to a homeland have not felt targeted by Abdulhadi's rants; they run for safety by showing
loyalty to her Zionophobic movement and do not need my encouragement.

6.15.19 @2:12pm - Excited towards keynote addressing the UCLA Jewish Graduation Ceremony on Sunday.
Hoping to inspire the very students who were recently labeled "White supremacists" by a super BDS ideologist
(still at large, still not condemned).

6.15.19 @1:28pm - (1/2) If instead of criticizing the way SEM's ARE used we take the trouble to explicate how
they SHOULD be used, we will, I believe, end up with SCM's (Structural Causal Models) like those described in
#Bookofwhy or https://ucla.in/2Nbq7W2 . But,
6.15.19 @1:28pm - (2/2) now that we are discussing classical SEM, I am curious to know whether there is still a
large community of SEM users who have not switched to SCM ??? If there is, can someone act as its spokesperson
and explain what, aside from habit, prevents them from making the transition?

6.15.19 @2:35am - Sewall Wright had much harder time communicating with statisticians. He did not have the
RCT metaphor to open the first window to their hearts, and he did not have the logic of counterfactuals to

defend the structural assumptions behind his diagrams. A true hero. #Bookofwhy

6.14.19 @3:39pm - This is truly a gem -- thanks for posting. I have not read it for many years. And now that I re-
read it, it covers so many topics that we have discussed here on Twitter. It even discusses modern hangups with
"mimic RCT or else your causal effect is "not well defined". #Bookofwhy

6.14.19 @11:44am - The correct link to "Causal Foundations of SEM" is this https://ucla.in/2Nbq7W2 There was
a missing space. #Bookofwhy

6.14.19 @3:43am - Regarding the causal foundations of SEM, I have found another article that addresses this
issue head on: https://ucla.in/2Nbq7W2and . It also explains why teachers of SEM are so helplessly confused
about a problem so simple. #Bookofwhy

6 14 19 @2:12am - Thank you Rabbi Dunner for immortalizing our meeting at the Algemeiner Gala I felt truly
6.14.19 @2:12am Thank you, Rabbi Dunner, for immortalizing our meeting at the Algemeiner Gala. I felt truly
fortunate to be able to contribute to our three common ideals "Truth, humanity and Jewish peoplehood" - a
spec of sanity in the age of BDS madness.

6.14.19 @1:37am - Many have had the same SEM experience. The answer is that no magic dust is needed; the
coefficients were always causal, and the association just help us to quantify those causal relationships. Bollen
and I discuss it in "Eight Myths of SEM" https://ucla.in/2QnG9dr #Bookofwhy

6.11.19 @8:15pm - Thanks, Lisa, for letting us know that you will MC tomorrow's Gala. Looking forward to
seeing you again. And may the angels bless you and the other warriors in the trenches.

6.11.19 @6:25pm - @causalinf Have you looked into the way Peter Steiner represents RDD in DAGs? I have a link
to his 2017 paper: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6117124/ ... #Bookofwhy @PHuenermund ,
@eliasbareinboim , @yudapearl , @Jabaluck @EpiEllie , @paulgp

6.11.19 @6:11am - (Replying to @TuomasPernu) I feel the same way, except that I found solace in the
metaphysical assumption that reality acts as a "society of listening agents". So far, I have not found a more
satisfactory metaphysical theory in the philosophical literature. Would be curious to examine alternatives.

6.10.19 @3:49pm - (1/2) "difference-making" or "counterfactual dependence" are various names philosophers
used, but they fell short of operationalizing this relation through the simple mathematical concept called: "a
function". I dont know what book you have on shelf, but if it is #Bookofwhy or
6.10.19 @3:49pm - (2/2) https://ucla.in/2KYYviP or https://ucla.in/2KYvzau or Causality, then you will satisfy
you quest to see coherence among the assumptions and the conclusions. The philosophical literature,
unfortunately, is still hung up on "what do we REALLY, REALLY... mean by CAUSE". Hopeless

6.10.19 @2:40pm - (1/2) One gift modern causal inference has given to us is a clear answer to your question:
"what the assumptions and stages are, and how they relate to each other." See eg #Bookofwhy. Another gift is
seeing all those assumptions expressed in terms of one primitive relation:
6.10.19 @2:40pm - (2/2) "One variable listening to others". One may argue that this primitive is just a causal
relation in disguise. Perhaps, but it is still a gift to see ALL assumptions and ALL conclusions emerging from ONE
easily grasped idea. #Bookofwhy

6.10.19 @1:55pm - Not entirely unfair. It lends support to the idea that perhaps "listens to" is an irreducible
primitive in our mind which will stay irreducible until we can map the neural paths that are activated when we
ask "why".

6.10.19 @1:53pm - (Replying to @PHuenermund) Not entirely unfair. It lends support to the idea that perhaps
"listens to" is an irreducible primitive in our mind which will stay irreducible until we can map the neural paths
that are activated when we ask "why".

6.10.19 @5:21am - (Replying to @fpgil) And I understand the Portugal is celebrating today its independence day
-- the oldest independence day in Europe. Monday, 10 June: Dia de Portugal. I just heard it on Israeli TV, so,
from Israel with love: Two small countries with rich rich history.

6.10.19 @4:09am - (1/2) I an not sure whether Woodward is satisfied with the structural account, see
https://ucla.in/2MnawNT , but the philosophers that I read do (eg Spohn, hitchcock,Glymour). What is the
point of asking for "the nature" of causation withouot telling us what type of answer will be
6.10.19 @4:09am - (2/2) accepted? The structural account, based merely on the notion of "listening to" is as
satisfactory to me as any that I have seen in philosophy. And, in addition to its metaphysical satisfaction it can
also solve the Simpson's puzzle how can you ask for more? #Bookofwhy
also solve the Simpson s puzzle, how can you ask for more? #Bookofwhy

6.10.19 @3:30am - Portuguese!! My goodness!! Now I can start arguing with all my Brazilian students on equal
footing. Thanks for posting. #Bookofwhy

6.10.19 @3:21am - (1/2) The #Bookofwhy solution makes NO ASSUMPTION beyond what is specified by Lord
himself. If you prefer to fantasize a complex multi-hall version of the problem, and fail to collapse it into Lord's
distribution, do not blame Lord, nor #Bookofwhy . The key point is that, for
6.10.19 @3:21am - (2/2) the simple two-hall problem, each serving ONE diet, the solution provided in
#Bookofwhy p. 217 is valid, and resolves a paradox that still baffles many good minds, even today, including
many statisticians. Again, no extra assumption beyond those given by Lord and Fig. 6.9b.

6.10.19 @2:53am - (Replying to @stephensenn @analisereal) The distribution is given to us by the two ellipses;
This is Lord's construction, not ours. NOTHING ELSE is assumed about variances and covariances. Mixed models
were invented to help in the construction of distributions, not for handling a fully specified
distribution.#Bookofwhy

6.10.19 @12:14am - The less background the better. Except, if you have not taken any stat-101 or econ-101
classes you will miss the fun of asking: "How come my professor never told us that causal inference is easy?"
#Bookofwhy

6.10.19 @12:02am - (1/3) The #Bookofwhy is not about "what causal calculus cannot do" (eg, play chess,
translate languages etc) but about the many miracles it CAN do. Among them resolving the simple version of
Lord Paradox, with two dining halls, each serving one diet, and a very large sample. So,
6.10.19 @12:02am - (2/ ) believe it or not, but this simple version is still paradoxical to most mortals, and has
been paradoxical for half a century. It is now resolved by causal calculus. Your multi-hall version may be of
interest in a certain context, but I cant understand why you are insisting
6.10.19 @12:02am - (3/3) that this idiosyncratic version is essential for resolving the simple version, and that he
who does not attend to your version is guilty of neglecting the foundations of statistics or worse. I do not buy it.
Lets focus on the simple version -- are you happy with the solution?

6.9.19 @11:16pm - Good news for fans of #Bookofwhy. The book is now available in paperback ($14) from
amazon: https://amzn.to/2WqMGG4 and, more importantly, friends tell me they got it delivered next day and
it is much easier to read, handle and carry on the train. Tell your grandchildren.

6.9.19 @8:22pm - (1/2) I find VanderWeele's decomposition somewhat artificial. The two clinically meaningful
components of mediation are: (1) The extent to which observed effects would be PREVENTED by disabling the
mediating path and (2) The extend to which observed effects would be SUSTAINED
6.9.19 @8:22pm - (2/2) with the direct effect disabled. These two components (necessary vs. sufficient)
collapse in linear systems, but are distinct when interactions are present. Each can be estimated using the
mediation formula, https://ucla.in/2QmZrjc , and https://ucla.in/2N5Gdk2 , #Bookofwhy

6.9.19 @5:19pm - Congratulations, again, and I am re-tweeting your vision statement here, because it coincides
so perfectly with my conception of where causality-land is heading. There may be some hidden cause in action
here, but reality prevails regardless of models. #Bookofwhy

6.9.19 @5:00pm - (1/2) (Replying to @jwbelmon @jon_y_huang and 3 others) Some people use an "arrow on
arrow" notation to indicate effect modification, but I find it unnecessary and confusion, because, if A modifies
the effect of B on C then, from basic logic, B must modify the effect of A on C. In linear systems we can just
use a product term
6 9 19 @5:04pm (2/2) (Replying to @yudapearl @jwbelmon and 4 others) and in non parametric system every
6.9.19 @5:04pm - (2/2) (Replying to @yudapearl @jwbelmon and 4 others) and in non parametric system, every
parent is by default assumed to modify the effect of all other parents. I you need to find the degree of effect

modification, you can use counterfactual logic, as VanderWheele is doing, or as we do in causal mediation
#Bookofwhy

6.9.19 @3:26pm - (Replying to @stephensenn @analisereal) "The sample size is only two" My goodness!! we must
be thinking about a different universe. I do not deny the existence of your universe but, given that I naively
assumed that the two ellipses show a very large sample size, please describe your universe slowly slowly
#Bookofwhy

6.9.19 @2:37pm - (1/n) Now that we a research question, we go to the next step: What data we have available.
The answer is given by the two ellipse, showing (observational) measurements taken on 3 variables W_I W_F and
D (Diet), no separate data on Hall. It is quite conceivable that each Hall #Bookofwhy
6.9.19 @2:37pm - (2/2) serves several diets, but the data does not provide separate measurement on H vs. D, so
we assume Hall determines Diet unambiguously. Can we continue from this assumption? Or you prefer to
introduce distinction between H and D.

6.9.19 @2:24pm - I second it! It is a great summary of the field, and makes the debate about "in statistics" or
"extra-statistics" hollow and irrelevant. #Bookofwhy

6.9.19 @7:46am - (Replying to @stephensenn) Ring around the Rosie .... and we still do not what your research
question is. A hint, perhaps? or a clue?

6.9.19 @7:20am - Perhaps Fisher/Nelder/Bailey/Speed/@PhilDawid can articulate what your research questioin
is? I would gladly give them full credit for the ideas of symmetry/exchangeability if they can use them to solve
your problem w/o knowing what you aim to estimate. #Bookofwhy

6.9.19 @7:09am - I cannot understand the issue without knowing three essential elements: (1) What you wish to
estimate, (2) what assumptions (if any) you make and (3) what data you have available. Once specified, all
"issues" are resolved mathematically. #Bookofwhy

6.9.19 @6:23am - (1/2) DV Lindley was a devout Bayesian, but he was also the first to understood that if (in
Simpson's paradox) the same data leads to two different conclusions depending on the story, then Bayes analysis
is helpless, because Bayesian methods are propelled by fitness to data. I have
6.9.19 @6:23am - (2/2) written about this in "Why I am only a Half-Bayesian" https://ucla.in/2nZN7IH . It's a
good paper because many Bayesians harbor the illusion that, if you only spray priors on models and take
sufficient data, the posterior will peak around the correct model. #Bookofwhy

6.9.19 @4:54am - (Replying to @stephensenn) If you aim to analyze experimental studies from two different
studies, on two different group, lets do it correctly, along the theory of data fusion. eg https://ucla.in/2Jc1kdD
. Again, we need to articulate question, assumptions and type of data available. Ready? #Bookofwhy

6.9.19 @2:26am - (1/2) I am saying that the causal story behind the data determines which statistician is
correct, and that the story should be articulated in the language of diagrams (not as Rubin and Holland told it).
Thus, if we believe that diagram 3 is what Wainer etal meant, we go to diag 3.
6.9.19 @2:27am - (2/2) if we believe that Diet is different from Hall, and that another story describes how data
were generated, we can draw the corresponding diagram and interrogate it regarding the correct analysis. Lets
do it. Which diagram do you like? What is your research question? #Bookofwhy
6.9.19 @1:31am - I do not understand expressions such as "be all and end all"; I have not used any, nor implied

any. As to your question, the diagram on p217 already contains "Hall"; it coincides with "Diet", exactly as
presented in Wainer etal. Nothing changed. Just causal modeling. #Bookofwhy

6.9.19 @1:01am - (1/2) It is no longer secret. The lecture that @eliasbareinboim gave at Columbia last month
http://www.cs.columbia.edu/streaming/2019-Spr/elias_bareinboim.mp4 ... was actually a job-talk and I wish
to congratulate Elias on accepting a faculty position at Columbia, starting July 2019. I wish also to
6.9.19 @1:01am - (2/2) congratulate my colleagues at Columbia for strengthening their ranks with a top
innovator of causal-inference research. May this marriage lead to major breakthroughs and smarter machines.
Amen! #Bookofwhy

6.9.19 @12:26am - (1/4) I believe this set of slides reinforces what I tweeted earlier:
https://twitter.com/yudapearl/status/1137335718580183041 ... -- its hard to cut the embilical chord to
Mother-Stat. It occurred to me that this urgency to stay in Stat-womb was also the motivation behind the
potential-outcome framework. The benefits
6.9.19 @12:26am - (2/4) were obvious, nothing is new, Y_1 and Y_0 are ordinary variables, with some missing
values, so what? Everything else is ordinary statistics. The price, of course, was (1) Everything was tied to
experimental "treatments", not to "events" or absence of events, and (2) we need
6.9.19 @12:26am - (3/4) to express knowledge in the language of {Y_1, Y_0} , namely, in the formidable
language of "conditional ignorability". Some would argue: What's wrong in letting statisticians broaden the scope
of statistics and then believe that "it is all statistics"? I believe PO is a good
6.9.19 @12:26am - (4/4) of what could go wrong. Distinctions play a role in science. Treating the Ladder of
Causation as one chunk, in the name of "It is all statistics", while ignoring theoretical barriers between the 3
levels creates more confusion than the stat-womb warmth can sooth. #Bookofwhy

6.8.19 @9:08pm - (1/1) You seems to oscillate between your desire to validate the model and your desire to
answer some causal question, eg. "is there a non-zero effect of Z on Y". Which of the two you wish to start? As
@EpiEllie explained, conditioning on the mediator may lower or amplify
6.8.19 @9:08pm - (2/2) the association between X and Y, so this test is out. (Using your notation) If you can put
down two competing models, one can tell you right away if there is a test that can distinguish them apart. The
name of the game is to confess and let the mathematics work #Bookofwhy

6.8.19 @8:15pm - (1/2) Whether a problem area is causal or not depends not on the diagrams we can draw or
intervention we can perform. It depends only on the research question we aim to answer. In my l last encounter
with fMRI researchers I was disappointed by their lack of understanding of causal

6.8.19 @8:15pm - (2/2) inference and blind allegiance to potential outcome vocabulary. See
https://ucla.in/2N9nhBb . Are these authors representative of the field? If not, please describe what Dynamic
Causal Models in fMRI are aiming to estimate, and one simple example of such a model #Bookofwhy

6.8.19 @7:57pm - (Replying to @jwbelmon @sjalexander @EpiEllie) @jwbelmon. I will try to help, but I do not
understand what you want to do. If you can specify: your research question + your scientific assumptions, then
whether it can be done or not can be answered mathematically regardless of what the field has been doing.
#Bookofwhy

6.8.19 @7:49pm - @EpiEllie has done a good job describing the assumptions behind IV methods and their wide
applications. I commented recently https://twitter.com/yudapearl/status/1136818007315849216 ... that
epidemiologists are now proficient in describing IV methods in DAG's language. The next step is to REPAIR bad
IV's #Bookofwhy

6.8.19 @2:50pm - (Replying to @y2silence @SALubanski) Most importantly, do not miss Carlos's final comment,
@ p ( py g @y @ ) p y, ,
which puts things in the proper perspective.

6.8.19 @5:29am - (1/n) (Replying to @jd_wilko) It is not really "provocative", just a gentle way of luring

statisticians to cut the umbilical chord from Mother Stat. Instead of surgical do-operator, you condition on a
variable F (force) which does the surgery for you. I used it in 1993 https://ucla.in/2pJqtW3 when
6.8.19 @5:43am - (2/n) (Replying to @yudapearl @jd_wilko) I thought that statistician are not prepared for a
surgery. Other researchers, too, labor to create the illusion of remaining in the stat-womb. E.g., Heckman etal
created a fix-operator to enforce this illusion https://ucla.in/2L8OCyl . The folks at @harvardEpi still teach CI
6.8.19 @5:56am - (3/3) (Replying to @yudapearl @jd_wilko @HarvardEpi) 3 by "imitating RCT's". It is a tough
umbilical chord to cut. Denis Lindley was the only statistician I met who said: CUT! In all these schemes we still
need to import information from outside the data, which is the key to realizing that we are out of the stat-
womb. #Bookofwhy

6.8.19 @4:18pm - Good luck and Happy landing! I am not exactly sure what @AVORA does or how, but when I
see people taking causal modeling seriously, I know that #Datascience has a future beyond curve fitting. I will
be looking forward to seeing example applications. #Bookofwhy

6.8.19 @3:51am - (Replying to @dlmillimet @PHuenermund) Great blog to teach us how "real world
econonomists" think and work. Perhaps it is a good way of slowly introducing econ. readers to advanced
methods, using graphs. E.g, the "IV with Endogenous Control" is solved here https://ucla.in/2MqNBRU Others in
http://bayes.cs.ucla.edu/csl_papers.html

6.8.19 @2:49am - (Replying to @mimblewabe @viktorklang) Doubly agree. And this calls for a theory of
explainability, namely, what is it about the CONTEXT of the conversation that makes us interpret "why" one way
or another, say, "what for" vs. "by what means" #Bookofwhy

6.7.19 @7:09am - The closest would be the talk I gave at USC a few months ago:
https://www.youtube.com/watch?v=ZaPV1OSEpHw&t= ... Some poetry, but mainly somber causal wisdom.
#Bookofwhy @thinkmariya

6.6.19 @11:46pm - @_MiguelHernan A different take on Cochran's contributions to #observational studies is


available here: https://ucla.in/2mohEis , which emphasizes his "explanation of the mechanism by which the
effect is produced" rather than "imitating" RCT's, which #Bookofwhy advises against.

6.6.19 @9:40pm - (Replying to @rakotesh) The flyer says: CIO UCLA Health Sciences IT TOWN HALL
Friday, June 7, 2019 10:00 AM to 12:00 PM
Pauley Pavilion, 301 Westwood Plaza
Arena Floor / East Bleachers
Enter through the NORTH EAST doors.
I will ask them about streaming, but you may have more clout.

6.6.19 @7:11pm - @EPIDEMIOLOGYY @mendel_random Now that epidemiologists are proficient using DAGs to describe
selection bias:
https://journals.lww.com/epidem/Fulltext/2019/05000/Selection_Bias_When_Estimating_Average_Treatment.8.aspxthey
... they will be happy to know that methods of repairing such bias are also available at the nearest DAG's brewery:
https://ucla.in/2OgMF4g , #Bookofwhy

6.6.19 @2:33pm - (Replying to @Abraham_RMI @_MiguelHernan @davidlederer) Not reporting would have been
an option a few decades ago. Not these days, when we can explicate the assumptions behind our ES estimates
d b it th t f th t If b ti XYZ tb l i ABC C diti l
and submit them as part of the report: If you buy assumptions XYZ, you must buy conclusions ABC. Conditional
conclusions have value #Bookofwhy

6.6.19 @2:13am - If you happen to be at UCLA on Friday, 11 am, you are invited to attend a talk of mine

https://ucla.in/2Mw7lZM in which I will try to summarize #Bookofwhy to health scientists, in 20 minutes of


slides and poetry. See you there.

6.6.19 @1:33am - (1/2) Glad you plan to include precise conditions in your paper and, hopefully, relate them to
identification theory. For example, https://ucla.in/2QpSZb9 gives examples of when surrogate experiments can
decide Q and when they can't. It would be nice to simulate your method
6.6.19 @1:33am - (2/2) on such examples and see if the existence of h coincides with identifiability of Q (using
the corresponding interventional data). Another interesting comparison would be to see how your h works when
Q is identified from the (set of) graphs inferrable from the data #Bookofwhy

6.5.19 @11:36pam - (Replying to @mekarim7) Thanks, this was fixed already in our errata sheet.

6.5.19 @11:07pam - Suppose our target quantity is Q=P(y|do(x)), how would we get it from h(x)? Moreover,
suppose our datasets are P_i(y|x1,..do(z1..) for many many i's, x's and z's, Is Q necessarily identifiable from the
available datasets? What happens if it is not? #Bookofwhy

6.5.19 @6:01pm - (Replying to @wgeary) The most inhumane atrocities in the history of mankind were
committed by those who saw humanity on one side, and one side only. Now watch BDS puppets chanting:
"human rights", "human rights", they can's see humans on the other side. Puppets can't see. They can only chant.

6.5.19 @4:15am - You are not alone, Petros. It is really the best introduction to causal inference that I know.
And even the fear of sounding self-promoting will not deter me from saying so. One must be honest when so
many communities are still under the boot of traditional education. #Bookofwhy

6.5.19 @2:04am - (1/2) Colleagues keep asking for my opinion on Bottou's method of revealing how the world
works: https://medium.com/mit-technology-review/deep-learning-could-reveal-why-the-world-works-the-way-
it-does-9be8b5fbfe4f ... I refer them to my May 12 Tweet, where I pleaded with followers to explain how he
extracts invariances using training, but to explain it in the language
6.5.19 @2:04am - (2/2) of conditional, or interventional probabilities, P(y|x,do(z)...) , because, all we can get
from training are such probabilities. So far, I have not received an explanation, so I remain speechless and
mighty curious. Does anyone understand it? #Bookofwhy

6.5.19 @12:45am - @jjz1600 Hi James, Excuse my innocence, but what exactly is wrong with the NYT AD? I was
there, in 1948, and I can verify every sentence in the AD as factual and well documented. I would welcome the
opportunity to tell you personally what Tlaib's ancestors did to mine. @DrMikeH49

6.4.19 @9:02pm - (1/2) My puzzle is: How many of those econ seminars are aware of the fact that, in addition
to "touching" on issues, a large chunk of them can actually be "solved"? I am not asking this question to prove
anyone wrong. I REALLY DO NOT KNOW. It is already a year after #Bookofwhy, so
6.4.19 @9:02pm - (2/2) I am sure some econ. students are doing more than "touching", perhaps even with
encouragement of their professors. But I really do not know the extent of this enlightenment. I don't see it in
the universities that I watch, and I have no way of judging, unless you help me.

6.4.19 @4:34am - I got a review of Ian Stewarts forthcoming book Of chaos, storms and forking paths: How does
statistics help us to understand the world? by Andrew Gelman https://www.nature.com/articles/d41586-019-
01680 M t l ti t ti ti d " d t di th ld" t diff t thi #B k f h
01680-y ... My natural reaction: statistics and "understanding the world" are two different things #Bookofwhy

6.4.19 @3:15am - @jo_mendelson Glad you posted this quote from


https://jewishjournal.com/news/los_angeles/299482/adl-ucla-hillel-call-on-ucla-to-issue-a-more-full-throated-
response-to-anti-israel-anti-semitic-incidents/ ... It is so easy for university administrators to hide behind
"academic freedom" and theological definitions of anti-Semitism, and forget that thousands of creative
academics have been criminalized on their campus

6.4.19 @2:50am - Remember Ang Li paper on how to select customers (or patients, or voters) that are most
likely to respond to your request/action, though "responsiveness" is an unobserved, counterfactual notion? A
revised and improved version is now posted here: https://ucla.in/2EgRzvl #Bookofwhy

6.4.19 @12:27am - (Replying to @yudapearl @dbweissman) 2/2 namely, in econometric, but are afraid to do so
even when needed? Or, to put it more mildly, what economics department is likely to prepare students who
KNOW how to solve such problems? Does #Bookofwhy underestimate what these students actually know?
Genuinely curious.

6.4.19 @12:14am - (Replying to @dbweissman) 1/2 OK, I am willing to reconcile with the existence of the gap
between what people know and what they must do by their "fields". Fine. Can we conclude then that in the top
economics departments in the US, students DO KNOW how to solve elementary problems in causal analysis

6.3.19 @5:00pm - (1/3) #Bookofwhy from the viewpoint of an enlightened economist. As you can probably
guess, I am particularly interested in your comment: "His [Pearl] grasp of what economists, for example,
understand ..
6.3.19 @5:00pm - (2/3) and don't understand about causal relationships is incomplete,". What is it that
economists DO understand and that I assumed they DON'T? Has this part of "econ. understanding" been
expressed formally in the econ. literature since Haavelmo and Cowell's Com.? Can economists
6.3.19 @5:00pm - (3/3) solve the toy problems posed to them here: https://ucla.in/2mhxKdO ? I am genuinely
trying to understand what they know that they are laboring to hide from us. E.g., Do they know which
parameters can be identified by OLS? Which models have testable implications? etc. etc,...

6.3.19 @12:51am - (Replying to @BD_Zumbo @AlinaVdav) I salute your courage in joining commonsense, it is a
dangerous road, as you must have felt already, but your students will thank you forever for saving them from
outdatedness. #Bookofwhy

6.3.19 @12:30am - (Replying to @chandra1250) I really do not know, I am just a scribe. Perhaps the publisher
knows.

6.3.19 @12:28am - (Replying to @KumailWasif) Thanks for you kind words. I will keep writing, but I need your
help in expanding its potentials beyond my limited horizon. Thanks #Bookofwhy

6.3.19 @12:22am - (Replying to @NandoDF @dlowd and 5 others) I have not read Pinker's new book, but I
thoroughly enjoyed his "Enlightenment Now". Perhaps because I am a born optimist, or because it brought to my
attention data and insight I was not previously exposed to. #Bookofwhy

6.3.19 @12:17am - Your recommender must have been an extremely keen book-reviewer. I happened to have
read this book already and, strangely, I enjoy reading it again and again. No dog in the fight, just the pleasure
of a child listening to Aesop's fables for the tenth time. #Bookofwhy
6.3.19 @12:00am - (Replying to @riazshahzain1) Thanks for writing. And may that it inspires you to write an
even better one.

6.2.19 @11:57pm - If I have encouraged even one professor/student at NYU to voice his/her knowledge and
convictions about Israel, or to speak against the way Zionism is maligned on some BDS-occupied campuses, I
would consider this Award the greatest honor of my life. Thanks!

6.2.19 @1:02pm - (Replying to @RevDocGabriel) Yielding to friends' wisdom. Will move to "incomplete". Thanks

6.2.19 @2:09am - (Replying to @DanielWhibley @EpiEllie @HarvardChanSPH) Note the difference. In #Bookofwhy
you think: "who listens to whom?" In @HarvardChanSPH you think: "What if I intervene?". Luckily the difference
does not affect practice until you get a variable like "Blood-pressure" in your model. Now begins the fun:
https://ucla.in/2LY8M2X

6.2.19 @12:44am - (Replying to @VDimitrakas) Willing to change adjective. "primitive"? "incomplete"? "limited" ?
"insufficient"? "weak"?. Open to suggestions.

6.2.19 @12:28am - Unfortunately, what I said about statistics was misconstrued to mean "bashing" or
"contempt". It was not. It was an objective assessment of what can be achieved with the impoverished language
that statisticians had to work with. I hope #Bookofwhy is taken this way.

6.1.19 @10:13pm - I've never seen a book so cheap -- 2 pounds on Kindle -- a real bargain. I assumed readers
lost interest, but it is still #1 in "probability and statistics" (after everything I said about statistics!) and "AI and
Semantics". No explanation! Causal or otherwise #Bookofwhy,

6.1.19 @3:43am - Another supplement to Didelez etal paper on Mendelian randomization can be found in this
re-posted blog discussion: http://causality.cs.ucla.edu/blog/index.php/2019/06/01/graphical-models-and-
instrumental-variables/ ... It took place in 2014, when Bryant and Elias attempted to show Imbens how DAGs
can help him live a fuller IV-life @mendel_random #Bookofwhy

5.31.19 @11:48pm - Some readers of Didelez etal paper http://ncbi.nlm.nih.gov/m/pubmed/17715159/ are


surprised by the absence of "d-separation" from the discussion. No need to worry; it is substituted by a
surrogate called "Moralized graph", the way it was first proven in Europe by Lauritzen etal (1990)

5.31.19 @11:11pm - It all started when Nancy Pelosi elevated the bigotry of Ilahn Omar to the pulpit of the US
Congress and no Democrat dared ask her to consider the consequences. I tried
https://jewishjournal.com/news/nation/295245/judea-pearl-calls-on-pelosi-to-remove-omar-from-foreign-
affairs-committee/ ... Left unchallenged hate is infectious and my fellow Democrats are now stunned helpless.

5.31.19 @12:56am - (Replying to @Basucally @edwardhkennedy) Not merely "to define", but also to derive
properties that follow immediately from the definition. For example, consistency, ignorability, testability etc,
See https://ucla.in/2QpcGzS #Bookofwhy

5.31.19 @12:49am - This paper by Didelez etal is indeed a VERY good start. It is written in DAG's language and
contains extensions to non-linear systems. It need be supplemented with new results on "conditional IV" and
"Instrumental Sets", plus specific tasks where MR must listen to DAGs #Bookofwhy

5 30 19 @5 31 (R l i t @ k i 7) Th k f t hi thi t #B k f h
5.30.19 @5:31am - (Replying to @mekarim7) Thanks for catching this typo. #Bookofwhy

5.30.19 @5:27am - It turned out to be more educational (for me) than I thought, and I will soon look into
whether Mendelian randomization can be the next beneficiary of DAG's power. Bullet proof vests are back in the
closet, awaiting a "mostly harmless" economist. #Bookofwhy

5.29.19 @4:24am - (Replying to @maximananyev @rodrikdani) I believe our conclusions are almost identical.
Though I have hard time assessing the over all benefit of the "credibility revolution," as I explain here:
https://ucla.in/2mhxKdO . #Bookofwhy

5.28.19 @6:37am - (Replying to @bnielson01) The memory and traces of alternative options are missing from
this definition. Does a thermostat have free-will? It seems to satisfy your "error correction" definition.

5.28.19 @4:27am - Finding yourself in this stellar company makes you really humble. What? Who? Me? You must
be kidding!! #Bookofwhy

5.28.19 @2:14am - (Replying to @strangecosmos) Agree, yet moralists jump out of their wits if you ask them to
accept that free will is just an illusion. This is why the ancients insisted on saying: Its not an illusion, its
actually there, 'freedom of choice is always granted'. It seems to clash with determinism -- no more.

5.27.19 @11:37pm - (Replying to @zakkohane @segal_eran and 3 others) Feeling jealous being unable to be with
you at the Weizmann Institute (I spent one sabbatical there) and discuss new ideas on AI/Big data with
colleagues and students. Have a great meeting.

5.27.19 @11:08pm - (Replying to @strangecosmos) No difference, assuming that the organism that does not
have free will has the computational facility to generate the illusion of having free will.

5.27.19 @7:56pm - It is program synthesis with an additional ingredient: the synthesizer module leaves traces in
short term memory, so that it can go back to where it was before the incident, re-simulate itself, and perform
the repair in accordance with the instruction. #Bookofwhy

5.27.19 @4:28pm - This is how "You should have slowed down" should be interpreted. The key point is, "I, the
teacher, do not know your software, so I cant tell you exactly what to tweak, whatever it is, make sure that
after tweaking, the resulting action should compute to "slow down". #Bookofwhy

5.27.19 @4:27pm - (Replying to @bnielson01) This is how "You should have slowed down" should be interpreted.
The key point is, "I, the teacher, do not know your software, so I cant tell you exactly what to tweak, whatever
it is, make sure that after tweaking, the resulting action should compute to "slow down". #Bookofwhy

5.27.19 @1:22pm - Fixing the program would fix the criminal robot, but will not tell others what was wrong with
the offensive program. When we advertise widely "for bad behavior of type-I" an entire community of robots
undergoes fixing; namely, reassignment of priorities to software #Bookofwhy

5.27.19 @1:09pm - It is the illusion of free will that allows us to supplement punishments with specific repair
instructions, eg. "You should have slowed down" , as if the robot had an option to act differently. It's an
effective communication trick [ Chapter 10, #Bookofwhy]

5.27.19 @2:33am - Readers remind me that this is precisely the motto quoted below the title of chapter 10 of
p y q p
#Bookofwhy. It reads: "All is pre-determined, yet permission is always granted." -Maimonides (Moshe Ben
Maimon) (1138-1204). Note that he does not pose it as a dilemma, but as a fact.

5.27.19 1:28am - (Replying to @mendel_random @UCLA) Barring quantum noise, I am planning to be there. And
if you have a list of discussion items, it will give me a chance to prepare the right bullet-proof vest. In

gratitude, I will hand you a signed copy of #Bookofwhy and a personal invitation to join the "causal revolution."

5.26.19 @7:52pm - (Replying to @hlprmnky @TheAnnaGat) Godel is right! Our robot cannot have a detailed
wiring diagram of its software, but this does not preclude having a blue-print of its software, which may be
sufficient for conducting free-will flavored conversations, as we do with the rough blue-print of our
software.#Bookofwhy

5.26.19 @6:25pm - (Replying to @TheAnnaGat) Once we understand the neural wiring behind the FW illusion
there will be no FW problem. The sentence "I have an option" will be interpreted in terms of the wiring, an
interpretation that precludes having an option. #Bookofwhy

5.26.19 @6:03pm - In the old days the FW dilemma was: "God can predict what you are going to do, how can he
punish you for what you did?" Modernity replaced God with the equations of physics, messed with quantum
uncertainties. Our problem is a robot which surely "has no FW", should we punish HIM?

5.26.19 @4:43pm - (Replying to @jrwill9) Compatibilists (like me) believe causation IS compatible with free
will. The question is, does it make sense to tell a robot (driven by deterministic algorithms): "You SOB, you
ought to have known better" same way as we talk a child who, presumably, does have free will.#Bookofwhy

5.26.19 @7:50am - (Replying to @chandra1250) I would not be surprised if it has. But I would love to see a
reference. I did not call in "emergent property" because of interest to us is the CLASH between two levels of
description, not the appearance or disappearance of a property. #Bookofwhy

5.26.19 @6:55am - (Replying to @DebyNavarroR) Zionophobia is an obsession, not a mode of thinking. The only
logic it recognizes is the mechanics of slander. They are incapable of realizing that the fate of real human
beings may be affected by their barking.

5.26.19 @6:32am - (Replying to @KerrShip) I dont take those attacks to be "nasty" or personal. Zionophobes
truly believe they have a monopoly on truth, human rights and social justice. They are deeply shocked to find a
thinking organism challenging their bubble. It was important to expose their moral blinders.

5.26.19 @6:19am - An interesting take on free-will. https://mindmatters.ai/2019/05/younger-thinkers-now-


argue-that-free-will-is-real/ ... They are distorting somewhat what I say, but no harm. I say "there is no free
will but it is beneficial to have this illusion and to act as through it is real". They say: "Pearl says there is free
will" Not much different

5.26.19 @5:25am - (Replying to @Spinozasrose) One good thing about Zionophobic posts, they sometime link to
gems of articles. I have found this gem of Amos Oz https://www.vox.com/2014/8/2/5962103/amos-oz-to-
israel-s-critics-what-would-you-do ... referenced in an article by a Zionophobic supremacist writer named
Litvin. Its worth reading to empower the armies of sense and co-existence .

5.26.19 @4:35am - (Replying to @joshua_saxe @DebyNavarroR) An important clarification is needed: by "Jewish"


we mean Jewish in the peoplehood sense, not religion. Otherwise the Zionophobes will jump at you with: "Aha,
you want a theocratic state! Like Iran or Egypt!" Luckily the majority of Israelis are secular, bonded by heritage.

5.25.19 @10:45pm - (Replying to @droverbytrade) Are the actual slides available?

5.25.19 @9:14pm - Murray Gell-Mann (Nobel 1969) died yesterday at 89.


https://www.sciencenews.org/blog/context/murray-gell-mann-gave-structure-subatomic-world ... I happened
to meet Murray twice and, inspired by the IC algorithm, he showed keen interest in causal discovery. He was
fun to be with, always curious and always offering an opinion, often blunt #Bookofwhy

5.25.19 @7:05pm - (1/3) As I tweeted earlier, I must terminate this strange discussion with enemies of co-
existence who are genuinely SHOCKED to hear that some people hold different views on Zionism. This is typical
of people in the self-righteous bubble of the far-far left. They see themselves as
5.25.19 @7:05pm - (2/3) God's anointed priests of human rights and anti-racism. So, when someone reminds
them that other people have rights too, and that denying those right is RACISM, they undergo a traumatic
mental SHOCK, "We? Racist?" they ask, "Unheard of!" Someone must tell them:
5.25.19 @7:05pm - (3/3) "Yes, look yourself in the mirror!" I hope the mirror convinces them. In the meantime, I
going back to science, but not before offering readers another glimpse at "Who is indigenous?" here:
https://jewishjournal.com/opinion/226997/balfour-declaration-100-redefined-indigenous-people/ ... and
liberal definitions here: http://bayes.cs.ucla.edu/DPF/anti-zionism-on-campus-jp-ch16.pdf ...

5.25.19 @4:27pm - Seeing that you are starting to put ugly words in my mouth, I have to cut off this
conversation. For the record, what I have articulated is that Zionism is a home-coming endeavor - a restoration
of human rights to indigenous people: https://jewishjournal.com/opinion/14791/yom-haatzmaut-2007-israel-
means/ ... https://www.meforum.org/2001/early-zionists-and-arabs ...

5.25.19 @4:22pm - (Replying to @SamerAbdelnour @Vieroe @jvplive) Seeing that you are starting to put ugly
words in my mouth, I have to cut off this conversation. For the record, what I have articulated is that Zionism is
a home-coming endeavor - a restoration of human rights to indigenous people;
https://jewishjournal.com/opinion/14791/yom-haatzmaut-2007-israel-means/ ...

5.25.19 @3:59pm - (Replying to @SamerAbdelnour @Vieroe @jvplive) I dont think you read carefully what
Zionism is. It is POLITE to let people who uphold an ideology define what it is, not those who labor to defile it.
An ideology is defined by the way it is taught in kindergartens, not the way diplomats vote, then change their
votes, then.....

5.25.19 @3:42pm - (Replying to @KerrShip) The settlement are built on wheels -- some will be uprooted, and
some will remain as tolerated Jewish minority in a Palestinian State, once an agreement is reached. The
obstacle is the agreement, not the settlements.

5.25.19 @3:29pm - (Replying to @jmourabarbosa) Sorry to disappoint you. But being a scientist means being
honest, worship truth, and constantly tune to new evidence. It may disappoint people who have formed an
opinion by reading hateful literature about Israel, but I was there, I am well informed and craving for co-
existence

5.25.19 @3:12pm - (Replying to @SamerAbdelnour @Vieroe @jvplive) You are right! BDS is not anti-Semitic, it is
Zionophobic, a more dangerous form of racism: See why: https://jewishjournal.com/columnist-2/editors-
note/298066/why-anti-zionism-is-more-lethal-than-anti-semitism/ ... and also:
https://jewishjournal.com/opinion/11481/ Note: No BDS supporter has ever said: "I am not anti-Zionist" so, the
label "Zionophobic' should be a badge of honor
5.25.19 @7:49am - (Replying to @Vieroe @SamerAbdelnour) You forgot that I am also a student of
counterfactuals, so lets do a little exercise: Assume the occupation is lifted today, do we have any evidence

that Palestinian's violence will subside rather than increase, given their improved position and what they
promise their children?

5.25.19 @7:39am - (Replying to @SamerAbdelnour @mszargar @AsraNomani) They tell me that German's
fatalities in WW-II were even more lop-sided. What does it prove? That the maximizer fails to accomplish its
aims? Oh, I forgot to ask: can you now say: "equally indigenous"? In a hundred years from now? Two hundreds?

5.25.19 @7:28am - (Replying to @wgeary) That Israel tries to minimize fatalities is clear even to her enemies,
otherwise they would not use civilians as human shield. After all, even if Israelis are inhuman monsters,
fatalities are bad for public relation. Lets stop this nonsense, I know Israel, and you know it too.

5.25.19 @7:15am - (Replying to @SamerAbdelnour @mszargar @AsraNomani) When moral grounds are shifting
we invoke "international law". Surely, no one is obligated to accept anyone's neighbor. But those who cry
"oppressed" could strengthen their case by showing some commitment to a permanent peace, once oppression is
ended, by agreement.

5.25.19 @6:58am - (Replying to @SamerAbdelnour @mszargar @AsraNomani) If I were to smell racism it would
be in the glaring asymmetry between: "equally indigenous" on one side and "Me, Me, Me" on the other, followed
again by empty slogans from BDS dictionary.

5.25.19 @6:46am - (Replying to @wgeary) Body-count says nothing about the conflict, especially when one side
tries to minimize fatalities and the other brags on maximizing them. In statistics we call this a typical case of
"selection bias". If you seek peace, support the minimizer, not the maximizer.

5.25.19 @6:32am - (Replying to @belial42) Never in human history was a nation threatened with extinction
called an "oppressor". India never questioned the legitimacy of Great Britain, nor did Algeria threaten the
sovereignty of France. Eliminationism has its price.

5.25.19 @6:21am - (Replying to @SamerAbdelnour @mszargar @AsraNomani) Calling people "settler colonialists"
does not make them less indigenous than your family. Those "colonialists", remember, recite poetry written in
that land, in the language spoken in that land, by heroes ruling that land. By dismissing their claims you
diminish yours.

5.25.19 @6:05am - (Replying to @belial42) If I were under the boot and I knew that one word would set me
free, I would say the word, even if I dont mean it. Trouble is, its hard to cheat, because if you really believe
that Israel is temporary, you can't tell your child "Israell is permanent." Its tough to be eliminator.

5.25.19 @5:51am - (Replying to @Vieroe @SamerAbdelnour) Nice theory, except it does not match with facts.
The besiegement started in 1936 (no occupation) when Haj Amin al Husseini declared a boycott on Jewish
products, and a genocide by entrapment on European Jews seeking refuge from Hitler (My grandparents were
among the entrapped).

5.25.19 @5:38am - (Replying to @KerrShip) How about one nation globally? How about just US and Mexico, share
the land with equal rights to all, and happiness ever after? HMM! Now we begin to see some difficulties. Well,
multiply those difficulties by 100 and we end up with "two states for two peoples". i.e., Zionism-101
5.25.19 @5:01am - (Replying to @Vieroe @SamerAbdelnour) So let us work together to allay this perception of
threat. I will do it by telling my Israeli friends "It is not 150,000 rockets, it is only 130,000" and you do it by
saying aloud: "equally legitimate and equally indigenous ". Do we have a deal?

5.25.19 @4:52am - (Replying to @SamerAbdelnour) All people, oppressed and not-oppressed, must accept each
other right to freedom and dignity. Those who deny freedom to their neighbors must accept some responsibility
for the consequences. What I hear on Twitter however is continuous one-side denial, "Me, Me, Me," not one
"Us".

5.25.19 @4:35am - (Replying to @SamerAbdelnour @Vieroe) Truth can be shocking, agree. But anyone who
aspires to co-existence needs to examine the state of mind of both sides, including Israelis, 95% of whom
believe they are under siege by, first, rejecting neighbors, second, 150,000 Hezbolla rockets, third, Iranian
proxies, more?

5.25.19 @4:22am - (Replying to @belial42) Sorry, but if you use populist slogans such as "apartheid, shooting
children, bulldozing," it is hard for me to believe that you are genuinely "interested" in a comparison. It sounds
like you really think Israel "shoots children" for pleasure. This is BDS thinking, tell us more.

5.25.19 @4:05am - (Replying to @Vieroe) I will describe it by, first , avoiding populist slogans (eg oppression,
violation, illegal) which connote sadistic intents and, second, a temporary and unwanted predicament imposed
on Israel by neighbors who openly promise her demise if she withdraws.

5.25.19 @3:50am - (Replying to @belial42) It is not a matter of "being nice". It is a matter of "as soon as they do
not wish us dead, and openly say so". Some difference!

5.25.19 @3:46am - (Replying to @Vieroe) The Israeli government is a product of 71 years of beseigement under
existential threats. "Give us one year of normalcy," say my Israeli colleagues "and we will show you what our
government can do.

5.25.19 @3:35am - (Replying to @mszargar @NYTimesCohen) I have met literally hundreds of BDS supporters.
None of them can utter the words "equally legitimate and equally indigenous". It is against their DNA. I even
challenged some of them to

5.25.19 @3:28am - statehood once they grant this right to their neighbors. My friends in Israel are tuned to
Palestinian schools and mosques for an INKLING of acceptance -- none thus far, just plans of elimination.

5.25.19 @3:06am - (Replying to @SamerAbdelnour @mszargar @AsraNomani) This is what BDS calls for:
http://bayes.cs.ucla.edu/DPF/anti-zionism-on-campus-jp-ch16.pdf ...

5.25.19 @3:52am - (Replying to @SamerAbdelnour @mszargar @AsraNomani) I was right. You cannot bring
yourself to say the words "equally indigenous" and you admit that SJP and the entire Palestinian leadership still
deny the right of my people to a homeland in any borders. This is what I meant by "victim's cover." A cover for a
plan of elimination.

5.25.19 @2:01am - (Replying to @SamerAbdelnour @mszargar @AsraNomani) The principles are simple: "Two
states for two peoples, equally legitimate and equally indigenous". Your saintly "marginalized, non-violent
movement" (SJP) will not be opposed if it adopts these principles. But it can't, because it was created to fight
( ) pp p p p , g
them under victim's cover.

5.25.19 @1:17am - What has BDS to do with Zionism? Everything! See http://bayes.cs.ucla.edu/DPF/anti-

zionism-on-campus-jp-ch16.pdf ... What is Zionism? The right of a people to a homeland -- in some borders. A
right denied to my people by every BDS supporter. A genocidal denial that started the conflict 72 years ago, and
remains its core.

5.24.19 @7:06pm - (Replying to @causalinf @PHuenermund) I am speechless. What is the research question?
assumptions? data?

5.23.19 @10:42pm - (Replying to @robertwplatt) Hard to implement. A review may be 3-5 pages, and all you
want to tell the world is that seasoned reviewers in the 21st century still think that "causation is just a species
of correlation" or that "confounding is a statistical concept" etc. History needs to know that. #Bookofwhy

5.23.19 @10:28pm - If Israel is so evil, why shouldn't SJP be supported? NYU campus needs to hear that
President Hamilton's objections to BDS and SJP are based on moral principles and shared liberal values, not on
donors or alumni pressure. Details are here: https://jewishjournal.com/news/nation/299053/nyu-president-
calls-bds-supporting-graduation-speech-quite-objectionable/ @AsraNomani

5.23.19 @9:27pm - Every time someone posts this speech I feel an urge to share it with others, finding myself
re-identifying with these ideas, and with the realization that, unless re-enforced, they tend to fade into the
background. I worked a lot on this commencement speech - my first. #Bookofwhy

5.23.19 @4:16pm - I propose to re-run our poll, now that we have discussed the pro and cons of publishing juicy
reviews of submitted papers.
Poll:
-- Is it ethical to publish selected extracts from anonymous reviews of your submitted paper?
72% yes
28% no 407 votes. 5 hours left

5.23.19 @9:03am - (1/1) This is amazing!! Thanks for finding it. I can't believe that econometric students would
sit through this class and not laugh how hard the instructors sweat to avoid commonsense. I Imagine one of
these students meeting an epi friend in the library, or reading #Bookofwhy .
5.23.19 @9:03am - (2/2) or hearing about d-separation, and asking: "Wow! And all this sweat just to prove that
economists can do things independently and differently, home- grown style? It is still great progress for eco.
education compared with PO w/o dags. See https://ucla.in/2L8OCyl

5.23.19 @7:08am - (Replying to @JicMic) Truth and justice trump slogans and herd mentality.

5.23.19 @6:53am - (1/2) Thanks for posting this confession. I have felt the same when I had to renounce my
Award. I have also come to realize that NYU administrators do not know how to handle the monster created on
their watch, and no one tells them how. They dont realize that all they have to do is
5.23.19 @6:53am - (2/2) to tell the campus the reasons why Zionist students and faculty are welcome on
campus, the distinct contributions they make to the cultural tapestry of NYU and the inspirational power of
Israel's miracle on other minorities. They just need to be honest and tell the truth.

5.23.19 @4:09am - Sorry, typo! It is a Rung-1 task on the Ladder of Causation (see #Bookofwhy). Diagnosis goes
from evidence to beliefs not caring whether the beliefs are about causes (of the evidence) or about
consequences thereof.
consequences thereof.

5.23.19 @3:52am - An amazing investment in health science and data science: https://www.jpost.com/Israel-
News/Israel-announces-NIS-60-million-investment-in-precision-medicine-research-590176 . I am mighty proud to
see how the super-education I received in Israel is now spawning pioneering projects of such magnitude to
benefit science and mankind. #Bookofwhy

5.23.19 @3:36am - The paper aims at finding minimal Boolean functions that separate positive from negative
cases. A Rung-2 task. The function found explains why your diagnosis algorithm concluded that you have disease
D. It does not explain why you have disease D or what would cure it. #Bookofwhy

5.23.19 @12:29am - (1/2) (Replying to @theblub @twainus and 10 others) Let me end with one more remark, to
rule out any impression that I do not accept metaphysical questions as legitimate. Take the question: "what is a
cause". It is legitimate if you allow me to translate it into "how can we explain the consensus among Sapiens
about certain
5.23.19 @12:36am - (2/2) (Replying to @yudapearl @theblub and 11 others) cause-effect relationships. This now
allows me to further translate the question into a computational one: Find a parsimonious mental
representation of shared knowledge that explains how humans can access that representation and swiftly come
up with same answer. #Bookofwhy

5.23.19 @12:13am - (Replying to @Kleee @DorotheaBaur) They played this role effectively until mid 20th
century. They lost it when they rejected computational metaphors from entering the language of dialogue. My
humble prediction: Searle and Dreyfus objections will become curious anecdotes in the history of AI and HAI.
#Bookofwhy

5.22.19 @10:38pm - (1/2) (Replying to @theblub @twainus and 10 others) 1/2 I was not asking for a book or a
paper, but for an idea or a method. Whether an idea is useful or not depends on what your question is. eg.
Algebra is useful if my question is: what values of x satisfy certain conditions. The question "What is taste" on
the other hand
5.22.19 @10:44pm - (2/2) (Replying to @yudapearl @theblub and 11 others) does not strike me as a legitimate
question because I do not know what kind of answer would satisfy the questioner. Take: "taste is what unicorns
swear by" or "taste is a combination of neural activation that result in a scream 'SPICY'", which would be more
satisfying? Why?

5.22.19 @7:04pm - (Replying to @SiaKordestani @SFSU) No! Times are changing. It is not "anti-Semitism seeking
to cloak itself as political discourse". Rather, it is Zionophobia seeking to cloak itself as normative anti-
Semitism".

5.22.19 @1:30am - (Replying to @theblub @twainus and 10 others) The way I speak may betray years of
frustration to understand this literature. But if you examine my work your will see that I have never turned
away a useful idea because of a mind made up (at least I do not recall such mishap). What's the idea I cannot
afford to miss?

5.22.19 @1:17am - (Replying to @Plinz) Everything could be part of higher education if so presented. But
accusing you of child molestation in a class on human sexuality is not exactly conducive to higher education.
Abdulhadi did precisely that, and we, normative Zionists, now request to be publically decriminalized.

5.22.19 @12:56am - (Replying to @theblub @twainus and 10 others) My language is: English, math, physics,
programs and human cognition. Using this languages please tell me one idea that I should learn from Dreyfus ,
without which I am going to waste lots of time on dead ends. No ref to other philosophers please, no acronyms
nor Heidegger
gg

5.22.19 @12:17am - (Replying to @theblub @twainus and 10 others) I am dying to learn from them, but I have
not found anyone who can translate them to a language I understand. In particular, to convey an idea from
beginning to end without quoting five other phenomenological philosophers, so as to make you feel an outsider.
Substance please.

5.22.19 @12:08am - (Replying to @Plinz) Naive it may be, but it need to be aspired to, and if no one says the
word "racist" students may get the idea that Adulhadi is a normal person, perhaps even an "educator", so
perhaps I and my Zionist colleagues are indeed "white supremacists" . Sorry. She is just a Zionophobe.

5.21.19 @11:47pm - (Replying to @Plinz) It is not "sacralization" but common decency. There is such a term in
English called "racism", which is sometimes used improperly and sometimes properly. University administrators
often condemn hate speeches that take place on their watch, to set the norms right. No sacralization

5.21.19 @11:28pm - (1/2) Readers ask if I was not too harsh calling Abdulhadi "racist". My answer: What would
you call a guest lecturer who comes to your university saying "Muslims are terrorists, but I have nothing against
Muslims who disavow Mohammad". Evidently, archaeology professors think it's
5.21.19 @11:28pm - (2/2) "educational". The fault is not really with the instructor, but with a university that
does not internalize the equation: Zionophobia = Islamophobia, which means that religion does not have a
monopoly on human identity, and all symbols of identity should be equally respected.

5.21.19 @10:56pm - The Dreyfus video reminds me of Marvin Minsky who once said: When I hear someone
saying: "A computer cannot do Y", what he is really trying to say is: "I havn't got a clue what Y is". I will change
my mind if I someone translates Dreyfus into English, math or program. #Bookofwhy

5.21.19 @6:19pm - Replying to @twainus @yhazony and 9 others Any idea when this video was taken? Dreyfus is
repeating the same ancient arguments that I heard him state in the 1970's. Is he back in fashion? #Bookofwhy

5.21.19 @3:40am - This is amazing! So access to anonymous reviews IS available from some journals. Victory!! I
feel less inhibited to start my juicy selection. #Bookofwhy

5.21.19 @12:08am - (1/3) Our poll shows a slight preference - 58/42 - to allowing publication of juicy quotes
from anonymous reviews. I have hoped for a more decisive preference. Aside from the entertaining value of
such a collection, and its encouraging effects on young researchers,
5.21.19 @12:08am - (2/3) I am concerned with its historical value. Written under the shield of power and
impunity reviewers comments are the most honest and faithful reflections of the state of mind of a scientific
community at any given period. Such historical treasure should not be allowed to rot
5.21.19 @12:08am - (3/3) in the archives of outdated journals. There should be at least some status of
limitation before unveiling this information to the public. Anyone knows what happens to these archives? Can a
historian request access to the reviews of Turing's paper of 1937 ?? #Bookofwhy

5.20.19 @6:57pm - For all who requested a copy of my contribution to the book "possible minds" (edited by J.
Brockman, 2019), I now have a link to it: https://ucla.in/2wj4pox So let us continue our voyage, from Babylon
to Athens. #Bookofwhy

5.20.19 @4:02pm - (Replying to @68kirk) Judea Pearl Retweeted Judea Pearl Causality is all about invariance,
of course. But the relationships between these and machine learning should be articulated this way:
5.20.19 @1:47pm - My! My! Thanks for posting this picture --it made my day. I have enjoyed telling students my
personal story on surviving 1948. Too bad I did not meet Miss Iraq, my wife is from Baghdad. Go on spreading
the word about an inspiration called Israel. Happy Birthday!

5.20.19 @1:32pm - (Replying to @adviceonstock @IanCero) I have a strong issue with this statement. Machine
Learning and Deep Learning models do NOTsupport casual inference, unless by "support" one means "can be
used in", like "arithmetic supports classical mechanics." See https://ucla.in/2HI2yyx #Bookofwhy

5.20.19 @3:00am - (Replying to @glarange72) CI is Causal Inference, and by "CI + ML hype" I mean the noise
generated by the many who claim to be combining the two when they can barely do one. #Bookofwhy

5.19.19 @10:57pm - If you are inspired, as I am, by new perspectives on causal data-science, marinated in
substance and free of CI+ML hype, you will enjoy this lecture by Elias Bareinboim which was delivered at
Columbia last month http://www.cs.columbia.edu/streaming/2019-Spr/elias_bareinboim.mp4 Recommended
to readers of #Bookofwhy

5.18.19 @8:25pm - (Replying to @maximananyev) Terrific PhD topic: Write a program that automatically
generates such a news article. Input: voting data + polls + voters model. Output: a post-election article on "why
the polls got it so wrong". Criterion for PhD: article accepted for publication in a top newspaper. #Bookofwhy

5.18.19 @10:45pm - (1/2)


For many years now I have been playing with the idea of publishing a selection of juicy reviews that my papers
have received from anonymous, self-assured reviewers. Some colleagues advised against it, since it violates
reviewers' trust in eternal concealment. Here is a poll
5.18.19 @10:45pm - (2/2)
1. As a reviewer, would you mind finding portions of your review in print (anonymously, of course)?
2. As an author, do you think this would improve quality and accountability of reviewers?
33%(1) yes; (2) yes.
09%(1) yes; (2) no.
41%(1) no; (2) yes.
17%(1) no; (2) no.

5.18.19 @6:47pm - (Replying to @aesopesque) I do not think risk-averseness would change any of the results; we
can always assume that benefits are measured in "utiles" instead of "dollars". #Bookofwhy

5.18.19 @4:23pm - Good luck for the 2020 Eurovision, and happy singing. From Israel (2019), with love, to The
Netherlands (2020).

5.18.19 @1:14am - How do you select customers (or patients, or voters) that are most likely to respond to your
request/action when "responsiveness" is an unobserved, counterfactual notion? A new paper by Ang Li proposes
a method. https://ucla.in/2EgRzvl #Bookofwhy

5.15.19 @11:43pm - (Replying to @WLoosen @SZ) @Wloosen, Thanks for posting. Funny, I received the Dutch
translation of #Bookofwhy in the mail today, and could not tell if they mean it. And saintly Mother Theresa on
this article made me fall in love again. I love Dutch, the language of Huygens and Till Eullenspiegel

5.15.19 @10:10pm - When I first posted this birthday note, readers asked: Who are the 1948-deniers? Today we
have a super-denier in Congress, Rashida Tlaib, who never heard Azam Pasha call for "a war of extermination
and momentous massacre" against Israel; Sec. Gen of Arab League, Oct. 11, 1947
5.15.19 @7:46pm - Carlos etal have a new paper on sensitivity analysis https://ucla.in/2LHhi66 which, to the

best of my knowledge, is the first to recruit graphical models into this struggling enterprise. I expect to see
soul-searching among traditionalists and awakening in the field. #Bookofwhy

5.15.19 @3:38pm - (Replying to @AndersHuitfeldt) How about this: https://ucla.in/2vYMCm7

5.15.19 @2:02am - I was sent a new article on Evidence-Based policy.


https://zimbra.cs.ucla.edu/service/home/~/?auth=co&loc=en_US&id=701690&part=2 ... Perhaps it can
enlighten readers to tell us where this enterprise stands in the microcosmos of causal inference. #Bookofwhy

5.15.19 @12:38am - (Replying to @DickeySingh) Sure! All we need now is to understand how you decide that the
pressure causes the barometer reading, not the other way around. It may be possible, but to understand what's
behind it, we need an explanation cast in conditional probabilities, not in training algorithms.

5.14.19 @11:27pm - Kudos to Maayan Harel @maayanvisuals for reminding us of the #Bookofwhy anniversary,
and for inspiring us with her eye-opening illustrations. May 15 is also the general calendar birthday of Israel,
Maayan's (and my) birthplace - We have much to celebrate tomorrow. Ko Lechai

5.14.19 @5:38pm - (Replying to @mrgunn @vgr @StatModeling) 1.8.19 @11:59pm -- Gelman's review of
#Bookofwhy should be of interest because it represents an attitude that paralyzes wide circles of statistical
researchers. My reaction is now posted on https://bit.ly/2H3BH3b Related posts: https://ucla.in/2sgzkPZ and
https://ucla.in/2v72QK5

5.13.19 @2:45am - (Replying to @HenningStrandin) Correct. But this does rule out "probabilities of
counterfactuals" since those individuals may have different behaviors, with a distribution over the behavior
types.

5.13.19 @2:15am - (Replying to @eadeli) Wait, wait, you did not get to the part where I encourage students to
rebel against their textbooks. This will surely make you angry. #Bookofwhy

5.13.19 @1:58am - (1/ ) I was delighted to read your Causal Ladder blog-post, especially the way you explain
the necessity of the counterfactual layer and the vivid examples you used. (I literally forgot the great party we
had with Ann and Bob). A word about the exogenous variables U:
5.13.19 @1:58am - (2/2) These variables specify a "unit", be it an individual, an agricultural plot, time of day,
etc, whatever refinement is needed to make all relationships deterministic. I hope this clarifies the dilemma
posed in your last paragraph. #Bookofwhy

5.12.19 @3:39pm - I must fully endorse this recommendation . I did not realize how much I owe to my training
in physics (eg., https://en.wikipedia.org/wiki/Pearl_vortex ...) until I had to face the model-blind thinking of
modern statistics, and the difference between statistical and causal models. #Bookofwhy

5.12.19 @2:40pm - (1/ ) Communication between CI and ML folks will improve drastically if we can translate
sentences such as: "Bottou trains his NN under conditions ABC" into sentences of the form: "Given the
conditional probabilities P(y|x,do(z)...)". After all, what do we get from "training" if ..
5.12.19 @2:40pm - (2/ ) if not conditional probabilities, both observational and interventional. Another benefit
for the translation: theorems of impossibility. CI has developed a theory that tells us if certain tasks can be
accomplished given information in the form of probabilities P(y|x, do(z)...
accomplished given information in the form of probabilities P(y|x, do(z)...
5.12.19 @2:40pm - (3/3) We can use this theory to prevent disappointments from "training" schemes that lead
to impossibilities. As far as I know, theories of what's possible or not possible were not developed (yet) for
training schemes. Why not use what we have? eg,https://ucla.in/2Jc1kdD #Bookofwhy

5.11.19 @3:32pm - (Replying to @isli_amar) Your proposal is sincere and well-intended, but it crumbles against
the logic of Algerian children who are asking: If Israel is a colonial endeavor (like French rule in Algeria) why
should it be recognized in ANY border? Indeed, WHY? Are you prepared to tell them the truth?

5.11.19 @1:48pm - (Replying to @_fernando_rosas @KordingLab) Terrific question!! (which every student want,
but is afraid to ask). Answer: We learn quantitative "effect size", P(y|do(x)), while before we had only
qualitative information, eg. "X does not listen to Y", or P(x|do(y))=P(x). #Bookofwhy

5.11.19 @1:41pm - (Replying to @nlpnyc) "1948-deniers" are authors (eg Gelvin) who deny that the 5-army
attack on Israel in 1948 was GENOCIDAL IN INTENT: "a war of extermination and momentous massacre which will
be spoken of like the Mongolian massacre and the Crusades" (Azam Pasha, Sec.Gen. Arab League, Oct. 11, 1947)

5.11.19 @1:18pm - (Replying to @_fernando_rosas @shell_ki) The observed data is just "observed data" ie, a
bunch of correlations among variables. It is generated by causal relationships, yes, but it does not tell us what

those relationships are, unless we assume extra-distributional assumptions (eg graph) called "causal model"
#Bookofwhy

5.11.19 @3:32am - (Replying to @isli_amar) Yes. I remember 1967. The world was waiting for a plausible peace
process when, on August 29, the Khartoum Arab League Summit issued its famous: "Three No's"; No peace with
Israel, no recognition of Israel, no negotiations with Israel. They can still reverse it-- we are waiting!

5.11.19 @12:26am - Had a great celebration yesterday of Israel's Independence Day - speaking to students as a
"1948 eyewitness", feeling like an endangered species, and thinking gloomily: who will bear witness when my
generation is gone and the professional 1948-deniers take over?

5.10.19 @11:46pm - When I was a student, speaking against #racism was an obligation, not "courage." It's not
funny, the change happened on our watch!!

5.10.19 @2:28pm - (Replying to @sweichwald @atypical_me and 3 others) Your definition is a good one. But I
would like to embrace partially specified models under the label "causal model" . For example: X----->Y + X<---
U--->Y which does not identify P(Y|do(X)) is still a causal model; it tells us P(X|do(Y)) = P(X) #Bookofwhy

5.10.19 @1:40am - (Replying to @eigenhector @ericjang11 and 8 others) In CI, we call it "functional model" or
"completely specified structural model", where the response of each listening variable is functionally defined. In
such a model all counterfactual queries, conditional as well as marginals, are estimable. Why invoke quantum?
#Bookofwhy

5.10.19 @12:57am - (1/ ) (Replying to @ericjang11 @eigenhector and 8 others) In CI we classify problems along
3 dimensions: 1. what we know, 2. what we wish to know, 3. what type of data sources we have. We then ask:
Can we obtain (2) from (1)&(2)? We try to avoid giving agency to algorithms and acronyms. For example,
"Model-based RL" is too vague for
5.10.19 @1:03am - (2/ ) (Replying to @yudapearl @ericjang11 and 9 others) evaluation, because its capabilities
depends on the kind of "models" invoked, and because you can do everything with a highly refined "model".
Whether RL and its varieties can accomplish one task or another depends on (1) and (2) see
Whether RL and its varieties can accomplish one task or another depends on (1) and (2), see
https://twitter.com/yudapearl/status/1065867272072249344 ... The best way to
5.10.19 @1:12am - (3/ ) (Replying to @yudapearl @ericjang11 and 9 others) communicate about capabilities is
to use canonical examples. For example, how would acronym ACR handle the napkin problem? or Joe's "would
be salary" problem? Is the output guaranteed to be consistent? In these canonical examples, (1) and (3) are
formally specified,
5.10.19 @1:17am - (4/4) (Replying to @yudapearl @ericjang11 and 9 others) as in https://ucla.in/2Jc1kdD ,
leaving no room for ambiguity, thus enabling us to determine which query can be answered from a given
combination of knowledge and data #Bookofwhy

5.9.19 @10:50am - (Replying to @theAlexLavin) Great way to think, agree. But what principle need we assume
to show that y=f(x,e) is more "invariant" than say x=g(y,e').

5.9.19 @10:40am - Not so. Not when two sides agree to a win/win sharing partnership. Besides, we have not
heard this equation on Cinco de Mayo, last week. Reminds me of a parody I wrote on editors who feel the urge
to spoil birthdays with equations: https://www.haaretz.com/1.5002611 Happy Birthday Israel!!

5.9.19 @2:45am - (Replying to @PHuenermund) My, My! I did not realize Israel shares a birthday with EU, both
determined to prevent another genocide, and both succeeding thus far, to some extent. I wish however that
Israel's peace would depend on something so tangible as coal trade.

5.9.19 @2:31am - (Replying to @BreyonWilliams) Congratulations, Breyon. I usually add: well deserved, after
reading one's dissertation but, in your case, I am willing to bet it is. So, welcome to PhD-land, and I hope you
revolutionize econometrics.

5.9.19 @1:23am - Today, May 9, is Israel's Independence Day. I invite all readers to join me in celebrating the
71st birthday of the country where I learned to speak, and thanking her for what she has contributed to
mankind, and for redefining the meaning of "miracle." Happy Birthday Israel !!!!

5.9.19 @12:14am - (Replying to @arih1987 @Stanford and 2 others) The trick is not to dis-invite racist speakers
but to politely explain to them why they should dis-invite themselves, as I tried to do with Cornell West
https://jewishjournal.com/opinion/judea_pearl/170625/open-letter-cornel-west/ ... The host, the audience
and the public got the idea, and some say the speaker got it too.

5.8.19 @10:41pm - (Replying to @miquelporta @socestadistica and 5 others) @miquelprta, what brings you to
mention the NYT review of #Bookofwhy? True, it is one of the best written reviews, but how is it related to p-
values and to other topics discussed there, on Mt. Olympus?

5.8.19 @4:36pm - (Replying to @AngeloDalli @ylecun) A free (signed) copy of @Bookofwhy to the first person
who extracts the principle from the equations.

5.8.19 @3:36pm - (Replying to @babeheim @MPI_EVA_Leipzig) I vow for all equations, Keep us informed when
you dig into counterfactuals. Have fun!

5.8.19 @3:28pm - (Replying to @ach3d) good choice!!

5.8.19 @3:27pm - (Replying to @AngeloDalli) Did you get the principle? I can't get to the paper itself, can you?
Invariance is always at the heart of causation,; do we have a new method of interrogating the invariance? Let
k if di i
me know if you dig it.

5.8.19 @3:15pm - (Replying to @NevinClimenhaga) Interesting abstract but it hides the principle/assumption till
the thick of the paper. Can you summarize it for us in Tweeter length? As I tried to do here
https://ucla.in/2umzd65 , Tool # 7: causal discovery. #Bookofwhy

5.8.19 @4:39am - (Replying to @JDHaltigan @juli_schuess @doinkboy) I sure do, but not as a stand-alone
argument. Why? Because to model-blind researchers "confounding" is just "nonignorability" which is defined
relative to a "treatment," and requires no notion of "cause". See how Imbens and Rubin labor to explain
"unconfoundedness." #Bookofwhy

5.8.19 @2:21am - (1/ ) Depressing to see a once friendly campus (UCSD) consumed in slander. As I note here:
http://bayes.cs.ucla.edu/DPF/anti-zionism-on-campus-jp-ch16.pdf ... -"in the grand opera of BDS's slander
machine, it is not the libretto that matters but the stage and the megaphone. The charges may vary from
season to season,
5.8.19 @2:27am - (2/ ) the authors may ' rotate, and it matters not whether a resolution passes or fails, nor
whether it is condemned or hailed. The victory lies in having a stage, a microphone, and a finger pointing at
Israel saying, "On trial!" It is only a matter of time before innocent students
5.8.19 @2:27am - (3/3) mostly the gullible and uninformed, will start chanting, "On trial!" It worked in Munich,
and it has worked on some campuses. The effect will be felt among the next generation of policy makers.

5.7.19 @11:20pm - (Replying to @JDHaltigan @juli_schuess @doinkboy) I love this argument, but I dont think
manipulationists like Hernan will buy it. They would say: OK,"U causes Y" colloquially, but thou shall never say
"causal effect of U." Sounds a bit inconsistent? Agree. But we have not heard from them since
https://ucla.in/2sASU9V #Bookofwhy

5.7.19 @6:51pm - Faculty of New York University: Oppose Academic Boycott of NYU Tel Aviv - Sign the Petition!
http://chng.it/HZsJnJcX via @Change

5.7.19 @6:33pm - (Replying to @djinnome @juli_schuess @sweichwald) Thanks for noting the break. We will fix
it soon (I hope).

5.7.19 @1:47pm - (Replying to @martin_garcia_a @_MiguelHernan) Fine, but I am questioning the benefit of
separating "description" from "prediction", skipping "diagnosis" and lumping together "intervening" and
"retrospecting" under one opaque category "causal inference".

5.6.19 @8:56pm - (Replying to @thehuntinghouse @IsraelCampus and 6 others) So is the racist mentality of BDS
cronies, an extremely interesting object of academic study. A newly evolved aberration.

5.6.19 @8:31pm - As I wrote to NYU president Sexton here: http://bayes.cs.ucla.edu/DPF/anti-zionism-on-


campus-jp-ch16.pdf ... "When a group of self-appointed vigilantes empowers itself with a moral authority to
incriminate the academic activities of their colleagues, we are seeing the end of academia..."

5.6.19 @8:17pm - I was happy to read today that a petition to restore commonsense to NYU has gathered over
3,000 faculty, students and alumni signatures. https://www.change.org/p/faculty-of-new-york-university-
oppose-academic-boycott-of-nyu-tel-aviv ... Refreshed by this response, I feel proud again being a NYU alum.
https://www.newsweek.com/nyu-tel-aviv-campus-petition-1415921

5.6.19 @1:47pm - (Replying to @ehud) Worrall lives in the pre-causal era (ie, probabilistic causality) and so, it
@ p ( py g @ ) p ( ,p y) ,
does not satisfy my curiosity: What is "evidence-based-Med"? Is it a QUEST for principles, or a SET of principles?
If the latter, is there a simple example where I NEED one such principle?.#Bookofwhy

5.6.19 @4:07am - (Replying to @XiXiDu @wtgowers) We have to take into account that some observers, even in
academia, have not watched the "peaceful demonstrators" in action, and truly believe that Israelis shoot

civilians for sport. We tend to underestimate the power of Hamas propaganda because it sounds so absurd, but
it works

5.6.19 @3:38am - (Replying to @wtgowers) De-legitimization is a very good predictor of how one would act
given the chance. The fact it has not diminished over the past 83 years means that one is very very serious
about acting, given the chance.

5.6.19 @1:31am - (Replying to @ArashBroumand) You keep mentioning "politics" and "rhetorics" where I see
none. I see humanity and I see human rights and I see a conflict that can be resolved when each side
acknowledges the human rights of the other. This makes me an optimist, because I have found one such side
already.

5.6.19 @12:58am - (Replying to @ArashBroumand) I wish I could share your optimism. Unfortunately, I know
that peace can only prevail if both sides agree to the "equally indigenous" principle. I also know that the vast
majority of my Israeli friends agree to it - curious what you hear from your Palestinian friends. Optimist?

5.5.19 @11:30pm - (1/ ) (Replying to @mohomran) "Being in a position" is an interpretation. The fact is that
Palestinians have been denying Israel's right to exist RELENTLESSLY, for 71 years, 24/7. From school-teachers to
TV anchors from journalists to imams from academicians to intellectuals. These are the people whom my
5.5.19 @11:48pm - (2/2) (Replying to @yudapearl @mohomran) friends in the Israeli peace camp are tuned to.
None came forth with the word "coexistence", or "equally indigenous". Why 71? It is now 83 years of Arab "me,
me, me!" rejectionism. Wait!, how about an intellectual like yourself, can you say the words "equally
indigenous"?

5.5.19 @9:53pm - (Replying to @Doc_Yemen @aynumazi) We both know that Hamas/Rashida want more than
just "basic necessities". We both read the Hamas Charter (Israel destruction), and we know that Rashida can
never say the word "co-existence." We also know that Gazan can have all necessities were it not for what
Hamas/Rashida want.

5.5.19 @8:34pm - (Replying to @Doc_Yemen @aynumazi) I was there, in that counterfactual world. Jerusalem
was caged in 1947, no food no water and we were promised "momentous massacre" by the Arab League. Still, all
we asked our neighbors (+ UN) was a legitimate co-existence for two indigenous peoples. Not denial of the
other!!!

5.5.19 @7:08pm - (Replying to @aynumazi) I am VERY serious, and I know something about DAGs and about the
logic of cause and effect. So you are invited to join me in examining the logic of Goliath/Rashida/Hamas as they
deny their neighbor the very rights they demand for themselves -- freedom and self-determination.

5.5.19 @6:08pm - When will the world learn that Palestinians' conception of "freedom" is somewhat different
than the ordinary. It entails freedom to de-legitimize the freedom of their neighbors.. A strange conception
indeed, but some think it nevertheless deserves the Hall of US Congress.
5.5.19 @5:48pm - Goliath are those who deny their neighbors what they demand for themselves: right to self-
determination.

5.5.19 @4:54pm - When will the world stop dehumanizing Goliath who just want to be free, and who just
happened to forget that David and his shepherd brothers likewise, just want a day for freedom.

5.5.19 @2:06pm - (Replying to @nathansttt @ehud) Reading your wife's overview I get a strong suspicion that
EBM is an emotional appeal for principles by which we can integrate epidemiological studies with informal
opinions of individual physicians. Has it advanced beyond the appeal?

5.5.19 @1:53pm - The more I look at Hernan's classification of data-science: {description, prediction and
#causalinference} the more I prefer the Ladder of Causation {Association, Intervention and Counterfactuals} as
in #Bookofwhy. Perhaps b/c #causalinference is becoming "what everyone is doing"

5.5.19 @12:57am - (Replying to @ehud) I have never been able to figure out what "evidence based medicine" is
all about. Perhaps an expert can explain? Is there a Medicine not based on evidence? Are there requirements on
the methods? #Bookofwhy

5.4.19 @10:21pm - I totally agree with @_miguelHernan that the competition devised by DORIE et al. (2019) to
compare "methods for causal inference" provides no information on "methods for causal infererence" and should
have been titled "methods for estimating certain formulas" #Bookofwhy

5.4.19 @9:27pm - (Replying to @jim_adler @Toyota_AI_VC @PitchBook) I am trying to understand the task that
the CAT is asked to perform. Anyone can explain to the uninitiated?

5.4.19 @9:23pm - (Replying to @joaoeira @gwern @Jabaluck) We tried a blog-based discussion on the same
issues with Guido Imbens, see http://causality.cs.ucla.edu/blog/index.php?s=imbens ... -- to no avail. I actually
prefer Tweeting because it forces you to cut the baloney. #Bookofwhy

5.4.19 @7:58pm - (Replying to @MariaGlymour @danielwestreich @UCSF_Epibiostat) @MariaGlymour You say: "If
effect modifiers are differentially distributed, trial effect estimates won't match the target population effects."
Is this a new complication that is not captured by the standard conditions of https://ucla.in/2N7S0K9 or
https://ucla.in/2L6yTzE ?

5.4.19 @7:34pm - (Replying to @chophshiy) Anyone doing causation outside "conventionally qualifying
institution" deserves a free copy of #Bookofwhy. If you can drop by my office at UCLA it will be waiting for you.
[My sec. is on vacation this week, try next one]

5.3.19 @2:50pm - (Replying to @_julesh_) Not clear to me why the last item is "theory" instead of
"counterfactuals". The first two items specify what we can do, not how we do it, shouldn't the 3rd also tell us
what we can do with the "theory" that we cannot do with the other two.?? #Bookofwhy

5.2.19 @10:47pm - Confession: the juicy stuff in #Bookofwhy was partly inspired by the juicy stuff of modern
academia and its submission to egos and clans. Funny, graphs are still prohibited in certain departments, and
top PO researchers can't do the homework problems in https://ucla.in/2G2rWBv

5.2.19 @7:14pm - (1/ ) (Replying to @IsraelCampus @nyuniversity and 5 others) This is a perfect time for
President Hamilton of NYU to defend the Study Abroad program on moral grounds and expose the hypocrisy of
the department of Social and Cultural Analysis by stating: "A country whose existence is under daily threats
the department of Social and Cultural Analysis by stating: A country whose existence is under daily threats
cannot be expected to invite in
5.2.19 @7:16pm - (2/2) (Replying to @yudapearl @IsraelCampus and 6 others) people who openly seek its
destruction. I [Hamilton] advise members of the SCA department to spend their energy in support Israel's right
to exist before criticizing her protective laws or policies."

5.2.19 @2:32am - (Replying to @omaclaren @bjh_ip) This distinct notation is useful, because it allows us to
distinguish "; theta)" from "|X=x)". In the former, theta can be an arbitrary index; in the latter, X=x must be an
event in our probability space. #Bookofwhy

5.2.19 @12:50am - (Replying to @bjh_ip) Relatedly, most modern Bayesians define their craft as that of
assigning priors to parameters and computing their posteriors. Rarely do they examine Bayes' original dilemma:
How do we express probabilities that we WISH to estimate in terms of those we CAN estimate. #Bookofwhy

5.1.19 @4:09pm - (Replying to @bjh_ip) Modern Bayesians would do well to take a second look at Bayes' original
paper, this time from a causal inference perspective, and follow the evolution of the epistemological term:
"Given that we know". I believe Chapter 3 of #Bookofwhy does a good job of presenting this idea.

5.1.19 @2:48am - (1/2) (Replying to @AlexMGeisler @clibassi) Such a course is urgently needed for data
scientists as well. The reason I am calling attention to the foundations of counterfactuals is that I see even
noted champions of @causalinference frequently abandoning those foundations. For example, the principles
discussed here
5.1.19 @2:53am - (2/2) (Replying to @yudapearl @AlexMGeisler and 2 others) https://ucla.in/2G2rWBv are
hardly recognized by "experimental" economists, or decorated integrators of ML and CI, and I am still not sure
about Harvard epidemiologists. Supreme Courts need clarity, cohesion and consensus. #Bookofwhy

4.30.19 @6:40pm - While debating Gorshuch on how causes and #counterfactuals DONT WORK, let's take a look
at how they DO WORK, and what it takes to extract them from regression. Here is a gentle introduction, as
harmless as they come: https://ucla.in/2G2rWBv . #causalinference, #Bookofwhy

4.29.19 @6:53am - (Replying to @yoavrubin) Thanks, but don't forget the Primer https://ucla.in/2KYYviP ,
which stands between #Bookofwhy and Causality, and which I continue to recommend for people who want to
DO #causalinf rather than talk ABOUT it.

4.28.19 @4:28pm - (Replying to @causalinf @agoodmanbacon) Mistakes like this can make history. The more I
hear of such mistakes, the more I am tempted to forgo modesty and recommend this book honestly and in the
strongest possible terms: Read it! For fun and insight! Here is Chapter 4 for a bait https://ucla.in/2G2rWBv
#Bookofwhy

4.28.19 @2:26am - (1/ ) I have decided to retweet my last reply because the distinction between the
methodological "science of adjustment" and the substantive "science of diseases" seems to be vulnerable to
ongoing confusion. Moreover, one cannot over-emphasize the miracle of the former. Not only are
4.28.19 @2:26am - (2/2) the assumptions qualitative, they are also meaningful and natural, namely, judgments
about what variables determine the value of another are the easiest ones for a domain expert to articulate and
communicate. Conclusions: adjustment is not an "art," it is a science.#Bookofwhy

4.28.19 @1:51am - (Replying to @EngineerDiet @TuckerGoodrich and 4 others) It is healthy to separate the
"science of adjustment" from the "science of diseases". The former is valid whenever your assumptions about
the former are valid. Moreover, and miraculously, the assumptions needed are qualitative: who affects whom.
The rest is algebra #Bookofwhy
The rest is algebra. #Bookofwhy

4.27.19 @3:07am - (1/ ) (Replying to @sifogrante1) You are probably right about the effectiveness of MOOC
format. However, we lack the administrative/institutional support needed to launch it. All we have is truth and
commonsense, and we trust those who audit MOOCs to compare what they learn to modern ways of doing
things,
4.27.19 @3:15am - (2/2) (Replying to @yudapearl @sifogrante1) and share their experience with us on Twitter.
In particular, check if the tools taught are compatible with CI as defined in
https://twitter.com/yudapearl/status/1120163925109035009 ... or, better yet, if they permit you to solve the
toy problems of https://ucla.in/2G2rWBv - my ultimate litmus test. #Bookofwhy

4.26.19 @4:10am - I was sent this illuminating talk by David Gross (Nobel, Physics) on the meaning of "truth" and
the scientific method. https://www.youtube.com/watch?v=FdljxE-WPo4 ... The first part classifies scientific
questions as in the Ladder of Causation, and should be recognized by readers of #Bookofwhy.

4.26.19 @4:02am - (Replying to @a40ruhr) Bidirected arrows do not create cycles, the simply state that the
correlation is created by some hidden common cause. Still, I would turn this arrow into unidirectional and pose
it as a homework for IV enthusiasts. #Bookofwhy

4.26.19 @3:45am - (Replying to @a40ruhr @USF_Economics and 2 others) I fail to detect cyclicity. Can you point
to it? I think the structure is ideal to serve as a homework in modern econ. textbooks, asking students to
identify the causal effect using IV method. This should lure "experimental economists" to catch up with
modernity. #Bookofwhy

4.25.19 @11:17pm - I am intrigued by your DAG, can you flesh it out so we can see the labels? It seems to
demonstrate that, contrary to Yale's economists, even simple IV cases need DAGs for identification. IOW, I dont
see how Yale students would identify it using "mostly harmless" p.85 #Bookofwhy

4.25.19 @7:48pm - The article you posted realizes that data analysts need to: "understand the data from a
wholly new perspective". Fortunately, this realization is taking place while the field is still young, so re-thinking
will be less traumatic than, say, statistics or economics. Hope!#Bookofwhy

4.25.19 @1:24pm - (Replying to @Soroush_Saghaf @JoshuaSGoodman @oziadias) Eager to learn from your
teaching experience. What, in you opinion, was the most useful tool, or concept, that econ. students learned
from #causalinference, which they could not get from econ. textbooks, say Angrist or Wooldwridge. #Bookofwhy

4.25.19 @3:10am - (Replying to @HenningStrandin) It is definitely non-manipulationist, and that is why


Rubin&Co +Harvard&Co refuse to use it, despite the fact that it gives meaning to counterfactuals, which they
do use, but only when enslaved to a "TREATMENT" ; they are missing out on "undoing of past events. "
#Bookofwhy

4.25.19 @2:56am - (Replying to @yskout) The beauty of this formalism is that, for certain questions (eg
interventions), all you need is the graph structure; for others (eg retrospective counterfactuals) we need the
functional form as well. Its a miracle, dont miss. See https://ucla.in/2N9f28c #Bookofwhy

4.25.19 @12:41am - Here is a more intuitive definition of Structural Causal Model (SCM). It is a society of
LISTENING variables, and a specification of how each variable would respond to what it HEARS from the
neighbors. Though somewhat informal, I find it more meaningful than the formal.#Bookofwhy
4.24.19 @5:19am - (1/ ) (Replying to @totteh) Agree with the wisdom, and agree with the pioneering value of
this book. However, one weakness is that it does NOT give an example where POs and DAGs are used to
describe the same causal structure. More severely, POs and DAGs are treated as two separate frameworks. It
would be
4.24.19 @5:30am - (2/2) (Replying to @yudapearl @totteh) helpful to show readers that the two emerge from
the SAME mathematical object, Structural Causal Model, as is shown for example in Section 4.2 of
https://ucla.in/2G2rWBv , especially "The Fundamental Law of Counterfactuals" Eq. (4.5). #Bookofwhy

4.24.19 @5:06am - (Replying to @nagpalchirag) If you show me one example where identifiability is easier to

see in PO, I will convert to voodoo. #Bookofwhy.

4.23.19 @9:20pm - Sad news. Nils Nilsson passed away http://robotics.stanford.edu/~nilsson/ An AI pioneer,
and a mentor to many of us since the 1970's. Always encouraging and always insisting on understanding new
ideas, and how they fit together in the grand scheme. I will miss him immensely.

4.23.19 @8:58pm - (Replying to @nagpalchirag @DaphneKoller) First time I hear someone saying so. Could you
share with us the first place where you found PO illuminating do-calculus. Eager to learn. #Bookofwhy

4.23.19 @4:53am - (Replying to @dlmillimet @PHuenermund @causalinf) Personal choices -- I will salute to
that!! As long as we know what the options are. And, to accelerate progress, we can let our students too see
what the options are. #Bookofwhy

4.23.19 @3:20am - (Replying to @DorotheaBaur @bobehayes) Do you think Kalev Leetaru (Contributor) read
#Bookofwhy? Worth checking.

4.23.19 @3:09am - (Replying to @ingorohlfing @DToshkov) The do-calculus is "type"-level, and its relation to
Woodward's interventions is discussed here: https://ucla.in/2MnawNT

4.23.19 @3:01am - (Replying to @DToshkov) "Intervention" is type level, eg. careless driving causes accidents.
Counterfactuals are 'token' level. eg "this accident would not have occurred had you driven carefully." Both are
handled harmoniously in the SCM framework, for philosophers to rejoice #Bookofwhy

4.23.19 @2:49am - (Replying to @smhall97 @alienelf @elanvb) This is a beautiful paradox, thanks for reminding
me. In my former life I solved it using Bayesian analysis. Today I would approach it as a causal problem and
model the process by which the envelops are stuffed; the prob. of doubling A is not 1/2 but varies with A.
#Bookofwhy

4.23.19 @12:08am - (Replying to @PHuenermund @dlmillimet @causalinf) Cruel? Please compare to modern
treatments of measurement error here: https://ucla.in/2LcpmHz and here https://ucla.in/2N9icIX , in both
clarity and generality. Cruel??? #Bookofwhy

4.22.19 @11:02pm - (Replying to @UusitaloLaura @KirsiNorros @teemu_roos) I have not seen #Bookofwhy in a
picture frame before. Glad it was translated into Finnish. Happy voyage to you.

4.22.19 @9:51pm - (Replying to @jkrt168t) What makes experimental economists strangers in CI-land is the fact
that they do not start with the available causal information and use it (with data) to answer causal questions.
Instead, they start by applying a pre-canned procedure to the data and then ask: Does the estimate
4.22.19 @10:55pm - (2/2) (Replying to @yudapearl @jkrt168t) obtained have "causal interpretation?". Or, "Under
4.22.19 @10:55pm (2/2) (Replying to @yudapearl @jkrt168t) obtained have causal interpretation? . Or, Under
what conditions can we interpret the estimate obtained as a valid answer to our question?" Those conditions are
then articulated in opaque language (eg ignorability) far removed from ordinary scientific thinking. #Bookofwhy

4.22.19 @4:26am - (Replying to @RevDocGabriel) I am not familiar with multi-omics data.But I can refer you to
integrating data from heterogenous sources involving, for example, diverse populations, diverse samples,
experimental and observational data etc. See https://ucla.in/2Jc1kdD or #Bookofwhy chapter 10, pp 350-358.

4.21.19 @11:06pm - (1/2) For AI researchers still immersed in the debate between model-based vs. model-blind
AI, I am retweeting a response of @wellingmax to Sutton's blog. Max agrees by and large with "The Seven Tools
of CI", https://ucla.in/2umzd65 , albeit w/o noting the theoretical impediments.
4.21.19 @11:06pm - (2/2) Along a similar vein, I was asked to retweet the last line of my slides in the Why-19

symposium http://why19.causalai.net . Gladly; it reads: "Only by taking models seriously we can learn when
they are not needed." And I still vow for it. #Bookofwhy

4.21.19 @8:14pm - (1/ ) In view of persistent ambiguities regarding the definition of "causal inference" (CI) I am
sharing here the definition that has guided me successfully throughout my journeys. CI is a method that takes
data from various sources, as well as extra-data information, and produces
4.21.19 @8:14pm - (2/ ) answers to questions of two types (1) the effects of pending interventions and (2) the
effects of hypothetical undoing of past events. See Causality (2000) Chapter 1. A vivid and recurrent example of
a non-causal question is any question that can be answered from the joint
4.21.19 @8:14pm - (3/ ) probability distribution of observed variables, eg, correlation, partial regression,
Granger causality, weak and strong endogeneity (EHR 1983) etc. See https://ucla.in/2N9f28c . This definition
excludes Pearson's (1911) and Fisher's (1925) descriptions of statistical tasks
4.21.19 @8:14pm - (4/4) and I would reserve judgment on how "experimental economists" fit into this
definition. I believe that, in due time, "experimental economists" will manage to articulate formally what
"extra-data information" they use, and thus become bonified members of CI. #Bookofwhy

4.21.19 @3:16pm - (Replying to @ahmaurya) The vast majority of economists that I know would be offended if
labeled "regression analysts", in the same way that physicists would be offended if labeled "arithmeticians",
though they use arithmetical operations every hour of the day. Regressionists live and die w/o causes.

4.21.19 @1:45pm - (1/ ) (Replying to @ahmaurya) You do not understand me correctly. The economentric
literature is motivated by causal questions, and has pioneered modern causal inference. See
https://ucla.in/2mhxKdO . Traditional regression analysts, however, shun causation, which evoked my surprise
at the paper discussed
4.21.19 @1:59pm - (2/ ) (Replying to @yudapearl @ahmaurya) which starts and ends with regressional questions
and, surprisingly, invokes a causal diagram. "For what purpose?" I asked. But may I suggest that, instead of
putting words in my mouth, please articulate the research question you claimed I have been avoiding.
#Bookofwhy

4.21.19 @12:08am - (1/2) (Replying to @ahmaurya) Are you asking why I am surprised to find "regression
analysts seeking the wisdom of causal diagrams when they are not asking causal questions"? Ans. Because I have
not seen it done in the regression literature, not even in the econometric literature and its "tricks". But
4.21.19 @12:21am - (2/2) (Replying to @yudapearl @ahmaurya) I thought you have earlier accused me of
avoiding a burning research question, the answer to which is revealed in the econometric literature. Glad we
are no longer there. Or, if we are, what is that question that I avoid? #Bookofwhy

4.19.19 @11:20pm - (Replying to @ahmaurya) I wish I knew what I have done wrong to earn this Tweet. What
questions are you asking whose answers I avoid? Try me, I would love to learn from you and the books you read.
4.19.19 @4:54am - (1/3) I have read this paper with great interest, trying to understand what makes regression
analysts seek the wisdom of causal diagrams when they are not asking causal questions and labor merely to
assess the magnitude of measurement errors. The answer seems to be two fold.
4.19.19 @4:54am - (2/3) (1) The diagram allows them to use Wright's Rules https://ucla.in/2LcpmHz to
compute correlations among latent variables (X,Y) in terms of correlations among observed proxies (x',Y'). This
could be done, of course, w/o the diagram, but only at the cost of painful algebraic
4.19.19 @4:54am - (3/3) derivations, as in econ. (2) The problem is in fact causal in disguise. Why else would
anyone be interested in cov(X,Y) as opposed to cov(X',Y') which is estimable from the data and is sufficient for
all predictive tasks? Curious if other readers agree. #Bookofwhy

4.18.19 @2:09pm - (Replying to @y2silence) Amazing discovery!! Unveiling the origin of ideas. #Bookofwhy

4.18.19 @2:06pm - (Replying to @abesilbe) I just spoke with students from NYU Realize-Israel. Harrassment,
threats and intimidation is unfortunately the modus operandi of the NYU SJP chapter. Please speak to them.
Curious, what gives you the hope that they would be different?

4.18.19 @1:15pm - (Replying to @abesilbe) I fail to grasp the logic of this new theory of evidence. You attended
a symposium that was not interrupted by SJP, from which we should infer that SJP does not resort to disruption
tactics nation wide, and did not interrupt the May 17 "indigenous people" meeting at UCLA?

4.18.19 @4:33am - An update from NYU: https://nyunews.com/news/2019/04/17/nyu-president-sjp-awards/ ...


It seems that NYU administrators have discovered a new and courageous way of handling disruptive student
organizations: Give them awards and do not show up to the ceremony. Dont ask, dont tell.

4.17.19 @4:33am - (Replying to @arnoldroth @epavard) And a joyful Passover to you @arnoldroth, and to your
family. We will never forget your beautiful daughter Malki. She and my son Daniel are the torches to the light of
which this insane world may discover one day why the normalization of evil is twice as evil as evil itself.

4.17.19 @2:33am - (Replying to @HolgerSteinmetz) Strange, I feel the opposite. My impression is that, at least
on this Tweeter forum, the number of such causality-dead people is shrinking by the hour. But this impression is
infected of course by heavy (and hopeful) selection bias. #Bookofwhy

4.16.19 @4:19pm - (1/ ) This sad happening at NYU unveils the power of ignorance in the electronic age. I bet
my esteem colleagues at NYU do not know that their university is awarding a "president service award" to SJP, a
student organization that prides itself on crushing meetings of other student
4.16.19 @4:19pm - (2/2) organizations. How can you tell when your university administrators are embarrassed
by their own words? When they start lecturing you on "free speech" -- the ultimate blanket for inaction or lack
of courage. Any NYU alumni on this Tweeter?

4.15.19 @3:29am - (Replying to @AngeloDalli) Interesting. But we need to ask Miguel if this is what he meant by
"description". His examples do not match this interpretation. #Bookofwhy

4.15.19 @12:39am - In the interest of many readers of Primer https://ucla.in/2G2rWBv who requested to see a
derivation of Eqs. 4.13 and 4.14, I am re-tweeting here an earlier post by Ang Li. It may look complicated, but
it is a straightforward application of probability calculus. Thanks @Ang_UCLA

4.15.19 @11:51pm - (1/2) I believe it is a mistake to assume that business applications care only about
interventions, not about counterfactuals. An astute businessman wishes to spend his advertisement budget on
people who are "swayable potential customers" not on "captive buyers" The distinction
people who are swayable potential customers not on captive buyers . The distinction
4.14.19 @11:51pm - (2/2) between these two groups is counterfactual in nature, and requires counterfactual
logic for definition, analysis, identification and estimation. There is more to counterfactuals than meets the
eye, and it is all here: https://ucla.in/2G2rWBv , told and exemplified. #Bookofwhy

4.14.19 @3:33pm - (Replying to @DrJohnKang @cd_fuller @julian_hong) Precisely, except that we need to add:
"...would behave under conditions different from those Joe's encountered in the the trial." See
https://ucla.in/2G2rWBv . Also, the word "generalizability" is reserved for extrapolations across diverse
populations. #Bookofwhy.

4.14.19 @2:39pm - (1/ ) An important correction. I actually do NOT agree with Hernan's classification of
#datascience. First, I do not see substantive difference between "description" and "prediction'. Second, the
"counterfactual" layer should be split into two, intervention and counterfactuals, as

4.14.19 @2:39pm - (2/2) in the Ladder of Causation #Bookofwhy or https://ucla.in/2umzd65 . The reason is
that these two layers of the ladder require different types of knowledge. You can never tell if "Joe's headache
would have gone had he not taken aspirin" by conducting RCT on aspirin and headache.

4.14.19 @5:17am - (1/ ) I just finished Hernan's "Second Chance.." and I recommend it strongly to data
scientists, statisticians and even eocnomists. If #Bookofwhy has not convinced you that curve-fitting is not
sufficient for CI, Hernan's paper will. I mention economists b/c based on our Twitter
4.14.19 @5:17am - (2/ ) conversation, "experimental" economists will not buy Hernan's "the validity of CI
depends on structural knowledge". Worse yet, economists conception of ML & Econ, [Millainathan, J. Econ Pers.
cited by Hernan] confirms my suspicion that their resistance to structure is more
4.14.19 @5:17am - (3/3) than a passing fad. I differ from Hernan on the primacy of RCT as a conceptual model
for CI, https://ucla.in/2sASU9V as well as on the taxonomy of causal tasks. I am sure nevertheless that this
paper will sway more #Datascience enthusiasts to the #causalinference fold.

4.14.19 @3:21am - (Replying to @twainus @katecrawford and 17 others) I must be too late for this trolley, I
can't see how it relates to Turing machine, probabilities and artificial intelligence. Suppose probabilities were
invented to model us (the way we handle uncertainty), what then? are we closer to GAI? #Bookofwhy

4.13.19 @12:47pm - (Replying to @kareem_carr @EpiEllie) I believe that, deep in their heart, data scientists
dream of an "automated data scientist". #Bookofwhy

4.13.19 @12:39pm - (Replying to @rlmcelreath) Most profound observation. Suggested remedy: (1) Post a
"preliminary version" on your site, written for your students and friendly colleagues. (2) Cite the "preliminary
version" in your final article. (3) We, readers, will ignore the "final" and read the "preliminary". #Bookofwhy

4.13.19 @12:29pm - Note that Hernan's trichotomy is different from the Ladder of Causation in
https://ucla.in/2umzd65 , which raises 2 questions: 1) Where is "diagnosis" (or "abduction") situated? and 2)
Where is the barrier between intervention and (unit-level) counterfactuals situated?#Bookofwhy

4.12.19 @10:58pm - Gee, thanks. I have almost forgotten about this overview and, now, as I read it 10 years
later, I am amazed at how little I would change in it. Highly recommended!! Adding (1) Transportability and (2)
Missing data, and would I still stand behind it today. Amazing. #Bookofwhy

4.12.19 @4:19pm - (Replying to @cqfdee @edwardlandesber @ericcolson) Good point. I called it "a gift of the
Gods" in one of my slides. It is hard to believe that smart people like Heckman would go through all possible
acrobatics to avoid it. https://ucla.in/2L8OCyl Any theory for his motivation? #Bookofwhy
4.12.19 @3:44pm - (Replying to @elibressert @ericcolson) If you have a copy, can you share the Table of
Content? Amazon does not allow us to peek inside. Strange.

4.12.19 @12:40am - (Replying to @paulfkrause @QuantaMagazine) I am an AI researcher who discovered, in the


1990's that economists of the 1940's had some good ideas about causation which were abandoned since. I have
borrowed them for AI and tried to share with econ., but the forces of resistance turned out insurmountable.
#Bookofwhy

4.11.19 @5:19pm - (Replying to @the_dismal_tide @PHuenermund and 4 others) The real world is always beyond
us, all we can hope for is to capture faithfully our understanding of reality, which is depicted on the r.h.s DAG.

And I never resonated to Box slogan: all models are wrong. It's empty unless you can tell us WHY some are more
useful than others.

4.11.19 @12:44am - (Replying to @PHuenermund @WvanAmsterdam and 3 others) A beautiful depiction of the
whole discussion of "experimentalists" vs. "structuralists". The left part is the template the experimentalist
seeks to fit, the r.h.s is the mental DAG he/she possesses while fitting. We see clearly how the fitting process
can be automated.#Bookofwhy

4.11.19 @12:18am - @ale_martinello called my attention to an excellent 2014 article on causal inference,
history of ideas, in The New Atlantis:
https://www.thenewatlantis.com/docLib/20150110_TNA43Barrowman.pdf ... Rarely can we find such lucid
writing combined with profound understanding of this difficult topic. Highly recommended. #Bookofwhy

4.11.19 @12:04am - (Replying to @MPenikas @ale_martinello and 5 others) Agree! The New Atlantis paper is
excellent. Rarely can we find such lucid writing combined with profound understanding of this difficult topic. I
dont know how I missed it thus far. Thanks for posting. #Bookofwhy

4.10.19 @10:18pm - (1/3) Commending you on so skillfully navigating the waters of DAGs and PO. That the two
are compatible comes from the fact that both are derived from structural causal models (SCM). DAGs are used
to encode what we know and PO what we wish to know. However, I find it hard to
4.10.19 @10:18pm - (2/3) understand why you say that "PO are most useful for estimation". Assuming that we
have obtained an estimand using DAG-based identification, isn't the estimand itself sufficient for estimation? Do
we really need to dress it in PO cloths before proceeding with the estimation
4.10.19 @10:18pm - (3/3) phase? This dressing habit, I believe, is a remnant of bygone age when, lacking DAGs,
people attempted to identify queries of interest in the PO language. But why go through that tormented
experience today, when we do have DAGs.??? #Bookofwhy

4.10.19 @4:36pm - Primer's Errata has been updated: https://ucla.in/2GeKwV9 . Def. of front-door condition
corrected, to read "no backdoor path". Thank you @juli_schuess .Good catch. #Bookofwhy

4.10.19 @3:18pm - (Replying to @ale_martinello @edolaw76 and 4 others) Another Grazie !!!

4.10.19 @2:50am - (Replying to @juli_schuess @_MiguelHernan) The issue of "multiple versions" attains clarity
when cast as "disjunctive action" e.g. P(Y=y| do(X=a or X=b)) [paint the wall green or purple]. It is illuminated
in https://ucla.in/2N9nSCV and https://ucla.in/2Iz9myt , which hopefully resolve @IanLundberg dilemma.
#Bookofwhy

4 10 19 12 21 (1/2) (R l i ) Th k f i hi d h f i Th
4.10.19 @12:21am - (1/2) (Replying to @ycaseau) Thanks for posting this adventurous hunt for causation. The
only comment I have is on "Granger causality". In 1991, I had a quiet dinner with Clive Granger in Uppsala,
Sweden. Between the 2nd and 3rd glass of wine, he confessed to me that he feels embarrassed by the name:
4.10.19 @12:26am - (2/2) (Replying to @yudapearl @ycaseau) "Granger causality", since it has nothing to do
with causality, but he can't stop people from using it; they need some way to express what is needed. I think we
should honor him by echoing his distinction. #Bookofwhy

4.9.19 @12:12am - In all modesty, this is one of the most profound statement I read on Twitter. My heart goes
out to students who struggle with PO w/o first solving a toy counterfactual problem like Joe's score in Fig. 4.1
here: https://ucla.in/2G2rWBv . It borders on child abuse. #Bookofwhy

4.9.19 @12:12am - In all modesty, this is one of the most profound statement I read on Twitter. My heart goes
out to students who struggle with PO w/o first solving a toy counterfactual problem like Joe's score in Fig. 4.1
here: https://ucla.in/2G2rWBv . It borders on child abuse. #Bookofwhy

4.8.19 @8:25pm - (Replying to @annasdtc @causalinf and 2 others) Too much to read!! And not enough time!!! If
you were my student at UCLA I would tell you: "One toy problem is worth ten books - Play!" #Bookofwhy

4.8.19 @9:49am - (Replying to @TheShubhanshu @fhuszar and 5 others) Great that you are curating this
resource. Let us know how we can help you.

4.8.19 @1:31am - (Replying to @fuiud @omaclaren and 2 others) Yes, I have. Perhaps a later version. JP

4.7.19 @11:13pm - It smells like spring is rising in econometric education -- DAGs before regression. Keeping my
fingers crossed, and wishing Econ. students a safe voyage and a bright future. #Bookofwhy

4.7.19 @10:36pm - (Replying to @nickchk) The idea of postponing discussion of regression is what econ.
education has been waiting for. The confusion between structural and regression equations is THE major
hindrance for econ. students and pushing the latter to the estimation phase should be a life saver. #Bookofwhy

4.7.19 @10:25pm - (Replying to @omaclaren @eliasbareinboim @analisereal) To help me navigate through the
pages, can you tell me where I would fumble if I were to assume that all is fine and rosy in my naive conception
of identifiability and estimation. What difficulties I would end up facing, etc.

4.7.19 @10:20pm - (Replying to @pedrohcgs @omaclaren and 2 others) Is it fair to say that the functional form
of the estimand determines critical properties of the estimator, such a being regular vs irregular. ??? Or is it
some other factors in the problem, say the data?

4.7.19 @5:00pm - A tiny comment on "repackaging of things we know". What exactly do we "KNOW" that
#Bookofwhy has "repackage"? We KNOW the rules of chess. These rules dictate which board position is a WIN for
white. Do we KNOW which board position is a WIN for white?" Care to explain 'repackage'?

4.7.19 @4:46pm - (Replying to @RuochengGuoASU) Agree. Now imagine a whole book on IV, named "harmless",
with not a single DAG, and a whole movement in econometric named "credibility" cheering this book for
replacing DAGs with ignorability judgments: "Z is conditionally independent of the potential outcome of Y, had
X been....

4.7.19 @4:36pm - (Replying to @KevinDenny) Show me one who does not find it handy, and I'll show you one
h h tti dt diti l IV h t d t b bl t j d if "Z i i d d t f th
who has not tried to use conditional IV, or one who pretends to be able to judge if "Z is independent of the
potential outcome of Y had X been x, given W=w". Which one did you have in mind? #Bookofwhy

4.7.19 @4:24pm - (Replying to @y2silence @HannesMalmberg1 and 4 others) Bingo!! I was going to recommend
these two papers, and you beat me to it. Thanks @Bookofwhy

4.7.19 @4:15pm - (Replying to @HannesMalmberg1 @autoregress and 3 others) I must repeat and repeat the
CREDIBILITY aspect: It is not merely to "encourage formalization" (which is also the aim of PO, etal ) but to
encourage formalization at the level where knowledge RESIDES, as opposed to a level where knowledge is
transformed to appease ID-strategists

4.7.19 @4:02pm - (1/3) The simplicity of IV validity quickly disappears with nuances. see
https://twitter.com/yudapearl/status/1115005154074300416 ... But, a more important aspect of the

"repackaging" is CREDIBILITY, namely, judgments are recruited from where they reside, not from where they are
distorted to appease the identifier.
4.7.19 @4:02pm - (2/3) Consider the IV validity again, and ask yourself "what judgements were necessary to
execute this exercise?". Mark them. Now compare to the judgments required in the PO framework, which are
cast in ignorability language, (See Angrist etal). Finally, ask "What type of judgments
4.7.19 @4:02pm - (3/3) would be more CREDIBLE if I were the one to make them?" I am sure your assessment of
the value of "repackaging" will become one of greater appreciation, perhaps even one of necessity. #Bookofwhy

4.7.19 @2:35pm - (Replying to @HannesMalmberg1 @autoregress and 3 others) Theory is easy as long as you
have 4 variables in the canonical template. Add a nuance and simplicity is shaken. Add one "control variable" (ie
conditional IV) and theory is helpless, experts are embarrassed and must resort to: "how often has it happened
in practice". #Bookofwhy

4.7.19 @2:12pm - (Replying to @HannesMalmberg1 @Jabaluck and 4 others) Those who are less courageous
salute you.

4.7.19 @2:10pm - (Replying to @HannesMalmberg1 @Jabaluck and 4 others) Start with teaching, then time will
tell, and time will forever be indebted, till the and times are done. @Bookofwhy

4.7.19 @5:32am - Summarizing our discussion of IV's, here is how DAGs distinguish valid from invalid IV's:
https://twitter.com/yudapearl/status/1114819200143376384 ... Prudent economists may wish to use these
conditions in research, teaching and communication, then convert to a DAG-free language for publication.
#Bookofwhy

4.7.19 @2:35am - (Replying to @yudapearl @nickchk and 2 others) 2/exemplified in @nickchk DAGs, but it goes
further, to include cases with multiple unobserved variables, multiple paths between Z and Y, as well as
auxiliary "control variables". A good source for generalized IV is https://ucla.in/2MqNBRU while #Bookofwhy is a
conceptual primer

4.7.19 @2:16am - (Replying to @nickchk @Jacobb_Douglas @Fhanksalot) Discussants on this thread might enjoy
a glimpse at the general condition for Z to be a valid IV. The condition goes: 1) There is an unblocked path

between Z and X, and 2) Every unblocked path between Z and Y contains an arrow into X. This condition
confirms of course the cases

4.6.19 @4:17pm - (Replying to @EpiEllie @Jabaluck and 4 others) Judea Pearl Retweeted Nick HK And here is
4.6.19 @4:17pm (Replying to @EpiEllie @Jabaluck and 4 others) Judea Pearl Retweeted Nick HK And here is
the accompanying thread as well @nickchk https://twitter.com/nickchk/status/1111864179000934400 ... I am
dying to see an experimental economist, versed in PO language, 2sls, and "credit expansion" (but no DAG)
explain to his/her students what kind of proxies should qualify as "good IV's" . #Bookofwhy

4.6.19 @3:52pm - (Replying to @EpiEllie @Jabaluck and 4 others) Yes, that's the one. Thanks

4.6.19 @3:50pm - (Replying to @fuiud) Only DAG-averse researchers use this scare tactics, and talk about "an
enormous DAG", larger than absolutely necessary for scrutinizing the assumptions upon which the conclusions
rest. I am surprised people are still buying it. #Bookofwhy

4.6.19 @3:10pm - (Replying to @Jabaluck @autoregress and 4 others) The thread I tried to recover was much
longer, just two days ago. But we are not going anywhere on this track. I wish you could say: I, Jason, can

answer these questions correctly w/o DAG. Instead, we hear about @autoregress doing it, or "Go read about
credit expansion". I'm out.

4.6.19 @2:30pm - (Replying to @Jabaluck @autoregress and 3 others) I have tried to recover this thread, but it
disappeared under a hundred later "notifications". I believe @EpiEllie retweeted it, just two days ago. Any
recollection, @EpiEllie ? We are searching for a Twitter discussion on "Must an IV be a cause of the treatment?"
#Bookofwhy

4.6.19 @2:20pm - (Replying to @yudapearl @Jabaluck and 5 others) When I hear Mr. X saying: "A computer
cannot do Y", what Mr. X is really trying to say is: "I havn't got a clue what Y is". I differ with Minsky on many
issues, but not on this one. #Bookofwhy

4.6.19 @2:11pm - (Replying to @Jabaluck @robertwplatt and 4 others This is the first time I hear about a task
that (1) humans do, (2) computers are on their way to do, and (3) a professor says: "It can't be done." I heard it
being said about emotions, religion, free will, etc. not about a scientific task. Marvin Minsky once said:

4.6.19 @1:12pm - (Replying to @Jabaluck @robertwplatt and 4 others) Luckily, the "automated-Angrist" is on its

way, where hand-waving talks about "intuitive identification strategy" become "principled identification
strategies" #Bookofwhy

4.6.19 @1:04pm - (Replying to @Jabaluck @autoregress and 3 others) My blindness!! "discusses at length"!! And
yet, in 2019, dozens of economists argue (on Twitter) "What if I can only measure a proxy of an IV ?".. (and
other variants on the theme.) Should we send them to "any applied paper worth a salt"? I fail to name one.
Help?. @Bookofwhy

4.6.19 @12:38pm - (Replying to @Jabaluck @HannesMalmberg1 and 4 others) Beg to differ. Barring cycles, DAGs
are analogous to writing down mathematical equations for the first time! That part of the model which is not
captured by its DAG's abstraction entails assumptions that makes modelers uncomfortable, eg, parametric
forms... #Bookofwhy

4.6.19 @12:19pm - (Replying to @taragonmd @SF_DPH) Third time!!! We need to fix that!! What did we do
wrong? It was supposed to be like a jingle, you hear it once and you ham it the rest of the day. Please brief us
here, in Twitter's trenches, how the book club is going. #Bookofwhy

4 6 19 @3:41am - (1/ ) You keep sending us to decypher a subject-specific body of "applied papers" that are
4.6.19 @3:41am - (1/ ) You keep sending us to decypher a subject-specific body of applied papers that are
ostensibly "concerned" with certain questions, but where those questions are not asked explicity and where it is
very hard tell whether/how those "concerns" are pacified to the satisfation of
4.6.19 @3:43am - (2/ ) (Replying to @yudapearl @Jabaluck and 4 others) their authors. Lets take the question:
"Is X a good IV?" I have just seen a long Twitter discussion on whether a non-exogenous X can be a good IV. This
is 2019, 90 year after the invention of IV, and I do not think the applied economists who are tormented by this
4.6.19 @3:50am - (3/ ) (Replying to @yudapearl @Jabaluck and 4 others) 3/ very applied question would find
the answer in the "applied papers" you uphold as supreme oracles of economic wisdom. #Bookofwhy

4.6.19 @3:24am - (Replying to @HannesMalmberg1 @autoregress and 4 others) I did a similar experiment in
dozens of Econ departments nation-wide. My mistake was to publish the results. Do not repeat, even if you have
a tenure. Observe the ferocious efforts your colleagues are making to maintain the myth that "you can also do it
in PO". #Bookofwhy

4.6.19 @2:22am - (Replying to @TenanATC @NateSilver538 and 2 others) Tracking trolling discussions may be a
fun pass time but hardly a learning exercise. My mantra: Solving ONE toy problem in causal inference tells us
more about science than 10 debates, no matter who the debaters are. #Bookofwhy

4.5.19 @8:32pm - (Replying to @EllieAsksWhy @aesopesque) The SCM framework actually saves economists a
huge amount of effort, see https://ucla.in/2L8OCyl The arrow-phobic resistance in economics is driven mostly
by egos and insularity. It is changing now; students are beginning to see what the phobia has denied them of.
#Bookofwhy

4.5.19 @8:11pm - (Replying to @Jabaluck @autoregress and 3 others) Judea Pearl Retweeted Judea Pearl They
arise all the time. Unless practicing economists have ceased to ask themselves: (1) Is X exogeneous? (2) is X a
good IV? (3)https://twitter.com/yudapearl/status/1114108180504567808 ... namely, questions that need to be
answered from where knowledge resides i.e., DAGs, not trasformations of DAG. #Bookofwhy

4.5.19 @3:31pm - (Replying to @autoregress @eliasbareinboim and 3 others) I think what @eliasbareinboim
meant to emphasize is that the two systems are not equivalent in terms of representational transparency, ie,.
(1) deciding plausibility of assumptions, (2)their consistency (3) their completeness (4)deciding testability, etc.
After all,

4.5.19 @2:54pm - (Replying to @PauSchae @PHuenermund) True, in econ models A=B and B=A are two different
equations. However, the temptation to equate the two was so intense that generations of economists ended up
with blunder after blunder (citation withheld to spare embarrassment). Notational distinction explicates the
asymmetry.

4.5.19 @2:38pm - Good to watch all WHY-19 slides in one package. Thanks @eliasbareinboim and all speakers.
#Bookofwhy

4.5.19 @2:09pm - (Replying to @DToshkov @eliasbareinboim and 3 others) Thanks to all responders. I now
understand "agent-based models". Input: a fully specified SCM model for each bird, encoding its reaction to
environmental conditions as well as to behavior of other birds. Output: Aggregate behavior of the flock. Simple.
#Bookofwhy

4.5.19 @1:52pm - (Replying to @yoavrubin) Having survived a decade since authoring this footnote, I agree with
your assessment: "the best footnote ever". I especially like the end: "citations withheld to spare
embarrassments" , though it did not protect me from the wrath of the embarrassed. #Bookofwhy
4.5.19 @5:58am - (Replying to @robertwplatt @Jabaluck and 4 others) I thought you and Jay are
epidemiologists, contaminated by DAGs since 1995, and differing culturally from mainstream statisticians.
#Bookofwhy

4.5.19 @5:35am - (Replying to @robertwplatt @Jabaluck and 3 others) I judge by textbooks and history books
(eg. Stigler). But my last comment was meant more to economics. Statistics, I agree, has spawned some pockets
of acceptance, still not strong enough to penetrate education. As bemoaned in this interview
https://ucla.in/2v72QK5 #Bookofwhy

4.5.19 @5:22am - (Replying to @juli_schuess @autoregress and 3 others) It is interesting, indeed. I use this
proof in many of my lectures, including why-19, since PO researchers are embarrassed to discuss the
comparison with DAGs. #Bookofwhy

4.5.19 @5:12am - (Replying to @robertwplatt @Jabaluck and 3 others) An inertia that lasts for over 3 decades
borders on a pathology, and is indicative of the power of those "pockets of resistance" you mentioned, which
thrive in the absence of "pockets of acceptance". #Bookofwhy

4.5.19 @3:11am - (1/n) Thank you @PHuenermund for summarizing so vividly the Why-19 symposium. I agree
with most of your observations and recommendations, especially those pertaining to causal inference in
economics. Last week saw a huge interest on Twitter coming from economists, triggered
4.5.19 @3:11am - (2/n) possibly by the challenge to analyze a causal chain using PO. While it unveiled the
obvious advantages of DAGs in compactness, transparency and inference complexity, some bystanders might
still have gotten the impression that one can do
4.5.19 @3:11am - (3/ ) without them through a heavy investment in PO training. Only passive on-lookers could
come to such conclusion, not one who actually tries to analyze the chain using the two languages side by side. I
therefore continue to advise readers: Do not rely on on-lookers, try to
4.5.19 @3:11am - (4/ ) solve this problem yourself, from beginning to end, its not too hard, yet it reveals the
essential differences between the two representations, one a direct mapping of your knowledge, the other a
convoluted transformation of that knowledge. Next time an economist asks you:
4.5.19 @3:11am - (5/ ) "What do I get by using DAGs?" you will be able to assert first-handedly; you get the
ability to answer certain questions that you would not be able answer otherwise, and these are questions that
economists ask themselves 12 times a day: e.g., Is this variable exogeneous?
4.5.19 @3:11am - (6/ ) perhaps conditionally exogeneous? Is this parameter estimable using OLS? Does my model
have testable implications? Are these two models statistically distinguishable, and more and more... I listed
some in https://ucla.in/2mhxKdO , Try them out, for fun and profit. They are not
4.5.19 @3:11am - (7/ ) meant to prove that economists do not know X or Y, but to entice them to enjoy the
power of new tools, still absent from their textbooks. Conclusion: Do not rely on "On-lookers", listen to your
own experience. Good luck, and Tweet if any questions. #Bookofwhy

4.4.19 @2:23pm - (Replying to @causalinf) Reminds me of that math teacher who admitted: Yes, multiplication
can be useful, once you learn it, but we can do everything with addition. And all bystanders, who did not try to
actually solve a problem with and w/o multiplication concluding: HMM, its just a matter of training

4.4.19 @1:12pm - (Replying to @EpiEllie) Thanks @EpiEllie . But as an incurable engineer I can only absorb new
information when it comes in the format of: Input --> Output. ie, What information is needed to start the
analysis (or simulation) and what conclusions come out of it. #Bookofwhy

4.4.19 @3:08am - Illustrated slides for causal inference instructors. Marco Zaffalo was kind enough to share with
us slides that he made for a master course, based on the Primer book https://ucla.in/2KYYviP . They can be
downloaded here https://ucla.in/2Uvp4Di . Thank you Marco. #Bookofwhy
4.4.19 @12:56am - (Replying to @EpiEllie) Can you describe in Twitter length what agent-based models are? I
heard this term used occasionally but never explained. How about it? #Bookofwhy

4.4.19 @12:16am - (Replying to @KordingLab @robertwplatt and 3 others) I am suspect of having a dog in the
fight, so my credibility can't match that of the "credibility movement" which, by definition, is both home-grown
and credible.#Bookofwhy

4.3.19 @11:27pm - (1/2) (Replying to @robertwplatt @KordingLab and 3 others) Agree, but let's assume that an
entire literature of a certain field, by some mysterious force, adamantly resists using tools that neighboring
fields have found useful, for 3 decades. Isn't it reasonable for an insider to call out the entire field: Look what
WE are missing!"
4.3.19 @11:35pm - (2/2) (Replying to @yudapearl @robertwplatt and 4 others) Now assume that instead of WE,
the insider names the field, and explicates in details what students and researchers in that field cannot do, so
as to call attention to those mysterious forces that hold back progress. Isn't he/she justified calling out an
entire field?

4.3.19 @10:52pm - (Replying to @paulgp @autoregress @Jabaluck) Right, the effect of X on Y would no longer
be identifiable. Though the effect of Z1 on Y will be. It's all in the diagram, I am not making it up. #Bookofwhy

4.3.19 @4:07pm - (Replying to @autoregress @Jabaluck) This is even better that Smoking and Tar. Because the
decisions at Z and Y are "man made", ie, controlled by humans who follow a scrutinizable protocol, much like
the canonical IV examples, which dependent on man-made lotteries. Great step forward, #Bookofwhy

4.3.19 @3:58pm - (Replying to @a_strezh @autoregress @Jabaluck) This is indeed how it was first introduced to
economists in the 1990's, see https://ucla.in/2pJzGNK , it was named an "instrument", and for a good reason; it
is a variable that is instrumental in the identification. #Bookofwhy

4.3.19 @3:28pm - (Replying to @robertwplatt @Jabaluck and 2 others) The apple pie is "good science" and "most
of the econ seminars I've been to". So everything is fine and dandy in the cause of "good science". Shouldnt we
also identify by name the supernatural forces that resist progress? I am forbidden from doing so, how about
others? #Bookofwhy

4.3.19 @2:10pm - (Replying to @yudapearl @robertwplatt and 3 others) Such a "Hey Look" alert invariably
evokes the question: "What will we be missing if we ignore the Hey Look?" And then the answer: "You wont be
able to do X Y Z" creates professional indignation: We can't do X Y Z? Conclusion, do X Y Z before going back to
Apple Pie. #Bookofwhy.

4.3.19 @1:59pm - (Replying to @robertwplatt @Jabaluck and 2 others) How can anyone object to "Mother's
Apple Pie" and "Lets continue what we are doing"? Not sure T. Kuhn will agree. There comes a point when
totally new ways of "clarifying assumptions" are born, and someone has to say: Hey Look! At the risk of irritating
others. #Bookofwhy

4.3.19 @7:39am - (Replying to @robertwplatt @PHuenermund and 2 others) I do not see anything missing. More
and more people get the idea that you use DAGs to encode what you know and PO to express what you wish to
know, and that we now have the logic of going from the former to the latter. The rest is up to subject matter
experts. #Bookofwhy

4.3.19 @6:14am - (Replying to @robertwplatt @PHuenermund and 2 others) DAG/SCM does not commit to any
identification strategy, it accommodate them all. Likewise it is not GLOBAL, but accommodates incremental
gy
construction of the model, The idea that SCM starts with a huge DAG was invented by people who could not find
another reason for not using DAGs

4.2.19 @11:58pm - (Replying to @PHuenermund @Jabaluck) I love quasi-experiments because, as I argued in


https://ucla.in/2mhxKdO , they are neat exercises in structural economics, awaiting to unveil more power if
assisted by a "structure" . Glorifying them in their current structure-less state seems premature to me.
#Bookofwhy

4.2.19 @11:15pm - Good news for European passengers. Our UK publisher (Penguin) is informing us that a
paperback edition of #Bookofwhy is now out. See https://www.amazon.co.uk/Book-Why-Science-Cause-
Effect/dp/0141982411/ref=tmm_pap_swatch_0?_encoding=UTF8&qid=1554210478&sr=8-1 ... (Unfortunately
they are not allowed to ship to US.) Another step towards seeing students climb the Ladder of Causation.

4.2.19 @8:38pm - (Replying to @analisereal @autoregress @causalinf) There is another aspect to consider. The
most reliable question people can answer is "does X affect Y?" DAGs are built on this primitive question,
suggesting that causal knowledge is organized in DAG structure. Encoding it differently results in the mess we
see on the left side.

4.2.19 @7:34pm - (1/2) This discussion seems to be pitting PO vs. DAGs as two competitors. They are in fact
complementary. DAGs are used to encode what we know and PO what we wish to know. SCM is a bi-lingual
framework, https://ucla.in/2umzd65 , tying the two through structural equation semantics
4.2.19 @7:34pm - (2/2) The reason some researchers still attempt to encode what they know in PO is purely
cultural -- they were deprived of DAG education. Glancing at how a simple chain is encoded in the two
languages makes you wonder why any bi-lingual would choose PO over x-->y-->z #Bookofwhy

4.2.19 @4:54pm - Link repaired!!!) Readers have alerted me to a broken link to my interview on 3:am. It is
repaired now, and can be clicked here: https://316am.site123.me/articles/the-causal-revolutionary?c=end-
times-archive ... #Bookofwhy

4.2.19 @3:20pm - (Replying to @aricaroline @causalinf @Jabaluck) Very reasonable, and I hope @causalinf finds
the time and inspiration to do it. However, my golden rule is: One example "of it" is worth ten debates "about
it". I say so because I have seen many flip overnight as soon as they solved one toy example hands on using
DAGs. #Bookofwhy

4.2.19 @2:41pm - @y2silence Thanks for remembering this blog page, I forgot all about it and I find it so timely
(in view of latest debates) that I am retweeting it to all readers. Please review carefully and feel free to use
whenever someone claims that everything can be done in PO #Bookofwhy

4.2.19 @1:07am - (1/3) In the interest of keeping this Twitter conversation as a platform for genuine learning,
and saluting our Golden Rule: "One example outweighs ten debates", I strongly recommend that readers try to
work out this toy example: It calls for analyzing a causal chain X-->Y-->Z
4.2.19 @1:07am - (2/n) in two frameworks: 1. DAGs, 2. Potential outcomes. It has two stages: (a) specify the
model assumptions in both languages, and (b) decide if those assumptions have testable implications. The
example is extremely important for understanding the often-heard claim:
4.2.19 @1:07am - (3/3) "The two frameworks are "provenly equivalent"" and its counter-claim: "logical
equivalence ain't computational equivalence." It is a great opportunity to engage in a fun example that most
debaters have tried to avoid. Good luck. #Bookofwhy

4.1.19 @11:00pm - (1/2) You are making heroic effort to establish the superiority of IV over MIV (mediator IV);
it's commendable. But no effort can escape the need to establish no-confounding in BOTH cases. The adjectives
t s co e dable. ut o e o t ca escape t e eed to establ s o co ou d g O cases. e adject ves
"desparate", "fruitless" "not understood" "horrible" are equally applicable to
4.1.19 @11:00pm - (2/2) BOTH cases. No judgment of "good natural experiment" can be made without ruling out
billions of "unmeasured confounders". Likewise, an MIV is no less "natural " than IV, each may have billions of
determiners hence requiring a "no confounding" judgment. #Bookofwhy

4.1.19 @7:46pm - (Replying to @analisereal @autoregress @causalinf) Agree. Only he who has not tried it would
say "it is a matter of taste". I urge all passengers to try it before speaking "about it". Computer science students
learn that good "representation" is more than just taste, its the difference between doable and undoable.
#Bookofwhy

4.1.19 @6:49pm - (1/n) Tired of caricatures? Note that we never construct a dag by listing 150,ooo variables.
We start by asking: can you think of a variable affecting both X and Y? Is it measured? If not, is it significant? If
yes, lump it together with all other such variables and mark it U,
4.1.19 @6:49pm - (2/n) "unobserved confounders", ONE node. Next you ask: Can you think of a variables that is
either (1) on the X-Y path and shielded from U, or (2) affects X and is sheilded from U and not affecting Y
(except..)? The former is front-door the latter is IV. And so on and on. At each
4.1.19 @6:49pm - (3/n) stage the question arises: What is "shielded"? and the answer is given, again, in term of:
"Can you think of a variable that resides here or there...and has a property that can easily be verified in the
"mind's DAG" which is expert in answering only one primitive question:
4.1.19 @6:49pm - (4/5) "Who is listening to whom?". Caricatures are not helpful. Note that you need the "mind's
DAG" to certify any candidate IV. Note also that you need the theory of identification to interrogate your mind's
DAG toward identification templates, eg. backdoor, frontdoor`, IV, ..
4.1.19 @6:49pm - (5/5) conditional IV, etc. etc. Finally, once you construct the DAG in this incremental way,
guided by hoped-for identification templates you may end up with 4-10 variables, and become uncertain of
identification -- go to do calculus, the ultimate arbiter (for NP). #Bookofwhy

4.1.19 @5:54pm - (Replying to @autoregress) If you have a DAG, you can see many things immediately. The
problem is with people who are DAG-averse, and claim that they can do everything in the "provenly equivalent"
PO language, including model specification. It is fun to try #Bookofwhy @causalinf

4.1.19 @2:40pm - (1/2) Not really. Consider the causal chain X--->Y--->Z. Students of pictures can immediately
conclude that X and Z are independant given Y. I do not know ANY student of Greek symbols who can easily
come to same conclusion from a symbolic representation of the chain, say using PO
4.1.19 @2:40pm - (2/2) (potential outcomes). It is doable, of course, but it would take you a good 5-30 minutes
of derivations. You must try it yourself to appreciate the difference and, if you fail, you might wish to take a
look at the solution: https://ucla.in/2QpcGzS
4.1.19 @2:40pm - (3/3) or give it to a PO expert, for fun. #Bookofwhy

3.31.19 @9:33pm - (Replying to @Jabaluck @MariaGlymour) Aha!!! Now we are getting someplace. Listen
carefully: "searching for... to fit" what is that space in which we search? For me, it is a mental representation of
knowledge which shares many DAG's features. For others it is a nebulous cloud, not to be depicted. What about
you?

3.31.19 @9:21pm - (Replying to @Jabaluck) I keep on begging you to help me do so by describing (conceptually,
w/o acronyms) what lessons need be imported. All I hear is: read his papers, go to Sudan, or "add Angrist type
techniques" (which I find already in SCM.). Give me ONE to add, b/c I am dying to automate it.

3.31.19 @9:07pm - (Replying to @Jabaluck) I am tired of hearing that I need to study income inequality in Sudan
before I can appreciate the methodological lessons that someone learned. Just tell us what those lessons are
and we will incorporate them in SCM, dont send us back to Sudan please. One m. lesson, and formal.
3.31.19 @8:37pm - (Replying to @MariaGlymour @Jabaluck) I see "looking for good quasi-experiments" to be
within structural economics, namely, economics that explicates scientific knowledge in SE, or in DAG, so that
we can search for "good" quasi-experiment, namely one that our knowledge certifies as meeting the IV
conditions.

3.31.19 @7:07pm - (Replying to @metrics52) Josh, I think you would like your doppelganger's performance. He
goes by the name "Least Harmful" and can execute all kind of acrobatics; front-door, back-door, IV,... you name
it. A true partner in credibility. How about helping us shaping his motion? #Bookofwhy

3.31.19 @6:18pm - (1/4) I think the Garbage Theory is fundamentally flawed. Since "credible inference" is
subsumed by "structural economics", garbage generation is a logical impossibility. This is expressed clearly in
https://ucla.in/2mhxKdO , Section on "Experimentalists". Quoting:
3.31.19 @6:18pm - (2/4) Quoting: "to the extent that the "experimental?" approach is valid, it is a routine
exercise in structural economics. However, the philosophical basis of the "experimentalist" approach, as it is

currently marketed, is both flawed and error prone." (The Refs are illuminating)
3.31.19 @6:18pm - (3/4) Thus, Good news to all sailors and passenger on this unassailable ship: "The garbage
attack is over." Moreover, the more you hear about things "you dont even know" (eg. "what data you need") the
closer we get to an automated-Angrist, because, if this is what we need to know,
3.31.19 @6:18pm - (4/4) we are already there. See https://ucla.in/2Jc1kdD on how smart structural economists
can select data sources for identification. What else do they ("credibility" claimants) think "you don't even
know"? The more the better. #Bookofwhy

3.31.19 @9:10am - (1/4) Just woke up, to the sound of garbage flying. Wow!. Must pacify some deadlines, but
not before stating: The aim of causal inference is to automate the process of generating id-strategies, starting
with mental models of the domain. I do not see any theoretical impediment to
3.31.19 @9:10am - (2/4) automate the process by which Angrist&Comp are generating their identification
templates from their conceptual knowledge of the world, since the knowledge contained in the former is
derivable from the former. This is what the Inference Engine is all about in #Bookofwhy p.11
3.31.19 @9:10am - (3/ ) An "automated Angrist" is not a far-fetched dream, it is partially implemented already
in Elias software, which searches a model for nuggets such as frontdoor, backdoor, IV, napkins and, if none is
found, goes to do-calculus. Thus, freeing economists to engage in things
3.31.19 @9:10am - (4/4) they do best, accumulate knowledge (through empirical studies), refining knowledge
and encoding it in a natural and transparent way, without worrying about id-strategies, which computers can
perhaps do better. It is doable.

3.31.19 @8:39am - (Replying to @MartinSGaynor @Jabaluck and 7 others) qq

3.30.19 @8:42pm - We were fortunate to hear Yoshua's lecture at http://why19.causalai.net while he was
under a strict embargo, to conceal the news he received a week before. After the lunch he had to make an
important phone call... The rest is in NYT.

3.30.19 @6:43pm - (Replying to @steventberry @Jabaluck) It is refreshing to read Angrist again, especially after
our discussions; I recommend it to all Tweeters. What he does not get is that "identification" need not be
"transparent" if the model is. Knowledge comes from our mental model, not from our manufactured id-
strategies.

3.30.19 @6:14pm - (Replying to @a40ruhr @snavarrol and 4 others) Not surprised. The "giants" refuse to accept
that even straightforward judgments, such as exogeneity or exclusion", come from a mental model that, once
explicated, regardless how sketchily, would make judgments so much more reliable. No, "more credible" says
the flag. #Bookofwhy
g y

3.30.19 @2:52pm - (Replying to @steventberry @Jabaluck) Conditional independence was anointed "Principle-2"
in https://ucla.in/2mhxKdO . Do you also use "Principle-1"? i.e., "the law of structural counterfactuals"? This
would certainly turn Angrist against you. Where are your discussions with Angrist's camp aired? #Bookofwhy

3.30.19 @2:12pm - (Replying to @Jabaluck @snavarrol and 3 others) Heckman and Pinto critics should be read in
the context of https://ucla.in/2L8OCyl which defines non-parametric identification, explains its rationale, and
outlines how LATE IV can be incorporated within the SCM framework. It also bemoans graph-avoiding
derivations. #Bookofwhy

3.30.19 @1:21pm - (Replying to @snavarrol @PossebomVitor and 3 others) Yet at the conceptual level the
Nonparametric IV problem is well defined: We want to identify the effect of X on Y but we can only randomize
Z. When is it doable? A complete (ie, necessary and sufficient) condition can be found here:
https://ucla.in/2QpSZb9 . #Bookofwhy

3.30.19 @7:06am - (Replying to @PossebomVitor) The difference between the IV and CF condition will show up
if you add an arrow from U to Z. Under such a model Z is no longer a valid IV, but it becomes a valid IV upon
conditioning on U. See "Generalized IV" https://ucla.in/2MqNBRU (A good paper for economists) #Bookofwhy

3.30.19 @6:36am - (Replying to @Jabaluck @ho_ben and 2 others) I surely agree that those assumptions have
substantive content that requires PO notation in addition to DAGs, but I would not know what that content is,
or how to assess its plausibility if it were not for the SCM semantics that tells me what I am assessing.
#Bookofwhy

3.30.19 @6:21am - (Replying to @Jabaluck @ho_ben and 2 others) No need to reassure me that you are not
DAG-averse, I can see that, and I consider it an asset to econometric education. I wish more scholars like you
would approach the arrow-phobic cult of econometric and say: Come on, it shows, and it is embarrassing all of
us. #Bookofwhy

3.30.19 @6:05am - (Replying to @Jabaluck @ho_ben and 2 others) These "parametric" assumptions (eg
monotonicity) are sometimes "more plausible" than exclusion assumptions, not necessarily "weaker". They can
easily be incorporated into the DAG framework, using the same PO notation, which derives its logic from SCM
semantics, #Bookofwhy p.276

3.30.19 @5:44am - (Replying to @yudapearl @Jabaluck and 3 others) not of identification. Glad you concur.
IOW: this series of methodological papers all assume away the identification problem, invoking some template
ignorability assumption, and, for a fixed given estimand they provide new estimation strategies. Agree?
#Bookofwhy

3.30.19 @5:28am - (Replying to @Jabaluck @ho_ben and 2 others) The issue was " identification templates" and
the contested sentence was: "Athey and imbens have been providing a series of methodological papers that
push the boundaries of those templates. " My understanding too was that the boundaries pushed were
boundaries of estimation

3.30.19 @3:26am - If you are in NYC this coming Monday, please note that @eliasbareinboim will be speaking at
Columbia on "Causal Data Science", to the tune of https://bit.ly/2YwHnr2 . I wouldn't miss it!
For details: https://twitter.com/ColumbiaCompSci/status/1111367760867659785 ... @ColumbiaCompSci
@DSI_Columbia @ColumbiaMed @CUSEAS
3.30.19 @3:54am - (Replying to @ho_ben @causalinf and 2 others) Strange. I have heard dozens of people speak
about the "Use of basic ideas from machine learning" for identifying treatment effects, yet none was able to tell
me what those basic ideas were. Perhaps our Twitter discussants can weigh in and illuminate the uninitiated?
#Bookofwhy

3.30.19 @2:29am - 3 remarks on DGs (Directed Graphs) and simultaneity:


1. The 3-steps of computing counterfactuals are valid.
2. d-separation is valid in linear systems
3. The 3-steps of identifying counterfactuals are valid in linear systems.
REF: https://ucla.in/2G2rWBv p. 96. #Bookofwhy

3.30.19 @2:28am - (Replying to @steventberry @Jabaluck @causalinf)


3 remarks on DGs (Directed Graphs) and simultaneity
1. The 3-steps of computing counterfactuals are valid.
2. d-separation is valid in linear systems

3. The 3-steps of identifying counterfactuals are valid in linear systems.


REF: https://ucla.in/2G2rWBv p. 96. #Bookofwhy

3.30.19 @1:27am - (Replying to @yudapearl @steventberry @Jabaluck) Or was it perhaps my explanation of


"What kept the Cowles commission at bay?". in https://ucla.in/2mhxKdO that you object to? I hope you agree
with my claim that they lacked "syntactic machinery for reading counterfactuals from a model or solving simple
id-problems. #Bookofwhy

3.30.19 @12:44am - (Replying to @steventberry @Jabaluck) I am trying to find which "summaries of "econ
thought"" could have triggered your objection. Do you not agree that "The causal lens is badly obscured in
econometric history-writing."? Has any historian narrated Haavelmo's three key insights described here
https://ucla.in/2mhxKdO

3.30.19 @12:23am - (Replying to @steventberry @Jabaluck) Well taken, and internalized. But curious: Is
"Modern Structural Modeling" different from DAGs? I usually call what I do "Structural Causal Models" (eg

#Bookofwhy p.276). Of which DAGs and PO are 2 abstractions. Am I stepping on someone's toes? What does MSM
say about control?

3.29.19 @7:23pm - (Replying to @steventberry @Jabaluck) Do you mean "causal" or "casual"? If casual, I wonder
why "You could make the same mistake with DAGs," I can't recall seeing even one student making "control"
mistake with DAGs. Have you? And btw, did the ancient religion ever discuss the issue of adding "controls"?
#Bookofwhy

3.29.19 @4:44pm - (Replying to @causalinf @Jabaluck and 2 others) @caisalinf, I beg to disagree on one tiny
issue-PO. No amount of command over PO notation can make up for the (usually ignored) fact that PO is a
derivative of structural models, as in https://ucla.in/2G2rWBv . Have you met a PO expert who can estimate
Joe's would-be salary p.94 ?

3.29.19 @3:46pm - (Replying to @causalinf @rdahis @Jabaluck) I am glad @causlinf is undertaking the task of
summarizing this conversation. If I writes it, I would be accused of trying to sell "MY approach". Plus I am still a
firm believer in the Axiom of Choice: "Solving one problem in DAGs is worth 100 debates ABOUT dags."
#Bookofwhy
3.29.19 @3:31pm - (Replying to @ho_ben @causalinf and 2 others) Miscommunication. When I asked you to
share "new templates pushed by Athey and Imbens" I did not mean links to their papers, but sharing ONE
template that appealed to you personally, whose conceptual basis can be communicated on Twitter #Bookofwhy

3.29.19 @3:15pm - (1/3) @Undercoverhist , First, Thanks for doing your tweetstorm on Eco-history, which I have
found to be illuminating. Second, I do not believe @causalinf was joking. Eco-historians (including Morgan, Qin,
Epstein, Hoover) failed to distinguish "theory-data tension" from
3.29.19 @3:15pm - (2/3) "Causalily-data tension" which, in my opinion, accounts for most hurdles in eco. For
the school of Hendry (and Sargan), for example, "economic theory" meant the statistics that governs data, not
Haavelom's theory which he viewed as a collection of experiments conducted by
3.29.19 @3:15pm - (3/3) mother nature. I hope you make this distinction in your upcoming book. I tried to make
it in my paper on Haavelmo https://ucla.in/2mhxKdO but I do not think it penetrated mainstream eco thinking.
The causal lens is badly obscured in econometric history-writing. #Bookofwhy

3.29.19 @1:42pm - One can find good use of PO&DAG combination in every paper on mediation (eg
https://ucla.in/2QmZrjc ) because PO and DAGs are both derivatives of Structural Causal Models (SCM), DAGs
encodes what you know, and PO what you wish to know - no conflict. #Bookofwhy

3.29.19 @12:59pm - (1/2) My paper https://ucla.in/2mhxKdO contains a section: "What held the Cowles
commission at bay?". It may sound heretic to most economists, bc it attributes the decline to a stalemate in
notation and computation, yet I still believe that, lacking the ability to SEE (ie compute)
3.29.19 @12:59pm - (2/2) simple features of one's model, makes it impossible for an economist to go beyond
Marschak and F Fisher. Eco-Historians minimized the role of Haavelmo and Wold, bc the causes-data tension
was not their priority. It takes a chemist to write the history of alchemy #Bookofwhy

3.29.19 @11:14am - (Replying to @ho_ben @causalinf and 2 others) What conclusion should I draw? Sit down and
relax? Or assist the progressives in their efforts to educate the remaining islands of resistance? BTW, I am not
familiar with the new templates pushed by Athey and Imbens. Can you share one? #Bookofwhy

3.29.19 @4:00am - (Replying to @ho_ben) Not exactly. By studying relativity seriously we know the limitations
of Newtonian physics. By NOT studying DAGs economists are still unaware of the limitations of their template-
driven methods. Did you say "many of us are working on same"? Please share, eager to hear #Bookofwhy

3.29.19 @2:47am - (1/n) I see a spark of agreement looming from this conversation. It is based on (I hope)
everyone's agreeing that "we need a DAG for inference, bc it carries the info we need for id." Another spark is
the fact that everyone (I hope) is talking about at least TWO DAGs, one residing
3.29.19 @2:47am - (2/n) in the mind and tacitly stores your understanding of the relevant domain, and one
(called Full DAG) is what you eventually explicate when you decide to draw it on paper for full analysis. Call the
former "mental DAG" (or m-DAG) and the latter ex-DAG (for explicit).
3.29.19 @2:47am - (3/n) Scott also introduced a project-specific DAG, or a premade DAG defined by the id-
strategy one wishes to use. Call it t-DAG (for template). Barring two repairable exaggerations, I generally agree
with Scott's depiction of "practical economists" and CI-theorists.
3.29.19 @2:47am - (4/n) But I need to add another brush stroke to facilitate full agreement: The mental DAG
and the Full-DAG are the same, while the t-DAG is a fragment of the former, selectively extracted to match a
specific id-strategy.
With these points of agreement, we see that the Full-DAG is
3.29.19 @2:47am - (5/n) used in two different roles. First, mentioned by Scott, to alert us to new id-strategies
(eg front-door) which our ancestors have missed lacking complete id-logic. Second, and perhaps more
important, to validate matching between the postulated template-DAG and our mental DAG
3.29.19 @2:47am - (6/n) - the ultimate arbiter of plausibility. For example, if we selected an IV strategy, our t-
DAG would be the canonical IV-DAG, and matching involves checking whether the exogeneity and exclusion
properties assumed in that canonical t-DAG hold in our mental-DAG, which is waiting
p p , g
3.29.19 @2:47am - (7/n) passively to be interrogated. This is what we normally call "judging for plausibility".
Why then do we insist that even practicing economists learn to read DAGs before engaging in heavy empirical
work? Because the task of matching requires reading your m-DAG. What do we mean
3.29.19 @2:47am - (8/n) by READING a DAG? We mean taking an arbitrary 4-variable DAG and checking if
properties such as exogeneity, exclusion or conditional exclusion hold in it. It is a matter of checking the
plausibility of one's assumptions, not of discovering new id-strategies.
3.29.19 @2:47am - (9/9) This is why we get suspicious when leaders of "credibility movements" tell us that they
do not need to read DAGs, since they deal with "real life" problems." IOW: "We hate to show you how poorly we
do when things are explicit, trust us to do better in real-life, #Bookofwhy

3.29.19 @1:57am - If one believes that Economists SHOULD learn DAGs, how can one trust leaders of a so-called
"credibility movement" who proclaim they do NOT need to learn DAGs because they deal with "real-life"
problems. I would be a bit suspicious of their un-aided judgment. #Bookofwhy

3.28.19 @5:35pm - (1/n) This is new and interesting observation: You are 1 and we are many. I am a believer in
crowd wisdom, because it integrates many perspectives, each taken from a different angle. Can we say that
about econometrics? Are there many angles if NBER postings are by "members only"?
3.28.19 @5:35pm - (2/n) Moreover, the claim "causal inference is all we THINK about" is true for children,
homosapiens, scientists and alchemists, yet #Bookofwhy distinguishes "thinking" from committing our thinking to
some mathematics. And here, all my economist friends bemoan econ textbooks, see
3.28.19 @5:35pm - (3/n) https://ucla.in/2NRFn7e . Another moreover, my history book tells me that Alchemists
too had norms and practices, and they were quite effective in developing metals and alloys. The only thing they
failed on were "toy problems"; eg,they could not explain why substances often
3.28.19 @5:35pm - (4/n) combine in integer proportions. They dismissed the "toy problems" with great
indignation, probably saying "we are solving important practical empirical and large scale problems, show us
what we did wrong!". It was by taking those toy problems seriously that we have science.

3.28.19 @4:42pm - (Replying to @robertwplatt @Jabaluck @PHuenermund) I remember coming to the conclusion
that the word "design" could stand some formal clarification, otherwise it sounds like "if you do the right thing
nothing else matters". I love it. Every student (and robot) should remember this important principle -- a life
saver.#Bookofwhy

3.28.19 @2:38am - (Replying to @juli_schuess @PHuenermund @jim_savage_) @jim_savage, I am curious if you


understand how "credibility" can be attached this way. I'm truly curious, and would value your insight, because I
have the feeling that you did try to solve a problem with dags, so you know where PO's come from and why it is
crucial. #Bookofwhy

3.27.19 @8:33pm - (Replying to @causalinf @JeffDenning and 4 others) I have just received Mostlyharmless from
Carlos, three days ago, and I read it with all the curiosity that our tweeter discussions have evoked. Let's
compare its treatment of PO with that of Primer: ttps://ucla.in/2G2rWBv then say "credibility" to our students,
with straight face

3.27.19 @5:29pm - (Replying to @JeffDenning @causalinf and 4 others) I got a copy of Mostlyharmless and, if
this book is representative of the "credibility revolution", I find it hard to understand how anyone who tried to
solve a problem with dags can attached the word "credibiity" to a methodology that ignores our models of
reality. #Bookofwhy

3.27.19 @9:49am - (Replying to @juli_schuess @Jabaluck @yskout) Well put. Glad you took the time to read
mostlyharmless and, like me, tried to make sense of how they justify adjustments and IV using PO. Only those
who have not used DAGs to justify things would prolong the myth that such acrobatic constitutes a
"justification". #Bookofwhy
j y

3.27.19 @9:20am - Exciting news to all sailors and passengers in AI and ML: 2019 Turing Prize awarded to 3 ML
pioneers. https://www.nytimes.com/2019/03/27/technology/turing-award-hinton-lecun-bengio.html ... Our
Why-19 symposium http://why19.causalai.net was fortunate yesterday to host Bengio's brilliant lecture just
before hearing of the good news. #Bookofwhy

3.27.19 @8:31am - (Replying to @juli_schuess @Jabaluck @yskout) It is rather implausible that our mental
representation of what's going on in the world is tailored to any "identification strategy" that 20th-century
scientists have devised. Such representation simply tells us what affects what, prior to engaging is
"identification" exercises.

3.27.19 @2:06am - (Replying to @yskout @Jabaluck) Agree. But it is the DAG that tells you if an IV is "GOOD".
For simple cases we do not need a DAG, but in cases where exogeneity and exclusion are questioned, or when
they need to be created, eg https://ucla.in/2MqNBRU we need DAGs.

3.27.19 @1:24am - The identification problem is NOT about determining the logical consequences of
assumptions Rather, it is about determining if the estimability of a desired causal effect follows logically from
the assumptions and, if so, how it is to be accomplished. #Bookofwhy

3.27.19 @1:10am - Replying to @peter_mourfield @PooyanJamshidi Thanks for posting.

3.25.19 @1:52am - Taking a leave from Twitter for two days, on pilgrimage to Stanford, speaking Monday,
March 25, 2:00 pm, at Jordan Hall, 420-40, https://why19.causalai.net . Title: "The Foundations of Causal
Inference, with Reflections on ML and AI". Admission free for Stanford students.See you.

3.25.19 @1:08am - (Replying to @MartinRavallion @SteveRo48195125 @Jabaluck) Conceptually, I agree.


#Bookofwhy for example claims that the RCT is valid because it mimics backdoor, not the other way around.
Same goes for the validity of natural experiments.

3.25.19 @12:00am - I must add that, in SCM, talks about embracing and unifying are not just words. Examine
again this chapter https://ucla.in/2G2rWBv and judge for yourself how naturally counterfactuals (hence PO)
emerge from a structure, how they are used to solve practical problems and, and, and

3.24.19 @10:39pm - (Replying to @robertwplatt @Jabaluck and 2 others) The symmetry does not hold here.
Why? Because Structural Causal Modeling (SCM) embraces PO and IV and DAG (eg mediation and attribution are
all done with PO). May I take it that you recommending SCM? If so, hat off, welcome to the causal revolution
#Bookofwhy

3.24.19 @7:19pm - (Replying to @CarnaticPrior) I do not doubt that de Finetti theorem is central to some
statistical tasks. But that does not make it relevant to causation and, certainly, I do not see how even comes
close to deciding whether one should condition on M to get unbiased effect estimates. #Bookofwhy

3.24.19 @6:34pm - (Replying to @HannesMalmberg1 @SteveRo48195125 and 2 others) Well put! I could not say it
better. What IS quasi-exp? The insiders say: its some setup that mimics RCT. The ousiders say: No, it is a model
of reality which, after doing mental checking, is judged to satisfy exogeneity. Now come DAGs and say: Lets
replace mental with formal

3 24 19 @6:22pm - (Replying to @smilleralert @autoregress and 3 others) I have never been a fan of the
3.24.19 @6:22pm - (Replying to @smilleralert @autoregress and 3 others) I have never been a fan of the
expression "can help think through things" but after seeing how it can be abused I pledged NEVER to use it as
long as I live. DAGs actually compute what economists are struggling to THINK through. Analogy: algebra
computes solutions, not help think.

3.24.19 @5:15pm - (Replying to @SteveRo48195125 @MartinRavallion @Jabaluck) My understanding of backdoor


is that we do not ADD variables. THEY ARE THERE in nature, we just acknowledge their existence in our mental
picture of reality and check if they are needed. We ADD them to our identification strategy, if appropriate, not
to the model. #Bookofwhy

3.24.19 @4:16pm - (1/2) First time I hear this: "PO can be helpful for thinking about what to control for. " I am
retweeting this statement as an example of how dangerous the expression "helpful for thinking" can be, and why
students of causality should work out simple problems in both languages, as
3.24.19 @4:16pm - (2/2) as I have done in my book, and in my slides. In my adult life, I have not seen a SINGLE
case where "PO can be helpful for thinking about what to control for", but since I am biased, test it on PO-
experts, eg https://ucla.in/2Fq9zV2 or https://ucla.in/2EihVyD #Bookofwhy

3.24.19 @3:08pm - Thanks Scott for another candid narrative on the role DAGs play in your research and
teaching. I wish there were more econ like you who say, lets first TEACH it, then see what new practical
opportunities open up. This is how Epi started - I dont think they are sorry #Bookofwhy

3.24.19 @2:52pm - (Replying to @Jabaluck @analisereal and 4 others) If I were to sniff the econ literature for
cases of bad control I would be accused of negative attitude, if not "bashing econ." Instead, I am presenting
"seat-belt usage" as a case of bad control authorized by super statisticians https://ucla.in/2EihVyD . Still not
convinced?

3.24.19 @2:37pm - (Replying to @analisereal @Jabaluck and 4 others) Glad you emphasized "disciplining current
ID strategies" which has been almost forgotten in the discussion, as if PO+IV+RC do not require formal assistance
in visualizing, elaborating and strengthening the assumptions behind them. They do. #Bookofwhy

3.24.19 @1:47pm - (Replying to @RealKevinThmpsn) I appreciate your candid narrative of change. I am trying to

stimulate in economists both the feeling of "I am missing something" and "I could do so much more". A very thin
difference between the two, the latter being more productive. #Bookofwhy

3.24.19 @2:10am - (Replying to @diomavro @Jabaluck) proof of equivalence in Causality chapter 7. But this only
says that if we start with same assumptions we end up with same conclusions, it does not guarantee that
encoding the assumptions in one system will be comprehensible or compact. See http://ucla.in/2EihVyD and
slide.

3.24.19 @12:36am - My favorite M-bias example is "seat-belt usage" as described in Section 2 of this paper
https://ucla.in/2EihVyD . What's interesting about it is that "Seat-belt usage" was actually controlled for in
Rubin's study. #Bookofwhy

3.23.19 @9:57pm - (Replying to @yudapearl @Jabaluck and 2 others) one of us can whisper on Josh's ear that
the "difficult" problem was solved and made easy in 1993. It can't be me, because Josh will suspect that I want
to SELL him DAGs, and econometric will go through another decade without a solution. #Bookofwhy

3.23.19 @9:46pm - (Replying to @Jabaluck @EpiEllie @autoregress) This is an illuminating blog from
"mostlyharmless", not so much for discovering M-bias, but for confessing that the issue is "difficult" and may
ostly a less , ot so uc o d scove g M b as, but o co ess g t at t e ssue s d cult a d ay
depend on whether variables are "exogenous" or "mediating", namely, the economist is advised to interrogate
the mind's DAG. Perhaps

3.23.19 @9:21pm - (Replying to @Jabaluck @mathtick) Jason, I respectfully ask to be excused from answering
your Tweets, especially those that interpret or assess Pearl's "claims" or "quarrels"; assume I have none. If you
think econometrics has areas of potential improvement, please Tweet; I would be curious to read and

3.23.19 @9:24pm - (Replying to @yudapearl @Jabaluck @mathtick) occasionally add my two cents. But I would
rather not engage in theories on how "Pearl" can be a better salesman. I am interested in econometric
methodology, not in salesmanship. @StuartBuck1 #Bookofwhy

3.23.19 @7:14pm - (1/2) I actually believe economists have positive values and honest scientific attitude.
Unfortunately, they inherit dogmatic education and rigid cultural allegiance that stifle innovation. One hope is
their inquisitive students. Here, our tweeting readers can help. How about
3.23.19 @7:14pm - (2/2) petitioning the editors of the five top econometric journals to invite an introductory
review article on "causal inference among our neighbors", just to wet their students appetite. If I were an
Editor, I would consider it my duty to thus counter allegations of insularity.

3.23.19 @5:17pm - (Replying to @stewarthu) Compare econ literature to social science a la "Morgan and
Winship" (2007). The latter embrace diagrams, while the former bans them at all cost, denying students even
the rudimentary ability to read independencies in their own econ models. Not exactly "mathy", #Bookofwhy

3.23.19 @5:03pm - (Replying to @stewarthu) Econ used to be "mathy", searching for the right math to solve eco
problems. Then, invaded by statisticians, they started searching for problems to fit the math. It is indeed hard
to stomach the idea that a graph is a mathematical object, as honorable as an equation.#Bookofwhy

3.23.19 @3:45pm - (Replying to @yudapearl @stewarthu) Resistance to math can be justified when the math
seems totally unrelated to your research (eg abst. set theory to an economist). But resisting a math that helps
you solve problems that you yourself declared important and hard, that really takes cultish thinking.
#Bookofwhy

3.23.19 @3:30pm - (Replying to @stewarthu) What happened to Epi can be described as "bloodless" revolution,
namely, a dramatic change of perspective and tools. Economics is different. There, professional pride and
insular structure entail stiff resistance for every change. #Bookofwhy

3.23.19 @3:08pm - (Replying to @IgnacioPerezR @jreileyclark @_MiguelHernaa) It also has a nice title: "Draw
Your Assumptions Before Your Conclusions" which encapsulates the difference between CI and empirical
economists. The latter go by the principle: "First Decide on Identifying Strategy, then search your mind for
assumptions that support it"#Bookofwhy

3.23.19 @4:15am - (Replying to @DoctorActivist) I never took it "personal", I take it as tiring attempts to justify
the status quo, instead of helping to change it. My writing is indeed a "call out": econ. students need to know
how far behind their textbooks are, else the status quo will change incredibly slow. #Bookofwhy

3.23.19 @3:55am - Your question drove me to re-read https://ucla.in/2G2rWBv - the chapter on


counterfactuals in Primer. My gosh! So many debates and misconceptions could be avoided through this chapter.
I recommend it again to all readers and followers -dont miss! I have not found an alternative

3 23 19 @2 41 (R l i t @di )F t h i l d t il I dPi
3.23.19 @2:41am - (Replying to @diomavro) For more technical details, I recommend Primer
https://ucla.in/2KYYviP e.g., https://ucla.in/2G2rWBv

3.23.19 @2:35am - (Replying to @diomavro) I try not to say "directed graph" when talking to economists; this
totally shut their mind off. I start with a STORY, eg what would Joe's salary be had he stayed one more year in
college, given that currently his salary is 30K and that he won two prices for technical innovation.

3.23.19 @1:56am - (1/3) To be honest, Jason's repeated lectures on what I need to do to be taken seriously by
economists began to get on my nerves. I wrote #Bookofwhy having given up hopes of ever seeing old-guard
economists adopting of causal inference tools, so Jason's lectures
3.23.19 @1:56am - (2/3) sounded hollow to me, and I tried not to reply to him. I haven't given up on
econometric students though. So, if readers wish to help econometrics catchup with advanced tools, alert their
students to what they are missing and, as a pre-requisit, teach them how to solve
3.23.19 @1:56am - (3/3) a 4-variable problem before sending them to Kenya to study income inequality. Physics
students learn to solve equations of motion before they touch a telescope or visit an observatory.

3.23.19 @1:09am - (Replying to @Rodrigo20980033) The issue is not the citation but the narrative, which
conflicts with that of #Bookofwhy (p.334-5) on decomposition and identification. Check https://ucla.in/2R5Xqs1
and https://ucla.in/2N5Gdk2

3.22.19 @3:24am - (1/3) Thanks for posting this article on missing data and I am glad you found graphical
models to be beneficial for both visualizing and organizing the results obtained. I am not sure, however, if you
are familiar with recent works on missing data, as summarized for example in
3.22.19 @3:24am - (2/3) this review article and in which general graphical criteria are derived for recovery and
testability https://ucla.in/2LdEjZW . I hope you find it useful for
3.22.19 @3:24am - (3/3) would yield biased results. Your graphs seem to imply such criterion but I could not
derive one since MI is model-blind. Congratulations on an illuminating article in a hard field that was once
thought to be the sole province of statistical analysis. #Bookofwhy @IJEeditorial

3.21.19 @11:40pm - (1/2) Thanks for the link. The dippel etal paper may well be the first applied econometrics
paper on mediation, and it even has a DAG in the appendix: courageous, and structural equations with their
relations to counterfactuals. I would am tempted to forgive the authors
3.21.19 @11:40pm - (2/2) for not citing https://ucla.in/2R5Xqs1 where the mediation formula first appears (see
#Bookofwhy p. 334-5 for the story). This is progress, but note, contrary to ruling paradigm, they put down the
structural model on paper before searching for "identifying assumption". Bold strengthening your results. One
issue which was not clear to me is whether you have a criterion for detecting when MI

3.21.19 @11:15pm - (Replying to @autoregress @btshapir) I wish I could say, yes, I hope so too. But I am
stubborn believer in the Golden Axiom: He who has not solved a 4-variable problem, will not find areas where
more complex formalization can be "fruitfully applied. But your students will. It is the Law of lampposts.
#Bookofwhy

3.21.19 @10:03pm - (Replying to @RahelJhirad) Yes, it is free, according to AAAI headquarters, and a personal
mesg I received from Carol Hamilton, our Tzar. @eliasbareinboim is closer to the ruling party, so he should be
able to confirm it. I hope it is as free as I was given to understand #Bookofwhy

3.21.19 @8:40pm - (Replying to @quantadan @autoregress @btshapir) Love your analogy. Though I no longer
attempt to lead the horse to water. I am just hoping that econ. students are smart enough to notice that, if
they want water, they should and can get it on their own. And they will, for it is becoming more and more
accessible. #Bookofwhy
3.21.19 @7:00pm - (Replying to @ChrisSeveren @metrics52) I have "Master Metrics" (2015) on my Kindle. Is that
representative of the culture? or is "Mostly Harmless" more advanced? I am not surprise that the language in
which you "grow up" comes naturally. The question is how flexible the culture is to language enrichment
#Bookofwhy

3.21.19 @3:59pm - (Replying to @ChrisSeveren) I would really appreciate an example, so that I will understand
what they mean by "stating" the assumption, and 'providing" supporting evidence. Not that I doubt whether they
do that, I trust you, but as a computer scientist, I am interested in the language they use. #Bookofwhy

3.21.19 @2:05pm - (Replying to @PHuenermund @Jabaluck @autoregress And once you agree that the exclusion
restriction is not plausible, what then? Do you try to see if it can be remedied? eg by conditioning on an
intermediate variable. I have seen it in the literature, but I may have missed it.

3.21.19 @1:08pm - (Replying to @_scott_fleming_ @DavidnLang and 5 others) I understand that you just show up
and show your Stanford affiliation card.

3.21.19 @8:48am - (Replying to @autoregress @btshapir) I dont know how this thread turned me (again) into an
enemy of the people. I stopped trying to convince empirical economists to use DAGs - they wont. I'm still eager
to show their students how equivalent representations differ, which should be useful once they build a model.

3.21.19 @8:02am - (Replying to @autoregress) OOps, a little snug here. Normally one assumption is insufficient
to estimate any effect. Even in simple problems we need 10-20 assumptions (see slide) the combination of
which may be sufficient. Shouldnt we wait to validate 10 assumptions first, one at a time? then proceed?

3.21.19 @7:49am - (1/3) I would never use such language. Economists are not supposed to know about DAGs,
and I never expect them to view things through DAGs. I only expect them to commit to paper what THEY
believe is plausible, in a way that THEY think faithfully represents their beliefs, then ...
3.21.19 @7:49am - (2/3) continue to estimate what THEY think need be estimated. What I find missing in this
natural sequence of steps is "committing to paper". Indeed, this step is missing in the writings of empirical
economists, who prefer to jump directly from mental thoughts to estimation.
3.21.19 @7:49am - (3/3) This explains why they fail to appreciate the derivations of do-calculus, and always go
back to ask: "we need to think hard if this assumption is plausible". Without committing to paper they never get
to saying: assume for a moment my assumptions are valid, just for a moment.

3.21.19 @7:16am - (Replying to @autoregress) ok, we thought and we thought and we argued and argued and
we came to the conclusion that U has no effect on Z. Now everyone is in agreement. Next step =? Do we keep
the agreed story in our mind? or commit it to paper? Recall, we have 10 other things we agreed on. Please
advise.

3.21.19 @6:51am - (Replying to @autoregress) I am not clear what we disagree about. Perhaps you can state
your position in the form of "I think that enough thinking will enable us to......" or "We still need more thinking
to do before....". Sorry.

3.21.19 @6:38am - (Replying to @f2harrell) Not sure of the context. Mediation? Generlization? 3-languages?

3.21.19 @6:34am - (Replying to @autoregress) Agree, the two steps should be taken in sequel. The slide deals

with choices available in the second step: encoding our understanding of the world. Once we encode, there is
no more thinking necessary, we can sit back, relax, and let the do-calculus take over. #Bookofwhy

3.21.19 @5:52am - (Replying to @autoregress) Once we come to agreement on how to think about the world,
we need to decide how to represent our thinking formally, so as to combine it with data and estimate what we
wish estimated. Thinking alone is insufficient for deciding how to estimate things. #Bookofwhy

3.21.19 @3:31am - (1/2) With all the talks about "logical equivalence", "transparency" "defensibility" "testability"
and "empirical economics", I presume that only few students have had a chance to see an actual comparison of
a simple example articulated side-by-side in 3 different languages:
3.21.19 @3:31am - (2/2) 1.English, 2.Potential Outcomes and 3. Structural Models. Here is one for your
enjoyment: https://ucla.in/2Fq9zV2 a slide that I will be discussing at Stanford. The research question is of
course: Estimate the effect of Smoking (X) on Cancer (Y), given samples from P(X,Y,Z).

3.20.19 @9:05pm - This blunder permeates the literature, and leads to the mediation fallacy #Bookofwhy p.315
which econ. students should learn to avoid. Students interested in understanding mediation should go straight
to the model-based literature, eg https://ucla.in/2N5Gdk2

3.20.19 @7:40pm - I can identify the skinnerian level as a subset of Rung-1 (which embraces retrodiction.) and
Popperian level as Rung-2 (love the emphasis on "models"). The other levels are not described in term of
capabilities -- hard to place. And I dont find counterfactuals, strange.#Bookofwhy

3.20.19 @5:22pm - (Replying to @depistemology) I really haven't thought about the relations between the two
classifications. Where can one find the best accessible window to Dennett's levels? Something as concise and
exemplified as #Bookofwhy ?

3.20.19 @4:33pm - Confirming my pilgrimage to Stanford, I will be speaking Monday, March 25, 2:00 pm, at
Jordan Hall, see https://why19.causalai.net . My title: "The Foundations of Causal Inference, with Reflections
on ML and AI". Admission is free and no topic left behind. https://ucla.in/2umzd65

3.20.19 @2:28pm - (Replying to @akelleh @eprosenthal @EvanBianco) Great name "Causal Data Science" and
great blog too. How do you survive Columbia under such name? Last time I spoke at Columbia they almost

crucified me at the podium for suggesting they should name their Institute "Reality Science" instead of "Data
Science" #Bookofwhy

3.20.19 @12:27am - (Replying to @raymondshpeley @Jabaluck and 3 others) I would replace 2 and 3 with 2'.
encode that knowledge in a defensible model M 3a. check if your needed effect is estimable from M (and how)
3b. If so, estimate it, if not, elaborate M. [I would avoid "identifying assumption" like the mother of all
miscommunication.] #Bookofwhy

3.19.19 @2:15pm - Another role played by the orientation of the ellipse can be seen in Lord's Paradox,
#Bookofwhy page 213, elaborated in https://ucla.in/2JeJs1Q . But this takes us from regression to causality-
land. Great for class demonstration.

3.19.19 @12:40pm - (Replying to @TiernanRayTech) Thanks for posting. I would never have imagine that Pharma
folks read #Bookofwhy.

3 19 19 @4:13am - Trygve Haavelmo (1944 p 14) drew an analogy between two sorts of experiments: "those
3.19.19 @4:13am Trygve Haavelmo (1944, p. 14) drew an analogy between two sorts of experiments: those
we should like to make" and "the stream of experiments that nature is steadily turning out from her own
enormous laboratory, and which we merely watch as passive observers." Angrist-Krueger 2001

3.18.19 @6:34pm - (Replying to @gelbach @Susan_Athey) You must be wondering what makes me work so hard
on trying to understand where our communication failed. The story is a long one but, right now, I am trying to
understand if economists have a way of distinguishing "exactly-identifying assumptions" from other kind of
assumptions

3.18.19 @5:27pm - (Replying to @gelbach @Susan_Athey) After walking together through 20 Tweets, is it too
much to ask for help in refreshing by failing memory and granting me one last word: Is the task: trivial?
important? solved in the literature? taught in econ. classes? irrelevant? Outdated? Your word? #Bookofwhy

3.18.19 @5:12pm - (Replying to @raymondshpeley @eliasbareinboim and 2 others) Agree. But sometimes the
slow bites are the hardest ones of all. I truly do not know what "good causal empirical work" is, anyone knows?.
Note the hardship we go through to have economists opine if identification is important for policy analysis. All
small bites are giants

3.18.19 @4:43pm - (Replying to @gelbach @Susan_Athey) OK, no words. What about the TASK of "examining all
the assumptions conveyed by your econ. model and deciding whether they imply that a given variable X is
conditionally exogenous." Is that task trivial? important? solved in the literature? taught in econ. classes?
Opinion?

3.18.19 @4:05pm - (Replying to @gelbach @Susan_Athey) But what word should I use in the future?

3.18.19 @4:01pm - (Replying to @gelbach @Susan_Athey) Good. Lets not used the word "ascertain" any more.
What word do you use for the act of examining all you model assumptions and computing/determining whether
a variable X is conditionally exogeneous (given subset Z in the model).?? I promise to use your term only.

3.18.19 @3:16pm - (Replying to @ToreEllingsen1 @CarterPaddy) GMM is a totally different beast; it aims to
estimate parameters in statistical models instead of causal effects in nonparametric econ models. The best way
to see the difference is to try it on a simple problem (eg frontdoor #Bookofwhy p.224) Its essentially undoable

3.18.19 @2:44pm - (Replying to @gelbach @Susan_Athey) I take then it that you agree that ascertaining
"conditional exogeneity" is of primary importance for "policy evaluation", for it determines whether or not we
need to resort to bounds or other semi-identification methods economists have developed. I hope I read you
correctly?

3.18.19 @1:31pm - (Replying to @f2harrell @AndersHuitfeldt) I missed your reply, thanks. So you are
emphasizing the "reliability" of a parameter estimate, given a finite sample. As compared to our notion of
generalizability which emphasizes alternative models under infinite sample. Did I get it right?

3.18.19 @1:25pm - (Replying to @gelbach @Susan_Athey) You may be right. I would be eager to hear what kind
of engagement can lead to better communication, avoiding "shade-throwing"? What is your opinion? Can policy
analysis be conducted without one's ability to ascertain "conditional exogeneity"?

3.18.19 @1:02pm - (Replying to @f2harrell @AndersHuitfeldt) What about the generalizability problem I posed:
"Find if the OLS estimator of parameter beta in Model 1 estimates parameter gamma in Model 2 without bias." Is
this part of the generalizability problem addressed in your book? Should it be? #Bookofwhy
3.18.19 @12:59pm - (Replying to @f2harrell @AndersHuitfeldt) I get a double kick , because I still do not know
"where prob. models come from" and I am beginning to suspect people do not have such models in their minds.
The math models we have are attempts to patch up the fragments we do have in our minds. #Bookofwhy

3.18.19 @12:50pm - (Replying to @NicholasStrayer) Thanks for highlighting this important distinction

3.18.19 @12:49pm - (Replying to @elliottmcollins @CarterPaddy) You can always skip multiplication and
replacing it with addition -- just add a number to itself N times. The miracle of multiplication is that with a
meager investment of memorizing one multiplication table you can do science that would be left undone
without that investment.

3.18.19 @3:29am - (1/2) Readers of #BookofWhy may wish to attend the Inauguration of Stanford Institute for
Human-Centered AI (HAI), Monday, March 18, which will include a symposium and remarks by Bill Gates. A
livestream will begin 9:15am Pacific.
3.18.19 @3:29am - (2/2) I will not be able to attend, unfortunately, but as a Distinguished Fellow of HAI, I wish
the co-director Fei-Fei Li @drfeifei all the success that a Center for such critical and timely mission deserves.

3.18.19 @2:10am - (Replying to @yskout @eliasbareinboim and 2 others) A worthwhile endeavor! But be careful
to choose a paper that, in case your result differ from the author's, you will be able to convince readers that
yours are right. Perhaps by questioning one of the assumptions. So, choose an author who is explicit about
assumptions - Rare!

3.18.19 @1:50am - (Replying to @nyarlathotepesq @CarterPaddy) Totally irresponsible, I agree, even


delinquent. However, if someone demonstrates to us that lacking certain tools we cannot solve simple problems
that we OURSELVES declared to be important and fundamental to our field. Shouldn't we then try to expand our
tool set? #Bookofwhy

3.18.19 @1:30am - (Replying to @sarahmrose) The reason it is hard (if not impossible) to unveils the flaws in
"real examples" is that we do not have ground truth against which to gauge glaws in case we get different result
than our predecessor. Also, "toy examples" are quite "real" (drug, age,cancer) but are manageable.

3.18.19 @12:56am - (Replying to @sarahmrose) We would ask the same questions, but on larger problems,
problems that we either discarded as hopeless w/o DAGs, or that we treated by ad-hoc methods, not having the
tools to choose methods to fit the characteristics of our models.

3.18.19 @12:24am - (Replying to @maximananyev @Susan_Athey @joshgans) Good idea. Although it is hard for
me to envision an economist convinced by missed opportunities when he/she has no language to gauge the
amount missed. Plus, market efficiency disappears when trade embargoes are imposed. #Bookofwhy

3.17.19 @11:49pm - (Replying to @maximananyev @Susan_Athey @joshgans) Thanks for posting this discussion
http://bit.ly/2F8pJ3Y which illustrates nicely the M-bias which Rubin has denied since 2009
https://ucla.in/2EihVyD . Another thing I learned from it is that the word "design" is used for "data generating
model" (ie., SEM). Good to remember.

3.17.19 @10:25pm - (Replying to @f2harrell @AndersHuitfeldt) Here is a toy generalizability problem in


parametric models (Dunan, 1975): "Find if the OLS estimator of parameter beta in Model 1 estimates parameter
gamma in Model 2 without bias." ://ucla.in/2L8jUFg Is this the kind of generalizability problems addressed in
your book?
y

3.17.19 @9:55pm - (Replying to @maximananyev @Susan_Athey @joshgans) I have a different theory on why epi.
embrace DAGs and economists ban them; it has nothing to do with the research problems in the two fields (the
structures are identical.) In the 1990's Epi was led by Greenland and Robins, and Econ. by disciples of Don
Rubin. Simple.#Bookofwhy

3.17.19 @8:33pm - (Replying to @Ideal_Health18 @Susan_Athey @joshgans) Thanks for reaching out. Tell me
your research problem and I will try to find a nice introduction for you. Such introductions do exist, all we need
to know is the language in which you feel comfortable and ops we go. #Bookofwhy

3.17.19 @5:57pm - (Replying to @Susan_Athey @joshgans) How does one judge "better quality" in "empirical
applications" where assumptions are debatable, hardly articulable, and where we don't have ground truth
against which to gauge improvement. The Golden Axiom says: He who cannot do it on a 4-variable problem,
cannot do it period.

3.17.19 @5:18pm - (Replying to @paulnovosad @Jabaluck and 2 others) I think you meant "econs are good at
backdoor criterion" INFORMALLY, because I know only N<10 economists who know it formally, or who can use it
beyond 4-variable problems. If you know of N=11, there is hope.

3.17.19 @5:11pm - (Replying to @analisereal @Jabaluck and 2 others) Plus, it is always helpful to quote the
golden axiom: He who cannot do it on a 4-variable problem cannot do it on REAL WORLD EXAMPLES, where we
have no way of testing claims.

3.17.19 @4:47pm - (Replying to @elliottmcollins @Jabaluck @CarterPaddy) Ufortunately, First-graders solutions


would not improve with multiplication; they were perfectly sufficient and deeply sophisticated for the exam
given, which was selected by the teacher to cover addition, not problems requiring multiplication.

3.17.19 @3:04pm - (Replying to @Susan_Athey) What about the "treatment effect" literature, also going by
"policy evaluation" etc. Is CE of primary concern there? Does the DAG-avoiding description of that literature
make sense? or is the "avoiding" unintentional, or scientifically informed, or non-existing? #Bookofwhy

3.17.19 @2:55pm - (1/2) Nowhere has anyone presupposed that understanding "our" work is "as foundational for
doing causal inference as understanding multiplication is to doing mathematics." What IS foundational (as
multiplication) is deciding conditional ignorability in a scientific model
3.17.19 @2:55pm - (2/2) (ie, a model describing "how the world works") by ANY method. Plus the axiom that he
who cannot do it on a 4-variable problem cannot do it on REAL WORLD EXAMPLES where we have no way of
testing claims. But, hating to further insult the innocent, I'll quit this thread.

3.17.19 @2:26pm - (Replying to @f2harrell @AndersHuitfeldt) I am all for parametric models. But I have not
seen the problem of external validity, in the parametric context, expressed in layman terms, i.e., What's given,
what's assumed, what's needed. Does it take more than 2 Tweets? #Bookofwhy

3.17.19 @2:18pm - (Replying to @Susan_Athey) Would you consider inability to decide if a variable is
conditionally exogeneous (given an econ. model) to be a hindrance to identification of counterfactual
inference? Do you think that DAG-avoiding economists do not suffer from this inability? #Bookofwhy

3.17.19 @1:20pm - (Replying to @CarterPaddy @fuiud @Jabaluck) This is indeed my stance. See
https://ucla.in/2mhxKdO , where I expressed surprise at the heirs of Haavelmo, Wold and Maschak for not
pursuing these teachers, and not jumping enthusiastically at the opportunity to overcome the obstacles that
held the Cowell's Commission at bay.

3.17.19 @1:03pm - (Replying to @fuiud @CarterPaddy @Jabaluck) I am listening attentively to you (not to "our
field") Please describe to me your attitude to external validity. ie. what are wet trying to estimate, what
information is available to us, what assumptions we are making etc. I am eager to talk with econ who says "I
can" #Bookofwhy

3.17.19 @12:45pm - (Replying to @fuiud @CarterPaddy @Jabaluck) I dont buy the "reality of applied work" as an
excuse for not trying synthetic problems, whose correct answers can be obtained either analytically or by
simulation. I would love to learn new attitudes to external validity issues, explained from scratch. I'm listening.
#Bookofwhy

3.17.19 @12:27pm - (Replying to @Jabaluck @btshapir @CarterPaddy) In my humble field of computer science,
if an editor smells that new tools are being developed in neighboring fields, possibly applicable to another,
he/she sees it as an editorial duty to invite an introductory article to educate readers. Even stat journals do so,
not econ.

3.17.19 @12:08pm - (Replying to @Jabaluck @CarterPaddy) A passing word on those revered "top journals". The
editors of three such journals advised me not to submit papers because "they cannot find qualified reviewers". I
hope econ. readers advise those editors to retire if they want econ. to enter the age of modernity. #Bookofwhy

3.17.19 @11:56am - (Replying to @CarterPaddy @Jabaluck) There would definitely be a sweeping changes even
in published papers (ie, using addition). E.G. "conditional ignorability" would not be assumed apriori, but be
justified on scientific grounds, namely, by appeal to "how the world works" instead of "everyone makes this
assumption"

3.17.19 @11:47am - (Replying to @Jabaluck @CarterPaddy) True, if is not just "save time" but solving correctly
problems that were discarded by economists (multiplication) for being beyond the pale of their tools (addition)
hence they never appear in the "top journals".

3.17.19 @11:42am - (Replying to @CarterPaddy @Jabaluck) Not only "easier", but "possible". eg. Deciding to
treat or not to treat in Simpson's paradox is intractable to the unaided human mind. Same with choosing the

right set Z for "control" of confounding. Same with deciding if an RCT result is generalizable. #Bookofwhy

3.17.19 @11:33am - (Replying to @Jabaluck @CarterPaddy) Why would it be irresponsible for an observer to say
that first-grade students do not really understand arithmetic, if all the problems they solved thus far are
problems in addition? One can say so without showing that the solution would improve with multiplication.
#Bookofwhy

3.17.19 @6:33am - (Replying to @AndersHuitfeldt @f2harrell) Right. We are dealing with non-parametric models
which, by nature, allow unrestricted heterogeneity. I therefore presume that if "interaction" plays aany role, it
must be that one is working in some parametric model, which I missed. Can we clarify afresh what the problem
is?

3.17.19 @4:12am - (Replying to @CarterPaddy) The way we can handle feedback is to unfold the loop into a
sequential, time varying DAG and, unless we can assume linearity, this becomes quite messy. Luckily, when
linearity cab be assumed, d-separation holds, and some identification results can be obtained w/o unfolding.
linearity cab be assumed, d separation holds, and some identification results can be obtained w/o unfolding.

3.17.19 @3:22am - (1/5) Reading this justification of X||Y_x|Z, I was ready to plead ignorance of "cost sharing"
"copay" "actuaries" and "utilization trends" and quit before it gets too domain-specific. But out of respect to
your genuine attempt to capture the meaning of this statement,
3.17.19 @3:22am - (2/5) I offer my version, in generic terms. (1) The cryptic statement X||Y_x|Z, also named
"conditional ignorability" (CI) by PO folks, is a feature of the population under study and, when valid, provides a
license to estimate the ATE using regression, simply "controlling for Z"
3.17.19 @3:22am - (3/5) CI is the key assumption behind all works in PO. (2) Being a feature of the population,
it can be validated from our model of the world, without thinking about what we do or wish to do. It depends
only on how Z is related to X and Y in the presence of other variables if any.
3.17.19 @3:22am - (4/5) (3) If our model comes in the form of a DAG (a depiction of economic structural
equation model) we can validate CI by simply checking if all paths between X and Y are "backdoor-blocked by
Z". (4) The notion of "backdoor-blocked by Z" is a fun, game-like criterion on DAGs
3.17.19 @3:22am - (5/5) that can be mastered in 5-12 minutes by any economist who is serious about finding
out if ATE is estimable by regression. See #Bookofwhy or PRIMER https://ucla.in/2KYYviP . Shunning DAGs, PO
folks must assume CI apriori, unjustified, and some economists follow them blindly.

3.17.19 @2:14am - Very well put!! And there is an additional aspect to it. DAGs are not merely visualization
devices to help economists spot opportunities, they are also computational engines that spot those
opportunities and exploit them, while you relax and think only: "how the world works"

3.16.19 @4:49pm - Demanding "same treatment effects" is basically same as "genralizability". The next question
is: Given what we know about how the target and study populations DIFFER, decide if the latter still
"represents" the former. Examples: https://ucla.in/2L6yTzE https://ucla.in/2NIsI6B

3.16.19 @4:30pm - (Replying to @JWSBayes @f2harrell and 2 others) I know of two publications that tell it quite
nicely, perhaps different from @f2harrell and @stephensenn, but addressing a clear question: When can you
generalize and when you cant. Here they are: https://ucla.in/2L6yTzE and https://ucla.in/2NIsI6B

3.16.19 @3:22pm - (Replying to @Jabaluck) Forget DAGs. "scientific grounds" is anything that resembles the way
scientists communicate, or the way scientific knowledge is stored in the mind of a researcher who is asked to
judge if a given condition is plausible or not. You are now to justify X||Y_x|Z. You choose X,Y,Z

3.16.19 @2:50pm - (Replying to @Jabaluck) Great, if only PO researchers could justify those "sufficient
conditions" on scientific grounds. But, again, we should stop hypothesizing what other researchers can or cannot
do. It is time for us to hear one researcher saying: I CAN DO. Willing to volunteer? #Bookofwhy

3.16.19 @2:29pm - (Replying to @Jabaluck) Proving you wrong: It is time we hear "I am an applied researcher,
and I understand this and this, e.g., I can detect conditional exogenetity when it exists in my model, and other
features that are necessary for CI " instead of "I think so and so already understand... #Bookofwhy

3.16.19 @2:01am - (1/3) The criterion you propose (could econ. do better with DAGs) suffers from selection
bias. Could first-grade students do better with multiplication? No! The exam was only about addition. Lacking CI
tools, economists had to work on simple problems, where their lamppost methods
3.16.19 @2:01am - (2/3) (LATE, dif-i-diff, reg-disc.) seemed applicable. When I say "slow progress" I look at
areas such as: identification, mediation, transportability, sensitivity, which flourished in soc. sc. (eg. Morgan
&Winship, Imai etal) in the past 20 years, with almost no progress in econ
3.16.19 @2:01am - (3/3) And when we compare progress, let us discard hearsay eg., "so and so did it" and count
only those who say: "I can do it". Eg. "I can detect conditional exogeneity" or "I can repair violations of
exclusion". We hear very few "I can" in the econ. literature. We need more.
3.15.19 @12:56am - (Replying to @Jabaluck @analisereal) I wish I could believe it. But watching econometric
research progress over the past 25 years, it is hard to believe that laziness alone made them slip behind other
CI disciplines, eg Epi or Soc Sci., especially in identification, mediation, transportability, selection, more ..

3.14.19 @12:59am - (Replying to @Jabaluck @analisereal) Curious: Do you believe economists' systematic
avoidance of DAGs reflects informed scientific judgement or cultish dogma? #Bookofwhy

3.14.19 @6:45am - (Replying to @Jabaluck @analisereal) This paper is an example of how assumptions are made
not because they are defensible on scientific grounds, but because they allow certain mathematical derivations
to go through. If this is "more commonly used analysis in econ" I would be really worried. Arn't you?. #Bookofwhy

3.14.19 @5:49am - (Replying to @thosjleeper @analisereal @causalinf) I fail to see why we learn more
"empirically" when we tacitly assume exogeneity and exclusion for IV, as opposed to representing these same
assumptions explicitly in a DAG?

3.14.19 @5:16am - (Replying to @thosjleeper @analisereal @causalinf) DAGs are perfect for representing
absence of prior information, just fill them with arrows and unobserved confounders. #Bookofwhy

3.14.19 @3:47am - (Replying to @analisereal @thosjleeper @causalinf) I would phrase it a bit differently.
Instead of building models that represent their problems, they impose assumptions that permit identification.
The former are defensible on scientific grounds, the latter are fabricated for convenience, hence titled
"identifying assumptions."

3.13.19 @4:58am - (Replying to @TPA_Debray) So suppose I am dying to estimate the probability the Unicorns
have purple horns. It is quantitative, because I asked for "probability" and it is surely "uncertain". Is it a
statistical question?

3.13.19 @4:12am - (1/3) Imagine a modern chemist traveling back in time to a 16th Century alchemist
laboratory. This is how I felt upon reading this paper http://goo.gl/RHRKoE on "suppression" [ie. where adding a
new variable to a regression model, totally unrelated to the outcome,
3.13.19 @4:12am - (2/2) increases the predictive power of the model]. Being an incurable whiggish historian, I

have found it fascinating to watch how a phenomenon that has baffled social scientists since the 1940's is
unfolding gracefully through the lens of causal analysis. #Bookofwhy

3.13.19 @12:39am - (1/3) Jeff has a point. If we start with "colliders" and "confounders" economists can say:
"yes that makes sense but I didn't need a DAG to understand that." For this reason I usually start with "look at
your own model, the one you authored. Can you tell me a few things about it?"
3.13.19 @12:39am - (2/3) For example: "which parameter can be identified if we assume correlated
disturbances between Z and W?" Now, the economist feels embarrassed not to answer simple questions about
his OWN model, and, if he is curious, he would say: "can you do it with DAGs?". If he is not, ....
3.13.19 @12:39am - (3/3) he/she would get defensive, and say "How can you belittle the whole field of
econometrics...? " [ familiar?]. The trick is to educate the former without offending the latter. Tough but
doable. There is hope to Eco. #Bookofwhy BTW the paper by West is pure correlation, unfit.

3.12.19 @11:36pm - (Replying to @boback @EpiEllie) What is Kardashian ? A treatment? or an alternative


universe?
3.12.19 @11:33pm - Many readers have questioned whether "Causal Revolution" is an appropriate name for what
we are currently seeing in @causalinference . Here is a short note https://ucla.in/2JgpQ2S touching on Kuhn's
"Structure of Scientific Revolution", which puts things in historical perspective

3.12.19 @11:12pm - (Replying to @l__ds) Pearson and Yule observed the disappearance of correlation upon
aggregation, not reversal. I checked their papers years ago. See Pearson's quote in #Bookofwhy

3.12.19 @11:04pm - (Replying to @deaneckles @metrics52) This is great, because it sounds exactly like Pearl.
Did @metrics52 join the revolution? Or is he going to clash with Pearl in the next chapter, where students need
to learn how to actually compute counterfactuals? Lets wait and see. #Bookofwhy

3.12.19 @10:22pm - (Replying to @ghoshd @kareem_carr @EpiEllie) It means that, unlike "obesity", people
agree on how to measure it. If wealth is thus measured then, indeed, the query is just "probability of necessity"
as defined in #Bookofwhy and even better in PRIMER

3.12.19 @9:08pm - (Replying to @EpiEllie) Yesterday I did not know if Kylie Janner is a person or a tarantula.
So, I would not use this example to learn about SCM. I dont see what makes this example different than asking
for the "probability that factor F is necessary for outcome Y", which is in Primer Chp 4 & #Bookofwhy

3.12.19 @8:50pm - (Replying to @chrisalbon) If Causality flashbacks are too painful try #Bookofwhy, you will
forever swim in "flashforwards"

3.12.19 @8:13pm - (1/2) A few notes on the history of Simpson's Paradox. (1) The first reversal was noted by
Cohen and Nagel (1934, p.449). (2)Simpson was indeed the first to note that what we judge as "sensible"
depends on the STORY, not on the data [did not say "causal"] (3) In Causality (p. 177)
3.12.19 @8:13pm - (2/2) I attributed this discovery to Lindley and Novick (1981), but Hernan etal (2011)
corrected me; Simpson already noted it in 1951. (4) Not all examples of the paradox invoked confounding.
Lindley etal used mediation as in Fig. 1(b) of https://ucla.in/2Jfl2VS and #Bookofwhy

3.12.19 @3:09pm - (1/2) I did not realize that Ed Simpson's was still alive, and I feel really sad with his
departure. He will be remembered as an intellectual family member of all students of causality. The paper you
cite https://ucla.in/2Jfl2VS is a timely citation in light of recent discussions
3.12.19 @3:09pm - (2/2) on whether causality is IN or OUT of statistics. I dont believe anyone would claim IN, in
good faith, after reading the history of Simpson's Paradox and noting the persistent, century-long effort by
leading statisticians to avoid its causal dimension. Amazing! #Bookofwhy

3.12.19 @5:37am - (Replying to @EpiEllie @kareem_carr) By causal model I mean qualitative model specifying
who listens to whom. Surely, if we have good measurements on all relevant variables and data on how
successful people are in Jenner category, then the question translates to prob. of necessity #Bookofwhy, which
can be bounded

3.12.19 @12:16am - (Replying to @kareem_carr @EpiEllie) We can do the moon without data because the theory
of gravity gives us a fully specified dag, with functions attached to the arrows. Lacking functional specification,
we need data. #Bookofwhy

3.12.19 @12:09am - (Replying to @dccozine @ale_martinello @eliasbareinboim) I could not agree more. But if
they were only visualization-communication devices DAGs would end up where SEMs are today, ie, nice
diagrams all over the articles, yet when it comes to identification and testable implications.....no bad
mouthing..... see Journal of SEM. #Bookofwhy
3.11.19 @11:48pm - (Replying to @the_dismal_tide @kareem_carr) I would join one of the dozens Data Science
Centers now erected across the country and keep on bugging my mentors: "Hey, and where is the Science in our
building?" . Simultaneously I would take one of the obstacles of ML, say transfer learning and overcome it using
"science".

3.11.19 @10:08pm - (1/2) In the wake of on/off discussions of external validity vs generalisability vs
transportability, I am glad to note that Boston will soon get a glimpse at aunifying science of "data fusion", as in
https://ucla.in/2Jc1kdD . Elias Bareinboim will be speaking at Harvard on
3.11.19 @10:08pm - (2/2) March 21 https://twitter.com/HiTSatHarvard/status/1105146725377945600 and will
illuminate the topic with conceptual, mathematical and algorithmic results.@HarvardHealth @Kennedy_School
@HarvardChanSPH @_MiguelHernan @harvard_data @HarvardEpi

3.11.19 @9:16pm - (Replying to @kareem_carr) I mean "data science' as defined in one of my ancient papers:
The use of data to interpret the world. Not as it is used today, which is "the use of data to summarize data"
#Bookofwhy

3.11.19 @8:49pm - (Replying to @mathtick) I would not use the word "fraud" because it connotes intention to
deceive. Econ leaders are just lazy to buy (or even try) a new pair of eyeglasses. No fraud committed. But who

would volunteer to tell their students? #Bookofwhy

3.11.19 @8:37pm - (Replying to @RussellSPierce @balexanderstats) Natural experiments are Rung-1, i.e.,
observational studies in which the experimenter makes an assumption of exogeneity on one or more variables.
There is no need for continuum, because the the presence/absence of physical intervention is usually crisp.
#Bookofwhy

3.11.19 @8:30pm - (Replying to @kareem_carr @EpiEllie) You suspect wrongly about UCLA. Reading
https://ucla.in/2sASU9V carefully, we find that to estimate P(Y|do(X)), X needs (1) to be unambiguously
measured (e.g. blood pressure, or sex) and (2) population data and (3) a causal model must be available.
#Bookofwhy

3.11.19 @8:13pm - (Replying to @kareem_carr) Great question!! It will eventually be housed in Data-Science,
embracing both Data (now housed in statistics) and Science (now scattered among various scientific disciplines,
the common denominator of which will come under the second arm of Data-Science). #Bookofwhy

3.11.19 @8:07pm - Newly announced: Admission to the Why-19 Symposium http://why19.causalai.net will be
FREE for @Stanford community. I will be talking Tuesday, 3/26 2-4 pm on "What is Causal Inference" and will be
delighted to see you there. @StanfordMed @StanfordEng @StanfordAILab @drfeifei

3.11.19 @2:57pm - (1/2) How can anyone say that economists "never cared about such questions"??
https://ucla.in/2mhxKdO characterizes Haavelmo as the first to define causal effects in economic models. Still,
one can expose the current stalemate of econometrics and its insular structure without
3.11.19 @2:57pm - (2/2) being accused of "belittling others" How else can one convince Econ students that they
are being taught outdated methods, if the champions of those methods refuse to tell us how they solve a toy
problem (eg. your DAG)? Shouldn't students take this refusal as outdatedness?

3.11.19 @2:11pm - (Replying to @ale_martinello @_limbs_) @_limbs_ is asking a humble question: "shouldn't we
question the usefulness of traditional vs modern methods?" Why is he being accused of "belittling/denigrating
others"? When/how can we compare methods and display their faults and merits without being thus accused?
#Bookofwhy

3.11.19 @12:58pm - (Replying to @80Data @kareem_carr) @learnfromerror Interesting. And now (2019) do you
still believe Granger causality has anything to do with "causality"? #Bookofwhy

3.11.19 @12:55pm - (1/2) On the other extreme, one can takes the position that statistic is EVERYTHING,
subsuming all scientific endeavors because, whenever you ask a scientific question the answer is "a missing data
problem" - Bingo. To avoid such extremes I have found it useful (Causality p.38)
3.11.19 @12:55pm - (2/2) to delineate the boundaries of "statistics" accordance to its practices: "the study of
relationships governed by distributions of observed data", which also coincides with Fisher's definition of
"reduction of data", and makes causal relations "extra-statistical." #Bookofwhy

3.11.19 @11:57pm - (Replying to @RussellSPierce @balexanderstats) This discussion revolves around


observational studies. Interventions take us to Rung-2 of the ladder, and takes a different logic. For example,
"Would an intervention be sufficient for answering our research question?" is derivable in that logic. See
https://ucla.in/2Jc1kdD

3.11.19 @4:47pm - (Replying to @balexanderstats) I am going to use statistics, but only where it is useful,
namely in the estimation phase of the causal exercise. See the structure of the Inference Engine in
https://ucla.in/2umzd65 and also in #Bookofwhy

3.11.19 @4:43pm - (Replying to @Hl60464759) Some examples of extra statistical information are: Mud does not
cause rain, symptoms do not cause diseases , the drug does not change patients gender. In short, all the missing
arrows in the examples of #Bookofwhy

3.11.19 @4:13am - (1/5r) No offense, and I appeciate your sharing impressions with other readers. I am even
more grateful for mentioning @StatModeling which should give readers a glimpse at how some 2019 statisticians
think. I quote: "I find it baffling that Pearl and his colleagues keep taking
3.11.19 @4:13am - (2/5) statistical problems and, to my mind, complicating them by wrapping them in a causal
structure." This quote from Gelman's blog should enter the archives of scientific revolutions as proof that my
depiction of the inertial forces paralyzing statistics is not made up; and my
3.11.19 @4:13am - (3/5) description of causal inference as a "revolution" is not a fantasy. The resistance to
accepting needed assumptions as "extra statistical" is alive even in 2019. Moreover, readers of this quote take it
at face value that problems solved in #Bookofwhy can also be solved by
3.11.19 @4:13am - (4/5) Gelman's students w/o "wrapping them in a causal structure". This is the power of
blogs; no one asks you how: eg, "Can you show us how 'traditional statistical methods" would solve problems?"
People assume they somehow do. For otherwise, the all powerful science of statistics
3.11.19 @4:13am - (5/5) would be deficient, which is inconceivable, because someone would have noticed.
Well, the truth is "traditional statistics" IS deficient and her obedient students cannot solve those problems
without "wrapping them in a causal structure". And challenging them is no "bashing".

3.11.19 @2:30am - Glad you like #Bookofwhy. But where is "bashing of statisticians"? This is what the book says
about statisticians: "...they declared those questions off limits and turned to developing a thriving, causality-
free enterprise called statistics." This is not "bashing", it is history.

3.11.19 @2:11am - (1/2) Your DAG above can help us understand how classical economists (i.e. your mentors)
used to approach the problem of identification before graphs. From your recollection, did they use matrices?
method of moments? LISREL? PO? Imagine one of them addressing a class today.
3.11.19 @2:11am - (2/2) How would he/she convince the class that all the betas are identified even when IQ_0
and C are unobserved. Today we can do it no hands but, as a part-time historian I am curious how the world
y p
looked like before windshield wipers. Do you remember? Does anyone? #Bookofwhy

3.10.19 @7:27pm - (1/2) Well phrased! And I am glad you mention our paper on Econ texts
https://ucla.in/2NRFn7e . Intriguingly, all my econ. colleagues agree: "our testbooks are an embarrssment,
they dont represent our research, but who has time to write good texts". Moreover, some of the authors
3.10.19 @7:27pm - (2/2) of the reviewed textbooks wrote to me (in secret) for advice on how to repair things in
the next edition. So, no reason to despair, these new editions will be out in this and next year, econometrics
can be reclaimed and its current gurus bypassed. I am convinced #Bookofwhy

3.10.19 @3:49pm - (1/2) I have corrected you twice before, first on what I said about economists and, second,
on what we can believe about the research that some economists produce. I do not understand your need to
distort what I say. I have also explained (above Tweet) why economists conservatism
3.10.19 @3:49pm - (2/2) is a natural phenomenon, somewhat more glaring than in Computer Science, but still
forgiven. "No hard feeling if they dont, gold medals if they do." But, Hey, you owe us an answer on N>4. "No
hard feeling if you dont, gold...." #bookofwhy

3.10.19 @3:26pm - (1/4) Well put. Referring to a scientific paradigm committed to certain tools by the name of
the professions that have adopted that paradigm IS NOT AD HOMINEN. But, to pacify the offended, let me
declare publically that by saying "economists cant solve problem X" I mean:
3.10.19 @3:26pm - (2/4) "Researchers who have acquired ONLY the tools taught to them by economics
textbooks and economics professors adhering to those textbooks cannot solve problem X, unless they enrich
those tools by their own initiative". No hard feeling if they dont, gold medals if they do.(cont
3.10.19 @3:26pm - (3/4) We cannot ignore the fact that scientific research and education are organized into
tightly shielded paradigms (or cults) laboring to keep members from defection. In some extreme cases the tools
that an individual uses in research are perceived as proof of allegiance to
3.10.19 @3:26pm - (4/4) the cult or its gurus, an allegiance that entails significant academic benefits. The
sooner we admit this cultish structure the better we can correct for it. In summary, it takes courage to defect:
No hard feelings if you dont, Kudos and better science if you do. #Bookofwhy

3.10.19 @7:52am - (1/2) You and I make such claims all the time, and no one gets offended when we assert that
certain tasks cannot be performed absent certain tools. It does not reflect on the professional honor or
creativity of those lacking the tools. But I am still curious about your own personal
3.10.19 @7:52am - (2/2) experience among researchers who are offended by my observation that certain tasks
require certain tools. Have you really met anyone who can judge the plausibility of ignorability assumptions in
problems of size N>4? Please share your experience with us. #Bookofwhy

3.10.19 @7:12am - (Replying to @pophealth3) The analogy seems valid to me: Even with 3-dimensional objects,
we still need to calculate volumes using 2-dimensional surfaces.

3.10.19 @5:47am - (Replying to @ale_martinello) I am truly curious to know how economists did it w/o DAGs,
even in linear systems, in the presence of unmeasured confounders. What was the prevailing technique?

3.10.19 @4:57am - I am careful NOT to pose questions in terms of DAGs, that would be cheating. The questions
I asked are generic, e.g., which parameter in YOUR model can be estimated using OLS? I dont care if they use
DAGs or not to answer such important questions - I am surprised they dont care.

3.10.19 @4:27am - (1/2) In some simple cases, people who make ignorability assumptions can translate them
(mentally) into scientific assumptions and judge their plausibility. In most cases they can't and, then, the
conclusions are as opaque as the assumptions. Interestingly, last I checked they
3.10.19 @4:27am - (2/2) have refused to take a test to evaluate their ability to judge plausibility of their own
ignorability assumptions. But I am eager to learn from your experience. Have you met anyone who can judge
the plausibility of ignorability assumptions in problems of size N>4? #Bookofwhy

3.10.19 @3:50am - This is beautiful quote, thanks. First, note that I specifically refer to "causal question posed
in https://ucla.in/2mhxKdO ". Second, how many economists can you name who can answer those fundamental
questions? Seriously, don't leave us in suspense., more than 10? #Bookofwhy

3.10.19 @12:40am - (Replying to @thosjleeper) Again, I am from a small village and cannot make sweeping
statements like the above. In this paper https://ucla.in/2mhxKdO I was very specific about the kind of
questions most economists cannot answer. If you can answer any, tweet, and I will gladly note the exception
#Bookofwhy

3.10.19 @12:25am - (1/2) @petemohanty brought up Casella and Berger for a good reason, they are "pure"
statisticians who never claimed explicitly to deal with causal questions. Still, some statisticians today believe
that, if you dig deeply into the foundations of statistics in C&B book, you will be
3.10.19 @12:25am - (2/2) able to answer such questions. The people you cited (Angrist, Imbens, Rubin,
Gelman/Hill, Athey, Freedman, etc.) are not "purists", since they do talk explicitly about CI, and some even
admit to be doing so with extra-statistical assumptions (eg. "ignorability") #Bookofwhy

3.10.19 @12:03am - (Replying to @thosjleeper) I come from a small village. I do not speak of "the vast majority
of empirical research" nor of "all current statistical evidence". I deal with one claim at at time. Which claim
would you like me to evaluate? #Bookofwhy

3.9.19 @11:45pm - (1/2) I have invited you to take a leisurely journey into a book that you chose, search for
one toy example that you deem to be "causal" and decide if the book solves it to your satisfaction. So, why am I
unfair? I have extended this invitation to every colleague who tells me:
3.9.19 @11:45pm - (2/2) "Statisticians have done it decades ago." Thus far w/o success. So, where have I been
unfair to "decades of work..."? Fisher, Lord, Kruskal, and Lindley acknowledged that certain questions reside
outside the province of statistics; what would it take to convince you?

3.9.19 @10:59pm - (Replying to @BDataScientist) It reminds me of the finding (in the 1970's) that men earn a
higher salary than equally qualified women, and simultaneously, men are more qualified than women doing
equally paying job. It is described and treated in https://ucla.in/2LcpmHz under "reverse regression"
#Bookofwhy

3.9.19 @7:30pm - (1/2) The disconnect between statistics and causality is unfortunately not a matter of
"imperfection" and no amount of effort "to improve existing work" could bridge this disconnect. My analogy: it is
like asking a 2-dimensional creature what "volume" is. To prove my point,..
3.9.19 @7:30pm - (2/2) let's take an exploratory journey through the toy examples of Casella and Berger and
examine if ANY of them answers a causal question. If we find one, I will take back the 2-dimensional-volume
analogy and go study statistics. #Bookofwhy

3.9.19 @2:16am - Confessing negligence. I have been so fascinated by how much millage we get from the
primitive relation "is a function of" that I have neglected to explore other types of semantic networks. With the
exception of "Is part of", in 1986, see
https://www.sciencedirect.com/science/article/abs/pii/0004370286900275 #Boodofwhy

3.8.19 @6:46pm - (1/3) There isn't really great need to differentiate external validity vs generalisability vs
transportability, since we now we have a unified framework to handle them all, as in https://ucla.in/2Jc1kdD .
The most important distinction one needs to make is about the disparities
3 8 19 @6:46pm (2/3) between the study and target populations i e whether such disparities are "man
3.8.19 @6:46pm - (2/3) between the study and target populations, i.e., whether such disparities are man-
made" (as in recruiting subjects) or "nature made" (eg age differences). The interplay between the two is
described in https://ucla.in/2L6yTzE . Still, however we taxonomize these subproblems,
3.8.19 @6:46pm - (3/3) I would be very weary of any theory that does not provide you with playful solutions to

at least some toy problems, for example, the three toy problems in Fig. 3 of https://ucla.in/2N7S0K9 .
#Bookofwhy @BrownUniversity #kolokotrones @HarvardEpi @harvard_data

3.8.19 @5:59pm - (Replying to @djvanness) I am familiar with Manki's writings, and I am curious to hear from his
other readers HOW he represents that "extra statistical information", and whether he does that very well. My
litmus test for "very well" will be (as usual) the abilitiy to solve a toy problem #Bookofwhy

3.8.19 @3:46pm - (Replying to @malmyros) Causal diagrams in #Bookofwhy are a special (primitive) case of
semantic networks described in Minsky's book. The nodes are all "variables" (not "objects") and the arrows
represent ONE relation: "is a function of". Next generation AI will extend CI to full semantic networks.

3.8.19 @3:30pm - (1/2) If statisticians understand that "you cannot learn much of anything from a statistic
without `extra statistical' information," then one would expect statistics textbooks to spend at least a few
pages on how that "extra statistical information" can be represented ...
3.8.19 @3:30pm - (2/2) mathematically, and how it can be combined with data in order to obtain what
statistics alone cannot provide. I have not found such a textbook yet. But would be eager to learn. #Bookofwhy

3.8.19 @3:12pm - (Replying to @jwalkrunski) The statistical and causal notions of identification are compared in
footnote 7 of https://ucla.in/2N7S0K9 . The unique character of Q(M) is discussed in the other sources sited.

3.7.19 @6:02pm - (Replying to @omaclaren @eliasbareinboim) Any link to Stark?

3.7.19 @4:42pm - (Replying to @jwalkrunski) This concept of Q(M), and the identifiability definition that it
entails are the foundations of CI and appear in many works: from 1995 https://ucla.in/2Iz2LnH to 2014
https://ucla.in/2N7S0K9 to 2018 #Bookofwhy. I even called it "Principle-1" in https://ucla.in/2mhxKdO

3.7.19 @5:36am - (Replying to @ad_pickering) I havn't checked SEMNET lately. Are they still using regression?
Havn't they heard the bells of the revolution? Avanti Popolo .... The yoke of tyranny is broken... How do we
bring them into the fold? Do they still read The Journal of SEM? #Bookofwhy

3.7.19 @2:25am - (Replying to @iamconscious2 @RebekahBaglini) I wish I could be there. But, seriously, is
anyone doubting that we are having a causal revolution? I have an untested theory in mind: Those who doubt do
not know how to solve Simpson paradox (or any toy problem), and those who do not doubt can solve it. Worth a
test. #Bookofwhy

3.7.19 @1:41am - (1/1) The tools of Transportability analysis are almost unknown to researchers most badly
needing them, namely traditional trialists, probably because these tools are often buried in opaque ignorability
jargon. (see https://ucla.in/2NIsI6B). Students attending the
3.7.19 @1:41am - (2/2) transportability symposium at @BrownUniversity would undoubtedly be introduced to
modern methods, invoking graphs, which should turn every trialist to an expert in transporting study results
(see #Bookofwhy p. 354, and https://ucla.in/2umzd65 Tool 5) . Merry Transporting!

3.6.19 @10:45pm - (Replying to @omaclaren @analisereal) I understood your argument to be that allowing "all
functions" behind the nonparametric model is too permissive, some functions are pathological and would
prevent estimation. I proposed excluding such pathologies the same way we exclude non-positive distributions.
Why not?

3.6.19 @10:05pm - (Replying to @omaclaren) You make it sound like I disagree with your conclusions. I dont. I
disagree however to taking syntactic similarity of two definitions and calling it "exact equivalence". In one, the
model is defined by the thetas; in the second, Q is a function computable FROM the model.

3.6.19 @8:33pm - (1/2) Surely for every model M there is one distribution P(M). But this is where the 1-1 "index
set" analogy ends. The key feature of causal models, not share by statistical models, is that for every M there is
one Q(M), which the answer to our causal question Q. Identifiability
3.6.19 @8:33pm - (2/2) does not require inverse mapping M(P). Instead, it requires that whenever two models
agree on Q they also agree on P. Allowing "arbitrarily indexed family of distributions" is like allowing
arithmetical operations; thanks, but it misses the essence of causation. #Bookofwhy

3.6.19 @4:01pm - (Replying to @omaclaren) We normally attached the assumption of "positivity" to non-
parametric analysis. Would the sun shines brighter if we augment "positivity" with "non-pathological"? Can we
protect ourselves against pathological cases? #Bookofwhy

3.6.19 @3:38pm - (Replying to @cmirzayi @staci_risman) I agree. Authors are often surprised by who finds their
books useful or entertaining. I get letters such as: "I dont understand any of the math, but I admire Sewall
Wright and Barbara Burk" or "Now I understand what they tried to teach me in Econ. 101". ..etc. #Bookofwhy

3.6.19 @3:03pm - (Replying to @omaclaren) Questions. Suppose there is a class PD of pathologial distributions.
Can one recognize them from M? from the estimand? from the data? Can you demonstrate one such PD using 3
variables, 4 ? Suppose PD exists, does this minimize the importance of identification? #Bookofwhy

3.6.19 @2:53pm - (Replying to @LMcHugh_Russell @UnlearningEcon and 2 others) It seems to me that


#Bookofwhy should provide what you are looking for, especially when coupled with PRIMER
https://ucla.in/2KYYviP for fun and homework.

3.6.19 @7:24am - (Replying to @statwonk) I thought someone would rise to defend the honor of statistics. But
we are dealing here with "identifiability" not with causal intuition, nor with RCT. Take the simplest problem of
selecting covariates for adjustment, how many statistics professor can do it TODAY? #Bookofwhy

3.6.19 @7:03am - Correcting a link. The relation between the statistical and causal notions of identifiability is
explicated in footnote 7 of this paper https://ucla.in/2N7S0K9 .

3.6.19 @1:33am - I second @eliasbareinboim . The WHY-19 Symposium will give me, personally, the pleasure of
meeting readers and followers and profess/confess what I believe Causal Inference is all about, at least from
my humble, totally unbiased perspective. #Bookofwhy.

3.6.19 @1:11am - (1/3) Now it is my turn to say: No way! The "identifiability analysis" conceptualized as a
"property of a family of arbitrarily indexed probability measures," is good for statistical identifiability but is
totally useless for #causalinference. This conception is precisely what
3.6.19 @1:11am - (2/3) #Bookofwhy asks readers to abandon, and it may explain perhaps why statistics has not
made ANY progress in identifying causal effects for over 150 years. To do that we need extra-statistical
information. Footnote 7 in https://ucla.in/2N7S0K9Some explicates the relation between
3.6.19 @1:11am - (3/3) the two notions of identifiability. Some economists tried to import this "indexed
probability" conception to econometrics with known results. Readers of #Bookofwhy: Watch out for this trap; a
causal model is NOT a set of distributions, one for each choice of parameters.
3.5.19 @6:24pm - (Replying to @omaclaren @edwardhkennedy and 2 others) Once a causal parameter is
identifiable you forget about it being "causal". Dress it up as a "statistical" parameter, and ask "Is estimation
possible?" as if it came from a respectable statistical textbook. Causality plays no role in making the estimation
hard or easy. #Bookofwhy

3.5.19 @3:09pm - (Replying to @edwardhkennedy @omaclaren and 2 others) Pearl spends more time on
identifiablility, because he believes that the other task (estimation) can be handled by traditional, 150 years
old statistical methodology, and that it is our new understanding of identifiability that will make
#causalinference a reality. #Bookofwhy

3.5.19 @2:43pm - (Replying to @kareem_carr @generativist and 5 others) Not Typical??? Perhaps this was the
reason I thought: Eurika! I found someone at Harvard who can explain to me what is, and what is NOT a g-
method. Now I am back at the mercy of my inability to get it the official way. Hard! #Bookofwhy

3.5.19 @5:17am - (1/2) (Replying to @kareem_carr @generativist and 2 others) Your bite is an invaluable
resource for the outside world to understand the workings of Harvard culture. I will add a couple of comments
to those made by @analisereal. 1. SCM is more than DAGS+do. It contains all seven wisdoms of
https://ucla.in/2umzd65 . #Bookofwhy

3.5.19 @5:17am - (2/2) (Replying to @yudapearl @kareem_carr and 4 others 3. If g-methods are taught at
Harvard as "effects which epidemiologists care about." then what is NOT a g-method? i.e., what kind of effects
epidemiologists do NOT care about? Front-door perhaps? Napkin? Or mediated effects? Or is everything a g-
method? Why not? #Bookofwhy

3.4.19 @2:31pm - (Replying to @kareem_carr @generativist and 2 others) There is no room for embarrassment
in trying to taxonomize a new science, where every new trick gets a fancy new name. As the Mishna (220 AD
says: lo habayshan lamed, v'lo ha kapdan l'lamed. -- A bashful person cannot learn, nor can an impatient person
teach. #Bookofwhy

3.4.19 @8:40am - (Replying to @kareem_carr @generativist and 2 others) Three comments:


1. I do not find SCM (Structural Causal Models) on your list. See eg #Bookofwhy Does SCM unify all the rest? (my
school)
2. Are these schools independent of each other? Or nested within each other?
3. What are G-methods? Or, what is not a G-method?

3.2.19 @8:36am - (Replying to @wgrosso @shakir_za and 2 others) The point is that the limitations articulated in
http://ucla.in/2nZN7IH are not merely "interesting" but challenge the foundations of Bayesian epistemology.
Bayesian practitioners, however, continue to speak as if those limitations do not exist: This strikes me as very
bizzare.

3.2.19 @8:33am - Agree. And those who think that CI is a missing data prob. should read
https://ucla.in/2LdEjZW and compare the transparency of the assumptions needed, the theoretical guarantees
obtained, and the tools produced by each paradigm. It is a streetlight vs. flashlight comparison.

3.2.19 @3:10am - (Replying to @wgrosso @shakir_za and 2 others) Thanks for noting that the expression "waiting
for their posteriors to peak" may be given grotesque interpretation, with "posteriors" being the rear parts of
one's body, rather than "posterior probabilities continuously updated by data." Begging Bayesians for
forgiveness.

3 1 19 @1:57pm - (Replying to @shakir za @blei lab @fhuszar) I am glad you mentioned my "half Bayesian"
3.1.19 @1:57pm (Replying to @shakir_za @blei_lab @fhuszar) I am glad you mentioned my half Bayesian
paper https://ucla.in/2nZN7IH which makes me more convinced in "halfness" each time I read it. I wish I could
get serious feedback from the "full Bayesians", but they are busy waiting for their posteriors to peak on the
truth.#Bookofwhy

3.1.19 @6:34am - Paul, I agree with your observation. However the issue is not parametric vs. nonparametric as
much as it is causal vs. associational. ML vocabulary (at least in 2019) is purely associational or predictive, while
econometric parameters are causal, as you noted, residing out of ML

3.1.19 @6:24am - (1/2) @KoelleM While admiring your attempt to bridge #MachineLearning and #economics I
must add a note of skepticism. Economic parameters are CAUSAL, while the vocabulary of ML is statistical. The
two do not mix. E.g., Beta is defined causally as in https://ucla.in/2mhxKdO , Eq. (3)
3.1.19 @6:24am - (2/2) True, not many econometric books define beta this way. But those that do, enjoy
clarity and consistency, and escape the embarrassment of those that dont. See for example
https://ucla.in/2NRFn7e [The authors confessed to embarrassment] #Bookofwhy

3.1.19 @2:16am - (Replying to @fhuszar) To be honest, I meant the students in the audience but, now that you
mentioned the panelists, I would appeal to model-blind extrapolation. #Bookofwhy

3.1.19 @1:40am - I came across this interesting article about Causality in Machine Learning
https://blog.fastforwardlabs.com/newsletters/2019-02-28-public.html ... which also links to a recent panel on
causality https://www.youtube.com/watch?v=ynVr_zzUXtw ... It is interesting to see how young ML researchers
react to the new hype. Very encouraging #Bookofwhy.

2.28.19 @5:44pm - (1/1) The theory of transportability is not only a "nested version" of Transfer Learning (TL),
but the SOLUTION to the set of problems coined TL. It is listed as Tool-5 in https://ucla.in/2umzd65 and proven
complete in https://ucla.in/2PaSSyC . To the best of my knowledge
2.28.19 @5:44pm - (2/2) the enterprise of TL has yet to pass Pearl's Litmus test, of solving one TOY PROBLEM,
as those solved in https://ucla.in/2N7S0K9 and https://ucla.in/2Jc1kdD . Reason: TL is a probabilistic
enterprise while "transfer"requires causal assumptions. #Bookofwhy.

2.28.19 @12:38am - (Replying to @t_matam_t @audibleDE) Glad you found one. Can you share the link with
other readers -- I can use one myself. Thanks.

2.28.19 @1:30am - Kudos! And an overdue closure! I am extremely happy to see @DAGophile bridging modern
mediation analysis with the long, tormented and hopeless struggle that philosophers have had trying to capture
intuitions about direct and indirect effects using inadequate languages. #Bookofwhy

2.27.19 @11:28am - (Replying to @ConiByera @stephensenn @deaneckles) You hit it on the nail. Lord Paradox is
paradoxical in finite as well as infinite populations. It was introduced in the latter form, and has challenged
statisticians in that form. I do not see what insights can be gained by introducing finite sample complications.
#Bookofwhy

2.27.19 @2:14am - (Replying to @HenningStrandin @kareem_carr) I would take X=x to mean "the gravitational
pull at the molecule location is a vector x." And I would not even think about changing the moon location. I
would compare rushing under X=x to rushing under X = x' and that would give me the causal effect of X on Y.
#Bookofwhy

2.27.19 @2:05am - (Replying to @stephensenn @deaneckles) True, I did not know Nelder. But I looked up the
paper you cited and I did not see any potential outcomes to define the question nor DAG or PO to define the
paper you cited and I did not see any potential outcomes to define the question, nor DAG or PO to define the
assumptions. Perhaps he invented his own causal language. If so, can you translate?

2.27.19 @1:43am - (Replying to @stephensenn @deaneckles) If Nelder defines his research problem and encodes
his assumptions we can compare his solution to that of #Bookofwhy and decide if it is identical or elegant. But I
doubt he can do that. Why? Because both problem and assumptions are causal, and Nelder did not speak
causation.

2.27.19 @1:15am - (Replying to @kareem_carr) You seem to have forgotten an alternative C : Forget all this
torment about defining interventions. Imagine water molecules rushing to create a tide. Do they worry about
intervening with the moon's location? They just LISTEN to the gravitational pull and rush. Effect=listen+rush

2.26.19 @11:21pm - (Replying to @stephensenn @deaneckles) #Bookofwhy provides a solution to Lord's Paradox
given any DAG, including ones where initial weight has an effect on "Hall". See examples here:
https://ucla.in/2JeJs1Q . The trick is to 1. Define you question, 2. Encode your assumptions (DAG); the rest
follows (+fun), why resist?

2.26.19 @10:26pm - (Replying to @deaneckles) There is no longer any disagreement. We agree to talk about the

same DAG, authored by Miguel, with whatever humility and verisimilitude Miguel attributes to it. We see that
DAG carrying a variable named "gender" and we agree that it has causal effects. No trepidations. #Bookofwhy

2.26.19 @3:22pm - (Replying to @thosjleeper) To resolve disagreements, we must address them one at a time.
Disagreements about the structure of the DAGs are one kind, and those about whether a variable in the DAG has
a causal effect are of different kind. We are dealing with the second kind. #Bookofwhy #causalinference

2.26.19 @2:42pm - We must stick to the rules. We do not talk about "your DAG"; only about "Miguel-authored
DAG". If Miguel ever includes a variable "gender" or "blood pressure" in a DAG, then the causal effect of that
variable is quantified and interpreted with no trepidation. #causalinference

2.26.19 @2:29pm - (Replying to @omaclaren @ildiazm and 2 others) That's the whole point. Because I know that
you don't really believe "mud causes rain" I expect you to rebel against a formal theory that does not enable you
to articulate this simple belief. And, as a statistician, guess what this deficient theory is #Bookofwhy.

2.26.19 @2:20pm - Good point.,Thanks. I meant to write MOOT point. But, as a result, manipulability should
also become MUTE, namely, stricken from scientific discourse.

2.26.19 @2:09pm - (Replying to @kareem_carr @cshalizi) @kareem_carr, where is this quote from? It is full of
things I cannot agree with. But it is all a MUTE question because (I hope) Miguel and I have just reached a
consensus: If "gender" appears in your DAG, then it has a causal effect . see https://ucla.in/2sASU9V

2.26.19 @1:54pm - (1/2) Miguel, we have never been closer to agreement than this very moment --just replace
"your DAG" with "Miguel's DAG". Do you agree that the causal effect of any variable in any "Miguel-authored DAG"
can be quantified (if identified)?? If I hear a YES, we can jointly pronounce
2.26.19 @1:54pm - (2/2) the problem of manipulability a MUTE problem, to be stricken from the scientific
literature, past and future. Agree? And note that https://ucla.in/2sASU9V explains why we need to resolve
"blood pressure" first, before we go to "Obesity". So good to be in agreement #Bookofwhy

2.26.19 @1:27pm - (Replying to @ildiazm @_MiguelHernan @deaneckles) Ivan, thanks for defending
commonsense If it were not for you no one would dare ask Miguel: Come on! You do not really believe Pearl
commonsense. If it were not for you, no one would dare ask Miguel: Come on! You do not really believe Pearl
would put Unicorns/Cinderellas in a DAG, do you? So what's the point of evading the issue as articulated in
https://ucla.in/2sASU9V #Bookofwhy

2.26.19 @3:30am - (Replying to @GuillermoBurr) I am tempted to agree with your "three body" constellation,
except that I consider theories to be part of the solution, instead of the problem. #Bookofwhy

2.24.19 @2:22pm - Readers might enjoy the way "The Seven Tools of Causal Inference" appears in the March
issue of CACM https://dl.acm.org/citation.cfm?id=3241036 ... and the accompanying video
http://bit.ly/2GUEyJW . Too bad the Ladder of Causation was not illustrated in artistic colors #Bookofwhy

2.22.19 @11:50pm - (1/3) Thanks for defending my reputation. I have been thinking about why I sound
provocative to statistics-trained researchers doing #causalinference. I believe the reason is that I take the
difference between the two domains to be a "clash of civilizations", not of cultures,
2.22.19 @11:50pm - (2/3) and definitely not of a meeting of two "approaches". It is like trying to explain the
notion of "volume" to a two-dimensional creature. That creature will never forgive your provocative insistence
on adding another dimension to a comfortable 2-dimensional world. #Bookofwhy
2.22.19 @11:50pm - (3/3) Even in 2019, Gelman "finds it baffling" that we take comfortable statistical problems
and "wrap them in a causal structure" -- It is unforgivablly provocative. Yet you will never understand "volume"
unless you internalize the provocative idea of escaping from a 2-dim. world

2.22.19 @3:24pm - I would be very weary of anyone who argues pro or con any approach who refuses to solve
any toy problem in that approach. I spoke about the limitations of PO here: https://ucla.in/2v72QK5 only after
solving problems, side by side, in three different approaches. #Bookofwhy

2.22.19 @2:23pm - (1/2) Since #Bookofwhy is aimed not at the establishment but at free-thinking economists,
what prevents these economists from ignoring the establishment and communicating the indispensibility of SCM
to other enlightened economists? Fear of retribution? Lack of conviction?
2.22.19 @2:23pm - (2/2) As a student of Heckman, do you believe Jim now teaches his students some of the
indispensible elements of #causalinference, eg (1) structural counterfactuals (2) ways of deciding NP-
identification of causal parameters, (3) ways of deciding testability of causal assumptions ?

2.22.19 @12:32pm - (1/2) (Replying to @ghoshd) I fail to see why saying that "until the 1980s. The rest of
statistics, ... remained in the Prohibition era," would "alienate" any group. If I were a statistician, I would
rejoice hearing that something great happened to my field in the 1980's. As to Gelman's post
2.22.19 @12:32pm - (2/2) (Replying to @yudapearl @ghoshd) it is tainted with years of tormented attempts on
my side to have Gelman solve ONE toy example before speaking about @causalinference in statistics. And I
would appeal to your honest assessment: How many PhD statisticians do you think can solve such an example?
#Bookofwhy

2.22.19 @12:11pm - (Replying to @hangingnoodles) I am wondering, are accusations of "alienating people"


typical to scientific paradigm shifts? Thomas Kuhn does not elaborate on this aspect of shifts. Does anyone have
more pointed references to historical parallels? #Bookofwhy

2.22.19 @11:51am - (Replying to @learnfromerror) I have hard time relating "sample & parameter" to the issue
of knowledge organization. Curious; what do Bayesian statisticians say about the way humans organize
probabilistic knowledge in their heads? #Bookofwhy

2.22.19 @11:40am - (Replying to @cmirzayi) I believe I substantiated these characterizations with hard facts.
What do we call someone who excludes Israeli women from participating in a women's movement? Isn't "outright
bigot" a mild characterization? There is a sharp difference between "disagreement" and moral deformity.
2.22.19 @11:25am - Very keen observation. The reason you have not seen it is because Bayesian statisticians do
not feel comfortable making assertions about how people store probabilistic information; they leave it to
psychologists. They forget that priors come from people not from tables.#Bookofwhy

2.22.19 @10:34am - (Replying to @_limbs_ @loudquack @shravanvasishth) Well put! Here I did not imply
deliberately avoiding causality for some udlterior motives. I attributed the missed opportunity to (1) their
justified excitement over the discovery of correlation and (2) lacking a language to go beyond it. Deliberate
avoidance was practiced later.

2.21.19 @4:49am - (1/2) Two comments on this thread. First, Pearl does not "go after people". He goes after
outdated ideas and, occasionally, people invested in those ideas feel threatened. Second, while it is true that
the threatened tend to cling to their street lamps and totally ignore advances
2.21.19 @4:49am - (2/2) in #causalinference, I think you are underestimating the curiosity and critical thinking
of their students, and their ability to read the forbidden literature and to understand the limitations of what
they are being taught. I count on them. #Bookofwhy.

2.20.19 @10:17pm - (Replying to @Andrew___Baker) No problem. But if not too much trouble I would still like
to read at least one (accessible) article where a legal decision hangs on which of the two causal type we weigh
more heavily. It should be illuminating for #Bookofwhy 2nd Ed.

2.20.19 @8:39pm - (Replying to @jim_savage_ @Andrew___Baker) I dont get it. (Born naive). Can you be more
specific? What is it about "nobody has had a thought before".? #Bookofwhy

2.20.19 @7:29pm - (Replying to @Andrew___Baker) Blatantly wrong? #Bookofwhy it does not claim philosophers
did not distinguish necessary from sufficient causes, it claims that "what weight" to assign to each component
was left open. Am I wrong? Any counter-citation? Can we discuss weights w/o having quantities to weigh?

2.20.19 @10:33am - (Replying to @albertocairo @mendel_random) Agree with @albertocairo because, to me (as
a computer scientist), the litmus test of "taking them seriously" is commitment to NOTATION. As #Bookofwhy
argues, it was only in 1920 that slogans were broken with new notation for causal relations, escaping the
dryness of tables.

2.20.19 @1:43am - (Replying to @MHC_UNC @lsarsour and 2 others) It is sad for me to see an esteemed
university and a reputable School of Public Health fall for the charms of an outright bigot.
https://forward.com/author/judea-pearl ... @UCCpublichealth

2.20.19 @12:35am - (Replying to @maqartan @aecoppock and 4 others) But is it true that at the end of the
analysis all we can get is the ATE in a sub-population that is both unidentified and uninterpretable. I am not
minimizing the importance of the analysis, I am just checking if I understood its limitations. #Bookofwhy.
#causalinference

2.19.19 @12:27pm - (Replying to @aecoppock @SonjaASwanson and 4 others) Can't we just ignore the presence
of U1 ? According to the classical definition (eg Causality fig. 7.8(b)) Z remains a perfect IV with or without U1.
#Bookofwhy

2.19.19 @2:12am - Just received a Kindle version of http://bit.ly/2UGCJDU . Eager to see if my esteemed
colleagues contributed earth-shaking ideas to the discussions we have had here on Twitter. #Bookofwhy
2.18.19 @7:15pm - PRIMER https://ucla.in/2KYYviP is full of DAG puzzles designed carefully to build skill and
intuition (supported by Daggity.) See sample solution manual: https://ucla.in/2G11xUE . (Available upon
request.). My biased opinion: PRIMER is THE BEST introduction to CI. #Bookofwhy

2.18.19 @4:28pm - Statisticians themselves recognized that they can't be true to their own slogans and the "no
opinion" exclusion was officially struck out from the Royal Statistical Society manifesto in 1858. I wish we had a
record of the debate leading to that decision. #Bookofwhy

2.17.19 @8:40pm - Readers and passengers who happen to be in LA on Thursday 2/21 are invited to attend this
lecture by Stuart Russell, entitled: Human-compatible Artificial intelligence https://ucla.in/2GN4SWq
#Bookofwhy

2.17.19 @8:03pm - (Replying to @mayfer) We saw a microsoft resource tweeted around, but usually the
direction goes the other way: neural network modules are assisting a causal inference engine. #Bookofwhy

2.17.19 @7:47pm - (Replying to @zarzuelazen @Grady_Booch) I wish I could resonate with your architecture,
but seeing "Information Theory" next to "Action" my antenna rises: Information theory is purely probabilistic,

how can it govern Action ? It does not sit. Note: #Bookofwhy does not mention entropy, or Shannon's mutual
information.

2.17.19 @2:42pm - (1/2) Toy examples are educational devices absent of which a researcher cannot assess the
limitations of any given design, be it observational or experimental. It is necessary therefore for understanding
the limitations of RCT, whether they can be cured or replaced by obs. studies.
2.17.19 @2:42pm - (2/2) Here is an example of how researchers who have not mastered toy examples missed an
adequate assessment of how RCT limitations can be cured: https://ucla.in/2JfdQcv . I am speaking of
limitations such as selection bias and changing target populations. #Bookofwhy

2.17.19 @2:25pm - (Replying to @albertocairo) Glad they confirm the observation of #Bookofwhy, and I love
"The dryer the better" even more than "exclude all opinion" (ie. no priors) which I found in the inauguration
manifesto of the Stat. Royal Society.

2.17.19 @6:27am - (Replying to @thosjleeper) Agree!! But I see the simplicity of economists models a symptom
of their inability to think about more complicated ones. Solving ONE toy problem will open their eyes to what
they can do, especially in generalization and selection bias tasks, where they are so painfully behind.

2.17.19 @5:40am - (Replying to @thosjleeper) Disagree!! "It is only by taking models seriously that we learn
when they are not needed or not useful." (Must be Aristotle, 384-322 BC)

2.17.19 @5:29am - (Replying to @sidinusofaiii @intensivemargin and 4 others) https://ucla.in/2BDDTLK speaks 3


languages. Also, some folks are working on a Glossary of terms, to improve the translation. But they need your
help. Which concept, tool, or assumption from #Bookofwhy do you find needing a translation to your vocabulary
of choice? Please help.

2.17.19 @5:00am - (1/2) (Replying to @thosjleeper) I am proposing toy examples as educational, not persuasive
device. Folks who need tables of numbers to be persuaded about methods can easily find them in the
epidemiological literature and abstract back to economic problems -- the methodology is identical. Moreover
2.17.19 @5:10am - (2/2) (Replying to @yudapearl @thosjleeper) Those who have difficulties setting up the DAG
for their problems will not benefit from seeing how a problem was formalized in a totally different application
for their problems will not benefit from seeing how a problem was formalized in a totally different application.
Toy problems can help those folks by highlighting the commonality of structure. No substitute for toys.
#Bookofwhy

2.17.19 @1:47pm - (1/2) @PHuenermund was kind enough to remind us of his "not-invented-here" article
https://p-hunermund.com/2017/10/25/econometrics-and-the-not-invented-here-syndrome-suggestive-
evidence-from-the-causal-graph-literature/ ...where he also provides the quote from "Mostly Harmless
Econometrics" which you cited. As you can see, a model with 4 variable requires @metrics52 to spend pages of
informal
2.17.19 @1:47pm - (2/2) arguments. Imagine what an economist would go through in models of 10-20 variables.
(eg. fig.2 of https://ucla.in/2mhxKdO ). It is cognitively intractable and should be deemed impossible. This, I
know, will not convince the "not-invented-here" folks, but outdatedness will.

2.16.19 @10:10pm - (1/3) You ask if there is a "shorter" #Bookofwhy, and I assume you want to get the technical
meat w/o reading the stories. Yes, there is. If you take a look at Section 2 of https://ucla.in/2mhxKdO , you
will find the whole book summarized in 3 pages. But it must be supplemented
2.16.19 @10:10pm - (2/3) with the toy problems of Section 3. No matter how many books one reads ABOUT
economics, shying away from solving toy problems would leave one where econometrics is today -- two decades
behind the time. Plus, it is fun to see important methodological problems escaping their
2.16.19 @10:10pm - (3/3) textbook handcuffs and rejoicing game-like solutions. I therefore recommend: do not

skip the toy problems in https://ucla.in/2mhxKdO and their solutions. Try one - its better than reading a whole
book. Among the easy ones: Can your research question be answered using OLS?

2.16.19 @8:07pm - (Replying to @jehosafet @CarterPaddy) As #Bookofwhy narrates, epidemiology was fortunate
to have the enlightened leadership of Greenland and Robins (see https://ucla.in/2MpjsT6 ). Econometrics is
still begging for such leadership, and the gap gets wide in time. It's strange how much leadership counts.

2.16.19 @7:32pm - (Replying to @sidinusofaiii @intensivemargin and 4 others) ALPHA is any parameter in your
economic model. Most economists have trouble swallowing that elementary questions in Econ 101 (eg. about
ALPHA) require tools developed elsewhere; they see it as an insult to the profession, so they hope @metrics52
has it. Does he? Please check.

2.16.19 @6:10pm - (Replying to @mkessler_DC @CarterPaddy @t0nyyates) I think it should not take longer that
a 10-item glossary to do the translation. Let me start: Exogeneity = Nonconfoundedness, iv = iv, Control for =
condition on, Error term = Unobserved factor..Care to continue?

2.16.19 @5:24pm - (Replying to @intensivemargin @CarterPaddy and 2 others) That question of estimating
ALPHA is identical to the issue of "bad control" being addressed in MHE. If you can distinguish "bad control" from
"good control" you also know whether ALPHA is estimable by OLS. Thanks for bringing up the relation between
these two issues. #Bookofwhy

2.16.19 @5:02pm - (Replying to @intensivemargin @CarterPaddy and 2 others) I don't have a ready copy of
"Mostly Harmless Econometrics", but I am nevertheless willing to bet that the authors cannot answer the
question: "Given a general economic model, can parameter ALPHA be estimated by OLS?" How come? Because it
is super-human to answer it w/o graphs.

2.16.19 @4:32pm - (1/3) Your question: "why I wasn't taught the graphical approach" was raised by many
economists on this Twitter, and I have partially answered it in https://ucla.in/2mhxKdO and
https://ucla.in/2L8OCyl . Without going too deep into Psychology, the answer is "Not home grown!". (cont.
2.16.19 @4:32pm - (2/3) Your second question: "Would I be [taught it] today" is tricky. @Susan_Athey says there
is no need, because economists already have the answers. (see
https://twitter.com/yudapearl/status/1095979075674357760 ...According to others (eg @marcfbellemare,
@PHuenermund @causalinf), economists are beginning to rebel
2.16.19 @4:32pm - (3/3) against the tyranny of outdatedness. This workshop: http://why19.causalai.net will
provide an opportunity for both rebels and conformists to present their cases before Clio, the Muse of history.
#Bookofwhy.

2.16.19 @1:13am - (1/2) The DAG can only give us a partial preference-order on the functionals, as discussed in
https://ucla.in/2PORDX2 . In https://ucla.in/2ocoWqq however you can see a separate analysis, which declares
estimator # 2 superior to the other two. Interestingly, although functional #3
2.16.19 @1:13am - (2/2) is identical to # 1 (for all distributions) I am not sure the corresponding estimators are
equally powerful. Can the theory of influence functions shed light on these questions? @autoregress
@LauraBBalzer @edwardkennedy #Bookofwhy

2.15.19 @11:56pm - (Replying to @autoregress) 2.15.19 @1:23am - Correction: the three estimands are:
I corrected it. It should read: 1. E[Y|x]
1. E[Y|x] 2. SUM_z E[Y|z] P(z|x)
2. SUM_z E[Y|z] P(z|x) 3. SUM_z E[Y|x,z] P(z|x)
3. SUM_z E[Y|x,z] P(z|x) Which is most powerful?
And the answer is: ??????? See https://ucla.in/2ocoWqq

2.15.19 @12:24am - (1/2) Here is a simple puzzle for estimation experts: Consider the chain model X--->Z-->Y.
Below are three valid estimands of E[Y|do(x)]:
1. E[Y|x]
2. SUM_z E[Y|x,z] P(z)
3. SUM_z E[Y|x,z] P(z|x)
Each estimand defines a consistent estimator if we decide to take the MLE
2.15.19 @12:24am - (2/2) of each factor and combine them by the formula. Question: Which estimator is the
most powerful? (1), (2) or (3) ?? Once we agree on the correct answer, we will ask whether TMLE can help us
decide correctly, and how. #Bookofwhy

2.15.19 @12:12am - (1/2) I am glad to see this paper getting closer to publication. It was written in response to
a flood of Epi papers on generalizing study results, all from the potential outcome perspective, suffering of
course from the basic limitations of that perspecive (see page 1).
2.15.19 @12:12am - (2/2) I hope this paper opens new vistas for Epi researchers seeking to generalize study
results. #Bookofwhy

2.14.19 @1:32am - The sentence that gave me the chuckle is "Economists were responsible for asking and
answering all of these [causal] questions." It is 2019, and I can hardly name a handful (<6) of economists who
can answer even one causal question posed in https://ucla.in/2mhxKdO #Bookofwhy

2.14.19 @4:12pm - (Replying to @charleswangb) It is because A B E F and G are the parents of X, so they block
all backdoor paths from X to Y. #Bookofwhy

2.14.19 @5:48am - (Replying to @ildiazm @CalebMiles16) Agree. But where/when is the decision made that one
estimand is better than the other? Does TMLE examine all estimands? Before looking at the data, or after? Is
there an algorithm to extract optimal estimands from a given DAG? #Bookofwhy

2.14.19 @12:48am - (Replying to @CalebMiles16 @ildiazm) I do not believe the information provided by the
mediator Z is a convincing explanation as to why the product of the estimators is better than the one-shot
estimator. If we look at the saturated case (Appendix II https://ucla.in/2ocoWqq ), observing Z does not help.
#B k f h
#Bookofwhy

2.14.19 @12:16am - (Replying to @LauraBBalzer) I thought that once the TMLE obtains an estimand, she never
looks back at the DAG, she simply takes the estimand and finds the best estimator for it, given the data. If so,
what would tell it that the product of the estimates is better than the estimate of the product? #Bookofwhy

2.13.19 @2:06pm - (1/2) It was an inspiration to hear Laura speak at UCLA, and see TMLE dissected to its logical
roots. But nights bring second thoughts: Is it always true that optimizing overall estimand is better than
optimizing its individual components? Consider the chain X-->Z-->Y, where
2.13.19 @2:06pm - (2/2) the regression coefficients obey: R(YX) = R(YZ)*R(ZX). Guess which estimator is better?
The MLE of R(YX) or the product of the MLE of R(YZ) times the MLE of R(ZX)? Surprise! It is the latter! In blunt
defiance of what TMLE would suggest. See https://ucla.in/2ocoWqq #Bookofwhy

2.12.19 @11:52pm - I plan to stick around for another 2-3 decades, so just tweet if you want to dissolve any
disagreement. Lucky for us, modern #causalinference leaves no room for lingering disagreements, they can now
be examined under the microscope and breed poetry. #Bookofwhy #epitwitter

2.12.19 @9:32pm - (Replying to @abrahamnunes @GunnarBlohm and 4 others) I have not read "anticipatory

systems", but if you can summarize "what a model is" in Twitter's length I will try to comment on whether it
matches my understanding of what a model is.

2.12.19 @8:48pm - (Replying to @jamessseattle @DavidDeutschOxf) I am not familiar with Constructor Theory,
and the little that I read about tells me that it does not provide explanation that we, mortals, would call
"explanations".

2.12.19 @6:05am - (1/3) Dear readers and followers, I noticed that our Twitter audience has swelled to over
15K, which prompts me to thank you for energizing me with your comments and questions and for giving me the
illusion of being somewhat useful in this game of cause and effect.
2.12.19 @6:05am - (2/3) I have tried to keep my Tweeting focused on the science, minimizing foods, cute pets
and funny cartoons, yet entertaining enough to have you try causation's new tools that do what we used to think
we can't. I hope I can continue this experiment and make
2.12.19 @6:05am - (3/3) and make Tweeting exchange a valuable learning experience. As they say in the Mishna
(220 AD, anticipating Twitter): "Speak briefly and act fully" [In Hebrew: "Emor Meat Veaseh Harbe"] #Bookofwhy

2.11.19 @11:23pm - I am surprised that 51% thinks RCT has anything to do with the DEFINITION of causal effect.
Chapter 4 of #Bookofwhy labors to convince reader that it is just a means for "interrogating nature", that is,
nature has the answer before the interrogation. Plenty of work for educators.

2.11.19 @10:47pm - Stefan Conrady, Managing Partner of Bayesia, was kind enough to send us an interesting
selfie he took with the Lion Man that is featured in Chapter 1 of #Bookofwhy . Here it is
https://ucla.in/2SxKxLW

2.11.19 @7:46pm - (Replying to @Jabaluck @PHuenermund) I am also speaking about combining observational
methods and experiments in the same population, yet I find it hard to read Angrist's paper, because I cant
quickly see the principle: What is taken from each study? and how is it combined? Principles suffer when models
are avoided.

2.11.19 @6:05pm - (Replying to @Jabaluck @PHuenermund) If you are working on a fusion paper then I am sure
you will find the do-calculus to be helpful if not indispensible, as in https://ucla.in/2Jc1kdD and
https://arxiv.org/pdf/1902.01073.pdf ... . This is something economists cannot use, given their model-avoiding
mindset. #Bookofwhy

2.11.19 @1:17pm - (Replying to @Jabaluck @PHuenermund) This is indeed the key question. And the answer
depends on whether the RCT was conducted on 12K or 12 subjects, whether they were diabetic with likely
allergy to the drug, etc. etc. Some of these issues are discussed here: https://ucla.in/2JfdQcv #Bookofwhy

2.11.19 @1:07pm - (Replying to @charleswangb) No relation between the three layers. The assumptions could
be interventional or counterfactuals. Same with the Query. #Bookofwhy

2.11.19 @5:50am - (Replying to @NikHarmon @PHuenermund) On the other hand, there are (infinitely) many
unobserved Zs that can spoil the exogeneity of your IV, and infinitely many unobserved Ws that can spoil the
exclusion of your IV. Thus, we need to see the model before deciding on identification strategy #Bookofwhy

2.11.19 @2:46am - While I was bugged down in debates on "obesity" and "design", my faithful Google Alert
noticed that the field of #causalinference has advanced forward by this comprehensive paper on generalized
identification: https://arxiv.org/pdf/1902.01073.pdf . It's worthy of close attention #Bookofwhy

2.11.19 @12:40am - (Replying to @TheLeanAcademica) It is the first time I hear about it. Who would be a
typical beneficiary of the list?

2.10.19 @5:52pm - (Replying to @KordingLab @djinnome and 2 others) Model blind is not impossible if you
assume auxiliary conditions (eg ignorability) that turn your statistical routines into causal estimators. If pressed
hard, one can label these assumptions "a model", but then the causal story is gone, with all the fun of "seeing" it
working.

2.10.19 @4:45pm - (Replying to @KordingLab @djinnome and 2 others) #Bookofwhy is about causal MODELING.
The course, as I read it, is about MODEL-BLIND methods. There is some overlap there, once we extract from the
model all the information it can provide (e.g. estimability, testability) and we no longer need it. What remains
is statistics.

2.10.19 @4:15pm - (Replying to @robertwplatt) If you "get it" then you do not need my help. The distinction
between "analysis' and "design" follows logically from their respective characters. But I am open to learn what
language a person who "gets it" should choose to describe mathematical derivation of causal quantities.

2.10.19 @1:23pm - (Replying to @KordingLab @djinnome and 2 others) @KordingLab, I am unable to recommend
this course, except to seasoned #Bookofwhy students who wish to see with their own eyes what an "alternative
approach" looks like.

2.10.19 @12:50pm - (1/2) Agree. Look how careful and accommodating I am speaking to statisticians
https://ucla.in/2v72QK5 or publishing in http://Stat.Sc . But there are other folks there in #causalinference,
e.g., ML folks, philosophers, and students yet unharmed by traditional education.
2.10.19 @12:50pm - (2/2) These folks will take it as pathology if I defy logical standards and name a simple
"mathematical derivation" as "design". But, by all means, if you have conduits to traditional "designers", use
their language to show them that "design" is no longer "thinking hard".#Bookofwhy

2.10.19 @6:14am - (Replying to @thosjleeper) As an enlightened reader of both Causality and #Bookofwhy,
would you respect me if I play diplomacy and call formal mathematical derivations "DESIGN", just because many
y p p y p y ,j y
people were educated in an era when these derivations had to be done by informal "thinking hard". ???

2.10.19 @6:00am - (1/3) Causally speaking, it is the other way around. Those who "get it" cannot bring
themselves to call mathematical analysis "design" or "thinking hard". Moreover, speaking of "implementation of
these methods" sounds like there are "alternative methods". I do not see any; (cont.
2.10.19 @6:00am - (2/3) The choice is between doing thing systematically or retreating to the days of "thinking
hard", namely, between survival and staying behind. At the same time, if you and others think that mild
language would create more survivors, by all means, you have my blessing. But
2.10.19 @6:00am - (3/3) please do not ask me to call formal mathematical derivations "DESIGN", just because
many people were educated in an era when these derivations had to be done by informal "thinking hard".
Readers of #Bookofwhy will not forgive me, because they do "get it".

2.9.19 @10:36pm - (Replying to @thosjleeper) I just answered Robert Platt on this very question: No, The only
researchers who truly get it are those who are shaken by shifting vocabulary to realize that "the times they are
a-changin" and, to survive, they must re-think old molds. Diplomacy dupes them them into slumber .

2.9.19 @10:25pm - (1/2) The distinction between "statistical analysis" and "causal analysis" is made visibly and
consistently in all the #causalinference literature. The little confusion that remains dwarfs in comparison to the
benefit of shaking the old "designers" to understand that much of their
2.9.19 @10:25pm - (2/2) torment and "hard thinking" can be accomplished today by "analytical methods" The
only researchers who truly get it are those who are shaken by shifting vocabulary to realize that "the times they
are a-chanin" and, to survive, they must re-think old molds. It's #Thebookofwhy

2.9.19 @7:52pm - (1/2) In the old days, estimation was the only task that folks could submit to formal
treatment, so they called it "analysis". All the rest was entrusted to "hard thinking", so they called it "design".
Today, many of the latter tasks can be treated analytically and should (cont.
2.9.19 @7:52pm - (2/2) therefore be called "analysis". But traditional "designers" refuse to hear the bells of
change and prefer to retain their posture and terminology. Awaken them, "for the times they are a-changin'"
#Bookofwhy

2.9.19 @7:26pm - (Replying to @MariaGlymour) I am yielding to your epi experience if you feel that way. But
before quitting I would just leave you with an open question: Are you sure the clarity you obtain from thinking
RCT is not a placebo effect induced by our statistically dominated education? #Bookofwhy

2.9.19 @6:01pm - (Replying to @MariaGlymour) I get clarity and specificity by thinking about the actual policy
that is about to be implemented Monday, by so and so, who is about to issue an instruction to so and so. Why
should commonsense thinking need the help of hypothetical trials, distorted by randomization? #Bookofwhy

2.9.19 @4:26pm - (Replying to @MariaGlymour) I dont see the relation between specificity of policy and RCT.
There is no "randomization" in my thinking about minimum wage policy, not even "trial", there is only

implementation details and some knowledge of the forces involved. RCT is one way of testing, not a way of
thinking.

2.9.19 @3:53pm - (Replying to @MariaGlymour) But what if I do not have any RCT in mind, not even
hypothetical. All I have is a burning desire to predict how a 1$ increase in minimum wage would affect
unemloyment, if enacted tomorrow. Must we think RCT when we can think what variables would change, how
and why? #Bookofwhy

2.9.19 @3:07pm - (1/n) Of course I am not using "analysis" the way applied social scientists do. Recall, the
.9. 9 @3:0 p ( ) O cou se a ot us g a alys s t e way appl ed soc al sc e t sts do. ecall, t e
boundaries between "design" and "analysis" were drawn before people realized their research question can be
written down, submitted to analysis, and receive an answer. (cont.)
2.9.19 @3:07pm - (2/n) My definition of "analysis" is simple: Everything that the inference engine can conclude
without human input is "analysis". Any input needed from the researcher is "design". See #Bookofwhy . True,
most applied social scientists are not aware that inference engines exists.
2.9.19 @3:07pm - (3/n) Some even see the engine as a threat, namely, not in their textbooks. But living in an
exciting period of paradigm shifts entails redrawing boundaries, redefining outdated concepts and re-
formulated the fundamental question "what is the study goal?" I am excited! #Bookofwhy

2.9.19 @11:49am - (Replying to @pophealth3) The "graphs are bad" argument (and culture) explains why Rubin
had no model, and had to resort instead to a "think hard" recommendation. Unaware of this misguidance, some
people still think that "mostly harmless econometrics" is not harmful. #Bookofwhy".

2.9.19 @11:22am - Analysis is what tells you what confounders are relevant. Lacking tools for this analysis
("graphs are bad") Rubin recommended "think hard". Analysis means "Stop thinking so hard", let the mathematics
think for you. U "think hard" only once, when you commit to a model. #Bookofwhy

2.9.19 @4:18am - (Replying to @thosjleeper) By "imitation" I mean e.g. watching for placebo effects, or
selection bias. Why do we take those precautions? To create conditions which are similar to those created by

the do-operator -- the definer of "causal effects". See #Bookofwhy pp. 143-9

2.9.19 @2:44am - (1/4) Glad we agree on content. But if I recall correctly you said (on Target Trials) "this is a
design issue, not design" So what in your opinion is the dividing line between design and analysis? My dividing
line: whatever the inference engine can infer mechanically is "analysis"

2.9.19 @2:44am - (2/4) And whatever requires human judgment (on top of the inputs to the engine) is "design".
I would be curious to know if you agree. This means that when AI succeeds in producing an "automatic
scientist", everything will be "analysis". An interesting anecdote, I believe Don Rubin
2.9.19 @2:44am - (3/4) was first to pit "design" against "analysis" in a 2008 paper "FOR OBJECTIVE CAUSAL
INFERENCE, DESIGN TRUMPS ANALYSIS." But the number of mistakes in that paper does not speak much for
"design". That paper also sounds like the scientific manifesto of "Target Trials" #Bookofwhy
2.9.19 @2:44am - (4/4) I think the idea that an observational study should imitate some RCT is totally
misguided. It is RCT that should imitate things, not an observational study conducted in the natural habitat of a
population, where causal effects are derived properties of the model. #Bookofwhy

2.9.19 @1:53am - (Replying to @TunnelOfFire) If all variables are observed then the causal effect (of X) is
identified by the adjustment formula, with the parents of X as the "controlled for" covariates. See #Bookofwhy
p.220, or Causality, Eq. (3.13) https://ucla.in/2BDDTLK

2.9.19 @12:00am - (Replying to @austinvhuang) I am not after winning debates. I am here to show researchers
that "they need not distinguish manipulable from non-manipulable variables; both types are equally eligible to
receive the do-operator and both benefit us by evaluating its effects." https://ucla.in/2sASU9V #Bookofwhy

2.8.19 @10:10pm - (1/3) Yes, I have done the "science thing" and have shown here https://ucla.in/2EpxcNU
that E[Y|do(X=x)] is empirically testable even when X is non-manipulable, which was one of your requirement
to be "part of science". This still far from computing the effect of Obesity, but
2.8.19 @10:10pm - (2/3) as we know, science progresses incrementally. Obesity presents TWO obstacles, non-
manipulability and lack of agreeable measurement. Progress demands that we tackled one obstacle at a time,
so I took on blood-pressure which is well-measured but non-manipulable. I am satisfied
2.8.19 @10:10pm - (3/3) with the way the non-manipulability obstacle was removed -- are you? If not, then I
would really like to know how your team treats such variables when they show up in the dag. No trap here, just
y y y p g p ,j
innocent curiosity. #Bookofwhy

2.8.19 @8:40pm - The arguments on "Well-definedness" , pro and con, are laid out here https://ucla.in/2sASU9V
and here https://ucla.in/2EpxcNU . The latter contains a subsection on "off policy" actions (in RL) and how they
can be evaluated using "in policy" actions, once we have a causal model.

2.8.19 @4:35am - (Replying to @BL4PublicHealth @mathtick) Let's take the question "who r u going 2 recruit?".
How do you make this decision? Dont you take into account what question you need to answer, whether the
character of the recruitees impedes or assists your ability to answer that question? etc., This is what "analysis"
is about.

2.8.19 @4:22am - (Replying to @robertwplatt) I suggest that having this notation, and having the machinery to
infer the answer to our research question from this notation is "analysis". Moreover, design options that matters
(eg what to measure) can be decided mechanically using this notation, and knowing the question.

2.8.19 @2:29am - (Replying to @thosjleeper) Please examine the "design choices" you have and ask if you can
make them by whims, or you need to think twice before making them. If the latter, who is guiding you in
"thinking twice"? What knowledge do you consult in this thinking? Can't this thinking be replaced by analysis?

2.8.19 @1:56am - (Replying to @mathtick) It all becomes clear when we focus on a concrete example. Lets say
we have a "design choice" what measurement to take: Z or W, or both. The choice is not arbitrary, it takes
analysis to tell us if we can get by with Z only, W only or we need both to answer our research question.

2.8.19 @1:36am - (Replying to @stephensenn) Its OK to tell your students that such a notation was provided by
Adam and Eve, as long as you can also teach them how to combine observational and experimental studies to
get results that none can provide in isolation. See https://ucla.in/2Jc1kdD or #Bookofwhy page 357

2.8.19 @1:26am - Moreover, I cannot think of any "design choice" that does not require a close guidance of
analysis. I question, therefore, whether the word "design" still conveys respectability or, rather, escape from
responsibility. #Bookofwhy

2.8.19 @1:05am - (Replying to @Profinit_EU) I love this robot. Amazing how asking "why" can make you look
smart. #Bookofwhy

2.8.19 @10:48pm - Today we have mathematical notation to describe how data are generated, how treatments
are assigned and how subjects are recruited. See https://ucla.in/2Jc1kdD . Armed with this notation, do we
still need to talk about "design" as competitor to "analysis"? #Bookofwhy

2.7.19 @9:52pm - (Replying to @iwashyna @eliasbareinboim and 6 others) The English language offers us many

ways of expressing Y_x informally, so as to engage informal audience. For example, using counterfactuals:
"what if X were x," or simply "how X affects Y". These linguistic utterances precede the RCT metaphor by 2000
years, see #Bookofwhy

2.7.19 @9:43pm - (Replying to @eddericu @eliasbareinboim and 6 others) Speaking of "well-definedness", I am


waiting to hear from @_MiguelHernan how they handle a measurement of blood-pressure. Do they delete it
from the model? or leave it in the DAG and mark it as "dangerously non-manipulable"? If so, what do they do
with the markings? #Bookofwhy
2.7.19 @6:43pm - Related quotation: "If our conception of causal effects had anything to do with randomized
experiments, the latter would have been invented 500 years before Fisher." --- #Bookofwhy (page 135)

2.7.19 @9:36pm - (Replying to @eddericu @eliasbareinboim and 6 others) No way!. It is "covariate-specific


effect" because the authors keep on talking about the intended population (or a person) for which the study
result will eventually apply. "Well definedness" is a red herring as shown here https://ucla.in/2sASU9V and here
https://ucla.in/2EpxcNU

2.7.19 @6:11pm - (1/2) The danger of introducing Latin terminology without warning, is that there are dozens
of students anxious to tackle new problems with causality's new tools. They are wasting hundreds of hours only
to discover that "Target Trials" is just Latin for "causal effect". #Bookofwhy
2.7.19 @6:11pm - I would also question the premise that researchers speak RCT more fluently than they speak
causal effects. Most mortals are born with causal intuition, we then indoctrinate them to speak only RCT and,
now, we justify Latin as if they were native RCT speakers. Back to #Bookofwhy

2.7.19 @4:40pm - (Replying to @EpiEllie @eliasbareinboim and 6 others) I am not familiar with any causal effect
or estimand of causal effect that is not "ideal" in the same sense, namely, it is what RCT is trying to emulate.
So, I do not understand the point of introducing a new adjective into the crowded glossary of causal vocabulary.
#Bookofwhy

2.7.19 @3:45pm - (Replying to @robertwplatt @EpiEllie and 4 others) Suppose I never heard about the word
"design" and, naive me, all I know is how to estimate the causal effects of treatment X on a given subpopulation
from a combination of experimental and observational data on other subpopulations. What do I gain by singing
"design"? #Bookofwhy

2.7.19 @2:54pm - (Replying to @eddericu @iwashyna and 5 others) Delving more and more does not tell us
whether the research question answered by "Target Trials" is the same as the one answered by "covariate-
specific effect". Beside, why should we care about what RCT emulates if we can answer directly the research
question at hand? #Bookofwhy

2.7.19 @2:33pm - (Replying to @robertwplatt @EpiEllie and 4 others) The word "design" is sometimes used to
avoid analysis, I hope this is not the aim of #TargetTrials. For example, the question: "What is it appropriate to
compare that person to?" is answerable by analysis, once you explicate what the research question is. What is
it? #Bookofwhy

2.7.19 @2:50am - (Replying to @HenningStrandin) But my dissatisfaction with the causation-avoiding statistics is
aimed primarily at the inadequacy of their language (ie probability), hence their methods. What is it about
trad. empiricism that you dislike? #Bookofwhy

2.7.19 @2:44am - I second @analisereal idea. We have proposed a summer school in causal inference to IPAM
(at UCLA). So, the more evidence of interest we get from readers the greater our chance to have it approved by
their (rather conservative) board. #Bookofwhy

2.6.19 @11:30pm - (Replying to @edwardhkennedy @EpiEllie) I wonder if this quote should still holds in the era
of causation. Is it the case that statistics has studied the way scientists think about and process scientific data?
Something to ponder about, but I can feel the doubts crawling all over me. #Bookofwhy

2.6.19 @11:22pm - (Replying to @iwashyna @EpiEllie and 3 others) Fascinated by the term "#TargetTrials I asked
if it can be translated into a problem familiar to my students. I believe the answer is: Yes,it is the problem of
p y , p
evaluating "covariate-specific effect" as in https://ucla.in/2KYYviP , p. 70 (w identification conditions)No?
#Bookofwhy

2.6.19 @2:49pm - (Replying to @PWGTennant @babylonhealth) That's an interesting scenario. And while the #ML
and #DL and #RL folks are busy assuring us that: "Yes, we are all in favor of #causalinference", the business
community will outdate them by just doing it. Interesting. #Bookofwhy

2.6.19 @8:15am - (Replying to @PHuenermund) Well put. The puzzle is not so much "Why do economists not
trust ..." but "Why do economists authors avoid the question you raised"? Embarrassment? Lack of answer?Fear
of exposure? Love of darkness? #Bookofwhy

2.6.19 @6:24am - (Replying to @KordingLab @RivasElenaRivas @manuelbaltieri) If you organize such a workshop
in LA, I will be happy to participate.

2.6.19 @4:51am - (Replying to @manuelbaltieri @KordingLab) There are several advertised. But I would watch
for impostors. My litmus test: Can the instructor solve a toy problem in causal inference? For example, the
problems illustrated in #Bookofwhy

2.5.19 @6:10pm - For ML folks who asked: "what's in it for me?', the paper on non-manipulable variables,
https://ucla.in/2sASU9V , is now updated with a sub-section on untried actions in reinforcement learning (RL)
#Bookofwhy #causalinference

2.5.19 @2:35am - (1/2) Thanks for a beautiful summary of the two perspectives. It re-ignites an ancient
question I once asked: Do the Harvard Epi people really distinguish manipulable from non-manipulable variables
in practice? How? What do the do with a variable like blood-pressure that happens
2.5.19 @2:35am - (2/2) to appear (as obs. covariate) in their DAG? Do they delete it entirely? Leave it in the
DAG and mark it as "dangerous"? What do they do with the markings? Do they play with the incoming and
outgoing arrows? Anyone knows? #Bookofwhy #causalinference #Epitweeter @epiEllie

2.4.19 @3:09am - (Replying to @NeuroStats) I am not aware of a different word, though "leveraging prior
knowledge" is implicitly assumed in all philosophical accounts of causation. In Lewis account of counterfactuals,
for example, one has to specify prior knowledge about which worlds are closest to ours. #Bookofwhy

2.3.19 @7:47pm - Seriously, @learnfromerror , can you share ONE ingredient that you think my characterization
of Bayesianism is lacking. Hopefully it can be described in Tweeter grammar, same as I tried to do, without
going through volumes of intricate literature. #Bookofwhy

2.3.19 @7:20pm - (Replying to @cubic_logic @learnfromerror) It is always desirable to run more stringent test to
sharpen our knowledge. But what do we do if we need to bet before sharpening? Do we have a coherent

methodology to do so? Bayes said: YES. Cast you suspicions in priors and treat those priors as if they were
proportions.

2.3.19 @4:18pm - (Replying to @FamedCelebrity) If Bayes formula "isn't per se needed" can you mention an
alternative? In my experience of the 1980's all those who offered alternatives (eg. Dempster-Shaffer, Fuzzy
logicists) were careful (and proud) to proclaim themselves "non-Bayesians". @Bookofwhy

2.3.19 @4:10pm - Good question, but "use prior information" is not an afterthought; it is an empirically testable
statement: Would our frequentist bet differently knowing that the coin was obtained from the neighborhood
statement: Would our frequentist bet differently knowing that the coin was obtained from the neighborhood
grocery as opposed to a shady gambler? See #Bookofwhy page 90

2.3.19 @3:45pm - I beg to differ. "Must be assigned a probability" is vacuous unless it is a commitment to obey
the laws of probability after the assignment. Bayes' innovation was to translate "given that we know X" into
"conditioned on X" as well as all the laws of conditioning #Bookofwhy p. 102

2.3.19 @3:31pm - This is seriously all that I meant, with all the epistemological commitments implied by the
statement "assign prior probability" (#Bookofwhy bottom of p.102). But from your question I gather that you
require additional commitments from a red-blooded Bayesian. Can you share ONE?

2.3.19 @12:35pm - (Replying to @amine_ouazad @shell_ki) Right! "Ignorability" is not a statistical concept
despite the fact that it is dressed in probability outfits and sings the songs of conditional independence. The
causal character of this concept is betrayed by the counterfactual variable Y(1), which is not an observed
variable.

2.3.19 @11:57am - (Replying to @learnfromerror) Leveraging prior knowledge is the essence of Bayesianism.
When appropriate, that knowledge can be encoded through conditional probability, or priors on parameters,
and when it is not appropriate it should be encoded differently. I'm eager: What is your definition of
Bayesianism?

2.3.19 @11:44am - (Replying to @alexpghayes @f2harrell and 3 others) Causality without probability is
formulated nicely in structural equations, prior to assigning probabilities to the U terms, also called "error
terms" or "units". See the firing squad example in #Bookofwhy (chapter 1) or https://ucla.in/2Nbq7W2

2.3.19 @11:03am - Nothing makes me happier than to hear a leading statistician say: "convincing". Nowadays I
hear it only from PhD students and post docs, which is OK for the future of statistics but leaves me with the
feeling that I could have explained things better. Perhaps I could. Thanks.

2.2.19 @5:20pm - (Replying to @learnfromerror @elementary_peng @richardtomsett) At leaset people who eat
"hot dogs" agree on what "dogs" are. Do Bayesians agree on "who is Bayesian"? I gave a humorous definition:
"assigning priors to parameters." Do we have a more technical definition that would satisfy the high priests?
#Bookofwhy

2.2.19 @1:29pm - A friend sent me a mini-review of #Bookofwhy which appeared on The Verge's "AI Reading
List" https://www.theverge.com/2019/1/29/18200585/understand-ai-artificial-intelligence-reading-list-books-
scifi . Written by Rumman Chaudhury, it is a most perceptive and accurate description of what the book is
trying to say, in the context of current discussions of AI.

2.2.19 @12:49pm - (Replying to @AdrianBauman @JamesSteeleII and 5 others) Having seen painful problems
induced by vague terms such as "interrelatedness" I am now more cautious, to the point that my mind demands
to know what makes direct relatedness different from indirect relatedness. Surely your algorithm treats them
differently, should it?

2.2.19 @12:30pm - (Replying to @HenningStrandin @_MiguelHernan) If you include me among "people in causal
inference" then I would differ. My claims do concern objective properties of causation, as encoded in the
"listening" metaphor. Causation is the algorithm by which Nature assigns values to variables.

2.2.19 @11:36am - (Replying to @JamesSteeleII @SamueleMarcora and 5 others) What I would truly prefer is an
explication of what the arrows (or edges) mean The caption says "drive activity" which sounds causal and
explication of what the arrows (or edges) mean. The caption says drive activity , which sounds causal and
directional. More important, what do we want to convey when we connect X to Y but NOT to Z, even though Y
is connected to Z. #Bookofwhy

2.2.19 @12:54am - Sorry for the wrong url. My Toronto speech is linked here:
https://www.youtube.com/watch?v=9MOtUvZ-tPc

2.2.19 @12:03am - My college route was from Rutgers to Brooklyn Poly, but I only took it three times: 1.Written
exam, 2.Oral exam. 3.Thesis defense. Classes were boring. I publically confessed, and apologized for this unruly
behavior in a commencement speech at Toronto: https://www.youtube.com/watch?v=9YMHqO6Z7AI

2.1.19 @10:50pm - (1/2) This passage demonstrates the monumental linguistic hurdle that statisticians had (and
still have) in dealing with causation. Lindley was the only one to recognize the need for statistics to lift itself
from the two-dimensional plane to a 3-dimensional view of the world....
2.1.19 @10:50pm - (2/2) Even today, few statisticians are able to acknowledge this barrier. Gelman, for
example, writes: "I find it baffling that Pearl and his colleagues keep taking statistical problems and, to my
mind, complicating them by wrapping them in a causal structure." And this is 2019!

2.1.19 @10:16pm - Thank you, Gerry, for remembering Daniel.

2.1.19 @12:21pm - (Replying to @SFMagus @schmarzo and 2 others) I was quite involved in DST, late 1980's, and
even wrote two summary papers: https://ucla.in/2RetYPS and https://ucla.in/2Ranb9U . Readers of
#Bookofwhy would have fun trying DST on the Monty Hall problem and seeing it fail. Lesson: always try your
theory on a toy example.

2.1.19 @11:57am - (Replying to @SFMagus @schmarzo and 2 others) I am not familiar with DST-type algorithms,
and have chronic forgetfulness to acronyms in general. Would be glad to look into it if you can provide a link to
a concise description, hopefully in input-output terms. #Bookofwhy

2.1.19 @4:32am - (Replying to @f2harrell @tvladeck) Just downloaded? This is the fate of papers submitted to
"handbooks" or "antologies" -- no one reads them except librarians. But https://ucla.in/2nZN7IH is truly worth
reading. I could not have said it better, except perhaps expanding on the potential outcome framework.

2.1.19 @4:19am - Causality (2009) is the mathematical expansion of #Bookofwhy. See book page:
https://ucla.in/2BDDTLK . Sitting between them, in smiles and playfulness, is Primer https://ucla.in/2KYYviP .
All three are dealing with human intuition; each amplifying that intuition in its own way.

2.1.19 @4:07am - To the best of my knowledge a Kindle version of #Bookofwhy is available from Amazon (I have
one), and so is an e-audible version.

2.1.19 @3:56am - One pleasure that people in my age and my feeble memory are blessed with is the joy of re-
reading and get inspired by their own writings, as if they see them for the first time: My! This is so true! Please
post it again. "The siren song of curve-fitting." #Bookofwhy #AI #ML

1.30.19 @8:20pm - (Replying to @tvladeck @f2harrell) The causal revolution and bayesianism are not at odds.
They are orthogonal to each other. What is odd is for Bayesians to claim ownership on prior knowledge and
insist on expressing that knowledge in only one way: assign priors to parameters, else you ain't a "Bayesian".
1.30.19 @3:38pm - (Replying to @f2harrell @davidmanheim) Before asking about "putting a prior", which is an
outdated stat habit, lets ask what "probability of causation" is, how to define it formally, and what we want the
answer to tell us. The #Bookofwhy does it through PS (prob. of sufficiency) and PN(necessity). No prior needed.

1.30.19 @1:50pm - (Replying to @The_RickMc @dccozine) What you are saying is that "the vast majority of work
in political science and economics" is stone-aged and in crying need for re-tooling. I fully agree. Indeed, see the
Three Bullets" https://ucla.in/2NIsI6B - easy to remember, and easy to apply. #Bookofwhy #causalinference

1.30.19 @1:03pm - One can spray priors on any set of mathematical objects, turn the Bayesian engine on and
wait, as do most Bayesian mechanics, for the posteriors to peak on something meaningful. In most cases they
won't, and http://ucla.in/2nZN7IH tell us why. Let's think 'why' #Bookofwhy

1.30.19 @12:22pm - Jim, you have a good eye for good papers. I just re-read http://ucla.in/2nZN7IH and I find
it truly informative and convincing (no dog in fight). Yes, here is the link to all our papers, new and old
http://bayes.cs.ucla.edu/csl_papers.html . Some of the least popular are my favorites. @Bookofwhy

1.30.19 @11:44am - (Replying to @F_Vaggi) If I have a full causal diagram, why would I rush to spray priors
around, unless I was brainwashed by a professor who feels uneasy unless he/she sprays priors. And this holds
even when I am unsure about the diagram; spraying priors does not repair lack of causal knowledge.

1.30.19 @11:34am - (Replying to @The_RickMcs) Correction: Modern social scientists do use Pearl's framework,
eg Morgan&Winship. And while mud and rain is not their main concern, they do need to somehow encode the
knowledge that the price of beans in China does not effect the election in LA. #Bookofwhy

1.30.19 @10:56am - This paper https://ucla.in/2nZN7IH explains why priors on parameters is not a good way of
encoding the BULK of human knowledge, which is causal, not statistical. For a quick example, let us try to
express our prior belief that mud does not cause rain. My prior is high #Bookofwhy

1.30.19 @2:09am - I side with you. This paper https://ucla.in/2MjxxTq explains why @_MiguelHernan definition
of "well-definedness" is overly restrictive. Naturally, when one starts with narrow definitions, every scientific
question appears "ill-defined", and our job is to broaden them. #Bookofwhy

1.30.19 @12:42am - @ruchowdh My!My!"Humble writing style" is the most flattering compliment I heard from a
reader. The statisticians protesting "We knew it all along" should hear it. And you may be right, perhaps the
writing style IS humble, considering all the misconceptions #Bookofwhy had to undo

1.30.19 @12:21am - (1/2) Bayesian Networks are Bayesian in the sense described in #Bookofwhy (Chap 3) and
further defined in https://ucla.in/2nZN7IH . It is also the sense that makes some territory-minded statisticians
have adopted ...
1.30.19 @12:21am - (2/2) a narrower view of Bayesianism: "if you don't assign priors to parameters you ain't a
Bayesian." I dont buy it, do you? Priors on parameters is just one way of encoding prior knowledge, and not a
very good one.

1.28.19 @8:35pm - (1/2) I thank readers for congratulating me on this award, which carries special meaning to
every Israeli-born and to every history-minded observer. For context, the Hebrew University in Jerusalem was
inaugurated in 1925 by Albert Einstein and Sigmund Freud among other
1.28.19 @8:35pm - (2/2) scientific giants, and came to symbolize both the historical revival of Jewish
peoplehood and its age-old commitment to the pursuit of learning. I feel truly honored to be given the
opportunity to express admiration and show support for Israeli scientists and academicians.
1.25.19 @2:26am - (Replying to @Miles_Brundage) Very interesting article, thanks for pointing. The fact that
they are using the Ladder of Causation already indicates that the authors are on the right track. Now I need to
understand if the connection to RL is substantive or just in name. #Bookofwhy

1.25.19 @12:11am - And for readers who enjoy artistic science, here is a non-numeric counterexample:
M = X XOR U, Y= M XOR U

This yields Y = X and P(M|do(X))=1/2 and P(Y|do(M)) =1/2


So, the composition formula gives zero effect while the true effect is 1. #Bookofwhy

1.24.19 @12:17pm - (Replying to @omaclaren @JohannesTextor) We cant end this beautiful discussion on a
negative note. It sure DOES guide us on what to do. First, completeness tells us that there exists no np solution
if Y---M are confounded -- great help!!!. Second, it suggests creative remedies. E.g., measure an M2 mediating
M-->Y, etc.

1.24.19 @3:14am - (Replying to @4lertus @nsaphra @KaiLashArul) Sorry, I usually refer to the book Causality
(2000, 2009) as "causality". #Bookofwhy

1.24.19 @2:10am - (1/3) Your composition formula is right ONLY if we assume no M--Y confounding. But in
general, even if M is the only mediator between X and Y, the formula does not hold. This is in fact a beautiful
example how modern #causalinference deconstructs strongly held intuitions #Bookofwhy
1.24.19 @2:10am - (2/3) Intuition says: We have two RCT studies, one gives P(Y|do(M)), the second gives
P(M|do(X)), So, if we have a new study with X--->M--->Y the ONLY directed path from X to Y, we should forget
confounding and chain the two causal effects together to get P(Y|do(X)). WRONG! We do
1.24.19 @2:10am - (3/3) need to worry about confounding, but only Y--M, not X--M. It is also beautiful because
it is not easy to show that the intuition is wrong, so I bet most traditional researchers would go wrong here.
Funny, some researchers still refuse to believe in the revolution #Bookofwhy

1.24.19 @12:38am - (Replying to @omaclaren) The way to answer your question is to try and prove your formula
through do-calculus in the same way that front-door was proven (Causality p.87-88) and see if you get stuck. If
you get stuck (and you will) then you know the formula is wrong, because do-calculus is complete.

1.24.19 @12:14am - Contrary to intuition, the answer to your question is NO. When Y-M are confounded, the
chain-rule does not hold. We can get:
P(y|do(x)) = SUM_M P(M|do(x)) * P[Y|do(x), M]
but we cannot write P[Y|do(x), M) = P[Y|do(x),do(M)] because Rule 2 of do-calculus is not satisfied.

1.23.19 @12:56am - (Replying to @PauSchae) Because "main effects" are properties of "race" itself, and "side
effects" are properties of those "nature intervention" (eg. cosmic radiation, genetic ancestry, ...) that
determine why one person is black and the other white. #Bookofwhy

1.23.19 @12:42am - SUTVA excludes "race" because it insists that treatments obey the "consistency rule",
namely, those "nature's interventions" which determine "race" should be free of side effects. We can't even
conceive of those race-determining interventions, can we speak of their side effects?

1.22.19 @1:51pm - (Replying to @AnalyticExec) I can imagine that "why" would be of great importance to the
insurance industry, but I am not aware of a related technical literature. Have they attempted a formalization of
"why" comparable to "but for" and "NESS" that we find in the legal literature? #Bookofwhy
1.22.19 @4:39am - (1/3) I am surprised and delighted that economists found interest in my paper on non-
manipulable variables. Why delighted? Because economists deal with such variables all the time. Why surprised?
Because they do not have an expression comparable to E[Y|do(X=x)].(cont.) #Bookofwhy
1.22.19 @4:39am - (2/3) Even economists who drifted to Rubin's camp are stuck. True, they can write Q=E[Y(x)]
and pretend that it is the same as E[Y|do(x)]. But is it? Y(x) comes from Rubin's PO framework where X=x is a
name of a manipulable treatment that has sworn allegiance to SUTVA. (cont.)
1.22.19 @4:39am - (3/3) and SUTVA forbids race, gender and even "minimum wage" to enter the privileged
company of "well-defined treatments". So what do economists do? Same as other stuck people. They speak
SUTVA in writing and use commonsense in practice. I hope my paper helps them speak straight.

1.20.19 @5:13pm - (Replying to @NeuroStats @analisereal @paulportesi) Thanks @NeuroStats. Great window
into the mind of a 1952 statistician who had the courage to admit: "I did not know anything about path
analysis...Why?" And this is 30 years after Sewall Wright introduced path analysis and was left to defend it
alone. See #Bookofwhy chapter 2.

1.20.19 @2:25pm - (Replying to @mesuturkiye @Cambridge_Uni @CambridgeUP) Sorry, can't travel out of
California these days. Wish Twitter could carry my warm signature across the Atlantic. Warm e-signatures - the
next challenge for AI.

1.20.19 @1:53pm - (Replying to @MariaGlymour @PHuenermund and 3 others) Should the water molecules wait
for "legitimacy" before they "perceive" the gravitational pull of the moon and rush to join the tide? I wrote this
paper https://ucla.in/2MjxxTq to allay the hesitation of those poor molecules, as well as other natural beings
(eg humans) #Bookofwhy

1.19.19 @6:17pm - (Replying to @IFindAnomalies) Anxious to ask: Why Y ?

1.18.19 @11:18pm - (Replying to @statwonk @bajwa_jamal @nntaleb) I don't think even Gelman would classify
#causalinference methods as "qualitative". Yes, the input knowledge is qualitative (ie, "who depends on whom")
but the output is an "estimand", namely, a RECIPE for processing data to get a QUANTITATIVE answer to our
question. #Bookofwhy

1.18.19 @10:51pm - (Replying to @sarahkmels @stephensenn) So the hard question is: If two half-Bayesians co-
author a paper, will it be frequentist or Bayesian? The #Bookofwhy says: it all depends on their half-priors.

1.18.19 @7:21pm - (Replying to @TheLeanAcademic @StatModeling) What seems to be the problem. Please
share with us the voice of practical virtue.

1.18.19 @12:01pm - (Replying to @eddericu) Good catch. It will be corrected soon.

1.18.19 @11:59am - But someone needs to explain to stat students why their professors speak P when the world
speaks P+M. Someone needs to explain to them that it is temporary, that their professors will catch up. Now,
when you say that, the enlightened professors get offended: "I'm P+M!" #Bookofwhy

1.18.19 @12:46am - (Replying to @MaartenvSmeden) You must be lucky to be in an enlightened stat


environment. But look up the last ten presidential addresses on the RSS website. Not one touches on M. Talk
with all chairs of stat departments within a 300 Km radius from you - not one would know an object outside P.
Try it, I did.
1.17.19 @11:51pm - Very interesting! But can't we really account for path-dependence through side-effects.
Surely my headache would depend on whether I take aspirin from my cabinet or go buy it from a North Korean
pharmacy. But the difference should show up in M, if the latter option is serious.

1.17.19 @11:14pm - What seems to be the problem? Please share. I used to be a Bayesian myself, see why I
half-quit https://ucla.in/2nZN7IH . So my heart can't allow a Bayesian to remain frustrated, especially with
counterfactuals, the gems of scientific thinking. #Bookofwhy

1.17.19 @11:06pm - (Replying to @MaartenvSmeden) There is no "blame" here. Missed opportunity? Yes. But no
"blame". Statisticians chose to focus on P, not M, (90% of them still do) and so they have not encountered the
need to define "estimand" with the extra provision of being "a property of P". No blame, just facts #Bookofwhy

1.17.19 @10:54pm - (1/3 Indeed #Bookofwhy gives Neyman an honorary mention as a pioneer of the revolution.
However, his counterfactual subscripts are not events but indices of treatments. Thus they cannot be related to
nonexperimental data. (See Rubin's Fisher Lecture) The first formal connection
1.17.19 @10:54pm - (2/3) to data was made by Sewal Wright in 1920, and that is why a whole chapter is
devoted to him. Rubin and Rosenbaum also made this connection in 1983, through the consistency rule Y=x Y_1+
(1-x)Y_0 and, of course, the PO framework is part of the revolution. But, let us examine
1.17.19 @10:54pm - (3/3) ....examine what replaces P in the PO framework, so we can articulate and answer
causal questions. The revolution is defined by the result of this examination. Why resist the title "revolution"?.
Can't your students do today ten times more than anyone could two decades ago?

1.17.19 @8:23pm - (1/2) The primary reason some authors are reluctant to use the do-operator in
#causalanalysis is the oddity of saying do(X=x) when X is nonmanipulable (e.g ., sex or blood-pressure). To
diffuse this reluctance I am posting a new paper https://ucla.in/2MjxxTq that explains why EVERY
1.17.19 @8:23pm - (2/2) EVERY variable should be given equal right to invite the do-operator. The paper
explicates the scientific meaning of this invitation as well as its benefits and testable implications. Conclusions:
Manipulativity prohibitions - a relic of the past. #Bookofwhy

1.17.19 @3:44pm - (Replying to @MaartenvSmeden) I see no "bashing" in explaining to people why they cannot
find the item "property of the distribution" in their dictionary definition of "estimand". I see no "bashing" in
explaining that, prior to the revolution, EVERYTHING was a property of P. Isn't it true? So, why "bashing"?

1.17.19 @2:30pm - (1/4) Since the word "estimand" seems to be giving readers problems, let me share my
definition. First, it is indeed that which needs to be estimated, ..However, in the era of #causalinference, one
needs to add something, to avoid confusion. #Bookofwhy
1.17.19 @2:30pm - (2/4) One needs to add that an "estimand" is a property of the distribution that governs the
observed data,e.g. E[X], var[Y], E[Y|X=x],. We need it because causal quantities e.g., E[Y|do(x)] or E[Y_x|Z=z]
may also need to be estimated, but they are not expressed as properties of P.
1.17.19 @2:30pm - (3/4) hence they are not "estimands" unless they are "identified", i.e., reduced to properties
of P, and when they do, they can be regarded as "recipes" of what to do with our data (see #Bookofwhy) to get
a valid answer to our causal question. This important distinction is not in
1.17.19 @2:30pm - (4/4) favorite text, or encyclopedia, or handbook. Why? Because those were written by
statisticians prior to the causal era. For them, everything was properties of P, including the research question.
The were not interested in causal questions which are not defined by P but by M.

1.16.19 @7:08pm - (Replying to @Grady_Booch @IntuitMachine and 4 others) We call this facility "causal
modeling" and, miraculously, the model gives you both "what if I see x" and "what if I do x" from the same
compact representation. (Not to mention "what if did x' instead of x"). It is really a miracle. #Bookofwhy

1 16 19 @6 00 (R l i t @I t itM hi @tdi tt i h d 4 th )F h bl l D RL i
1.16.19 @6:00pm - (Replying to @IntuitMachine @tdietterich and 4 others) From my humble lens, Deep RL is
semi-interventional, for it cannot infer consequences of untried actions (eg ban cigarettes, raise minimum
wage). I would also hesitate making a big fuss about abduction; it has become a technical term, meaning
different things to different folks.

1.16.19 @5:44pm - (Replying to @ceptional) Well summarized.

1.16.19 @5:43pm - (Replying to @tmorris_mrc @kareem_carr) Tim, If you understand how statisticians think,
please tell us how 20K followers of Gelman's blog do not rise up to demand an explanation for the logic of doing
estimation before identification. Is it out of politeness, lack of curiosity? or what? #Bookofwhy

1.15.19 @11:44pm - Yielding to readers' interest in my exchange with Andrew Gelman, I have posted a portion
of it on my blog and am offering it here for Twitter inspection: https://ucla.in/2FCjtTL . I have learned a lot
from it, but still not quite sure how many statisticians think. #Bookofwhy

1.15.19 @11:22pm - (Replying to @sksq96) Wishing you smooth sailing and a poetic sense of accomplishment.

1.15.19 @11:16pm - (Replying to @IntuitMachine @tdietterich and 3 others) Beautiful lecture by Hassabis. I
could not resonate his the interpolation-extrapolation-inventiveness ladder, perhaps because curve-fitting gives

both interpolation and extrapolation. But I was struck by his repeated calls for counterfactual thinking. He
would like #Bookofwhy

1.15.19 @5:39pm - (Replying to @austinvhuang @omaclaren) Agree. In this case, instead of yes/no refutation,
we have almost-yes/almost-no refutations. This is always the case in noisy worlds.

1.15.19 @5:04pm - (Replying to @austinvhuang @omaclaren) You do not need to guess the functions that drives
your model. You can falsify the structure without knowing the functions. #Bookofwhy

1.15.19 @4:24pm - (Replying to @omaclaren) No, you do not need to guess the function. You can falsify w/o
guessing. Overall I cant figure out what you are after. If you want to convince me that there are other tasks in
physics and engineering that use physical models, and do not need identification, I am convinced.

1.15.19 @3:43pm - I assume many readers are wondering about the origin of those funny "doors". They come
from an ancient (1993) paper "Graphical Models, Causality, and Intervention" https://ucla.in/2pJqtW3 and,
since no one complained, I kept using them https://ucla.in/2Iz3OnD and even #Bookofwhy

1.15.19 @3:29pm - (Replying to @omaclaren) Labels aside, no one is excluded. If you know the functions, you
are welcome to use them but, then, you will not benefit from the magic of "nonparmetric identification", which
miraculously estimates your query of interest WITHOUT knowledge of those functions. #Bookofwhy

1.15.19 @1:53pm - (Replying to @omaclaren) Oliver, the Pearl I know has not done "parameter estimation" since
WWII. On the other hand, "statistical falsification" is something he is very proud of: It's Tool #1 here:
https://ucla.in/2umzd65 . For more about his work, try #Bookofwhy and (dont miss!) https://ucla.in/2KYYviP

1.15.19 @6:52am - If you only heard #Harvard side, Primer will poison you irreversibly. For the first time you
will meet free counterfactuals, liberated from their leashes to "treatments" and RCT's, smiling in their natural
habitat, ready to be estimated and help others get estimated. Irreversible
habitat, ready to be estimated and help others get estimated. Irreversible

1.15.19 @4:31am - (Replying to @LydiaManiatis) Good catch. It should read: "Glad I had the time, in 2012, to
prepare lectures."

1.14.19 @11:09pm - (1/2) My favorite toy problem is Simpson's paradox: Would you or would't you recommend
the drug to a randomly chosen person. Just asking yourself: "What do I need to know to decide?" already
elevates you two levels above most data analysts. Next, I would go to #Bookofwhy
1.14.19 @11:09pm - (2/2) and to https://ucla.in/2KYYviP especially https://ucla.in/2G2rWBv , with dozens of
worked out examples. Leave the philosophers to argue what causality REALLY is, and enjoy solving causal
problems they can't. I had a professor who said: "From doing comes the understanding".

1.14.19 @6:25pm - I do not understand this hunger for Twitter fights, when solving one toy problem can teach
us so much more, both conceptually and operationally. If you watch any such debates you can tell immediately
who solved a problem and who just talks "about" things from a smelling distance.

1.14.19 @3:01pm - (Replying to @quantadan @NeuroStats) Quanta, you are so so right! And I say it at the
frightening risk of sounding self-serving: "There is no better introduction to causal inference." I dream of a
world in which data scientists communicate informedly on models, actions & counterfactuals see
https://ucla.in/2G2rWBv

1.14.19 @1:52pm - Worth following, agree. In the past two days, the discussion culminated in a conditional-
convergence based on a division of labor, and on new evidence that Pearl is not as thoughtless as Gelman's
review portrays him to be. #Bookofwhy

1.14.19 @1:32pm - (Replying to @adibzaman @eliasbareinboim) If we were to ask this question on every ML
paper published we would have had some AI in the output. #Bookofwhy

1.14.19 @1:22pm - I like this lecture, thanks for re-posting. It goes quickly over the fundamentals (including the
two basic laws of causal inference), and delves into novel applications. Glad I had the time, in 2002, to prepare
lectures. #Bookofwhy

1.14.19 @8:27am - (Replying to @DimDrandakis @ipfconline1 and 4 others) Not really. Folks doing prediction
and object-recognition are not really trying, and those doing model-blind deep-learning are trying the
impossible. My explanation: https://ucla.in/2umzd65 . #Bookofwhy

1.14.19 @7:58am - (Replying to @JohannesTextor @mendel_random and 2 others) Thanks for pointing to, and
correcting the ije summary, its really misconceived, almost as bad as the "pluralists" whining. My take:
https://ucla.in/2Pnj22p

1.12.19 @11:04pm - (Replying to @omaclaren @eliasbareinboim and 2 others) Gelman's statement: "Statisticians

use models all the time, statistics textbooks are full of models" is plain wrong --I just examined all stat
textbooks on my shelf. We could perhaps excuse engineering/physics books that estimate parameters in systems
of equations of known form.

1.12.19 @7:17pm - (Replying to @omaclaren @PHuenermund) DAGs are to regression as fire is to water. But this
is secondary. I mentioned the problem of "deciding which covariates to adjust for" because it remains an
elementary problem in @causalinference that 90% of statisticians cant do. Can BHM folks do it? Can anyone that
k ?
you know?

1.12.19 @6:39pm - (Replying to @omaclaren @PHuenermund) This is indeed a causal model that, in some deep
sense, can compute the effect of rain on mud. Thus, presumably, one can use HCM to do every task in
#causalinference, ie, decide which covariates to adjust for, decide if an ACE is identifiable, etc. Would you like
to try on a toy?

1.12.19 @6:24pm - (Replying to @omaclaren @PHuenermund) But does it present any method of estimating ACE?
I do not question the quality of the book, I just want to know if it is relevant to works which normally fall in the
category of #causalinference. #Bookofwhy

1.12.19 @5:51pm - (Replying to @omaclaren @PHuenermund) I was craving for example of a causal problem,
solved by standard statistical means, (or by hierarchical models,) not just any example. Does the book has ANY
causal example? Say finding an average causal effect, or effect of treatment on treated, or counterfactual,
etc... ?

1.11.19 @8:59am - (Replying to @pophealth3) Clinical trials became textbook material in 1935. This is one place
where statistics allows for causal considerations (albeit informal), and #Bookofwhy acknowledges this
allowance. What about observational studies, Haavelmo, Rubin and Robins? Were not the ten presidents aware?

1.11.19 @8:12am - (Replying to @pophealth3) (1/2) What ten presidents say over a period of 20 years tell us
something about how the community as a whole viewed causation; a high priority challenge or a nuisance that
should better be handled by others? It also reflects on Gelman's claim to have been doing CI for 30 years.
1.11.19 @8:15am - (Replying to @yudapearl @pophealth3) (2/2) Was he a rebel or a leader in the eyes of an
establishment that failed to even mention CI as a worthy challenge? Was he really doing CI or just imagined so?
#Bookofwhy

1.11.19 @1:00am - Sorry, the link to the 10 Presidential Addresses is here: https://t.co/uH7ilQDrNt covering
the period 1999-2017, i.e., the heat of the "Causal Revolution", with not an echo from the statistical leadership.
#Bookofwhy

1.11.19 @12:49am - An interesting document came to my attention, linking to 10 RSS presidential addresses
https://t.co/C5LfAR7w1d . It illustrates the amazing disinterest of statistics in #causalinference, as noted in
#Bookofwhy and protested/denied by leading statisticians (eg Gelman, Senn...)

1.10.19 @12:35am - It is only after tweeting "SUTVA is for PO folks who speak no structure" that I came to
realize, indeed, how clearer #causalinference classes could become by skipping discussions of SUTVA and going
straight to structure. Thanks for raising the SUTVA issue. #Bookofwhy

1.10.19 @10:32pm - (Replying to @cdsamii @eliasbareinboim and 3 others) (1/2) I confess to be among those
who are trying to understand your new framework, and can't. Since you are assuming away ID through
ignorability, there remains a new framework for estimation. Can we start from here, specifying what is needed
and what is given? I will try ...
1.10.19 @10:45pm - (Replying to @yudapearl @cdsamii and 4 others) (2/2) (1)Needed: a measure of program
effectiveness in the target population (2)Given: information coming from a source population. But since we
assumed the two populations are essentially the same (Mod. ignorability), what makes this estimation problem
unique? The estimand? How?

1.10.19 @3:01pm - (Replying to @SueMarquezR @jim_savage_ and 4 others) Thanks Sue, this is a thoughtful and
precise paper on a deeper-level "structural economics" literature. However, it appears to me that anyone doing
this deeper analysis should be able to do counterfactuals in structural equations. Yet the latter is oddly missing.
#Bookofwhy

1.10.19 @2:39pm - (Replying to @cdsamii @PHuenermund and 3 others) Can you help us understand what "old
stuff" you are using? I for one am not familiar with any "old stuff" that would solve even a simple
transportability problem (unless the two environments are (conditinally) the same, ie ignorability). One "old
stuff" would do. #Bookofwhy

1.10.19 @2:20pm - (Replying to @_limbs_) Agree. The arrows in hierarchical models are set-subset
relationships, not cause-effect as in DAGs. In this sense, it is still fair to say that Gelman does not use DAGs in
his writing. E.g., he cannot infer conditional independence in his model, not to mention identification.

1.10.19 @1:57pm - (Replying to @jim_savage_ @JimMinifie and 4 others) Appreciate the warning, good point.
But I am not saying "must use my method". I am genuinely not aware of any definition of counterfactuals in
structural equations, or any method of identifying counterfactual statements in the economics literature.
Concrete pointer would help.

1.10.19 @1:41pm - (Replying to @cdsamii @PHuenermund and 3 others) Paul is right. Even if the issue is
estimation, readers need to know what identification method is used, what assumptions are made, and perhaps
what asymptotic results prevail. Your paper sounds like solving an identification problem. Why not help readers?
#Bookofwhy

1.10.19 @7:17am - (Replying to @dailyzad) Good question. I know how. Give me a toy example, and I will not
try to avoid it.

1.10.19 @12:57am - (Replying to @mendel_random @StatModeling Some folks think differently, e.g.,
https://www.atsjournals.org/doi/pdf/10.1513/AnnalsATS.201808-564PS They understand that solving toy
problems is pre-requisite to understanding causation, and avoiding them cast doubts on the validity of "real
world answers" #Bookofwhy

1.10.19 @12:35am - It is only after tweeting "SUTVA is for PO folks who speak no structure" that I came to
realize, indeed, how clearer #causalinference classes could become by skipping discussions of SUTVA and going
straight to structure. Thanks for raising the SUTVA issue. #Bookofwhy

1.10.19 @12:15am - (Replying to @JimMinifie @jim_savage_ and 4 others) @JimMinifie I am not sure who those
stat/econometricians are. Heckman ? Angrist? Imbens and Rubin? If the latters, I have related their work to SCM
here: https://ucla.in/2mhxKdO , I did not repeat it in #Bookofwhy (only in passing) because it would have
sounded too negative.

1.9.19 @11:08pm - (Replying to @_limbs_) Impossible. Gelman has never used DAGs in his work (no Rubin's
student ever has) so his notion of "model" is different. For him postulating a parametric family of distributions
(normal, Bernoulli, logistic etc) constitutes "a model". The idea of "causal model" is unknown there.

1.9.19 @9:46pm - (Replying to @jim_savage_ @Andrew___Baker and 2 others) @jim_savage I am willing to take
the test and would like to find out how "policy counterfactuals" can be constructed without the help of DAGs.
Can we start with a simple "policy counterfactual" problem with 3-4 variables, so that we can see our way
through it? I am ready

1 9 19 @8:15pm - John is an economist who never heard about SUTVA He simply learned to compute
1.9.19 @8:15pm - John is an economist who never heard about SUTVA. He simply learned to compute
counterfactuals from structural equations, as in Primer: https://ucla.in/2G2rWBv Will John ever go wrong in his
scientific career? He won't! SUTVA is for PO folks who speak no structure. #Bookofwhy

1.9.19 @5:35pm - (Replying to @JadePinkSameera) It is hard to illustrate "reluctance", but take a look at
Stigler's "The Seven Pillar of Statistical Wisdom", which is a thorough account of the achievements of 20-Century
statistics, and judge for yourself how proud he is in the contributions of statistics to #causalinference.

1.9.19 @2:55pm - An ounce of advice to readers who comment on this "debate": Solving one toy problem in
causal inference tells us more about statistics and science than ten debates, no matter who the debaters are.
#Bookofwhy

1.9.19 @11:24am - (Replying to @PHuenermund @eliasbareinboim @heckmanequation) Heckman loved the do-
operator so much that he renamed it "fix-operator" and went on to praise "fix" and criticize "do". He forgot that
the operator comes with the benefits of graphical models, and left economists hungry for transparent causal
inference. https://ucla.in/2mhxKdO

1.9.19 @4:04am - (Replying to @rquintino) Simpson's is the litmus test of causal thinking. You want to know
what a statistician knows about causation? ask him/her about Simpson's paradox. [And you will be mighty
surprised at the majority of the answers. Try it.] #Bookofwhy.

1.9.19 @3:54am - (Replying to @mgaldino) Oh, No, Pearl does get it, for he has heard the "less convincing"
argument used many times in the past. But today his request is aimed not at "convincing" Gelman but at
establishing credibility and a common language of communication. One toy problem please, forget "convincing".

1.9.19 @1:09pm - (Replying to @f2harrell) Would love to, but the wealth of topics overwhelms me. I can only
contributes to discussions where I know something about. Do you want to start with Gelman's views on
causality? Or his concept of a "model"?

1.8.19 @11:59pm - Gelman's review of #Bookofwhy should be of interest because it represents an attitude that
paralyzes wide circles of statistical researchers. My initial reaction is now posted on https://bit.ly/2H3BH3b
Related posts: https://ucla.in/2sgzkPZ and https://ucla.in/2v72QK5

1.8.19 @8:37am - (Replying to @PHuenermund) My point is that had he understood the #Bookofwhy he would
have realized that the toned is justified. e.g. there is NOT a single model in any statistics text if "model" is
taken in the sense defined by #Bookofwhy, namely, assumptions about the world external to the data.

1.8.19 @8:09am - The qualitative/quantitative distinction is forgivable. When we write down a model we are
only expressing qualitative relationships. But the rest of Gelman's review entices me to pose a new discussion
topic: Which is Gelman's main obstacle to understanding/accepting #Bookofwhy?

1.7.19 @7:07pm - In view of recent interest in front-door estimators and their relation to regression, it would
be interesting to note that front-door actually hatched from regression analysis, as seen in this 1993 paper
https://ucla.in/2pJzGNK . #Bookofwhy

1.7.19 @6:14pm - (Replying to @fhuszar @paulwillemjvr and 3 others) Much relieved.

1.7.19 @2:47pm - (Replying to @paulwillemjvr @shakir za and 3 others) Anti-causal in 2019? Can't be!
1.7.19 @2:47pm (Replying to @paulwillemjvr @shakir_za and 3 others) Anti causal in 2019? Can t be!

1.7.19 @5:39am - (Replying to @juli_schuess @SLMsociology and 2 others) It is really beautiful and short, but it
takes an artist to appreciate beauty, and a philosopher to appreciate the benefit of unification. Glad beauty is
still a virtue among readers of #Bookofwhy. Most PO enthusiasts get upset when you show them how Y_x is
computed from a model.

1.6.19 @5:38pm - (Replying to @omaclaren) My copy says: . "Here B can precede X; yet we should still not
condition on B, because that would violate the back-door criterion." [#Bookofwhy Readers: If in doubt about
typos, please check "errata sheet" on https://ucla.in/2KYvzau . All typos were corrected on 4th printing]

1.6.19 @4:41pm - (Replying to @omaclaren) I checked my copy. I see no typo, nor reverse causation. Are you
still having problems with the M-bias? #Bookofwhy

1.6.19 @4:26pm - (Replying to @seanmcarroll @littlebode) I am eager to read you expansion of this idea, which
I mimi-leveraged to explain compatibilstic free will (page 364 of #Bookofwhy). More needs to be done there,
especially to show that the thermodynamic direction of time remains in tact under determinism. Please keep us
informed.

1.6.19 @12:34pm - (Replying to @littlebode) What does Sean Carroll say about causation? I hope good things,
perhaps new ideas? new problems? new tools?

1.6.19 @12:29pm - I have hoped the #Bookofwhy would shake up stagnation by emboldening students, but you
found a more effective way -- embolden writers. Brilliant! And may progress be your bounty.

1.6.19 @11:39am - (Replying to @yskout) The relationship is not clear to me. When I learned factor analysis, it
was presented as a purely probabilistic problem (eigen values of covariance matrices), hence, totally divorce
from causality. But perhaps my teachers failed to motivate it from a more causal angle.

1.6.19 @11:33am - (Replying to @thedmca @jnd1er and 3 others) Thanks for the kind mention.

1.6.19 @6:27am - (Replying to @ChrisAdamsEcon) I dont see why guidelines for economists should be vastly
different from those issued in epidemiology. The mantra "No causation without clear communication" should
rule both. #Bookofwhy

1.5.19 @10:29pm - This is an incredible vote of confidence in the methodology of #causalinference, especially
in observational studies. The standards that regulate scientific writing shape scientific thoughts and practice. I
wish editors of economics journals generate similar guidelines.#Bookofwhy

1.5.19 @8:41pm - (Replying to @SLMsociology @erikdbwestlund @stevepowell99) Why should anyone disagree? If
readers can solve the toy problems on page 93-100 of Primer https://ucla.in/2G2rWBv , kuddos to you and your

book. If you feel they could stand some help, it is easily done. We have enough disagreements with the arrow-
phobes. #Bookofwhy

1.5.19 @7:20pm - (Replying to @dccozine) Strange, and I thought no one would notice that in the insular bubble
called "econometrics". Evidently, some folks read the details -- there is hope to Eco. #Bookofwhy.
1.5.19 @7:06pm - (Replying to @erikdbwestlund @stevepowell99 @SLMsociology) @SLMsociology and Winship
book is the greatest thing that happened to social science since Duncan and Blalock. The only missing ingredient
(reparable in one page) is to show how counterfactuals emerge organically from SEM, thus unifying the
chapters. #Bookofwhy.

1.5.19 @6:43pm - (Replying to @erikdbwestlund @stevepowell99) From the last lecture I heard (by Imbens, NIPS
2017) it appears that many folks are still using "matching methods informed by Rubin" but, nowadays, someone
who read #Bookofwhy (page 274) is likely to ask: "Is it valid?".

1.5.19 @5:18pm - Your summary amounts to a whole new book, or a valuable R-supplement to Primer
https://ucla.in/2KYYviP . Why not publish it? One comment: ETT should be introduced as another research
question, rather than an estimation technique. See p. 106 of https://ucla.in/2G2rWBv #Bookofwhy.

1.5.19 @1:37pm - (Replying to @jon_y_huang) Thanks for the pointer. But though they explain how to get
Stochastic IDE, I am unable to grasp what research question it answers and whether it can be used to
approximate NIE, whose meaning I do grasp. Can anyone articulate it in Twitter language? BTW, is it a do-
expression?

1.5.19 @1:24pm - Reading #Bookofwhy before Athey's lecture is a good way of getting perspective of the field.
You will be puzzled though by the almost complete overlap of topics and zero overlap of methods, bringing us
back to a cultural enigma called "econometrics" https://ucla.in/2mhxKdO

1.5.19 @2:48pm - (Replying to @CaitLamberton @linzcpage) Yes, there has been some change in causality-land,
and I am glad it came to the attention of folks in your field. I dared call it "causal revolution" in #Bookofwhy, to
highlight the sudden outflow of results. Enjoy.

1.5.19 @1:56am - (Replying to @dccozine) Methodological allegiance, I have found, is the greatest impediment
to scientific progress. Behold how robust and long-lasting it is. You can easily guess an author's PhD advisor
decades after graduation. Scared to offend the boss and be charged with treason. #Bookofwhy

1.4.19 @9:56pm - (Replying to @MartinRavallion) The "evaluation" community was an honest revelation. I knew
of course that every researcher should and is concerned with evaluating the impacts of programs, but I did not
realize that the task of evaluation presents unique problems that justify a separate discipline. #Bookofwhy

1.4.19 @7:37pm - (1/2) Thanks for a positive and informative review. I like it because (1) I learned from it
about a community called "evaluation", new to me, and how it has been dealing with causation. (2) The
reviewer is not scared to say that this community can learn something from #Bookofwhy (a
1.4.19 @7:37pm - (2/2) (a hard thing for a statistician to say, and high treason for an economist). (3) I learned
what I could have done better in #Bookofwhy to make the reading obstacle-free for evaluation researchers. If
you have not read #Bookofwhy yet, I receommed reading this review first.

1.4.19 @11:07am - (Replying to @MarcoIppolito) Your conclusion is right (if I properly understood you question).
Path-0 is not blocked by U and Path-1 is blocked by U. Conclusion: All backdoor paths from B to E are blocked
by U.

1.3.19 @4:44pm - (Replying to @david_colquhoun @EpiEllie @stephensenn) Quite the opposite. DAGs are the
perfect language to express the possible existence of unanticipated confounders, in case you really fear them.
The double arrows <--> were invented to express this belief, why not use it? #Bookofwhy
1.3.19 @1:24am - (1/2) A reading group of both @_MiguelHernan's book and Primer would be of great value,
especially to economists, and I would start with Primer. First, it's much shorter. Second, it is good to have a 3-
dimensional view of the world before examining it through a microscope. Specif..
1.3.19 @1:24am - (2/2) Specifically, seeing how counterfactuals emerge naturally from an economic model
before they are hooked to treatments and RCT's, and how the notational distinction works between
interventions and counterfactuals (do(x) vs. Y_x) before it disappears in one big PO. #Bookofwhy

1.3.19 @1:04am - But the most gratifying news came yesterday, when #Bookofwhy became #1 on Amazon
bestsellers list for BASEBALL:
https://www.amazon.com/gp/bestsellers/books/16315/ref=pd_zg_hrsr_books_2_3_last And economists are
telling us there is deep wisdom in crowd.

1.2.19 @6:11pm - (Replying to @iamnemo8 @amychua @skdh) Thanks for the favorable mention #Bookofwhy

1.1.19 @11:44am - (Replying to @LaurentFranckx) Now that you brought it up, I see that it needs repair. It
should read: "... using naive statistical methods". Thanks. #Bookofwhy

1.1.19 @5:03am - (Replying to @mathtick) I also had difficulties understanding Wang&Blei. They claim
identification, but do not present an estimand. I could not tell if they leverage external information, as in here
https://ucla.in/2N9icIX . A hard to decypher enigma. #Bookofwhy

1.1.19 @3:38am - (Replying to @mathtick) Please refresh my memory on the "blessing of multiple causes"

1.1.19 @2:24am - (1/n) Dear readers and followers. This is a brief end-of-year report that I would like to share
as 2019 is about to enter our space. I am very gratified with the warm and wide acceptance of #Bookofwhy and
its key messages: the inevitability of causal modeling and the logic ...
1.1.19 @2:24am - (2/n) ... the logic imposed by the Ladder of Causation. The need to marry ML and CI has
become a popular banner even in communities that are strangers to CI. The introductions of free courses on
graphical models at @HarvardEpi and Microsoft (DoWhy) [there may be others] promise to.
1.1.19 @2:24am - (3/n) ...promise to end decades of confusion and bewilderment that were perpetuated by
sluggish educational and editorial establishments. Some disagreements do remain, among them the debate over
manipulability https://ucla.in/2EpxcNU and the need for a three-level hierarchy.
1.1.19 @2:24am - (4/n) But these differences are bound to disappear, I believe, as soon as the full power of
structure-based tools become familiar to the community. To that end, I would like to provide free access to
Chapter 4 of Primer (https://ucla.in/2KYYviP ), which develops and exemplifies ...
1.1.19 @2:24am - (5/n) harmonious symbiosis of counterfactuals, structural models and the do-operator --
https://ucla.in/2G2rWBv . Enjoy, and may 2019 make our collective understanding dwarf that of 2018.
#Bookofwhy

12.31.18 @7:58pm - (Replying to @NeatWitTweet @chriskroiss) I recommend this course very strongly,
especially to economists who are already two decades behind the time @HaravardEcon Do not wait for your
professors to recommend it, they won't. Once you acquire these tools, the hardest problems in your Econ
classes become a child play.

12.31.18 @7:30pm - (Replying to @AlvaradoMoutin1 @HarvardEcon) In such hopeless cases it is doubly important
to know what simplifying assumptions are capable of moving you towards a contingent solution. Blindly applying
textbook routines (eg IV) is not wiser than searching under the lamppost. #Bookofwhy

12.31.18 @6:49pm - (Replying to @theute_ @junpenglao @generativist) Ask the Stanford professor if he can
teach any of the seven tools described here https://ucla.in/2umzd65 and, if not, what he would suggest to
creative students who wish to acquire these tools #Bookofwhy
creative students who wish to acquire these tools. #Bookofwhy

12.31.18 @3:52am - (1/2) I am not familiar with the two courses you mentioned, so I am basing my opinion on
the advertised content. Sobel's course is more dogmatic than Rubin's. It is likely to be harmful in its religious
aversion to graphs, to structural models, and to modern way of thinking. (con.
12.31.18 @3:52am - (2/2) Roy's course stands between @HarvardEcon and #EpiTweeter. It makes limited use of
graphs, but no unification with PO. Mostly harmless. #Bookofway

12.30.18 @11:53pm - (Replying to @jaiz30) Glad the #Bookofwhy made it to another smart collection.

12.30.18 @1:07am - I'm not sure what we can infer from your job market list; (1) that #HarvardEcon has chosen
jobs over science, (2) that some applied works do not require science, or (3) that #HarvardEcon students
acquire scientific tools by independent reading? I pray it's (3), not (1) #Bookofwhy

12.29.18 @1:45am - (Replying to @AlvaradoMoutin1 @HarvardEcon) I need more context to answer your
question. Note however that an IV analysis is not regression-based but an exercise in causal analysis and, so, an
economist is allowed to say whatever logically follows from the causal assumptions embedded in the IV
structure. #Bookofwhy

12.29.18 @1:36am - A note of appreciation to David Auerbach @Auerbachkeller for including #Bookofwhy in such
stellar company.

12.29.18 @12:09am - (Replying to @Awesomnomics) I would not go as far as "the first thing," but it should at
least be brought to their attention. The current isolationist attitude of educators and editors is breeding an
obsolete generation of econometric students, bound to perpetuate isolationism. #Bookofwhy @HaravardEcon

12.28.18 @11:47pm - (1/3) I am not searching for traces of one work or another. My question was whether you
believe @HarvardEcon students are prepared for the challenges of modern #causalinfernce. e.g., Can they
determine conditional exogeneity or testable implications in a given economic model? C....
12.28.18 @11:47pm - (2/3) Can they compute counterfactuals in a given systems of economic equations? More
pointedly, do they have the tools to solve simple problems like those listed in Sec. 3.2 of
https://ucla.in/2mhxKdO , which Epi students learn as basic alpha-bet? Or will they stay behind? (con.
12.28.18 @11:47pm - (3/3) Many economists on this Twitter look up to @HarvardEcon for leadership in the
reformation of econometric education into the age of modernity, among them seasoned educators who are
expressing an urgent need for such reformation. It's time for a bold step forward. #Bookofwhy

12.28.18 @4:24am - (1/2) In 2014, the editors of Econometric Theory asked me to comment on the
"experimentalists" vs. "Structuralists" debate in econometrics. I have hoped Sec. 4 of https://ucla.in/2mhxKdO
would convince economists that there is more to causation than IV. Isn't there? @HarvardEcon
12.28.18 @4:24am - (2/2) Speaking of "foundations of of causality". Do economists take IV analysis to be
foundational? ie, sufficient for deriving other tools and results in #causalinference developed in the past two
decades? (eg identification, mediation, generalizability) #Bookofwhy @HarvardEcon

12.26.18 @4:56am - (Replying to @totteh @EpiEllie) I agree, but we still need to explicate what makes the
causal effects of "infeasible interventions" scientifically meaningful, namely, informative to policy makers who,
like Harvard Epi students, care only about the impacts of "feasible interventions". #Bookofwhy #causalinference

12.26.18 @12:10am - It is encouraging that another "father of deep learning" acknowledges the crucial role of
#causalinference in AI and, remarkably, not for managing actions and counterfactuals but for facilitating
adaptation and generalization, as in Tool 5 of https://ucla.in/2umzd65 #Bookofwhy
12.25.18 @8:58pm - (Replying to @autoregress @PHuenermund and 2 others) Perfect question, closer to the
core, once we clarify what is meant by "infer from data". Surely, nothing can be inferred from obs. data alone;
But can we infer those elusive causal effects from data+ASSUMPTIONS,-- same assumptions we use in standard
obs. studies? #Bookofwhy

12.25.18 @6:51pm - (Replying to @autoregress @PHuenermund and 2 others) Perfect question, closer to the
core, once we clarify what is meant by "infer from data". True, nothing can be inferred from obs. data alone;

But can we infer those elusive causal effects from data+ASSUMPTIONS,-- same assumptions we use in standard
obs. studies? #Bookofwhy

12.25.18 @3:16pm - (Replying to @shell_ki @melb4886) Who are the gurus of social epi to dictate what one is
supposed to use or not to use? The whole purpose of #Bookofwhy is to examine the wisdom of those gurus,
unveil its pillars, and show how shaky those pillars are in the light of causal logic. #Bookofwhy #causalinference

12.25.18 @2:31pm - (1/2) The intenventionist debate demonstrates indeed why the Greeks invented logic, and
why it does not surface in SCM, where things are "meaningful" if they are authorized by the logic of "listening".
The debate rages in Rubin's PO, crafted to mimic RCTs, where "listening" is
12.25.18 @2:31pm - (2/2) meaningless, and non-manipulable variables, eg. "blood-count" or "gender" require
special handling (still to be explicated). Is there a coherent logic that properly accounts for such variables as
"effect modifiers" yet refrains from ascribing them "causal effects"? #Bookofwhy

12.25.18 @2:32am - (Replying to @igon_value) Much of the power of informal medicine stemmed from the
#causalinference engine that physicians carried in their heads, and its ability to conclude: "no point fighting
obesity if obesity is harmless", now prohibited by the interventionist school, since "harmless" is "ill-defined"

12.25.18 @2:18am - (Replying to @stevesphd) Not sure if you read #Bookofwhy. If you do, note that we, in
modern #causalinference, do not "refuse" tools, nor remedies, nor methods, whether we understand them or
not. We push towards understanding, yes, and we oppose those who "refuse" or "exclude" without the
understanding.

12.25.18 @12:43am - This sounds like medical practice in the prescientific age. I hope some dissidents on the
Harvard team are busy marking the limits of strict interventionalism. Even the Royal Navy rejoiced the late
(1912) discovery of vitamine C. #Bookofwhy #Epitweeter

12.24.18 @8:23pm - (Replying to @Niels_Bremen) Santa was graciously creative to you. Structural equation
models play key role in #Bookofwhy though always in conjuction with the do-operator, thus giving meaning to
SEM parameters which your friends on SEMnet are laboring to estimate. Enjoy.

12.24.18 @4:33pm - My last verse of Greek poetry : "People of Athens, I can't turn my migraine-headache on and
off, but it sure has well-defined causal effects, see why: https://ucla.in/2EpxcNU " Even sworn interventionists
(@EpiEllie) open their heart to poetry when logic compels. #Bookofwhy

12.24.18 @1:56am - (Replying to @AlfredoMorabia @IntJObesity and 3 others) Truth is in the survival of the
legend. Why else would this story survive in our textbook if it were not for scientists admission that truth can
be surprising, even very surprising, and that the scientific community is not always charitable to the surpriser.
#Bookofwhy
12.23.18 @10:53pm - (Replying to @benmbrew @EpiEllie @kareem_carr) Agree! What #causalinference needs
these days is 1 ounce of commonsense and 9 ounces of poetry.@kareem_carr proves them non-contradictory.
Kudos! #Bookofwhy

12.23.18 @4:30pm - (1/2) (Replying to @edwardsjk) I agree. There are times when addressing a problem in
some language is better than not addressing it at all. But my naive question deserves addressing too: Is the
language of analysis adequate, or outdated? Wouldn't researchers be more likely to deal with measurement
error..
12.23.18 @4:40pm - (2/2) (Replying to @yudapearl @edwardsjk) ... if given cognitively meaningful tools to
manage it? More specifically, does the PO framework offer cognitively meaningful tools for managing such
problems as: measurement error, missing data, generalization, and even confounding??? #Bookofwhy
#causalinference #epitweeter

12.23.18 @6:41am - (Replying to @_eleanorina) By non-manipulable I mean we do not have a "Well-defined


intervention", I , such that every member of HEPO will feel comfortable saying "the effect of blood-pressure on
Y in none other but the effect of I on Y". #Bookofwhy

12.23.18 @5:22am - (1/2) I am writing an addendum to https://ucla.in/2EpxcNU to further clarify the


difference between SCM and HEPO (Harvard Epi school of Potential Outcomes). Can anyone familiar with HEPO
practices describe how they treat a non-manipulable variable (like blood-pressure)
12.23.18 @5:22am - (2/2) (like blood-pressure) that happens to appear (as an observed covariate) in the DAG?
Do they delete it from the model entirely? or leave it in the DAG and mark it as non-manipulable? or play with
its incoming and outgoing arrows? #Bookofwhy #causalinference #Epitweeter ???

12.23.18 @12:32am - My perennial naive question: Why cast a problem in a lesser understood framework? When
we cast "missing data" as a causal problem, we gain tools and insight: https://ucla.in/2LdEjZW . Same with
measurement error: https://ucla.in/2N9icIX . Why go back? #Bookofwhy #causalinference

12.22.18 @11:39pm - Very useful overview of ML latest. Thanks. I would be interested in a pointer to those who
"argued that causality is somewhat of a theoretical distraction," and how they think we can achieve
interpretability or explanations without invoking causation. #Bookofwhy

12.21.18 @5:58am - (Replying to @timssweeney) Any discussion of consciousness is in danger of ending up in


metaphysical deadends. However, if we take the position that consciousness is a shallow blueprint of one's
software, we can embed do(x) in a very shallow blueprint with one actor-observer and one
environment.#Bookofwhy

12.21.18 @2:56am - (Replying to @Luis_de_Miranda @harari_yuval) Thanks for the pointer to Bergson's
"fabulation". Is there a good summary of his ideas? #Bookofwhy

12.21.18 @1:27am - I didn't realize data-scientists would view #Bookofwhy from these angles, but perhaps it is
good that I didn't, there is much to be learned from viewing the world through new windows. #causalinference

12.20.18 @7:17pm - I am thrilled #Bookofwhy is found useful by top experts in decision science and risk
analysis. It reminds me that, bottom line, rational decision is the ultimate goal of data science.
#causalinference

12.19.18 @8:58pm - (Replying to @BreskinEpi @EpiEllie and 2 others) I love it! It tops my current favorites: "I
dont trust graphs so let's assume ignorability" or "I don't trust untestable assumptions so let's use propensity
score" or "its dark where I lost it so let's search under the lamppost" Great! #Bookofwhy #causalinference or
score , or its dark where I lost it, so let s search under the lamppost . Great! #Bookofwhy #causalinference or

12.19.18 @5:55pm - Adding completeness to accuracy, EHPO has been examined recently here
https://ucla.in/2EpxcNU and compared to standard scientific thinking, in which exposures, blood-pressure, and
earthquakes do have causal consequences, separate from their stimulants #Bookofwhy #causalinference

12.19.18 @12:54pm - More accurately, the PO framework is split between Rubin-PO and Epi-Harvard-PO (EHPO).
The latter (correct me @EpiEllie) welcomes graphs but restricts all causes to be "well-defined interventions",
eg, disallowing sex, earthquakes and blood-pressure.#Bookofwhy #causalinference

12.19.18 @2:16am - (Replying to @HenningStrandin @mendel_random) Beautifully put. And I also prefer it to
other metaphors suggested by philosophers. After all, no matter how rigorous one wants to be, any
"understanding" must end up with primitives that are universally meaningful. "Listening" is super-universal.
#Bookofwhy #causalinference

12.19.18 @1:16am - (Replying to @mendel_random) Both ways. I use "causal character" for both "cause" and
"effect". When X listens to Y then both X and Y earn "causal character", X for being the "cause" and Y for being
the "effect". #Bookofwhy

12.18.18 @11:51pm - (Replying to @TJ_Kelleher @Scott_E_Page and 8 others) Its a miracle that you found time
for the #Bookofwhy. Thanks.

12.18.18 @11:32pm - (Replying to @DaveBrady72 @Eule_Geheule and 2 others) The differences are
encapsulated in "Three Bullets" https://ucla.in/2NIsI6B , Easy to remember, and easy to apply. #Bookofwhy
#causalinference

12.18.18 @11:26pm - (Replying to @Eule_Geheule @DaveBrady72 and 2 others) The notations are
complementary. The "frameworks" are miles apart. SCM uses both graphs and counterfactuals, PO is religiously
opposed to the former. It shows. #Bookofwhy #causalinference

12.18.18 @2:16pm - (Replying to @AlfredoMorabia @IntJObesity and 3 others) "Let's use all the tools ..." said the
Greeks. Then, when Hippasus of Metapontum proved that there are irrational numbers, they threw him
overboard; they could not stomach their own logical tools. Can we accept the conclusions of our own
assumptions? #Bookofwhy #causalinference

12.18.18 @2:02pm - (Replying to @deaneckles @hoktay) A paper published 3-4 years ago becomes "ancient"
because, at my age, I forget why for heaven sake I spent so much time on it. Reading the "introduction" reminds
me, and makes me feel less guilty.

12.17.18 @10:44pm - (Replying to @hoktay @deaneckles) This ancient paper https://ucla.in/2N9icIX was really
fun to do. The realization that if you only knew how bad your measurements are you could repair them threw
me off completely. Thanks for the nostalgia. #Bookofwhy

12.17.18 @4:24pm - (Replying to @robertwplatt) We have all seen models in which we agree on the assumptions
and in which we can formally identify P(y|do(x)) and still X is not controlled by any "well-defined intervention"
(say sex or race), thus leading strict-empiricists to claim: I do not know what P(y|do(x)) means.

12.17.18 @2:33pm - (Replying to @robertwplatt) Sure: Drug--->BP--->Cardiac arrest , with confounders all over
the place Again the objection is not to the model or its estimability the objection is to the audacity of
the place. Again, the objection is not to the model or its estimability, the objection is to the audacity of
claiming that the causal effect of BP is "well-defined" even if the model permits identification.

12.17.18 @2:08pm - (Replying to @robertwplatt) This assumption is needed to separate the issue of
measurement from that of control. Hernan's objections to the do-operator hold for blood-pressure too, which is
well-measured but escapes direct control. To focus on one issue at the time, can we speak about the effect of
BP???

12.17.18 @6:20am - True, but in our lucky case we have both: (1) a beautiful formalism and (2) it can be easily
applied to real-world scientific problems. I can't name any real world scientific problem that (1) HAS a solution
and (2) escapes the beautiful formalism. Note the cap HAS. #Bookofwhy

12.17.18 @5:27am - (Replying to @WC_Lab @johnsontoddr4 and 3 others) "Minimal surgery" means making the
least perturbation to the model and still ensuring that X be equal to x. If your model is a DAG, it translates to
removing the arrows entering X and setting X to x. It is illustrated in #Bookofwhy chapter 1 and other chapters
as well.

12.17.18 @4:20am - (Replying to @WC_Lab @johnsontoddr4 and 3 others) The do-operator is a mathematical
operation on a model. Of course it does not "exist" as intervention in the physical world. But it can be given an
interpretation as if Nature performs a minimal surgery. If you are bothered by the poetry, replace "do-operator"
by "Operation-23".

12.16.18 @11:55pm - (Replying to @OptimizingMind ???? unpack ????

12.16.18 @8:37pm - (Replying to @eddericu) They co-exist, because "consistency" is defined differently in these
two frameworks. In PO it is defined relative to the actual intervention that made X=x. In SCM it is defined
relative to an ideal intervention that establishes X=x without side effects. #Bookofwhy #epitweeter

12.16.18 @8:26pm - (1/2) (Replying to @talyarkoni) Let's assume an extreme case that all this gibberish about
BP and concentrations is just pre-scientific jargon that says nothing about the world, only about our perception
of the world. Still, as computer/cognitive scientists, isn't it worthy of formal analysis? #Bookofwhy
12.16.18 @8:30pm - (2/2) (Replying to @yudapearl @talyarkoni) What exactly does scientist 1 say to scientist 2
that enables them to swiftly reach this common illusion they call "understanding"? Should we program robots
with this powerful jargon, just for the sake of effective communication, never mind lowering BP or saving lives.

12.16.18 @7:31pm - (Replying to @deaneckles @LuGram12) Is this connected to Aronow's book? Are you
suggesting that pre-modern methodology (eg Angrist-Pischke) encapsulates wisdom that post-modern methods
(eg Morgan-Winship) ignore? Or that social scientists are better off arguing without graphical models? Like PO
folks? #Bookofwhy

12.16.18 @7:20pm - (Replying to @loudquack) Is #Bookofwhy bashing statisticians? See two blog posts
http://causality.cs.ucla.edu/blog/index.php/2018/06/15/a-statisticians-re-reaction-to-the-book-of-why/ ...Is
anyone bashing statisticians or are they making it up?

12.16.18 @5:22pm - (Replying to @loudquack) Why are you uneasy? If the criticism is justified it should be
welcome, if not, it should be refuted. You seem to side with the former, so why feel uneasy? #Bookofwhy

12.16.18 @5:08pm - (Replying to @deaneckles @LuGram12) In revolutionary times the "standards setups" are
changing daily. Is Angrist&Pischke the "standard" or Morgan&Winship? Is Little&Rubin the "standard" on missing
data, or the modern https://ucla.in/2LdEjZW ? If Aronow book stops at pre-modernity, it will be hard to
recommend.

12.16.18 @4:07pm - (Replying to @johnsontoddr4 @aruncann and 2 others) You got it on the nail. I can only
speculate on the cultural roots of Miguel's objections: Rubin's wedding vow to RCT. #Bookofwhy
#causalinference #Epitweeter

12.16.18 @3:49pm - (Replying to @LuGram12 @deaneckles) I see a section devoted to causal inference, page
256, which is a welcome novelty for a statistics book. But I cant tell to what extent it covers the "causal

revolution" ie, whether it stops at ignorability (1974-1983) or goes beyond. Perhaps @deaneckles can illuminate?
#Bookofwhy

12.16.18 @3:21pm - (Replying to @johnsontoddr4 @aruncann and 2 others) I have tried my best to write a very
clear exposition here: https://ftp.cs.ucla.edu/pub/stat_ser/r483-reprint.pdf I will try to improve if you only
tell us which point needs further clarification.

12.16.18 @2:15am - (Replying to @LuGram12) I believe the whole #Bookofwhy is about "contolling for
unmeasurable confounders". Pseudo-randomized designs is just one way of achieving this "control". #Bookofwhy

12.16.18 @2:05am - (1/3) In the past two days readers had a chance to carefully examine
https://ucla.in/2wpolWr (Appendix) and verify that health scientists communicate in terms of states of
variables, as opposed to manipulations of variables. They talk about agents and substances being ``present''
12.16.18 @2:05am - (2/3) or ``absent'', being at high concentration or low concentration, smaller particles or
larger particles; they talk about variables ``enabling,''``disabling,'' ``promoting,''``leading to'' ``contributing
to,'' etc. Branding these causal relationships "hopelessly ill-defined",
12.16.18 @2:05am - (3/3) "extrascientific" or "White Magic" does not do justice to the bulk of scientific
discourse. What for? Just because PO started with RCT perceived as the mother of all knowledge? #Bookofwhy
#causalinference #EpiTweeter

12.16.18 @12:42am - (Replying to @BD_Zumbo) This is a brave move, which will usher clarity in your field. The
behavioral sciences, from my reading, are still torn between the PO and the SEM perspectives. Unifying the two
would be a healthy relief to many researchers. See "Eight Myths.." https://ucla.in/2QnG9dr #Bookofwhy

12.15.18 @7:42pm - Many PO-educated folks draw comfort from excluding all but "empirical phenomenon." Still,
once they learn to derive a PO from a model, they draw extra comfort from answering same "research questions
on empirical phenomenon" (& same assumptions) using more powerful tools. Try it.

12.15.18 @6:59pm - (Replying to @_creepinatshirt @snavarrol) Thanks for noting. The second link is
https://ucla.in/2L8OCyl "Reflections of Heckman and Pinto", written after they sweated over 8 pages to prove
that we dont really need graphs to derive the front-door formula. Economists will risk everything to defend
their econ. textbooks.

12.15.18 @5:51pm - (Replying to @snavarrol) Unfortunately, Heckman is still reluctant to teach his students
(and many followers) how to derive a counterfactual from a toy model. See https://ucla.in/2mhxKdO . Care to
guess why? See https://ucla.in/2L8OCul . And because econ students are not rebelling (yet!) #Bookofwhy

12.15.18 @4:05pm - Not entirely "empirical phenomena". You rely on ignorability assumptions in observational
studies. Where do they come from? Ans. A model which resides in your mind, and which you suppress when you
studies. Where do they come from? Ans. A model which resides in your mind, and which you suppress when you
only allow it to imitate RCT's, instead of deriving PO's. #Bookofwhy #EpiTweeter

12.15.18 @2:54pm - (1/n) Your suggestion to start a conversation was taken seriously in
https://ucla.in/2EpxcNU (2018),which provides rigorous analysis of each of your 2016 arguments. In particular,
it analyzes the logic of "consistency," the difficulties of defining "well-defined interventions,"
12.15.18 @2:54pm - (2/n) of Rubin (1974) there is no such a beast as a MODEL. All he had were conditional
probabilities of potential outcomes {Y(0),Y(1)}. Subsequently, those who entered #causalinference through a
PO-education suffer from the same deficiency -- no model. #Bookofwhy. (cont....)
12.15.18 @2:54pm - (3/n) Instead, the potential outcome Y(1) is defined in the context of a real-life RCT, not a
model of how a population responds to RCT. Hard-core PO folks, including Rubin's disciples in economics
continue to operate in this model-blind conception. DAG-using Epidemiologists
12.15.18 @2:54pm - (4/n) have advanced towards model-hood, but not all the way. Since Robins and Greenland
works (1986) were rooted in Rubin's PO, many epidemiologists today still view DAGs as tools for serving the RCT
conception of potential outcomes, not as a mathematical object (cont.
12.15.18 @2:54pm - (5/n) that DEFINES potential outcomes. This explains why @_MiguelHernan depicts it as
black magic when I assert that an ideal intervention is defined as a property of one's model. This conceptual
barrier continues to impede communication until ..(YES)... a metamorphosis occurs...
12.15.18 @2:54pm - (6/n) Once a PO-schooled researcher learns to derive a counterfactual from a simple model
he/she is liberated for life, never to return to the darkness of model-void. So, derive ONE today, as in
#Bookofwhy https://ucla.in/2G2rWBv eg, estimate Mary's salary had she not quit school

12.14.18 @10:18pm - (2/n) the utility of defining ideal mathematical constructs, and the practical benefits of
attributing causal qualities to non-manipulable variables, from blood-pressure and temperature to gender and
obesity. #Bookofwhy #EpiTwitter

12.14.18 @10:10pm - (1/n) Your suggestion to start a conversation was taken seriously in
https://ucla.in/2EpxcNU (2018),which provides rigorous analysis of each of your 2016 arguments. In particular,
it analyzes the logic of "consistency," the difficulties of defining "well-defined interventions,"
12.14.18 @10:10pm - (2/n) the utility of defining ideal mathematical constructs and ideal manipulations, the
semantics of multi-version interventions, and the practical benefits of attributing causal qualities to non-
manipulable variables, from blood-pressure and temperature to gender and obesity.
12.14.18 @10:10pm - (3/n) The Appendix further examines ordinary conversations among health scientists,
points out the ubiquity of non-manipulable causes and their communicational benefits. I have received no
objection to any of these arguments and assumed their power and transparency convinced you
12.14.18 @10:10pm - (4/n) and your followers in #causalinference. I still believe they deserve serious
considerations. #Bookofwhy #EpiTwitter

12.14.18 @5:24am - (Replying to @_MiguelHernan @dingding_peng) It seems that both of us are looking forward
to a serious conversation. I suggest we start by agreeing on what a model is, for I need to make sure you do not
really believe that: "Pearl believes that any causal effect we can name must also exist." Do you? #Bookofwhy
#Epitweeter

12.13.18 @11:23pm - (Replying to @dingding_peng) Now that I read this article again, I find it hard to believe
that anyone could adhere to that ill-defined notion of "well-defined intervention". Science seems to be making
progress, even in the 21st century, when students are kept on short cultural leashes #Bookofwhy #EpiTweeter

12.13.18 @4:02am - (Replying to @IntuitMachine @raulincze and 2 others) Honestly, I am not sure any of the DL
researchers grasps today what I was talking about then. Why? They are busy reaping the fruits of success. To
understand the ladder one needs to take the time and solve ONE toy problem from A to B, eg Mary's income had
she not quit college.
12.13.18 @1:20am - Is #Bookofwhy bashing statisticians? Carlos posting on Cross-Validated has been updated
with two of my posts http://causality.cs.ucla.edu/blog/index.php/2018/06/15/a-statisticians-re-reaction-to-
the-book-of-why/ Is anyone bashing statisticians or are they making it up?

12.12.18 @4:06pm - My final Confucius proverb: What do we tell a statistician, an economist or a ML researcher
who insists on doing model-blind learning? Confucius says: "It is only by understanding how models should be
used when we have them that we can learn how to live without them!" #Bookofwhy

12.12.18 @1:49pm - (Replying to @raulincze @dennybritz @Waymo) Beautiful symbiosis between a model of the
world (map) and a local machine learning imitation. #Bookofwhy

12.12.18 @4:50am - Readers ask: "Why lonesome, and why elite force." Loneliness will strike when you discover
that your editor, reviewers and even respected colleagues know very little about CI. Force will emanate from
#Bookofwhy tools that enable you to see their weaknesses and do what they can't.

12.11.18 @11:44pm - (Replying to @prof_goldberg @BrookingsInst) As you might have heard, the #Bookofwhy is
advancing standards for proof of causation that are grounded in observational studies, and could serve as
alternatives to RCT's. These were criticized as endangering public safety. Would you say that they should
serious be considered?

12.11.18 @6:11pm - (Replying to @malmyros) "And he whispered to me try the do-calculus....and the problem
got solved in 3 steps." #Bookofwhy. Imagine what students are missing whose professors deprive them of the do-
operator, not to mention do-calculus. They do not even know when to complain. #EpiTwitter #causalinference

12.11.18 @5:13am - Welcome to the elite force of the army of commonsense and, at the risk of sounding like a
"sage" I will add that commonsense is a very lonesome road. Still, it is also fun, because history is on your side;
my high-school teachers told me so, see end of https://ucla.in/2Qg0Rfs

12.11.18 @2:58am - Thanks, Frank, for posting @DrewLevy illuminating review of #Bookofwhy. Even those who
read the book cover to cover will find it valuable to recapture the spirit of the book and its key messages. I
especially enjoyed "Why do the heathen rage?" on the psychology of the resistance.

12.10.18 @4:52am - (Replying to @sytelus @YosephBarash) I've found it, thanks for asking; nostalgia is a healthy
emotion, makes you feel young and useful. "How to Do with Probabilities What People Say You Can't," was first
presented in 1985 https://ucla.in/2qalldx and then published in 1988. Enjoy. #Bookofwhy

12.10.18 @3:18am - (Replying to @HenningStrandin) Very nice example. It highlights several dimensions of
"action" which I need to organize in my mind before answering. Will be back!. #Bookofwhy

12.10.18 @1:38am - (Replying to @Physical_Prep @_MiguelHernan @ChristophMolnar) Whose article did you
love? No, I have not read yet the book by Molnar but, from his description, it seems that he takes

interpretability to be machine's ability to show that it did what it was told, not why it was necessary to do. The
latter requires a causal model. #Bookofwhy

12.10.18 @1:12am - (Replying to @bensprecher @GaryMarcus and 2 others) Pleading a chronic illiteracy to
acronyms. But if GAN can help us predict the effects of banning cigarettes I would be curious to know the
PRINCIPLE by which it extracts the needed information from data. Once we understand the principle, you can
call me a "proud well-trained GAN"
call me a proud well trained GAN

12.9.18 @7:00pm - (Replying to @tdietterich @GaryMarcus @ylecun) And vice versa: "Symbolic causal models, in
principle and in fact, provide grounding for the semantics of sensory data that current deep learning methods
lack." Conclusion: "current methods" on both sides need be informed by the "current methods" of the other side.
#Bookofwhy

12.9.18 @3:41pm - (1/2) (Replying to @tdietterich @GaryMarcus @ylecun) In my God little acre of the physical
world, we do perceive and act on symbols, not on sense-data. For example, I am tweeting this text after I
imagine its impact on readers, which is an exercise in symbolic representation. Note, I have never tweeted this
mesg before, so...
12.9.18 @3:49pm - (2/3) (Replying to @yudapearl @tdietterich and 2 others) ... it could not possibly be stored
in the arsenal of DL functions. Another way of looking at it, in #causalinfeence we seek a grounding for data,
not for symbols; we want to find a real-world interpretation of sense data, not a summarization of data which
DL gives us.

12.9.18 @3:13pm - The debate about manipulability is summarized in https://ucla.in/2wpolWr and is


orthogonal to my latest tweet, which commends your nips talk and informs #causalinference students that a
more refined taxonomy of causal questions yields great benefits https://ucla.in/2umzd65

12.9.18 @1:01pm - (Replying to @jfeldman_epi @_MiguelHernan) #causalinference and #Bookofwhy is all about
the idea that "theory matters". I am surprised you found nothing substantive there. I for one find only substance
there. And I would not use the word "priors" which connotes prior probabilities; causal theories require more.

12.9.18 @1:46am - (Replying to @pfau @dileeplearning and 4 others) My personal journey into Belief
Propagation is narrated here https://ucla.in/2Qg0Rfs , a chapter we had to discard from #Bookofwhy for space
considerations. But many find it educational, especially the story about Bill Gates.

12.8.18 @11:17pm - (Replying to @GaryMarcus @tdietterich @ylecun) I wish I could join this discussion but I
can't parse Tom's "DL provides grounding for the semantics of internal representations that current symbolic
methods lack." In my simple world, it is symbolic representation that provides semantics, not DL. What am I
missing? My personal journey into Belief Propagation is narrated here https://ucla.in/2Qg0Rfs , a chapter we
had to discard from #Bookofwhy for space considerations. But many find it educational, especially the story
about Bill Gates.

12.8.18 @10:19pm - (Replying to @NeuroStats @analisereal) Thanks Manjari for posting this paper. We should
have mentioned Tukey in #Bookofwhy as a singularity (among statisticians) who stated publically that causation
is NOT a species of correlation and that statistics can learn something from an outsider. Any link to the whole
paper?

12.8.18 @9:51pm - Glad to see @_MiguelHernan carrying the torch of C-word to the lion-den of #NeurIPS2018.
But if I were asked: Do all students of #causalinference agree with this hierarchy, I would say: NO! Lumping all
of CI into one level creates confusion and worse See https://ucla.in/2umzd65

12.8.18 @7:13pm - (Replying to @foil26 @melodem_group) External validity has been mathematized and
algorithmitized. So, there is no longer any need to urge researchers to "pay attention to it"; it is an integral part
of any scientific study. What remains is to educate social scientists in the logic of causal inference. #Bookofwhy

12.8.18 @3:44pm - I am retweeting this post in the hope that people would pause before re-chanting "Design
trumps analysis", or at least explain what it means, and why we cannot create symbiosis between the two.
p y , p , y y
#Bookofwhy #EpiTwitter

12.8.18 @2:37am - (1/2) (Replying to @stevesphd @eliasbareinboim and 2 others) In light of our ecumenical
understanding, can we dispose of the slogan "Design trumps analysis" which was tweeted here as an
uncontested, divine dictum? I first saw this phrase in Rubin's "For objective causal inference.." (Ann. App. Stat.
2008) and, ironically, following...
12.8.18 @2:42am - (2/2) (Replying to @yudapearl @stevesphd and 3 others) 2/2 .. this paper Rubin makes three
glaring blunders for lack of analysis. Those who swear by the Dictum "Design trumps analysis" should take a good
look at where contempt for analysis may take us. [The blunders are discussed in this underground paper
https://ucla.in/2EihVyD ]

12.7.18 @10:33pm - (Replying to @jaimiegradus @EpiEllie and 4 others) Excited to see your DAG workshop
spreading truth all the way to Boston U. Dont forget to carry a do-operaor with you, for protection -- a
formidable weapon. #EpiTwitter.#Bookofwhy.

12.7.18 @4:25pm - I am grateful to readers for making the #Bookofwhy part of their thinking about science,
about artificial intelligence and, almost forgotten, about cause and effect.

12.7.18 @10:20am - (Replying to @stevesphd @eliasbareinboim and 2 others) If you believe in a division of labor
you would appreciate that after you think hard about a real-life problem with unknown DAG there is still some
work to be done, i.e, to store the fruits of your thinking, to interrogate it, to combine it with data, etc. this is
our challenge.

12.7.18 @8:20am - (Replying to @StuartReid1929) I happened to befriend Herbert Simon in the early 1990's.
Believe me, you would not want me after talking to a real intellect like Simon. #Bookofwhy

12.7.18 @3:05am - (Replying to @stevesphd @eliasbareinboim and 2 others) Thanks for correcting me. I did not
have an explanation why Imbens could not solve the 4 toy problems, so I blamed textbooks. The real reason is
linguistic; asserting conditional exogeneity in the language of ignorability does not give us a clue whether it
holds in a toy problem.

12.7.18 @1:45am - (Replying to @statnav @EpiEllie and 3 others) To this one can add my commentary on Dawid's
paper, https://ucla.in/2N92rBM in which I labor hard to convince him that questions about Causes of Effects
must invoke counterfactuals, and that, contrary to Rubin's teachings, this does not make them less scientific.
#Bookofwhy.

12.7.18 @1:12am - (Replying to @eliasbareinboim @stevesphd and 2 others) This conversation with Imbens was
really illuminating. It reveals how flexible IV analysts are to tiny variations on their textbook models. E.g.,One
can tell right away that the idea of turning a bad IV into a valid one never appeared in IV textbook, hence the
silence.#Bookofwhy

12.6.18 @6:39pm - (Replying to @Farial04) "The Curse of Free Will" appeared here https://ucla.in/2N6x36Q ,
then in a chapter discarded from #Bookofwhy : https://ucla.in/2OV2mPp . For introductory material, I must

resist modesty, succumb to honesty, and strongly recommend https://ucla.in/2KYvzau , in fact, very strongly.

12.6.18 @11:15am - (Replying to @statnav @EpiEllie and 3 others) You will go down history as the first health
scientist to openly admit the existence of an "obliged to follow convention", which is suboptimal, yet coerces
authors to follow, so as to pacify reviewers, funders, recruiters and gurus. I call it "cultural roots" to be polite
authors to follow, so as to pacify reviewers, funders, recruiters and gurus. I call it cultural roots to be polite
.....

12.6.18 @5:08am - (Replying to @statnav @EpiEllie and 3 others) I fully agree, with one exception. I find clarity,
not cloud, in notational distinction. In a commentary on Dawid's legal papers https://ucla.in/2L9JHNu I show
how attention to notation protects us from wrongly articulating a target quantity. See also
https://ucla.in/2BTa9Lc

12.6.18 @4:16am - (1/2) (Replying to @mendel_random @stevesphd @jwbelmon) You keep demanding "real-
world" conclusions, and I argue for division of labor: You discover a new drug and I will try to analyse your
method and the reasons for its success, and express it analytically, so that others could use it in a totally
different domain. Is it a deal?
12.6.18 @4:26am - (2/2) (Replying to @yudapearl @mendel_random and 2 others) Along the same vein, note
that THREE decades after the giants Cornfield and Hill made their contributions epidemiologists were still
confusing "confounding" with collapsibility. Why? Because the giants did not have an analyst to generalize their
thoughts in mathematical form.

12.6.18 @3:52am - (Replying to @stevesphd @mendel_random @jwbelmon) Good Ans. But IV too relies on the
assumption of ignorability, plus exclusion. Rubin's dictum? My rebuttal still stands. Supertransparent? What
other causal models have you analyzed? Most importantly, we need analysis to handle 12 interacting IV's, with
or without dictums.

12.5.18 @11:42pm - (Replying to @mendel_random @stevesphd @jwbelmon) What part of current fixation with
IV stems from cultural habit and populist mentality, as opposed to informed choice among alternatives? The
latter requires that researchers study causal inference and then choose a method. Will they return to IV? I am
eager to see. #Bookofwhy

12.5.18 @11:12pm - (Replying to @EpiEllie @robertwplatt and 2 others) I am not aware of a crispier separation
between "causal effects we can estimate from data and those we can't" than the do-calculus. Are you? So you
should be the first to ask: How does it work? What have I missed by ignoring it? Ask, and your "useful separation"
will be granted.

12.5.18 @9:52pm - (Replying to @jon_y_huang @robertwplatt and 3 others) Thanks for the papers. I have
(delinquently) skipped this literature because I could not find a simple interpretation of the quantity being
estimated, rNDE; I am still unable to find it in the papers. Perhaps you have one? Indeed its one way of
approximating R-3 by R-2 substit.

12.5.18 @7:05pm - (Replying to @EpiEllie @robertwplatt and 2 others) Sounds crisp, assuming readers
understand what Rubin meant. Your note implies there are counterfactuals that are NOT potential outcomes,
and they are "used", not "avoided". Next step is to welcome notational distinction between them, and you are
ahead with do-calculus - Welcome!

12.5.18 @6:43pm - (Replying to @jon_y_huang @robertwplatt and 3 others) I have not heard of "stochastic
mediation", any pointers? Rung-3 is not there to "avoid" but to welcome and use under full awareness of
assumptions and available tools. "Avoidance" sounds counterproductive.

12.5.18 @5:40am - (1/3) (Replying to @eddericu @EpiEllie and 2 others) Thanks for illuminating the Rung-2/3
discussion with pages from "PRIMER" https://ucla.in/2n3VkuU . Indeed, Rung-2 contains queries that can be
estimated by various "doors" while some Rung-3 queries (e.g., NDE, PN,PNS) may be non-estimable by ANY
graphical means. Moreover,..
12 5 18 @5 44 (2/3) (R l i t @ d l @ dd i d 3 th ) R 2 i GUARANTEED
12.5.18 @5:44am - (2/3) (Replying to @yudapearl @eddericu and 3 others) ...Rung-2 queries are GUARANTEED
to be estimable from experimental studies, not so RUNG-3; some of its queries will forever remain beyond
experimental reach (on populations) and require extra scientific knowledge on individual behavior (eg.
structural equations). We can see..
12.5.18 @5:50am - (3/3) (Replying to @yudapearl @eddericu and 3 others) ..the significance of reducing a query
Q to a do-expression. Once we do this, we are guaranteed (1) Q is experimentally estimable, (2) Q is
observationally estimable if do-calculus says so, (3) If it says so, it tells you how. Teaching from PRIMER makes
everything clear. Thx

12.5.18 @6:24am - (Replying to @stevesphd @mendel_random @jwbelmon) Transparency means being able to
judge whether your formal assumptions are plausible, given your knowledge of biology. Sensitivity analysis
cannot repair misjudgment due to lack of transparency, nor can it test the veracity of the assumptions or their
compatibility with data.

12.4.18 @1:49pm - (Replying to @EpiEllie @eddericu and 2 others) Everyone distinguishes these two beasts in
one's mind, as well as in practice. The question is do we distinguish them in the mathematical notation? e.g., in
the research question that we write down? The current confusion would have been avoided by good notation.
eg, do(x) vs. Y_x

12.4.18 @5:17am - (Replying to @jwbelmon @mendel_random) MR is the application of IV method to problems


in which genes act as instruments. As such, MR is part of CI, and it rests on untestable causal assumptions that
can be seen in the DAG by those who understand graphs and are taken at faith value by those who do not.
#Bookofwhy

12.4.18 @4:02am - A recent emotional attack on #Bookofwhy and the idea that observational studies can
sometimes be more trustworthy than RCT (e.g., when selection bias is unavoidable) behooves me to re-tweet
my answer (below) to the question: "Is RCT still our gold-standard?"

12.4.18 @3:46am - (Replying to @JDHaltigan @mendel_random) I believe the jury is still out, for we need to
factor in " the additional virtue of observing people's behavior in their own natural habitat instead of a
laboratory" as stressed in #Bookofwhy and ignored by its attackers. See a serious debate here:
https://ucla.in/2JfdQcv

12.4.18 @1:21am - (1/2) (Replying to @EpiEllie) The distinction between Rung-2 and Rung-3 is not between "the
observed and the counterfactuals," but between hypothetical interventions and hypothetical counterfactuals.
@_MiguelHernan paper https://arxiv.org/pdf/1804.10846.pdf --- for example, classifies tasks into the following
tricotomy:
12.4.18 @1:32am - (2/2) (Replying to @yudapearl @EpiEllie @_MiguelHernan) 1) description, 2) prediction and
3) counterfactual prediction lumping interventions together with counterfactuals, both expressed in subscripts
(not distinctly using Y_x and do(x)). I am glad you rose to the defense of the distinction; the next step: show it
in notation.

12.3.18 @2:08am - (1/3) Readers ask: Why is intervention (Rung-2) different from counterfactual (Rung-3)?
Doesn't intervening negate some aspects of the observed world? Ans. Interventions change but do not contradict
the observed world, because the world before and after the intervention entails ...
12.3.18 @2:09am - (2/3) ... time-distinct variables. In contrast, "Had I been dead" contradicts known facts. For
a recent discussion, see Remark: Both Harvard's #causalinference group and Rubin's potential outcome
framework do not distinguish Rung-2 from Rung-3. (cont.)
12.3.18 @2:09am - (3/3) This, I believe, is a culturally rooted resistance that will be rectified in the future. It
stems from the origin of both frameworks in the "as if randomized" metaphor, as opposed to the physical
"listening" metaphor of #Bookofwhy
12.2.18 @7:34pm - (Replying to @junzhez @faoliehoek and 3 others) Interesting. Do you know a paper where
the two representations are shown side by side on a toy example?

12.2.18 @4:45pm - (Replying to @eliasbareinboim @alexdamour and 2 others) It reminds me of Larry


Wasserman's comment: "It is my impression that the "graph people" have studied the Rubin approach carefully
while the reverse is not true." (2014, Gelman's blog). It is my impression (JP) that they are producing a 3rd-
generation model-blind PhDs. Non-Stop.

12.2.18 @4:11am - (1/5) For readers wishing to sharpen their understanding of RL and CI, recall how we
answered the question: "How can RL optimize policies without a causal model?". We said: RL is on Level-2 of the
Ladder because it receives data from controlled interventions. #Bookofwhy
12.2.18 @4:11am - (2/5) Indeed, the model created by such control dictates a simple estimand for the query of
interest -- the action-outcome association in the bare data; no adjustment is necessary. (Same as RCT). We also
remarked that this model-blindness is acquired at a cost: evaluated policies ...
12.2.18 @4:11am - (3/5) ... must deploy the same action-set as the one used in training. A causal model is
needed to go beyond, i.e., to actions not used in training. An interesting observation by by Elias notes that a
model would be required even if we want to narrow down the action-set. ...
12.2.18 @4:11am - (4/5) For example, suppose the action-set used in training was {A1,A2} and we wish to
evaluate, without retraining, a new policy over a smaller set {A1}. This may be needed either because physical
limitations prohibit the activation of A2, or because the optimal policy requires ...
12.2.18 @4:11am - (5/5) .. requires that A2 be left uncontrolled. The tools of CI permit us to evaluate policies
with either expanded or narrowed action-sets. They invoke do-calculus to convert the target policy to a data-
compatible expression, as shown here https://ucla.in/2Jc1kdD #Bookofwhy

12.1.18 @1:56am - (1/2) (Replying to @CsabaSzepesvari @eliasbareinboim and 2 others) Testability is a property
of the assumptions, not of how they are articulated. There is even a symbolic logic that unveils it
https://ucla.in/2QpcGzS . However, since ignorability is posited by a human scientist, what counts is the
cognitive ability to recognize it in ...
12.1.18 @1:56am - (2/2) (Replying to @yudapearl @CsabaSzepesvari and 3 others) ... in one's model of reality,
eg, the DAG. Here, Elias' example provides a glimpse into what it takes for an unaided human to do. Don't skip
it. It is important to understand what darkness our PO colleagues are working under when they write (publish)
"assuming ignorability"

12.1.18 @11:21pm - (Replying to @EricTopol @carlzimmer and 8 others) @EricTopol, Wondering: what has
elevated my humble book #Bookofwhy to the Parthenon of the immortals? This is a mortal "what", not a "why"
question.
12.1.18 @11:21pm - (2/3) (Replying to @yudapearl @CsabaSzepesvari and 2 others) things, and how. Folks in CI
have spent their energy on those other setups, and have gotten a formal handle of the when/how question. To
give you a glimpse, suppose the action set is {A1,A2} and we wish to evaluate, without retraining, a new policy
over a smaller set {A1}.
12.1.18 @11:21pm - (3/3) (Replying to @CsabaSzepesvari @eliasbareinboim @mendel_random) [The shrunk set
may be the optimal] a causal model is needed to do the evaluation and we can tell you what kind. The same
goes for augmenting the action set to {A1,A2,A3}, again without retraining. You can see the potential.
#Bookofwhy

12.1.18 @10:18pm - (1/2) (Replying to @CsabaSzepesvari @eliasbareinboim @mendel_random) Agree. One of


the reasons RL literature focused on estimation, not identification, is that it was not needed in engineered
setups, where one had full control over the action-set during learning. RL folks are probably wondering now in
what setups, if any, they need to change.
12.1.18 @10:23pm - (2/2) (Replying to @CsabaSzepesvari @eliasbareinboim @mendel_random) things, and how.
Folks in CI have spent their energy on those other setups, and have gotten a formal handle of the when/how
question To give you a glimpse suppose the action set is {A1 A2} and we wish to evaluate without retraining
question. To give you a glimpse, suppose the action set is {A1,A2} and we wish to evaluate, without retraining,
a new policy over a smaller set {A1}.

12.1.18 @4:16pm - (1/2) (Replying to @CsabaSzepesvari @eliasbareinboim and 2 others) But if the DAG is the
source of our knowledge, why not articulate knowledge using easily verifiables features of the DAG, eg, missing
edges, and get testability for free. Connditional ignorability of treatment is nasty beast for the unaided
intellect.
12.1.18 @4:14pm - (2/2) (Replying to @CsabaSzepesvari @eliasbareinboim and 2 others) Indeed, the test you
suggest is what one would have to go through if one insists on representing knowledge in ignorability language.
It is like solving an equation 3x-4= 2x+1 by trying out all values of x until equality is established. #Bookofwhy

12.1.18 @6:35am - (Replying to @HenningStrandin @mendel_random) Different views. This is the key. And this
is what I am waiting for when people say "there is room for different views". Mention some. @Mendel_random
tried and all he could mention were variants of SCM. What else? religion? Give me a candidate and you will see
openmindedness galo

12.1.18 @6:25am - (Replying to @mendel_random @HenningStrandin) came into being ecause


epidemiologitswere able to conduct their studies clear headed, unperplexed by puzzles like confounding,
mediation, and selection bias. Recall, it took decades after the death of alchemy for modern chemistry to
discover new elements. #Bookofwhy

12.1.18 @6:20am - (Replying to @mendel_random @HenningStrandin) The control of confounding, mediation,


external validity and selection bias are extremely powerful results. I know, I know, you are waiting breathlessly
for new drugs and new medical discoveries. They will come, and when they do, it would be hard to tell how
many of them 1/2

12.1.18 @4:03am - (Replying to @HenningStrandin @mendel_random) On the contrary. Understanding the


revolution and internalizing the power it unleashes would behoove you to join it -- you would not have time to
do anything else. Look at all the powerful results people obtained lately. Cults? You are kidding! What have
they produced ?

12.1.18 @2:52am - (Replying to @HenningStrandin @mendel_random) Constructive commentaries are always


valid. They unveil where one's mindset and priorities are; recognizing and leveraging the potentials of the
causal revolution, or hagglingwhen it started or who deserves less credit for its happening. #Bookofwhy

12.1.18 @1:01am - (1/4) Replying to @eliasbareinboim @alexdamour and 3 others) Elias example demonstrates
vividly that there are objective criteria for transparency and, more generally, there are objective criteria for
preferring some models over others. Potential outcome (PO) enthusiasts have been avoiding such
demonstrations like a plague
12.1.18 @1:05am - (2/4) (Replying to @eliasbareinboim @alexdamour and 3 others) for the past two decades,
and pluralists have accused me of monotheism for presenting them. But readers of #Bookofwhy expect me to
explain honestly why I prefer graphical models over alternatives. So, please examine Elias example seriously
and, in addition to answering
12.1.18 @1:07am - (3/4) (Replying to @eliasbareinboim @alexdamour and 3 others) whether the model has
testable implications, try to determine if the set of assumptions are consistent, whether they are compatible
with any story you know and, if so, whether they are sufficient for describing the story. It is examples of this
sort that prompted
12.1.18 @1:14am - (Replying to @eliasbareinboim @alexdamour and 3 others) the publication of
https://ucla.in/2NIsI6B , which highlights three incurable deficiencies of PO [below] and questions whether PO,
void of graphs is truly "an approach" or merely dismembered component of an approach. #Bookofwhy

11 30 18 @9:28pm - (Replying to @alexdamour @eliasbareinboim and 2 others) I agree with Elias One of the
11.30.18 @9:28pm - (Replying to @alexdamour @eliasbareinboim and 2 others) I agree with Elias. One of the
most common questions I get from students and newcomers is: What are the strengths & limitations of each
notation system & how do we combine them? The answers are not entirely subjective, but the objective
component is lost once people get annoyed.

11.30.18 @8:26pm - (Replying to @alexdamour @eliasbareinboim and 2 others) DAGs and Ignorability statements
are derivatives of SCM (Structural Causal Models) so, saying that they can be used together amounts to saying
that several derivatives of SCM can be used together. No contest! Indeed, for beautiful symbiosis see
https://ucla.in/2umzd65 #Bookofwhy

11.30.18 @8:00pm - (Replying to @mendel_random) I answered 7h ago. Good point! "returning..to its womb"
should be: "returning my fascination with causality to its womb in AI". It will be corrected in the next printing. I
reiterate my comment that understanding the revolution means joining it, not knit picking from the side.

11.30.18 @5:01pm - (Replying to @CsabaSzepesvari @eliasbareinboim @mendel_random) We still need to clear


two issues. 1) translation to causal graph is NOT ONE'S SPORT. It is going to where knowledge resides, which is a
MUST when knowledge is provided by humans. 2) Some say "RL literature" is "everything", it should be replaced
by the specific task analysized.

11.30.18 @3:16pm - (Replying to @alexdamour @eliasbareinboim and 2 others) Reviewers suggest confounders if
(1) the paper is about a specific domain and (2) they know how to translate "ignorability" to DAG condition. But
if the paper is methodological, starting with "assume ignorability" means one thing: "I want to deal with
statistical estimation only"

11.30.18 @1:21pm - (Replying to @mendel_random) Good point! The phrase "returning..to its womb" should
read: "returning my fascination with the causal revolution to its womb in AI". It will be corrected in the next
printing. I reiterate my comment that understanding the revolution means joining it, not picking from the side.

11.30.18 @12:56pm - (Replying to @eliasbareinboim @alexdamour and 2 others) I think the notion of
transparency will become transparent if we take an extreme case of an author stating "I assume identifiability"
instead of ignorability". It is equivalent to wishing the problem to belong to the huge class of dags that permit
identification. But...But..

11.30.18 @5:27am - (Replying to @CsabaSzepesvari @eliasbareinboim @mendel_random) I think by "assumptions


are hidden" Elias (and me) tried to say that, even if you declare 10 times that you "assume conditional
ignorability" the assumptions are still hidden because no mortal can decide if they hold or not, even in toy
problems. I have tested famous mortals too

11.30.18 @2:39am - (1/3) (Replying to @deaneckles @KordingLab @mioana) Notwithstanding 3 inaccuracies,


presenting these methods using DAGs is a great service to neuroscience. Seeing the assumptions so vivid,
readers will never go back to standard presentations, crafted for graph-averse audience -- an endangered
species. @chethan @sweichwald
11.30.18 @2:46am - (2/3) (Replying to @deaneckles @KordingLab @mioana) 2/2 Viewing your RDD model,
students of causality would immediately say: "Hey! this is a front-door model, so why not do it in full generality,
free of assumptions of threshold and linear regression?" Indeed, why not? Detecting similarity among problems is
graph's specialty.
11.30.18 @2:58am - (3/3) (Replying to @deaneckles @KordingLab @mioana) While DAGs do not display
functional properties, like monotonicity, they display assumptions that other representations have difficulty
with, e.g., exclusion, or exogeneity, not to mention conditional exclusion and conditional exogeneity.
Economists are beginning to dig it.
11.30.18 @12:52am - (Replying to @mendel_random @eliasbareinboim) #Bookofwhy designates 1920 as year-
zero for "the causal revolution"; the year when S. Wright dared express a causal assumption in mathematical
form. Researchers who appreciate the magnitude of this step are busy reaping the benefits of the revolution,
not picking on priorities.

11.29.18 @3:54pm - (Replying to @mendel_random) science benefits from a division of labor. Some work on
generating and substantiating "background knowledge", some work on representing and utilizing that knowledge
once it is substantiated and some, luckily a minority, engage in trivializing what others are doing. #Bookofwhy

11.29.18 @9:57am - (Replying to @theophaneweber) Is this the reason why @Bookofwhy suddenly jumped to # 1
on Amazon best seller list in neural networks?
https://www.amazon.com/gp/bestsellers/books/3896/ref=pd_zg_hrsr_books_1_5_last

11.29.18 @9:40am - (Replying to @decodyng) 90% of scientific writing is trying very hard not to call in causal
modeling.#Bookofwhy #causalinference

11.28.18 @10:24pm - (Replying to @Emaasit @akelleh and 7 others) Seminal papers may be extremely time
consuming to read. If you are in compiling mood, of an immediate benefit would be a map of sub-problems and
their associated tools, cast in one unifying vocabulary. #Bookofwhy

11.28.18 @12:39pm - (Replying to @alex_peys @CsabaSzepesvari and 8 others) The phrase "We know how
actions are generated" needs to specify what actions are those; in training or at the target policy evaluated. If

actions are generated differently in these two regimes we have a neat transportability problem on our hand
which needs be solved, not avoided.

11.28.18 @2:00am - (Replying to @fhuszar @eliasbareinboim @CsabaSzepesvari) Counterfactuals are never


terribly difficult to describe. If they were then causal inference would be difficult, but it is easy. I remember
the Botou paper and classified it as an instance of process control problems (Causality pp. 74-76). Again, not
hard to describe or solve .

11.27.18 @11:41pm - (Replying to @CsabaSzepesvari @junzhez and 8 others) The phrase: Imitation learning,
human acts, robot should copy" triggered an old paper https://ucla.in/2oNr6gF (Fig. 14 and 15), which I bring
here only to show how problems like that should be formulated. Nothing is left to guessing; from target
quantity, to information available

11.27.18 @10:21pm - (Replying to @RealUrso) Sure! It is all here http://bayes.cs.ucla.edu/WHY/

11.27.18 @9:56pm - (Replying to @fhuszar @eliasbareinboim @CsabaSzepesvari) It is not drawing or not drawing
the model that counts but whether you correctly interpret the assumptions in the model and what you do with
them. I am waiting for Csaba to clarify whether his graph represents the engineer or mother nature. Each leads
to a different conclusion.

11.27.18 @9:42pm - (Replying to @CsabaSzepesvari @AIforHI and 6 others) I think we are heading here towards
some mutual understanding. Can you exemplify the engineering problem in which "we know how actions are
generated"?? Just describe the setup in which "we know" and, then, what we can conclude from this knowledge,
and what the final result is.
11.27.18 @4:57pm - (Replying to @GaryMarcus @ylecun and 6 others) Appropos. I am sharing my interview with
Martin Ford, https://ucla.in/2KBBSRj and, as you can see, I am still hoping for ML folks to liberate themselves
from data-centricity and other mind-suppressing substances #Bookofwhy

11.27.18 @1:44am - Having noticed that my twitter account swelled to 12K followers and 640 tweets, I created
a searchable file of all my tweets https://ucla.in/2Kz0FoY . I find it useful for retrieving old discussions by
topics, and digging out old gems. You may also find it useful. #Bookofwhy

11.27.18 @1:28am - Good News! #BookofWhy has survived the paper shortage and is now back on the shelves.
Moreover, it contains all revisions and corrections detected by October 27. Make sure you get the 6th printing
(#6 on last line of copyright page). May your reading be enlightening and rewarding

11.26.18 @1:17am - (Replying to @AIforHI @dustinvtran and 6 others) A question from a future partner. To most
folks in causality research the words "assuming ignorability" mean stripping a problem from its causal content
and solving a standard statistical problem instead (#Bookofwhy page 283). Must you really assume that?

11.25.18 @4:05pm - (Replying to @CsabaSzepesvari @junzhez and 8 others) This feeling is mutual. Plus, I know
that RL is saturated with creative researchers who can recognize powerful ideas directly from a tool-map,
before building an ant-eating robot. I have unveiled my tool-map is here: https://ucla.in/2umzd65 and I do not
charge rental #Bookofwhy

11.25.18 @2:28pm - (Replying to @junzhez @CsabaSzepesvari and 8 others) HMMM, but if this is the case, where
is the "learning"? It sounds more like "copying". Are the environments different?

11.25.18 @1:37pm - (Replying to @CsabaSzepesvari @eliasbareinboim and 7 others) We are on the same page
indeed, and the next step is to draw a more complete map of problem types and tools available as I have

started to do in the last re-tweets. I hope we are in agreement on my initial draft, mapping problems and tools,
not labels. #Bookofwhy

11.25.18 @12:20am - (1/4) (Replying to @CsabaSzepesvari @eliasbareinboim and 7 others) I used to tell my
students: "How do you find out if your ideas are worth a damn? When your colleagues start saying they used
them all along, they just never called them like that". Today I tell students "We made it!" Csaba says: "RL is CI,
we just never called them like that"
11.25.18 @12:35am - (2/4) (Replying to @yudapearl @CsabaSzepesvari and 8 others) Reiterating the relationship
between RL and CI. Rl has a causal component, provided by the actions used during policy learning. In contrast,
CI starts with no actions, just observations, and tries to infer effects of future policies. This task was not
addressed by RL so far.
11.25.18 @12:47am - (3/4) (Replying to @yudapearl @CsabaSzepesvari and 8 others) In recent years, CI research
embraced new tasks. (1) What if we have both actions AND passive observations, (2) What if we have actions
and observations and we wish to predict effects of new policies, not those used in training, (3) many more
combinations. What about RL ?
11.25.18 @1:03am - (4/4) (Replying to @yudapearl @CsabaSzepesvari and 8 others) Evidently RL folks were not
interested in those other tasks (no hard feelings) and have not developed the logic needed for their solution.
We are told they could have done so if they really wanted, which I do not doubt. A prerequisite however is to
learn that new logic.

11.24.18 @1:43pm - (Correcting a crucial typo: "development in CI (causal inference) are way beyond
"theoretical discussion", not MI. Indeed, these developments are shining in proofs and demonstrations and, most
importantly, in the minds of the intellectually curious. #Bookofwhy.
11.24.18 @1:24pm - (Replying to @alexdamour @dustinvtran and 7 others) The causal revolution is happening
already, with or without me. I am glad that ML folks are waking up to ask: "Are we missing the boat?" No, you
are not. Because developments in MI are way beyond "theoretical discussion"; shining and ready for enlightened
researchers to catch up.

11.24.18 @4:05am - (Replying to @alexdamour @dustinvtran and 7 others) I checked Google recruitment ad - no
"causal inference" mentioned. UCLA graduates 1/3 PhD per year in CI. Who will shape future education in this
country if they go work for Google? AI counts on the smart scientists at Google to foresee where the next
revolution is heading

11.24.18 @3:42am - (Replying to @dustinvtran @eliasbareinboim and 6 others) The methodological revolution in
health and social sciences should make any foresighted scientist jump into generating this preliminary evidence,
not wait for it to show up in the NYT. Especially in a fad-driven era where 99% of PhD's hear nothing about CI.
#Bookofwhy

11.24.18 @3:25am - (Replying to @stuz5000) Caution. CI is not an instant of probabilistic graphical models, but
of causal graphical model. The difference is critical. The former cannot reason about interventions. See latter
of causation #Bookofwhy

11.24.18 @3:08am - (Replying to @viettran86 @larosaandrea and 8 others) There are no modeling assumptions
needed for RCT, except the randomness of the randomizing coin.

11.24.18 @3:01am - (Replying to @learnfromerror) Lord paradox is a causal problem. Each statistician proposes
a method of estimating the causal effect of diet on weight gain. To discuss the notion of "valid" we must invoke
causal vocabulary. We cannot do it in the vocabulary of variances and covariances no matter how intricate.

11.24.18 @1:00am - (Replying to @CsabaSzepesvari @eliasbareinboim and 6 others) Are you protesting the
notion that RL has limits? or perhaps that RL (as practiced) is different from CI? or from AI?.. I am willing to
declare CI a tiny subset of RL, if this would entice 5% of RL researchers to adopt 5% of the tools developed in CI,
eg., https://ucla.in/2umzd65

11.23.18 @6:15am - (Replying to @ukmlv) Instrumental variables are treated in #Bookofwhy page 249, in the
context of John Snow's investigation (1854) of the cause of Cholera.

11.23.18 @4:53am - (Replying to @PrannayKhosla @dustinvtran and 7 others) Good link (I hope) to solving toy
problems in causal analysis: http://bayes.cs.ucla.edu/PRIMER . #Bookofwhy. #causalinference. The key: start
with what you know and act lazy, let mathematics do the rest.

11.23.18 @2:32am - (Replying to @c_prohm) No way !!! Both RL and CI seek actions and policies. CI is a bit more
open-minded in that it leverages extra-data knowledge to infer the effects of new policies, while RL (thus far)
has excluded such knowledge, at the cost of being "knob-blinded". #Bookofwhy

11.23.18 @2:21am - (Replying to @dustinvtran @eliasbareinboim and 6 others) "Strong empirical success" will
appear when Google starts advertising for "PhD in ML with in-depth understanding of causal inference."
Education comes first, empirical success, second. How many ML Phd's can solve the toy problems in
https://ucla.in/2neVkuU ?? or even #Bookofwhy ?
11.23.18 @12:44am - The most common example is IV. We train by randomizing the instrument Z, and we wish
to infer the effect of action do(X) that was not accessible in training. For that we need a causal model, assuring
us that Z is not directly affecting Y (the outcome), an untestable assumption.

11.22.18 @11:33pm - The relation between RL (Reinforcement Learning) and causal inference has been a topic
of some debate. It can be resolved, I believe, by understanding the limits of each. RL authors call this limit "off-
policy". I like "off-knobs", imagining a machine with finite number of knobs.

11.22.18 @11:18pm - (Replying to @eliasbareinboim @jasonhartford and 6 others) Is RL an exercise in causal


inference? Of course! Albeit a restricted one. By deploying interventions in training, RL allows us to infer
consequences of those interventions, but ONLY those interventions. A causal model is needed to go BEYOND,
i.e., to actions not used in training

11.22.18 @10:57pm - (Replying to @jasonhartford @eliasbareinboim and 6 others) By invoking interventions,


RCTs provide us causal information that can be used, of course, in a restricted causal inference, i.e., inferring
consequences of those interventions. A causal model is needed to go BEYOND, and infer consequences of
actions that were not randomized.

11.22.18 @6:42pm - (Replying to @jtrecenti @analisereal @shell_ki) AN interesting question. To answer it I


would need more grounding, because different statisticians mean different things by the "likelihood principle".
Can you ground it in the form of a claim about two variables, X and Y.

11.21.18 @5:34pm - (Replying to @dustinvtran @zacharylipton and 5 others) I am not arguing, nor judging, just
trying to appreciate what you do. Thanks for explaining, rather than sending me to "all the papers of XYZ".

Given that you start with a model, do you compute an estimand before processing data? If not, what guidance is
provided by the model?

11.21.18 @4:51pm - (Replying to @zacharylipton @eliasbareinboim and 4 others) I would be very curious to
know how "all the recent papers" escape a theoretical impediment. We need conceptual guidance from
someone who read them (who?) Do they "assume ignorability" like Imbens&Co.? Do they invoke a causal model?
Do they discover equivalence model classes?

11.21.18 @6:53am - (Replying to @amine_ouazad @shell_ki) Keep going? And spoil the fun for readers who enjoy
finding out for themselves how controversial questions get resolved by simple definitions? Sure, ignorability is
causal and it's implied by backdoor (despite 2 pages of hand-waving in Imbens-Rubin book.) Now try PS for more
fun.

11.21.18 @6:39am - (Replying to @gchierico @MeaningLifeTV and 2 others) I also got lost in this conversation.
Compatibilists should focus their arguments on one question: How would I program a robot to have a sensation
of free-will and harness this sensation to exhibit moral behavior. Once the conversation quits the robot, it gets
lost in metaphysics

11.21.18 @5:06am - (Replying to @DavidAOliverJr @shell_ki) Confounding? Try to express it as a property of the
joint distribution. Generations of epidemiologists and philosophers tried (#Bookofwhy), some economists are
still trying, and potential outcomes seduce them a PC criterion called "ignorability" that no mortal can apply.

11.21.18 @4:53am - (Replying to @amine_ouazad @shell_ki) Right! But explaining this to an economist is a
lifetime endeavor. And, if you try, you create an army of outraged enemies; from "How dare you ...?" to "We
knew it all along...!" #Bookofwhy

11.21.18 @4:46am - (Replying to @shell_ki) But saying "they are all causal, because I use them" does not allow
you to trace back the assumptions and distinguish concepts that can be inferred from data from those that rest
on assumptions outside the data. E.g., generations believed "confounding" is testable from data.

11.21.18 @4:30am - (Replying to @IntuitMachine @shell_ki) We need it in order to exclude totally irrelevant
sentences from the label "causal". For example, "Cinderella's hair is blue", which cannot be verified from the
distribution of the observables, yet would hardly be deemed "causal" in a scientific context. #Bookofwhy

11.20.18 @11:54pm - (Replying to @yudapearl @shell_ki) The many "likes" received by this simple definition jolt
me to entertain readers with a few examples: Which of the following is a causal concept? Spurious correlation,
Granger causality, Confounding, Endogeneity, Control for (eg., age), Randomized, Instrument, Propensity score,

11.20.18 @6:04am - (Replying to @shell_ki) Contrary to expectations, the definition of "causal modeling" is
fairly easy to articulate. To me, "causal model" is a set of assumptions about the data generating process, which
cannot be expressed as properties of the joint distribution of observed variables. #Bookofwhy

11.20.18 @3:40am - (Replying to @DToshkov Not necessarily. Prospective "causal effects" can be defined and
identified at the population level, without ever thinking about an individual. E.g., (assuming RCT), #(cured
under treatment-1) vs. #(cured under treatment-2). This is level 2, invoking only do-expressions, no PO.

11.20.18 @2:20am - (Replying to @DToshkov) "Counterfactual" is short for "contrary to facts." If we take this
definition as a guideline, there is nothing counterfactual in RCT. Inferring Joe's behavior from Jane's is not
contrary to fact, because there is no logical contradiction between Joe's and Jane's treatments.

11.20.18 @12:21am - (Replying to @yudapearl @DToshkov) blindness and opacity in the past 2 decades (See the
"three incurable bullets" of https://ucla.in/2NIsI6B ). I am concluding with a question: How do you shake a
framework from stagnation, if not by a bold an honest rhetorical strategy? #Bookofwhay #causalinference

11.20.18 @12:16am - (Replying to @yudapearl @DToshkov) 4. True, "the potential outcomes (PO) framework has
been central to such concerns", and the #Bookofwhy (p.269-) salutes Rubin's achievement in chapter and verse.
But "concerns"do not produce progress, and the centrality of PO did not prevent it from stagnating into
blindness

11.20.18 @12:15am - (Replying to @yudapearl @DToshkov with those concerns? Virgil (29 AD) was also
"concerned" when he said: "Lucky is he who understands the causes of things". Now what? I did not find anything
about omitted variables developed in 1994 that was not known to Duncan in 1970. 4. Now to potential outcomes
(PO) ...`

11.20.18 @12:12am - (Replying to @yudapearl @DToshkov) Duncan, Blalock, Goldberger) who adopted structural
equations. Not very charitable to those who abandoned the causal interpretation of structural equations in the
1980's. Moreover, what good is it "to be concerned" with causal questions if you cant develop a methodology to
deal

11.20.18 @12:07am - (Replying to @yudapearl @DToshkov) 2. Glad we agree that statistics deserves a rhetorical
buff of neglect, though most statisticians would plead "not guilty" since "I have been concerned with causal
questions all my life". 3.#Bookofwhy is full of admiration of early social scientists (Burks, Simon, Haavelmo,
11.20.18 @12:04am - (Replying to @yudapearl @DToshkov) It does not tell us anything about whether joe, who
recovered under drug would have recovered without it. It is unfortunate that this critical distinction between
the intervention and counterfactuals is blurred in most textbooks - a cultural neglect. #Bookofwhy
#causalinference

11.19.18 @11:57pm - (Replying to @DToshkov) Appreciating your recommendation of #Bookofwhy and defending
my rhetorical strategies. 1. Can we have interventional knowledge w/o counterfactual knowledge? Yes, an RCT
provides one. It tells us how likely joe is to recover with and without a drug, but no more.

11.19.18 @4:10am - (Replying to @PHuenermund @PogrebnyakE and 8 others) The glory of this invention and
the pain of the wrong turns are narrated here https://ucla.in/2mhxKdO . The #Bookofwhy briefly decries, and
mostly repairs the latters.

11.18.18 @12:04am - (Replying to @GoAbiAryan @nandanpc Causal utterances do not have parallels in classical
logic. ~X->Y is a logical formula and stands for: If we find ~X we can conclude Y, which is not the same as "if we
ban".

11.18.18 @11:00am - (Replying to @GoAbiAryan @nandanpc) Not really. You cant have causal words such as
"affect" in an associational sentence. And you cant have counterfactual phrases such as "if there were no" in an
interventional sentence. See examples on the Ladder of Causation in #Bookofwhy

11.17.18 @11:43pm - This AAAI-why-19 workshop http://goo.gl/2wR337 promises to be unique. The interface
between machine learning and causal inference will be given an informed, in-depth examination, guided by the
Ladder of Causation and #Bookofwhy

11.16.18 @7:36pm - (Replying to @Lizstuartdc @Megtron9 and 3 others) This new paper on generalizability
identifies three incurable limitations of the potential outcomes approach: https://ucla.in/2NIsI6B . Questions
about study-population vs. target-population would yield immediate answers when cast in graphical language,
e.g., #Bookofwhy (p.354).

11.16.18 @8:25am - (Replying to @statwonk @0xeinar and 2 others) I would love to agree on everything, but I
need to understand what the grievances are: "blindspots" "systematic errors", "land grab""ridicuouls claims"
"absurd" "classic case" "pushing back". What is all this commotion about? Can we first do things right and then
debate sources?

11.16.18 @5:38am - (Replying to @0xeinar @johnmyleswhite @statwonk @Oseinar), I have had the same
question in mind. I tried to enter and calm down the emotional discussion about my land grab crimes, but I
could not quite make out what the grievances are. #Bookofwhy

11.16.18 @5:28am - (eplying to @stephensenn) Your example resembles Wright's guinea-pigs diagrams. Can we
draw one? I am having difficulty understanding what affects what and, more seriously, what the research
question is. (how can "variance" be a "source of variation"). Dempster never believed in causal analysis -it
shows.

11.15.18 @3:11am - (Replying to @stephensenn) What does it mean "overlooked"? Does it mean that some
results in #Bookofwhy are wrong? or that more opportunities are available if we take advantage of hierarchical
error structures? I would love to see how, but I am handicapped in my reliance on toy examples. Can you share
one?
11.15.18 @2:16am - (1/2) I find this discussion rather illuminating, for it faithfully depicts the state of mind of
21st Century statistics.. A guy writes: "regression model is just a causal models," and then complains: "Why is
Pearl bashing statistics?" Many rush to his defence because, they too, .

11.15.18 @2:16am - (2/2) when they do regression, they really think causation. Bashing does not help, because:
"How can a whole glorious field be so wrong for so long? Pearl must be missing something." My hope is that
#Bookofwhy reaches statistics students before it is banned as undesirable.

11.14.18 @1:12am - (Replying to @Megtron9 @eliasbareinboim and 3 others) Another discussion of "Generalizing
experimental findings," highlighting the basic limitations of ignorability-based thinking, is:
https://ucla.in/2L6yTzE . It also distinguishes transportability from selection bias. #Bookofwhy
@causalinference

11.14.18 @12:15am - In view of increasing attention to problems of generalizability, I dare claim that there is
no generalization without encoding differences, i.e. selection diagrams (#Bookofwhy p. 354). For a recently
posted summary and common confusions in the literature: https://ucla.in/2NIsI6B

11.12.18 @12:23pm - Remember Lord's paradox (#Bookofwhy, chpt 6)?? Stephen Senn just posted a "statistical"
solution here: https://errorstatistics.com/2018/11/11/stephen-senn-rothamsted-statistics-meets-lords-
paradox-guest-post/ to which I have added a commentary showing how it ought to be solved in our millennium.
Enjoy the contrast. #causalinference

11.11.18 @1:12am - (Replying to @VoxBec) Glad to have you join the army of commonsense; there are many
challenges ahead! #Bookofwhy #causalinfernce

11.10.18 @6:26pm - (Replying to @kerinalthoff @ProfMattFox) My mini-advice: Assure epidemiologists that they
can now express their research questions in their mother tongue, the language of cause and effect; they no
longer need to deform their questions to fit the molds of outdated mathematics. #Bookofwhy #causalinference

11.10.18 @2:57pm - (Replying to @jwbelmon) Regression analysis is the opiate of the masses: Think dirty, act
clean and believe no one would notice. #Bookofwhy #causalanalysis

11.10.18 @1:12pm - (Replying to @juli_schuess) Thanks for sharing your teaching experience. #Bookofwhy

11.10.18 @10:50am - (Replying to @jwbelmon) Are you suggesting that #bookofwhy is not serious? That it does
not offer useful tools for making medical decisions? That there is such a thing as "statistical causality" which
offers alternative tools? Anxious to hear what you would rather see that is not in #bookofwhy?

11.10.18 @12:25am - (Replying to @Research_Tim) The word "approaches" suggests that we have a choice. But
this is not the case. If we seek explanations we have no choice but to use explanatory models. Fortunately,
today we know how to tailor models to research questions, thus avoiding many decades of confusion
#Bookofwhy

11.8.18 @11:37pm - (Replying to @MelMitchell1 @adibzaman and 2 others) "Schema Networks" are quite
powerful languages. What happens if we run them on some toy problems in #Bookofwhy? Do you think they
would give the correct answer? Put differently, do you think they could be useful in answering some questions
in #causalinference ?
11.8.18 @12:21am - (Replying to @MelMitchell1 @adibzaman and 2 others) I look forward to reading about
causal models in your upcoming book. Can you share basic ideas on how you envision the building of those
models from data, and how they are encoded, once built. Have recent developments in #causalinference been
of some help to you? #Bookofwhy

11.6.18 @10:17pm - (Replying to @MelMitchell1 @adibzaman and 2 others) Granted that causation is a necessary
(not sufficient) ingredient of intelligence, I was surprised to see how few ML researchers were aware of the
Ladder of Causation (#Bookofwhy), the limitation it imposes, and of the tools now available to overcome them.
#causalinference

11.6.18 @1:20am - (Replying to @DrChandraFord @UCLA) I was as excited and delighted as the many students
whom I had the chance to address in this talk. Let me share my last slide here (edited) "The peak of this
revolution is still ahead of us, and you now have the essential tools to be at its epi-center" #Bookofwhy
#causalinference

11.6.18 @1:10am - (Replying to @adibzaman @eliasbareinboim) You touched on a peculiar side of model-blind
ML. They complain of a stomach pain (eg, "Barrier of Meaning"), ignore the medicine and justify it by saying
(still in pain): "No one heard of it! Read the NYT!" @melmarienitch @tdietterich #Bookofwhy #causalinference

11.4.18 @5:27pm - (Replying to @PHuenermund) True, but how would an open-minded economist find out about
these theorems? Publication on NBER is reserved for club-members only, and editors of econometric journals
lack leadership to import knowledge from outside the bubble. Its tough to be an economist these days.
#Bookofwhy.

11.3.18 @11:37pm - (Replying to @fennell_p) Glad you opened #Bookofwhy on page 98, showing Bayes' billiard
tables-- my favorite. I have read hundreds of papers on Bayes Rule and Bayesian statistics, etc etc yet, putting
ego aside, I recommend this chapter over everything I read,, including the learned gurus of Bayesianism.

11.3.18 @11:14pm - (Replying to @JaimieGradus @EpiEllie) When done correctly, meta-analysis pools samples
over areas of agreement and discard samples in areas of disagreement (See #Bookofwhy p. 355). Once
awakened, data-analysts will rejoice the option of doing things correctly, e.g., https://ucla.in/2N7S0K9 ,
https://ucla.in/2Jc1kdD

11.3.18 @5:47am - (Replying to @jtrecenti) I am delighted to re-read this quote and, if it was not mine, I would
have retweeted it over and over. I am especially delighted to see it coming from a Statistician -- they usually
deem this quote either trivial or wrong. Taking it seriously unveils how true it is. #Bookofwhy

11.1.18 @5:40am - (Replying to @smakelainen) Hillarious, thanks, a hymn to counterfactual thinking. Let me
know where I can get the lyrics, if available. If I could sing like him... Hmmm....I would not have written
#Bookofwhy, perhaps another book, perhaps a jingle ... an anthem! Yes! ..if only.....

11.1.18 @5:26am - (Replying to @NeatWitTweet @MadaGadol) Ein Ma Lefached, Chamudim, Bou V'Nikra. Hakol
Mada, Hakol Chadash, Akaveh SheTehanu, Yehuda , Hamechune Judea or Yuda

11.1.18 @5:19am - (Replying to @MereteKonnerup) We seem to be living in two parallel universes, one is model-
based (#Bookofwhy), the other model-blind (@DinaPomeranz). My preference for the former is based on
comparing the scope of questions each can answer, and the transparency of the answers e.g., Replying to
@NeatWitTweetol) https://ucla.in/2odocS7
10.31.18 @1:33am - (Replying to @IMourifie @PHuenermund) Thanks for the paper. It's nice to see a 25-year-old
problem come to a happy solution, round and polished. May all problems end up this way. #Bookofwhy

10.31.18 @1:09am - (Replying to @Chris_Auld @PHuenermund @HolgerSteinmetz) I would add "almost


everywhere" to the statement that overidentification tests are tests of instrument validity. Imagine two invalid
IV's which are identical twins (ie each having the same direct effect on Y). I believe they would pass the
overidentification test, #Bookofwhy

10.30.18 @3:53am - (1/2) (Replying to @PHuenermund) IV's were invented in 1928, by Philip and Sewall Wright,
It is amazing indeed that 90 years later many economists still refuse to accept the non-testability of IV's and,
more generally, refuse to learn what is testable in a model. See https://ucla.in/2mhxKdO #Bookofwhy

10.30.18 @3:59am - (2/2) (Replying to @yudapearl @PHuenermund) One extension is worth mentioning: IV's do
have testable implications when X is discrete, as captured by the Instrumental Inequalities:
https://ucla.in/2pFa9pc #Bookofwhy These are closely related to Bell Inequalities in quantum physics.

10.29.18 @2:53am - For LA residents and LA travelers, I will be giving a book-talk at UCLA on Nov. 5. See
https://ucla.in/2qfruFA . This time, to Law School, Economics and Computer Science, a partnership formed to
deal with ML, social systems, and the causal forces that drive the latter. #Bookofwhy

10.26.18 @10:59pm - (Replying to @ehsan_hoque @rochci) Glad to see #Bookofwhy reviewed by another
independent observer. Hoping to see a summary shared, when available -- it is important to protect simplicity
from the wrath of traditional mystifiers #causalinference

10.26.18 @1:26am - (Replying to @shyamal_chandra) My website will lead you to many slides, videos, lectures,
interviews and other goodies. But it is hard for me to guess what "material" you wish to understand and what
backgrounds your friends and family have. http://www.cs.ucla.edu/~judea/ #Bookofwhy

10.25.18 @1:02am - (1/3) (Replying to @shyamal_chandra) Your question sent me back to serious counterfactual
thinking. Factually, the option of making my course public did not occur to me in 2010, when I stopped teaching
the course; perhaps because public courses were not as popular at the time. #Bookofwhy
10.25.18 @1:10am - (2/3) (Replying to @yudapearl @shyamal_chandra) Fortunately, UCLA now offers two
courses in this direction, one by Prof. Arah, and one by Prof. Hazlett. I hope they would make them public -
causality needs to be learned. Intellectually, however, writing #Bookofwhy was more challenging that teaching
"Causality", which has
10.25.18 @1:23am - (3/3) (Replying to @yudapearl @shyamal_chandra) already penetrated the academic elite
(with 13K citations). It would have required mainly effort of dissemination, not of introspection. Writing
#Bookofwhy on the other hand, forced me to view all this work from a different angle -- historical, philosophical
and computational

10.25.18 @12:55am - (Replying to @XamilyGuy) We surely care about both controlled and uncontrolled under-
representedness. The former indicts Harvard, the latter indicts society. But the word "control" may be misused.
To some it means "condition on" to others "hold constant". Only the latter should enter mediation.#Bookofwhy

10.24.18 @1:32am - Enjoyed my talk #PyData, thanks, and now sharing my opening motto: "Data is our window
to reality, and data-science is the eyeglasses that enable us to see through that window. It is not a mirror, in
which data see themselves interpolated under makeup. #Bookofwhy #causalinference
10.21.18 @5:51am - Evidently #Bookofwhy is currently out of stock. Our publisher assures us this does not
reflect a sudden rush of buyers, nor a breakdown of the printing press but a temporary U.S. shortage of paper.
Delivery date now stands at Nov. 15. Apologies to all knowledge-hungry readers.

10.20.18 @8:57pm - (Replying to @gileshooker @tdietterich) I especially liked your question: "are you trying to
understand how a particular function arrives at a prediction, or are you trying to say something about the
underlying causes of that prediction?" It relates to the slide I shared earlier, motivated by Simpson
paradox.#Bookofwhy

10.20.18 @6:10pm - (Replying to @yudapearl @tdietterich and 4 others) In the wake of our discussion on
Explainability, I am sharing a slide that I used at USC, aimed to show the disparity between explaining why the
system arrived at a given decision vs. explaining the logic of that decision. #Bookofwhy #causalinference

10.20.18 @5:58pm - (Replying to @tdietterich @DKedmey and 3 others) Tom, my criteria is not who is using my
work, but whether the enterprise called "explainable ML" has produced criteria or principles that would allow an
outsider like me to enter the field and use those principles, vs. starting from scratch. I honestly dont know the
answer. Help?

10.20.18 @4:17pm - (Replying to @DKedmey @marypcbuk and 3 others) Terrific pointer! Thanks, David. What
we need now is for DARPA to issue a new BAA for "Science-based Explainable ML" and exclude old PI's from
reviewing new proposals. #Bookofwhy #causalinference

10.20.18 @3:54pm - (Replying to @MariaGlymour) I have no problem with this interpretation of "interpretation".
My shrinkage reflected a long debate with economists and Rubinists who kept on asking: "Under what condition
does a structural coefficient have causal interpretation?" or "When would a duck be a duck?"#Bookofwhy

10.20.18 @3:39pm - (Replying to @tdietterich) My helplessness reflects genuine yearning for a set of criteria by
which we can distinguish "explainable" from "interpretable" from "dumb". I am still not sure if such a set is
available to "explainable ML" insiders, or they are just very good at hiding it from outsiders.

10.20.18 @3:49am - (Replying to @marypcbuk @RandomlyWalking @tdietterich) I do not doubt the quantity of
work that has gone into it, not only recently, but for the past 30 years. I am craving now for a set of principles
that distinguish a useful explanation from a dumb one, like weights and parameters. Are we in possession of
such a set?#Bookofwhy

10.20.18 @12:17am - (Replying to @RandomlyWalking @tdietterich) I am not insisting on a formal definition of


"explainability" but there ought to be some common understanding why a copy of the code would not be
accepted as an "explanation" of how the output was produced from the input. Has anyone articulated why? or
what else is needed?

10.19.18 @11:28pm - (Replying to @tdietterich) I listened to her talk and could not understand a thing,
probably because no one has explained to me what "explainable" or "interpretable" is. I know something about
causal explanations, which made me ask a naive question: Is there explainability w/o a causal model?
#Bookofwhy

10.19.18 @9:53pm - (Replying to @MariaGlymour) I shrink when I read a sentence such as: "estimates from ...
can be interpreted causally". I thought only regressionists ask whether something has causal interpretation, not
scientists. If your causal question is estimable then its estimate must have causal interpretation.
10.19.18 @12:13am - (Replying to @DrewLevy) If statistical mis-thinking rules the bio-marker literature, causal
mis-thinking is even worse, especially when it comes to "surrogate end-points". See https://ucla.in/2N7wOIN ,
section 4.2, and https://ucla.in/2QkLD8O #Bookofwhy #causalinference #fharrell

10.17.18 @5:02am - (Replying to @ThomasVConti) Your idea of #Bookofwhy reading group to economics students
is of enormous importance -- you must catch them before they get contaminated by their textbooks. I would
take the textbook examples and ask questions from the three levels of the Ladder, where textbooks are dead
silent.

10.17.18 @4:48am - (Replying to @zacharylipton) You dont get any tinge of anxiety when you get a tinge of
thrill each time you can do something today that you could not do yesterday. #Bookofwhy #causalinference

10.17.18 @3:56am - (Replying to @CT_Bergstrom) It reminds me of the question: "Should I study arithmetic? No
application in mind" @Bookofwhy @causalinference

10.17.18 @2:17am - (Replying to @TomCrellen) Thanks for catching. I hope we can fix it before the revised
edition comes up, in November. #Bookofwhy #causalinference

10.17.18 @2:10am - (Replying to @oacarah) And thank you Onji for hosting this talk. I thoroughly enjoyed it,
seeing so many students with this spark of curiosity in their eyes. I must retweet what I told them at the talk:
"You are the elite force in the army of commonsense. You're fortunate!" #Bookofwhy #causalinference

10.15.18 @10:43pm - (Replying to @ResearchSmarter @JDKun @MediaMetrics) Thanks for the photos. I had great
fun today at USC, perhaps because the audience was 90% students and 10% professors. I felt like I was talking to
the next generation, free of cultural hangups, open-minded and curious. Like readers of #Bookofwhy.

10.15.18 @10:33pm - (Replying to @quantadan @hadleywickham @bearloga) The first, first thing I would like
data-scientists to learn: Data science deals with the INTERPRETATION of data, not summarization of data nor
re-visualization of data. "Interpretation" is what the data tells us about world outside the data. #Bookofwhy
#causalinference

10.15.18 @1:18am - (Replying to @jasonroy) Great photo! Showing how easy it is for street cleaners to
understand the do-operator, and how hard it is for statisticians. I heard it is still not taught in Stat 101 (not sure
about #causalinference.) Glad it is in #Bookofwhy

10.14.18 @4:48pm - (Replying to @BreskinEpi @KatieMollan and 3 others) The structural theory of
counterfactuals tells us precisely when individual level effects can be estimated, and from what types of data.
For example, probabilities of sufficiency and necessity (#Bookofwhy ch. 8) are individual-level attributes, and
can be bounded tightly 1/2
10.14.18 @5:14pm - (Replying to @yudapearl @BreskinEpi and 4 others) as shown graphically, in
https://ucla.in/2n3VkuU . These bounds might even collapse to point estimates. In short, the structural theory
gives us yes/no answers (not opinions) about what is possible or impossible about ICE. The de-mystification of
#causalinference is not a slogan.

10.13.18 @1:27pm - (Replying to @_eleanorina @EngineerDiet) The acolytes who draw up diagrams are doing so
to explicate and communicate the limitations of scientific claims. More such acolytes are needed. #Bookofwhy

10 12 18 @8:10pm (Replying to @zacharylipton) Agree and the few papers I read from the FAT literature do
10.12.18 @8:10pm - (Replying to @zacharylipton) Agree, and the few papers I read from the FAT literature do
not indicate that the authors know that the various shades of "bias" - statistical, causal, and counterfactuals --
can be given precise characterization, liberated from confusion, and even estimated from data. #Bookofwhy

10.11.18 @5:23pm - (Replying to @USCBiostat @malco_barrett) I am going to be there Monday, your part of
town, I am told there is room for everyone.

10.11.18 @4:04pm - If you happen to be in LA Monday or Tuesday, you are invited to attend two lectures of
mine. The first is at USC https://ucla.in/2OO0kn5 and the second at UCLA https://ucla.in/2QJPtYv . I will be
summarizing #Bookofwhy with slides and poetry. See you there. #causalinference

10.11.18 @3:43am - (Replying to @michael_nielsen) When it comes to individual cognitive development, I am


willing to adopt a more nuanced version of Whorfian theory. But speaking of scientific progress, I am a hard-
liner: no notation - no progress. Causal inference in 20th century is a living proof. #Bookofwhy #causalinference

10.11.18 @2:34am - (Replying to @blattnerma) For a conceptual starter, the #Bookofwhy will be ideal. For a
more mathematical starter, I would recommend the PRIMER http://bayes.cs.ucla.edu/PRIMER/index.html
#causalinference

10.10.18 @9:37pm - (Replying to @andreabellavia @PaoloHead88 @StefanoRenzett1 Not sure Stephen stigler
would agree with my account of the history of statistics, but he will agree that it's about time to write this
history from a causal perspective, and that it looks quite interesting (albeit less admirable) from this angle.
#Bookofwhy #causalinference

10.10.18 @8:30pm - (1/2) The do-operator is described in #Bookofwhy and in this public lecture
https://ucla.in/2oNr6gF For a deeper mathematical treatment, see https://ucla.in/2n3VkuU or Causality
(2009).It is, indeed, the basis of causal reasoning, shunned only by some PO's #causalinference 1/2
10.10.18 @8:40pm - (2/2) ... shunned only by potential-outcome enthusiasts, fearing light. Even Heckman
confessed that you can't do economics without it, so he took "do" and called it "fix", to make eocnomists feel
more comfortable with it. It didn't work; poor "fix" died of loneliness. #Bookofwhy

10.10.18 @1:44am - (Replying to @wlog @EpiDancer @EpiEllie) Even if we differentiate between "observed data"
and "underlying data generating process" there is still a need to distinguish the operator "see" from "do". As to
confusions in economics, see https://ucla.in/2mhxKdO No causation without notation (just coined, but
true)#Bookofwhy

10.9.18 @10:49pm - (Replying to @EpiDancer @EpiEllie) Glad to see consensus emerging. "Control" is a relic of
the days when we could not distinguish between see(X=x) and do(X=x), and the only formal language we had
was probability theory, with one operator: Bayes conditioning ie "see". #Bookofwhy #causalinference

10.9.18 @10:12pm - (Replying to @thoughtfulflyer) Hume's two quotes appear together, almost in same breath,
on page 265 of #Bookofwhy, and it is really amazing that he did not realize that they rest on two totally
different rungs of the Ladder of Causation. Strange, I can't recall a philosopher screaming
"Hey!!"#causalinference

10.9.18 @12:06am - (1/3) The publisher of #Bookofwhy informs us that a revised and corrected 7th printing will
be available early Novemeber. Hold your breath and, in the meantime, all errata are marked here
https://ucla.in/2NRof0D , in red, for a smooth and flaw-less reading. #causalinference.
10.9.18 @12:13am - (2/3) Special beneficiaries of the corrected printing will be champions of RCT, some of
whom were offended by my labeling the front-door method a "serious competitor to RCT" It's been changed to
whom were offended by my labeling the front-door method a serious competitor to RCT . It s been changed to
"useful alternative", which should invite everyone to examine its merits and to start 2/3
10.9.18 @12:23am - (3/3) to start "triangulating" RCT against observational studies. In SCM, "triangulation" is not
a romantic aspiration but a working methodology. #Bookofwhy #causalinference. 3/3

10.8.18 @11:38pm - (Replying to @dynarski @EpiEllie) As important and necessary as descriptive analysis is to
causal research, one must deploy causal models to bring the two together. Description and causes speak two
different languages and causal models (eg, SCM #Bookofwhy) take inputs from both to cook up a
#causalinference soup.

10.8.18 @4:06am - (Replying to @mendel_random @eliasbareinboim and 3 others) Articulate your causal model
and don't fit any schema nor dogma nor popular estimators, just see if your causal question is estimable. If it is,
you've got a "real-life plausible example," since you yourself proclaimed it to be YOUR model. #Bookofwhy
#causalinference

10.6.18 @2:18pm - (Replying to @learnfromerror @LucasJfriesen @stephensenn) Just tell us what relations you
wish to assign responsibility to, what data you have available and "mirror mirror on the wall" will tell you if it is
doable or not, what else you need to complete the assignment, etc, etc. Articulate what you WANT to know,
not what we dont know.

10.6.18 @2:10pm - (Replying to @learnfromerror @stephensenn) Ans. What Pearl says can only be understood by
trying to "represent what might be going on" in any mathematical language, since "what is going on" is beyond
statistics. "How do we use data?" Patterns of arrows do have testable implications, screaming loud when
violated in data.

10.6.18 @1:00pm - (1/3) (Replying to @mendel_random @MariaGlymour) Idea-driven students, grad-schools and
book clubs #causalinference, should view your quest for "actual real life example" as evidence of an opportunity
to transform an underutilized theory into a practical and powerful methodology. They will. Here is why:

https://ucla.in/2IHbBkW
10.6.18 @1:12pm - (2/3) (Replying to @yudapearl @mendel_random @MariaGlymour) The reason I included
http://ucla.in/2IHbBkW (first front-door paper) is that it outlines the requirements as an "isolated" mediation
system X--->M--->Y, shielded from external influences, with only input (X) and output (Y) exposed to
confounded influences. #causalinference 2/3
10.6.18 @1:28pm - (3/3) (Replying to @yudapearl @mendel_random @MariaGlymour) If I were swayed by "give
me one real-life example" voices in 1993, I would still be thinking regression today, perhaps sprinkled with
ignorability justifications, but I would be scared of graphs, just writing books on why they spoil the youth.
#Bookofwhy #causalinference 3/3

10.5.18 @9:15pm - (Replying to @UCSF_Epibiostat @MariaGlymour) My, My, I did not realize you started a book
club at ucsf. I've got a nice example for your front-door magic, co-written by your sister Madelyn. Study
Question 3.4.2 from our Primer book, https://ucla.in/2n3VkuU It should entertain and challenge most students
#causalinference

10.5.18 @3:37pm - (Replying to @swanderingf) A quick answer: We know what causality is, but we do not
exactly know what interpretability is, ie, we do not have an agreed on mathematical definition of the criterion
by which we judge interpretability. e.g. would the answer: "Because you programmed me that way" be
acceptable?

10.5.18 @2:34am - (Replying to @stephensenn According to Don Rubin (see his Fisher's Lecture) part of the
controversy was over notation, and Fisher paid dearly for refusing to use Neyman's potential outcome notation.
It could have spared him the blunder over the mediation fallacy. #Bookofwhy page 441. #causalinference
It could have spared him the blunder over the mediation fallacy. #Bookofwhy page 441. #causalinference

10.4.18 @8:37pm - (Replying to @rogierK @KristinaLerman @JohnHolbein1) A snippet remark on the re-kindled
interest in Simpson's Paradox (SP). Page 211 of #Bookofwhy reminds us what makes SP "interesting", namely,
what phenomenon SP is a symptom of, arguably worthy of being searched for -- one of the three trends is non-
causal. #causalinference

10.3.18 @11:32pm - (Replying to @analisereal @ItsAStatLife @eliasbareinboim) Agree! The discussion were
necessary for dispelling the myth that DAGs are good for single treatment problems but, when it comes to
treatments over time, potential-outcomes should take over. The cost of this myth : a safe ticket to outdated
results #Bookofwhy #causalinference

10.3.18 @3:55am - (Replying to @ItsAStatLife @eliasbareinboim @analisereal) SWIGS are an expansion of DAGs
to those who need to touch Y(a). FINE. But when marketed and taught as a replacement to DAGs, stripped of
do-calculus, they act as handcuffed mutations of DAGs, unable to tell even what they can or can't do with
varying treatments. #causalinference

10.2.18 @8:23pm - (1/3) Well said. Multiple angles are great to highlight different aspects of the problem, when
needed. This is the beauty of SCM: All tools are welcome, and each is given a tag of what it's good for. We use
SWIG, for example, in the 2nd Rule of do-calculus #causalinference 1/n
10.2.18 @9:06pm - (2/3) to authorize exchangeability, and we use another twist of the DAG when another need
arises. It sounds strange therefore that one who feels comfortable with SWig will feel uncomfortable with do-
calculus, unless he/she is not given true exposure to the latter. #causalinference 2/n
10.2.18 @9:30pm - (3/3) Finally, let's ask what one can get from SWIG which is not gotten by Rule 1. It's
important to tell in advance which problems can be submitted to SWIG analysis and which are too hard for
SWIG, but solvable by other means. Can we tell? Yes we can #causalinference @BreskinEpi

10.2.18 @7:08pm - (Replying to @analisereal) Important point which I dared call: "The first principle of causal
inference": Y(a)=Y(M_a) Strantely, students brought up with PO education are rarely aware of it, partly because
Rubin resists it vehemently and partly because it takes away from the glory of SWIGs. #Bookofwhy

10.2.18 @3:29pm - (Replying to @analisereal @ItsAStatLife) It is not only a matter of complexity. The obsession
to " touch Y(a) or else" may lead us to mis-formulate problems and miss solutions. In our case, a missing-data
problem was mis-formulated as a Swig. Reason? The pleasure of touching Y(a) has costs. #Bookofwhy
#causalinference

10.2.18 @5:56am - Sorry, the correct link to my Berkeley talk is here: https://www.youtube.com/watch?
v=zHjdd--W6o4 … I checked, and it is still active. #Bookofwhy

10.2.18 @4:19am - (Replying to @martisamuser) I am glad you linked to my Berkeley talk


http://simons.berkeley.edu/talks/shai-shalev-shwartz-2017-3-28 ... I have enjoyed it a lot. It was given to
theoretical computer-scientists, not my usual audience, but one that understands the roles of theory and
models in processing data. Thanks. #Bookofwhy #Causalinference

10.1.18 @9:06pm - (1/2) For time-varying treatments, I highly recommend a recent Letter in Epidemiology
https://bit.ly/2NfroXl which contrasts the simplicity and transparency of working with DAGS (where knowledge
resides) with the cost of craving to touch a node named Y(a). #causalinference.
10.1.18 @9:23pm - (2/2) This contrast reflects a deeper issue: What comes first? Causal connections in our
world, or dependencies of Y(a) which are derivable from those connections? Is Y(a)||A a more valid justification
for adjustment than path blocking? Habit? Cult? or Science? #causalinference
10.1.18 @3:50pm - (Replying to @_MiguelHernan @analisereal) The description of sequential backdoor in terms
of "action avoiding paths" (Causality p.352) requires inspection of a single graph, and avoids chasing
independncies of Y(a), which are implied by, rather than explain exchangeability. Once tried - always used.
#causalinference

9.30.18 @8:36pm - (1/n) Indeed, what happened to sequential backdoor (SBD)? It is a puzzle not only for
readers of Miguel's book, but all folks doing "time varying treatment". I am sure if they were reminded how easy
SBD is, they would never go back to the torture of chasing Y_g||Ak|Lk conditions. 1/n
9.30.18 @8:42pm - (2/n) The one fault Miguel's book finds in SBD is: "It does not DIRECTLY show the
connection...[to] exchangeabitlity. I deem it a great benefit to skip the laborious chase after this "direct"
connection when we can establish it effortlessly by backdoor. See https://ucla.in/2OYVznm 2/n
9.30.18 @9:03pm - (2/4) Moreover, I doubt whether the counterfactual independence Y_a||A|L contributes to
the understanding of how exchangeability comes about. It was formulated in this convoluted way merely
because Rubin (1978) knew no other language, but probability, for expressing dependencies. 2/4
9.30.18 @9:27pm - (4/4) Today we have more advanced languages to capture causal dependencies, so our
students can be spared the agony that Rubin went through. "Conservation of Agony" isn't a physical law.The
sequential backdoor can show #causalinference students how to turn pain into fun.#Bookofwhy

9.29.18 @9:25pm - (Replying to @_MiguelHernan) Looking at Chapter 19, I dont understand why it spends so
much energy checking exchangeability conditions when we can skip them altogether and solve the problems
directly from the DAG, using sequential backdoor. E.g., compare your Fig 19.5 to #Bookofwhy p. 241.
#causalinference

9.29.18 @5:06pm - (Replying to @KirkD_CO @muscovitebob) It is not clear to me what Braingap considers to be
"Having a clear understanding of the data". Simpson's paradox is a causal, not statistical dilemma,, the
resolution of which requires an understanding of the causes behind the data, not the data. #Bookofwhy
#causalinference

9.29.18 @9:21am - (Replying to @rationalexpec @stevenstrogatz @camelsincaves) I had a quiet dinner with
Clive Granger in Uppsala, Sweden, 1991. He confessed to me that he feels embarrassed by the name "Granger
causality", since it has nothing to do with causality, but he can't stop people from using it; they need some way
to express what is needed.

9.29.18 @8:58am - (Replying to @camelsincaves @stevenstrogatz) Scientists, like children, succumb to two
weaknesses: (1) infatuation with new toys, and (2) addiction to old toys. Today, progress in causal inference is
held back by the latter. #Bookofwhy #causalinference

9.28.18 @11:44pm - (Replying to @PHuenermund) Most economists understand that curve fitting is not causal
inference, so they excuse themselves with: "we assume ignorability" which makes them feel less guilty. In
contrast, Most machine-learning researchers do not see a reason to apologize. #Bookofwhy #econbookclub

9.28.18 @5:44am - (Replying to @juli_schuess @cdsamii and 2 others) All I want to establish at this point is that
the finding: "Sex affects hiring" has clear policy implications (not stylized), regardless of whether "sex" is
determined by employer's suspicion, birth certificate, an entry in a form, or genetic analysis. #Bookofwhy
#causalinference

9.28.18 @5:01am - (Replying to @cdsamii @eliasbareinboim and 2 others) Evaluating job-training and anti-
discrimination programs are costly. The decision of what to fund first is also a policy decision. Corollary: studies
concluding with: "sex has strong direct effect on salary" are not "ill-defined"; they DO have clear policy
implications.#Bookofwhy
p y

9.28.18 @2:41am - Adnan Darwiche article "Human-level intelligence or animal-like abilities?" just came out in
CACM https://dl.acm.org/citation.cfm?id=3281635.3271625&coll=portal&dl=ACM … and should be of interest
to ML readers, as well as to other soldiers of artificial intelligence. #Bookofwhy #causalinference

9.28.18 @12:03am - Replying to @mattmasten @_MiguelHernan) Heckman agrees with Pearl when Pearl praises
econometric models over model-blind statistics. Humorously, things change when Pearl presents economists
with tools to maintain their edge. See https://ucla.in/2mhxKdO , https://ucla.in/2L8OCyl #Bookofwhy
#causalinference #ecobookclub

9.27.18 @11:52pm - (Replying to @eliasbareinboim @cdsamii and 2 others) It seems that what really provokes
manipulationists objections are studies void of policy implications. But studies concluding with: "sex has strong
direct effect on salary" do have policy implications, e.g,. What to fund? Anti-discrimination programs? or job-
training?#Bookofwhy

9.27.18 @7:42am - A friend sent me this 3 years old video https://ucla.in/2R37RMY It gives a fairly faithful
overview of the landscape in which #Bookofwhy was hatched. Enjoy. #causalinference #ecobookclub

9.27.18 @6:20am - (Replying to @zack_weber) I do the same thing, but for a different purpose -- to escape from
debates and to make peace with truth. #Bookofwhy

9.26.18 @7:52pm - (Replying to @yudapearl @cdsamii and 2 others) Reminds me again of DeMoivre (1657-1764)
asking: Would I be permitted to TEMPORARILY write down exp(inx), just for the sake of calculating cos(nx), in
return for my solemn oath never never to say i =sq(-1) #Bookofwhy #causalinference

9.26.18 @7:33pm - (Replying to @cdsamii @eliasbareinboim @deaneckles) We are waiting to hear from
manipulationists whether they would permit us to TEMPORARILY write down do(blood-pressure), just for the

sake of calculating do(drug), in return for our solemn oath never never to say: "causal effect of blood-pressure".
#Bookofwhy #causalinference

9.26.18 @8:22am - (Replying to @_MiguelHernan) Recall provision (2). Causal effect of A is well defined IFF A is
in your model of the world. Otherwise, A could be the size of unicorn - non scientific. By inviting A into your
model you bestow upon it scientific meaning and, in our case, causal effects as well. #Bookofwhy

9.26.18 @7:10am - (Replying to @_MiguelHernan) Delighted to accept the invitation to estimate "the effect of
obesity on death" if you just tell me where in your causal diagram I can find the variable 'obesity'. If it is not
there, perhaps we can do it with "gender", which is surely in your model? #Bookofwhy #causalinference

9.25.18 @5:08pm - (Replying to @yudapearl @_MiguelHernan) Moreover,, this definition has an objective
element to it, modulo our world knowledge; i.e., all people sharing your model of the world will also share your
estimate of the causal effects as defined in the model. Great feature!!! Not hypothetical!! #Bookofwhy
#causalinference

9.25.18 @4:58pm - (Replying to @_MiguelHernan) Recall condition 2): Your model represents what you believe
about the world. This means that properties of your model reflect properties of the world. Corollary (and it
ain't circular): Once defined, definitions are, by definition, well defined. #Bookofwhy #causalinference 1/2

9 25 18 @2:10pm (1/2) (Replying to @ MiguelHernan Not exactly! What Pearl says is: "The causal effect of A
9.25.18 @2:10pm - (1/2) (Replying to @_MiguelHernan Not exactly! What Pearl says is: The causal effect of A
on B" is well defined if (1) A is a variable in your causal model, and (2) your model represents what you believe
about the world. (He sometimes add: why else would you add A to your model? - for show?) #Bookofwhy
9.25.18 @2:19pm - (2/2) (Replying to @yudapearl @_MiguelHernan) For two entertaining yet unambiguous
accounts of what Pearl says see: https://ucla.in/2wpolWr and https://ucla.in/2Qb1h6v espeially Appendix II:
Causes vs. Enablers. #Bookofwhy #causalinference

9.25.18 @1:18am - (Replying to @AllenLaneBooks @TheTLS) Very perceptive reviewer: "circumnavigating leading
figures who have somehow come to hold up progress". May the Gods of counterfactuals protect us from
becoming up-holders of progress. Any trick of accessing the whole article? #Bookofwhy

9.24.18 @3:48am - (Replying to @ProfMillennial1) Very interesting observation. To get this level of accessibility
I will need to write another book. Will try. But, honestly, when I picked up his book I fell asleep on every page,
perhaps because in my funny mind his many principles follow from just one -- structure. #Bookofwhy

9.23.18 @10:56pm - Agree. It was more than 6 years ago that we wrote this page, and every time I read it I feel
like thanking the authors. Evidently, Campbell was so highly revered, that his disciples never got to suspect that
threats can be disarmed #Bookofwhy #causalinverence

9.23.18 @8:15pm - (Replying to @melodem_group) Confessedly, having been trained in different trenches, on
different fronts, I often fail to appreciate difficulties faced by some of my comrades. I would therefore value a
typical example of an assumption you may fail to make if it were not for a helpful "check-list" #Bookofwhy

9.23.18 @4:30pm - (Replying to @melodem_group) Some also ask why #Bookofwhy does not highlight the
influential frame of Campbell etal. My short answer: I really tried hard, but could not find the "frame". Frames
solve problems -- Campbell lists them. My long answer: See https://ucla.in/2CAO2cD #causalinference

9.23.18 @3:40pm - (Replying to @AlloriMD) "How" has two connotations: 1. How nature works, 2. How to achieve
what we want. The former is perfect for mediation. The latter sits on rung-2. "Why" is often used to cover the
former, as in "Why does lemon cure scurvy", which is almost synanymous to "How does lemon cure,,"

9.23.18 @3:01am - (Replying to @raquelrguima @she_knows_a_key) If we have big data, we can estimate PS for
each stratum with high precision, with no modeling assumptions. At this point the c-equivalence Theorem kicks
in and says: Yes, the bias will be exactly the same, for any set of covariates. #Bookofwhy, #causalinference
Nice Theorem !!

9.23.18 @12:51am - (Replying to @she_knows_a_key) My Japanese is still rusty, but I hope you told readers
that, asymptotically, the bias-reducing capacity of PS is identical to ordinary adjustment for the same set of
covariates. https://ucla.in/2CAO2cD I can't understand why people still refer to PS as "design", not "routine".

9.22.18 @10:41pm - A recent paper by Steiner etal https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6117124/


sheds new light on how RCT, IV, RD and other model-free "designs" can be understood and unified from a
graphical model perspective. #Bookofwhy #causalinference . A must for quasi-experimentalists and potetial-
outcomists

9.21.18 @7:03am - Replying to @CateyBunce @WiringTheBrain) When done correctly, "meta analysis" was
renamed "data fusion", because we could not find any meta-analyst who could do it correctly, i.e., neutralize
causal disparities, not seen in the data https://ucla.in/2Jc1kdD I am curious if there are other ways of doing it
correctly.
9.21.18 @3:04am - My UCLA colleague (and Chair), Adnan Darwiche, gave a talk on AI education at Tsinghua
University (the MIT of China) a few days ago. It is only 12 minutes and ML researchers and educators should find
it interesting: https://www.youtube.com/watch?v=bzaMyYqJ030 #Bookofwhy #datascience

9.21.18 @2:51am - (Replying to @ESteyerberg @tmorris_mrc and 2 others) Kudos to Els for speaking truth to
power. I did not realize however that causal reasoning needs defending; in my little village, it is model-blind
learning that people try to defend. #Bookofwhy #causalinference #datascience #causalai

9.19.18 @9:53pm - (Replying to @bzaharatos) Here it is: http://www.sciencemag.org/news/2018/09/meta-


analyses-were-supposed-end-scientific-debates-often-they-only-cause-more . And, as you see, the author
complains about the results but does not address the basic issue: What is meta-analysis supposed to estimate,
and why is it a valid answer to the research question? #Bookofwhy #causalinference

9.19.18 @8:47pm - (1/2) (Replying to @ROrellana4 @EpiEllie and 4 others) Thank you Robert for summarizing
students' reaction to chap. 7 of #Bookofwhy and noting the interest in do-calculus. Epi classes de-emphasize it
because it is not easily translatable to PO notation, to which most Epi teachers are culturally wedded. It is a
forgivable neglect
9.19.18 @9:42pm - (2/2) (Replying to @yudapearl @ROrellana4 and 5 others) as long as you are satisfied with
the low-hanging fruits of backdoor identification. But as soon as you move to questions of transportability, data-
fusion, and even selection-bias, I dont envy your Epi peers when deprived of do-calculus.
#Bookofwhy,#causalinference

9.19.18 @8:19pm - When people ask how meta-analysis differs from data-fusion I say: Meta-analysis averages
apples and oranges to get properties of bananas. Here is why: https://ucla.in/2N7S0K9 ,
https://ucla.in/2Jc1kdD Science Magazine story, Sep 18, slips on the banana. #Bookofwhy #causalinference

9.18.18 @4:19pm - ML researchers looking into counterfactual reasoning should ask: "What caused the fire, the
match or the oxygen?". It is derived here https://ucla.in/2Qb1h6v and just been updated with two new
appendices, on abduction and enablers. Enjoy the oxygen. #causalinference #Bookofwhy

9.18.18 @1:38am - (Replying to @ranilillanjum @EpiEllie) OK, let's call it "ontology". Please identify your
currently most favorite ontology of causation; we will compare it to the "listening" ontology, and see which of
them answers the criteria that you think a good ontology should satisfy. Fair? #Bookofwhy #causalinference

9.17.18 @7:43pm - (1/2) (Replying to @lonnej) Thanks for an illuminating review of my article "Seven Tools.."
https://ucla.in/2umzd65 . I need to clarify why RL falls short of covering the entire level of interventions. The
effect of doubling our price depends on the intent behind recorded price increases in the data.
9.17.18 @7:52pm - (2/2) (Replying to @yudapearl @lonnej) Whether those increases were taken to compensate
for production costs or to spite the competition makes a difference on what a price increase will do today,
when production costs are different and we could not care less about the competition. We eloborate on it in
#Bookofwhy

9.17.18 @3:32pm - (Replying to @EpiEllie @ranilillanjum) When does a philosopher stop asking: "And what is the
meaning of XYZ?". Ans. When you ask him/her: "What is the meaning of meaning?" and you show him/er that XYZ
satisfies all the requirements. "Listening" would satisfy your requirements, please check.#Bookofwhy
#causalinference

9.16.18 @10:48pm - (1/3) (Replying to @EpiEllie) Thank you @EpiEllie for summarizing @ranilillanjum and
@SCMumford book so succintly and creatively. Great thread! I must confess though that from all the allgories,
parables analogies and metaphors that our ancestors have devised to capture causation I find the "listening"
parables, analogies and metaphors that our ancestors have devised to capture causation, I find the listening
9.16.18 @10:54pm - (2/3) (Replying to @yudapearl @EpiEllie @ranilillanjum) "listening" metaphor most useful
for modern #causalreasoning. It goes: "X is a cause of Y if Y listens to X and decides its value in response to
what it hears." http://goo.gl/2xnGjc The formal logic behind it is the same as Lewis logic for "counterfactual
dependence", but
9.16.18 @10:58pm - (3/3) (Replying to @yudapearl @EpiEllie
@ranilillanjum) but its metaphorical power is what drives our decision whether to draw an arrow between X and
Y in a causal diagram. Since the presence/absence of such arrows affects critical steps in the analysis,
"listening" must carefully be listened to. #Bookofwhy #Econbookclub

9.16.18 @1:01am - This paper just came to my attention (and I can't forgive myself for not seeing it earlier)
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4077670/ ... A friendly, comprehensive and unifying (no
hangups) roadmap to epidemiological methodology. Highly recommended. #Bookofwhy #causalinference
@EpiEllie @miquelporta

9.15.18 @7:15pm - (Replying to @ProfMattFox @EpiEllie) For freshmen class, I would use Ellenberg's "why my
handsome dates are jerks?" or the two coins experiment (#Bookofwhy Page 200). But for more advanced class I
would ask: "How come it is so hard for us to come up with an example?" The answer is on page 199, Reichenbach
dictum.

9.15.18 @6:54pm - (Replying to @ClimentQD) Thank you for retweeting this simple truth, so often forgotten by
its beneficiaries, so often ridiculed by the priests of RCT, and so liberating in its clarity when "experimentalism"
hide the sun. Even structural economists (eg. Heckman) tend to forget it. #Ecobookclub #Bookofwhy

9.14.18 @11:05pm - (Replying to @Noahpinion) The fit-to-data practice that you are mocking is long gone from
modern structural modeling, surviving only in econometric textbooks and few outdated circles. Modern
structural models, unlike "natural experiments' tell us precisely how to test compatibility with data.#Bookofwhy

9.14.18 @9:05pm - (Replying to @yudapearl @Noahpinion) When you see a paper that relies on a natural
experiment, your first question should be "Could the assumptions be depicted in a structural model?" If the
answer is yes, check their plausibility. If the answer is no, recommend a major revision. #Bookofwhy
#ecobookclub.

9.14.18 @8:38pm - (Replying to @Noahpinion "Natural experiments" are studies in structural models where some
of the assumptions are justified by semblance to RCT and others are kept implicit, for lack of language, at the
mercy of the "experimentalist". See https://ucla.in/2mhxKdO #Bookofwhy #causalinference #ecobookclub.

9.12.18 @8:17pm - (Replying to @DKedmey) I would start with a simple story, like the firing squad, and ask the
five questions that we asked here: https://ucla.in/2oNr6gF , going from beliefs to actions to counterfactuals.
For desert, I would also ask an explanation question, e.g, does oxygen explain the fire? #Bookofwhy

9.12.18 @7:34pm - (Replying to @wattenberg @GoAbiAryan and 3 others) Readers of #Bookofwhy would believe
that counterfactuals are impossible without "opening the black box". So, either #Bookofwhy is wrong, or the
explanations generated by this system are not the same as those labeled "counterfactuals" in the Book. E.g.,
Does Oxygen explain Fire?

9.12.18 @3:34pm - (Replying to @LauraBBalzer @_MiguelHernan @societyforepi) My favorite source on


#DataScience is https://ucla.in/2umzd65 (no dogs). It makes claims similar to @_MiguelHerman, but it
separates interventions from counterfactuals. It deals with the "dynamic knowledge" issue via transportability
accross time-shifting environments. #Bookofwhy
9.12.18 @3:08pm - (Replying to @FaustoBustos) "Reasoning by cases" underlies both decompositions, and it
would be interesting to investigate the class of concepts that permit such decomposition. Recall also that
humanoids are causal, not probabilistic machines, so the causal version would be more intuitive. #Bookofwhy

9.12.18 @2:47pm - (Replying to @JenMandelbaum @quantadan @MyaLRoberson) Arrows do not address, they
describe. The question before us is how, after unpacking and re-unpacking all the multitude of factors, we
describe in our model that some employers will let hiring decisions be influenced by the candidate race? An
arrow can do that. Any alternative?

9.12.18 @3:46am - (Replying to @GoAbiAryan @wattenberg and 3 others) Anyone can translate this tool into the
language of principles and ideas? What is the input? where does it come from? Data only? or some external
knowledge? if the latter, in what form? etc. Thanks for helping. #Bookofwhy

9.12.18 @1:57am - Readers who asked to see a derivation of "What caused the fire, the match or the oxygen?"
can find it here; https://ucla.in/2Qb1h6v and connect to PO-orthodoxy, Rothman's PIE and other topics of
discussion. Enjoy the oxygen. @EpiEllie @_MiguelHernan #causalinference #Bookofwhy

9.12.18 @12:50am - (Replying to @DBetebenner @EpiEllie @_MiguelHernan) Complex numbers are math
constructs that help us answer questions about real numbers (eg, deMoivre Theorem). They are "ill-defined" to
my 12-year old grandson who needs to finish his math homework. The priests of "manipulation-based" causality
are similarly impatient: act or else

9.12.18 @12:19am - (1/4) Here is a vivid example where we can estimate effects of a manipulable variables
(eg, Smoking) only if we apply the do-operator to a nonmanipulable variable (eg, Tar). Sounds strange? Please
glance at the derivation of the front-door formula (#Bookofwhy, page 236) #EpiEllie
9.12.18 @12:23am - (2/4) Watch how do(t) sneaks in, creates havoc, pushes Smoking under the do-operator,
and leaves the scene unscratched, as if it has never been manipulated. Orthodox PO vice-squads would probably
charge us with crime against "well-definedness" (more),,,
9.12.18 @12:27am - (3/4) Others will forgive this temporary violation of prudent science as a harmless
mathematical gimmick that leads to a well-defined effects. I imagine that infinitessimals were forgiven in the
17th century on similar grounds.
9.12.18 @12:30am - (4/4) Gimmick or no gimmick, never tell your students that do(Tar) is "ill-defined". Complex
numbers are also "ill-defined," but where would science be without them? @EpiEllie @_MiguelHernan
#causalinference #datascience #Bookofwhy

9.11.18 @11:41pm - (Replying to @quantadan @JenMandelbaum) Surely, all our models are simplifications of
our complex world. But there comes a point when we need to stop elaborating and submit a model to analysis.
Would we be permitted at this point to draw an arrow from "race" to "hiring',keeping in mind everything said-
written on it?

9.11.18 @12:33pm - (Replying to @JenMandelbaum) An arrow going from "race" to "hiring" says exactly what you
expressed in words: preferential hiring based on race, and should not be forbidden from entering our model of
society. #Bookofwhy #causalinference @DavidSFink

9.11.18 @3:48am - (Replying to @BerkOzler12 @dmckenzie001) I think it is an encouraging sign. One more push
and Manski will be ready to join modern causal analysis and discover the "credible basis" that he and Goldberger
have sought for so long. #Bookofwhy

9 10 18 @5:52am - (1/3) (Replying to @ProfMattFox @DavidSFink @EpiEllie) More than politics Notation is what
9.10.18 @5:52am (1/3) (Replying to @ProfMattFox @DavidSFink @EpiEllie) More than politics. Notation is what
drives our thoughts and eventually our actions. If you forbid you students from saying: "race causes inequality"
you also forbid them from drawing an arrow from "race" to "hiring" and they would seek remedies elsewhere
e.g, rain-dancing.
9.10.18 @5:56am - (2/3) (Replying to @yudapearl @ProfMattFox and 2 others) To students of physics, semantics
commands awesome respect. The idea that an electron responds to the "field" around it, not to the particles
that created that field has had a revolutionary impact on science (eg Maxwell eqs). Why deprive your students
from such innovations
9.10.18 @6:00am - (3/3) (Replying to @yudapearl @ProfMattFox and 2 others) Finally, there is damage in loss of
opportunities (mentioned in a previous tweet). Forbidding some variables from receiving the do-operator, even
temporarily, might prevent you from identifying effects of manipulable variables (by surrogates). #Bookofwhay
#causalinference

9.10.18 @6:31am - (1/2) (Replying to @DavidSFink @ProfMattFox @EpiEllie) The reluctance of orthodox PO
framework to attribute causal power to non-manipulable variables (eg sex, race) is a harmful cultural hangup,
surviving like the monarchy, simply because Rubin spawned PO from RCT, and others did not shake it off yet.
#Bookofwhy #causalinference
9.10.18 @6:35am - (2/2) (Replying to @yudapearl @DavidSFink and 2 others) I can't see how/why liberating
epidemioligists from this hangup would prevent them from improving health. I tried to elaborate on it here
https://ucla.in/2MQF9MY #causalinference #Bookofwhy

9.8.18 @7:51pm - For our philosophically minded readers (and who isn't?), I am sharing an interview with 3:am
magazine on what/how/if the Causal Revolution contributes to the philosophy of science.
https://www.3ammagazine.com/3am/the-causal-revolutionary/ #Bookofwhy looks a bit different when viewed
from a philosophical lens.

9.8.18 @5:12pm - (1/2) (Replying to @statsepi @EpiEllie) Twitter may be a waste of time in mature fields such
as physics or statistics, but not in new fields such as #causalinference which are in their formative stage, and
must shake off conflicting traditions that labor to mystify things to survive. I find twitter useful for ...
9.8.18 @5:20pm - (2/2) (Replying to @yudapearl @statsepi @EpiEllie) ... useful for calling attention to areas of
demystification which traditional textbooks tend to suppress. Not out of malice, of course, but by lacking the
vocabulary needed for demystification. PS is a good example: https://ucla.in/2CAO2cD #Bookofwhy

9.8.18 @2:34pm - (Replying to @yudapearl @f2harrell and 5 others) I believe the discussion on PS could benefit
from a Section titled "Understanding propensity scores" from Chapter 11 of Causality (2009). Here is a link
https://ucla.in/2CAO2cD which should clarify why PS is smoother, not a bias-reducer. #Bookofwhy
#causalinference

9.8.18 @2:34pm - (Replying to @f2harrell @EpiEllie and 4 others) Moreover, let's not forget the PS is only as
good as the covariates that enter into it. Asymptotically, the bias-reducing capacity of PS is identical to
ordinary adjustment for the same set of covariates. That is, PS is a smoother, not a bias-reducer. #Bookofwhy
#causalinference

9.8.18 @6:23am - Readers asked for a conceptual summary of my book Causality (2009), free of the
mathematics that decorates most pages. I'VE FOUND ONE: https://ucla.in/2oNr6gF -- a lecture given to AI
audience in 1999, and its still fun to read. #Bookofwhy #Causalinference #ecobookclub

9.7.18 @7:45pm - (1/2) (Replying to @mendel_random) What would one call the mega-tons of research going
into "observational studies" from Cochran (1965) till today, if not "continuous erosion of the supremacy of RCT".
Cochran did not have the tools that the Causal Revolution is now providing us, the erosion IS continuing.
9.7.18 @8:00pm - (2/2) (Replying to @yudapearl @mendel_random) Suppose you have a choice between an RCT
on a risky drug, for which you can barely recruit 10 homeless volunteers, and an observational study, full of
y g, y y , y,
back and front door conditions, on 12,000 samples from the target population. Wouldn't you pause? Isnt your
pause "erosion"?

9.6.18 @10:30pm - OJ Arah has noticed that the non-manipulability issue is sipping into influential journals like
the Lancet.... http://dx.doi.org/10.1016/S2468-2667(18)30158-0 As an unbiased observer, I still prefer the
general discussion given here https://ucla.in/2MQF9MY #Bookofwhy #causalinference #ecobookclub

9.6.18 @10:06pm - (Replying to @preskill) It is not just you. Quite a few readers felt uncomfortable with my
songs. I assume they misinterpreted my personal joy of doing science from other songs they must have heard in
the past. For me, the whole thing was a miracle, and fun too. Enjoy #Bookofwhy

9.6.18 @3:07am - A new adage just came to mind: "Not all assumptions are equally indefensible." It came handy
in a comment I wrote on Lars P. Syll's "Causal Interaction and External Validity:
https://larspsyll.wordpress.com/2018/09/04/causal-interaction-and-external-validity/ But I see its usefulness
more broadly. #Bookofwhy #ecobookclub, #causalinf

9.5.18 @2:01am - (Replying to @aztezcan @ProfMattFox) I haven't given up, and twitter makes it easier: A DAG
is a "KNOWLEDGE AMPLIFIER"; in comes qualitative knowledge and out go three new creatures: 1) Vivid photo of
the input, 2) Logical implications of the input.3) Quantitative conclusions of data+input. Its a
Miracle!#Bookofwhy

9.4.18 @9:27pm - Explaining KIKKKO - "Knowledge In, Triple Knowledge Out." 1st K - Your raw input knowledge.
2nd K - The logical implications of 1st K, which are now explicit (e.g., independencies in data). 3rd K -
combined with data, 1st K turns from qualitative to quantitative. QED #Bookofwhy2:01am - (Replying to
@aztezcan @ProfMattFox) I haven't

9.4.18 @3:21pm - GIGO means GIGO! That's what the wise men said when the calculator first came out. They
forgot the other side of the coin: "KIKKKO means KIKKKO". "Knowledge In - Triple Knowledge Out." Always say
KIKKKO, when you hear a wise man say GIGO. #Bookofwhy #causalinference #ecobookclub

9.4.18 @7:16am - (Replying to @MagellanDeStato @Altea_Lorenzo and 2 others) Gee, and I thought it is David
Cox.

9.4.18 @6:53am - (Replying to @SchultenSimon @PHuenermund @eliasbareinboim) The 3 bullets hold in SCM;
their negations hold in PO. And it is not a matter of mere visualization It is a matter of doing things correctly
when you have no idea about "ignorability" (no mortal has). Complementary concepts? Yes! PO derives from
SCM. But you cant skip the source.

9.4.18 @6:33am - (Replying to @SchultenSimon @PHuenermund @eliasbareinboim) There is nothing really wrong
with the graph-less Potential Outcome framework, save for the three bullets at the end of this interview
https://ucla.in/2v72QK5 : *Interpretability, *Identifiability and *Testability. #Bookofway #causalinference
@EpiEllie

9.4.18 @6:13am - (Replying to @totteh @PHuenermund @eliasbareinboim) To detemine effect heterogeneity in


randomized experiments it is necessary to understand #causalinference, see for example
https://ucla.in/2uqDjdp . If you start by assuming ignorability, the problem becomes statistical: "Find a set of
regressors such that..." . #Bookofwhy

9 4 18 @5:47am (Replying to @Altea Lorenzo @XihongLin) Thanks for the slide !!! The RSS must have changed
9.4.18 @5:47am - (Replying to @Altea_Lorenzo @XihongLin) Thanks for the slide !!! The RSS must have changed
since the last time I visited their Proceedings. Recall, in 1833 they vowed to publish only data, data and data,
no opinion, no interpretation, and certainly no causation. A medal due to @XijongLin and @d_spiegel
#Bookofwhy

9.4.18 @4:20am - (Replying to @epi_viborg @evpatora) Joking aside, I think you will find these two
introductions to counterfactuals helpful in your PhD research: https://ucla.in/2wnb80c , and
https://ucla.in/2L8jUFg . If you are into #causalinference, you should also find them easy and intellectually
challenging. Goodluck. #Bookofwhy

9.3.18 @10:10pm - (Replying to @thoughtfulflyer @davidhumeinst @PhilosophyMttrs) True, I do follow the


thinking of David Hume, with one added twist: True, counterfactuals are behind causation, and behind most
scientific thoughts but, to my delight, they can also be algorithmitized. Thus, next generation robots will be
mini-scientists. #Bookofwhy

9.2.18 @11:36pm - (1/3) (Replying to @ngutten) RL can infer causal effects of those interventions only which RL
can control and randomize. But, like old PO, it cannot infer effects of events (X=x) or states which are not
directly manipulable. Whereas enlightened PO folks now accept some "hypothetical interventions", RL can't.
9.3.18 @12:41am - (2/3) (Replying to @yudapearl @ngutten) This comment on the role of reinforcement
learning (RL) in the ladder of causation also brings to light what conservative PO is missing by rejecting the do-
operator and insisting on physical, rather than conceptual interventions. @_MiguelHernan @Bookofwhy
@causalinference
9.3.18 @12:47am - (3/3) (Replying to @yudapearl @ngutten and 2 others) The most obvious loss would be the
use of surrogates experiments, i.e., finding manipulable variables that enable us to estimate the effects of non-
manipulable variables. See for example: https://ucla.in/2wH0Nfw @_MiguelHernan #Bookofwhy
@causalinference @causalityblog

9.2.18 @5:54pm - (Replying to @evpatora) Your summary of what you took from #Bookofwhy is very gratifying to
me, as co-author. But the item that truly warms my heart is #5) "Counterfactuals are derivable." Because it is so

crucial and so often ignored, even by causal analysts. Please post it high on your office door.

9.2.18 @3:43pm - (Replying to @autoregress) Note (Fig 1 #Bookofwhy) that the estimand is ALWAYS a statistical
procedure, yet this does not make every problem "statistical". The fact that we can solve a regression problem
using 2SLS, does not make it an IV problem. Its all in the description, not in the solution #ecobookclub

9.2.18 @5:24am - (Replying to @jonasobleser) Thanks for reminding me how words retain their truth even
months after they were written. It is also relevant to the discussion of "data science" which I defined as the a
'two body" enterprise -- data and reality. Data alone is hardly a science. #Bookofwhy

9.1.18 @11:54pm - (Replying to @ehud) You may be right. Our quote "noticed this sudden shift in Galton's aims
and aspirations: "What was silently missing was Darwin, the chutes, and all the `survival of the fittest.' . . ."

may have been taken from Stigler 2012 The title "The Solution to Darwin" is surely misleading me how words
retain their truth even months after they were written.

9.1.18 @5:40pm - (Replying to @dtweiseth) Not clear what you find objectionable in my treatment of Simpson's
paradox https://ucla.in/2Jfl2VS or here https://ucla.in/2PotANN . Is it in exposing its causal roots? or insisting
on answering: "What do we do if we find it? (as in https://ucla.in/2C8MVkg ) #Bookofwhy Puzzled ????
9.1.18 @5:32pm - (Replying to @jeremyfreese) I wish I could click five times on your "like" button. I normally
tell publishers that I am "password illiterate". But they are getting bolder and bolder in converting our time and
goodwill to profit and laziness. A "reviewers uprising" is in order.

9.1.18 @4:10pm - (Replying to @IntJObesity @_MiguelHernan) The Greeks invented formal logic because they
got tired of debating endlessly, a habit that democracy stirred up . Honoring my Greek roots I propose we
appeal to logic, rather than debated traditions in thinking obesity. @miquelporta @AlfredoMorabia
@causalityblog #Bookofwhy

9.1.18 @2:06pm - (Replying to @mendel_random) The limitations of "statistical techniques" lie not in the
"baggage of their history" but, rather, in their impoverished vocabulary. Sewall Wright extended that vocabulary
by adding a new symbol: "an arrow", without which we can't define MR. #Bookofwhy #econbookclub

9.1.18 @1:57pm - (Replying to @mendel_random) Do you have the reference to the initial exposition of MR? I
am curious to see how they justified the technique without causal logic, and without Sewall Wright.

9.1.18 @1:52pm - When a reviewer (of my paper in 2000) insisted that "instrumental variables" (IV) is a well-
defined "statistical technique" I challenged him/her to define IV without invoking causal vocabulary. He
couldn't! My paper was accepted. #Bookofwhy

9.1.18 @1:33pm - (Replying to @KordingLab @MariaGlymour and 2 others) What in the pre-causal days were
called "causal insights" are today "causal assumptions", expressed in some formal model. If you know of a "causal
insight" that has not yet been captured by SCM, let us know, and the next PhD project in my lab will be to
capture it. #Bookofwhy

9.1.18 @1:18pm - (Replying to @ehud) I was under the impression that Stigler agrees with our interpretation
that regression to the mean was not a solution but an abandonment of Darwin's question. Perhaps not as
explicitly as we tried to make it, but he hinted that it was a cop out rather than a solution. #Bookofwhy

9.1.18 @12:57am - (Replying to @IntJObesity @_MiguelHernan) Glad to see @IntJObesity interest in our recent
#Obesity debate. I wish to note that the link to my revised article is https://ucla.in/2wpolWr and that I believe
the disagreement to be temporary, emanating primarily from Rubin's cultural roots in RCT. #Bookofwhy
#causalinference

9.1.18 @12:15am - (Replying to @PHuenermund) Your discussion with Schaper brings to mind another useful
motto: How can you tell if someone never used DAGs? Ans. He/she repeats slogans of potential-outcome
loyalists; most typical: "The assumptions you find behind DAGs are stronger compared to those WE USE".
#Bookofwhy

8.31.18 @8:40pm - (Replying to @MariaGlymour @KordingLab and 2 others) My tweet got lost: Cartwright's
dictum "no causes in, no causes out" tells us that any approach that appears to be "better fit in "not really
understandable systems" must make causal assumptions about what we presumably do not understand and is, if
valid, a special case of SCM

8.31.18 @8:27pm - (1/3) (Replying to @KordingLab @alex__morley @dan_marinazzo) Cartwright's dictum "no
causes in, no causes out" tells us that any approach that appears to be "better fit in "not really understandable
systems" must make causal assumptions about what we presumably do not understand and is, if valid, a special
cases of SCM#Bookofwhy 8.31.18 @8:29pm - (2/3) (Replying to @KordingLab @alex__morley @dan_marinazzo)
The argument of "not really undersntandable systems" was used by "IV-experimentalists" to justify their model-
blind approach. They stopped using this orgument (to the best of my knowledge), when someone showed them
blind approach. They stopped using this orgument (to the best of my knowledge), when someone showed them
how easy it is in SCM to confess ignorance,. #Bookofwhy
8.31.18 @8:32pm - (3/3) One of my favorite motto in that discussion was: "It is only by taking models seriously
that we learn when they are not needed." It never fails.

8.31.18 @11:00am - Judea Pearl Retweeted @KordingLab Aug 31 Replying to @alex__morley @dan_marinazzo
The other problem is that @yudapearl #bookofwhy type causality approaches mostly help when there is hope for
you to get at the network that represents reality. Other approaches such as RDD may be a better fit in "not
really understandable" systems.

8.31.18 @7:07pm - (Replying to @eddericu @berglund_anita) Anyone who wishes to obtain the solution manual
for the beautiful examples in https://ucla.in/2n3VkuU should write to kaoru@cs.ucla.edu and indicate that it
will be used for self-study and will not undermine instructors who assign these questions to grade students.
#Bookofwhy

8.31.18 @1:54am - (Replying to @goodfellow_ian) The theoretical impediments listed in #Bookofwhy apply to
any model-free approach to learning, not necessarily DL. If we define DL as a function-fitter then it cannot rise
above rung 1 -- association. To be a causal reasoner DL must be guided by a model that tells it what to fit.

8.31.18 @1:37am - The Science magazine podcast containing our #Bookofwhy interview came out today! You
can find it at http://www.sciencemag.org/podcast/science-and-nature-get-their-social-science-studies-
replicated-or-not-mechanisms-behind . The book review segment starts at 18:48, and lasts a little over 5
minutes.

8.31.18 @12:53am - I was happy to find out today that my twitter has swelled to 10K followers.
@berglund_anita deserves special recognition and a signed copy of #Bookofwhy for crossing the 10,000 mark. I
will add an honorary plaque and a solution manual for all homeworks in https://ucla.in/2n3VkuU

8.30.18 @12:41am - Many readers of #Bookofwhy were kind enough to send us typos and errors for
consideration. We have now compiled them into one errata file https://ucla.in/2NRof0D For smooth reading,
please mark your copy, and alert others to do the same. #causalinference #Epibookclub #ecobookclub

8.30.18 @2:21am - (Replying to @goodfellow_ian) The problem with today's DL is not its depth but its model-
blindness. To appreciate, take your favorite DL program and run it for two weeks on data (X,Z,Y) generated by
the smoking-tar story of Fig.7.1 #Bookofwhy. Would it ever conclude that smoking does not cure cancer?

8.30.18 @2:09am - (1/3) Papers often cited as combining DL and causality fall into three main categories. (1)
Ignoring the impedements mentioned in https://ucla.in/2umzd65 . (2) Circumventing the impdiments by
invoking approximate causal models, ... #Bookofwhy
8.30.18 @2:09am - (2/3) elicited either exogenously or from data (Pillar 7), and (3) Stripping problems from
their causal content by assuming "ignorability," then using ML methods to estimate the left-over regressions.
Only (2) deserve the title "combining", and we should watch them carefully.
8.30.18 @2:09am - (3/3) Note however that is we do not learn how to infer actions and counterfactuals from a
given model, we are not likely to do so from data-extracted models. And this is the best commercial I can
compose for #Bookofwhy

8.29.18 @5:13am - (1/2) If two different changes in the environment give rise to same change in the data, each
calling for a different action (or a different way of handling the change) then it is hard to see how DL can look
at the data only and act correctly. #Bookofwhy #causalinference
8.29.18 @5:17am - (2/2) Examples of how different environmental changes result in the same distribution are
here: https://ucla.in/2N7S0K9 (Fig 3 a,b,c), which is also the gentlest introduction to transportability. Again,
the meat is in the examples, not in ideologies. #Bookofwhy #causalinference

8.28.18 @4:07am - (1/3) I am elated and encouraged to see a record number of 450 'likes" on my
recommendation to regression-minded data-scientists https://ucla.in/2L8jUFg . It tells me that regression
analysis is still the working horse of data science, that regression analysts are not too happy ...
8.28.18 @4:10am - (2/3) ...with the literature with which they were broght up, and that they are eager to enter
the age of causation with renewed vigor. One word of advice: do not under-estimate the power of toy
examples. If you really want to dig the new spirit of #causalinference and #Bookofwhy,
8.28.18 @4:17am - (3/3) solve the problem of poor Joe (Sections 4.4 & 4.5), who wonders what his salary would
have been like had he had one more year of education. It is a simple exercise in pure counterfactual analysis,
free of philosophical and cultural hangups -- dont skip. #Bookofwhy

8.28.18 @2:31am - (Replying to @LeaHilde) Glad to see that @Bookofwhy and @causalinference made it all the
way to Amsterdam. Yes, the exercises should be super helpful. I can't think of a more exhilarating experience
than solving in 2-3 lines problems that generations of philosophers deemed to be "meta-physical".

8.28.18 @12:49am - (Replying to @HolgerSteinmetz @PHuenermund @Noahpinion) This takes us back to my


heroine Barabara Burks (1926) (#Bookofwhy Chapter 9), who startled her colleagues with an innocent question:
"What makes us think that a partial regression brings us closer to effect size than pair-wise regression.?" She
could't get a job.#causalinference

8.28.18 @12:07am - (Replying to @HolgerSteinmetz @PHuenermund @Noahpinion) This paper has nothing to do
with causation. They talk about "effect size" (as do many wishful) but if you replace "effect size" with
"regression-coefficient size" every sentence will still be valid. My litmus test is notation: Diagram?, do(x) or Y(x)
? else it's pre-causal

8.27.18 @11:49pm - (Replying to @tdietterich @GaryMarcus) I have reasonable doubts that these problems will
be solved withing the DL paradigm. Why? Because generalization is a causal notion, needing a model of what
relations might be perturbed by changing environments. DL researchers are still treating it as a statistical
problem.

8.27.18 @5:35pm - (Replying to @Jabaluck @cdsamii @analisereal) The "overfitting" metaphor, whereby "the
more knobs there are the easier it is to get the desired result" is a relic of statistical thinking. (as is "letting the
data speak for itself") Things are quite different in causal modeling, begging for a drastic paradigm
shift.#Bookofwhy

8.27.18 @4:47pm - (Replying to @Jabaluck @cdsamii @analisereal) "Full models" do not exist. We are discussing
modeling very few variables that you perceive to be relevant to your problem of interest. For example, can you
convert a bad IV into a good IV by conditioning on another variable?. A repeated question that requires a tiny
model yet...

8.27.18 @3:51pm - (1/2) (Replying to @Jabaluck @cdsamii @analisereal) You just hit on another virtue that
economists can borrow from #causalinfernce and #Bookofwhy (it goes back to Haavelmo
https://ucla.in/2mhxKdO ): "Thinking" is done only ONCE - when the model is constructed - he rest is vividly
displayed. The model spares you the torments..
8.27.18 @4:00pm - (2/2) (Replying to @yudapearl @Jabaluck and 2 others) The model spares you the torments
of "thinking how X2 might be endogenous," and other thinking exercises that "IV-experimentalists" are
tormented by, all for refusing to put down a model of how variables are affecting each other. #Bookofwhy
#causalinference
8.27.18 @2:54pm - (Replying to @dtweiseth) I never assume "data" is just snapshot. See for example
#Bookofwhy Fig. 7. 6, which depicts sequential treatments. Temporal information, if available, can significantly
enhance our model's veracity, but it cannot replace causal information. eg the rooster precedes the sunrise.

8.27.18 @1:18am - (1/2) (Replying to @Jabaluck @ben_golub) @Jabaluck, I think you are right. The question:
"What can economists learn from DAGs given that they already excel in 1,2,3 ..." should not be answered with
litmus tests for 1,2,3,... This tends to offend professional prides and create eternal enemies. Instead ...
8.27.18 @1:22am - (2/2) (Replying to @Jabaluck @ben_golub) Instead, the answer should be: "This is where
DAGs excel and, if it has any merit for economics, your bright students will smell it from 5 miles away and
adopt it as obvious." The #Bookofwhy aims to do just that. Independenly, I invite them to enjoy:
https://ucla.in/2mhxKdO .

8.26.18 @8:11pm - (Replying to @eliasbareinboim @deaneckles and 6 others) The implication of Elias's note is: If
someone gets an answer different from the one obtained by do-calculus, it must be that someone either erred
or failed to solve it without the do-operator -- a risky practice of some potential-outcome circles.
#causalinference #Bookofwhy

8.26.18 @7:10pm - (Replying to @ravithekavi) Paragraphs that withstood the test of time, and in which I still
believe despite a stormy year in the trenches deserve a special status in my diary. Thanks for bringing it to my
attention, Ravi. #Bookofwhy

8.26.18 @4:47pm - (1/2) (Replying to @Jabaluck @analisereal) It is a deeply cherished belief among economists
that "econometrics has powerful results" for 1) 3) 5) ... The litmus test is to ask your brightest student to solve
any of the toy problems presented here: https://ucla.in/2mhxKdO . Have things improved since 2014? Lets see
8.26.18 @4:57pm - (2/2) (Replying to @yudapearl @Jabaluck @analisereal) I have chanced to read a very recent
survey paper on econometric identification. Things have only gotten worse by the "IV-experimentalism"
campaign, which now absolves economists from thinking about causal models, including about whether a
variable is a proper IV. #Bookofwhy

8.25.18 @10:51pm - (1/2) (Replying to @eliasbareinboim @jpirruccello @_MiguelHernan) Apologies! You are
absolutely right in objecting to my playful poetry about Harvard not welcoming DAGs. The poetry came from a
discarded chapter of #Bookofwhy describing the funny schism between the noth side of the Charles River
(Statistics) where diagrams are tabooed, ....
8.25.18 @10:56pm - (2/2) (Replying to @eliasbareinboim @jpirruccello @_MiguelHernan) ... and the south side
(HSPH) where diagrams are encouraged. My reply to Kevin Gray referred to the North side, and I will correct it
on the blog. http://causality.cs.ucla.edu/blog/index.php/2018/06/15/a-statisticians-re-reaction-to-the-book-
of-why/ ...#bookofwhy #causalinference #Epibookclub @MiguelHernan

8.24.18 @10:49pm - For data analysts who wish to understand causes and counterfactuals from linear regression
viewpoint, I recommend this tutorial: https://ucla.in/2L8jUFg . It illuminates key concepts in #Bookofwhy by
the light of linear models: effects, mediation, robustness and generalization.

8.24.18 @3:39am - Sorry for the wrong link. The gentle introduction to counterfactuals can be found here
https://ucla.in/2wnb80c which is Chapter 4 of the Primer Enjoy.

8.24.18 @2:53am - This gentle introduction to counterfactuals https://ucla.in/2BDDTLK should clarify for
everyone the perplexing questions that were tweeted here: do(x) vs. Y(x), structural vs. PO framework + more.
Enjoy, and don't let clarity surrender to tradition #Bookofwhy #causalinference

8.24.18 @2:21am - (Replying to @PWGTennant @eliasbareinboim and 3 others) Likewise, save for two problems:
(1) Where do we place "diagnosis" (eg Bayes' billiard Table) which is not "prediction" and it needs no
(1) Where do we place diagnosis (eg Bayes billiard Table) which is not prediction and it needs no
intervention? (2) Suppose we can run RCT on every variable, can we estimate countefactuals? eg ETT, "but-for",
CoE mediation, etc. #Bookofwhy

8.23.18 @10:24pm - (Replying to @rationalexpec) Causality https://ucla.in/2BDDTLK is more advanced,


requiring graduate level math, elementary probability theory and some logic. It goes through proofs, and
deeper philosophical discussions. It still stands correct though, like a rock, despite the dancing and prancing.
#Bookofwhy

8.23.18 @8:23pm - (Replying to @atatlas123) Sure! We can't let publishers slow down progress. If you write to
my assistant kaoru@cs.ucla.edu we'll make sure you get the solution manual (for personal use). Truly nice
questions and truly neat solutions! Enjoy. #Bookofwhy #Epibookclub #causalinference #causality

8.23.18 @7:49pm - (Replying to @KordingLab @melb4886 @fadinberg) An economist "careful about thinking
about causality" should re-think if he/she is in the right profession, b/c causal policy makers rely on his/her
thinking. As to Causal Revolutions, mine begins here: https://ucla.in/2o2EehE and here
https://ucla.in/2mhxKdO . Yours? #Bookofwhy

8.23.18 @6:49pm - The next reading to dive into would be https://ucla.in/2n3VkuU which is the technical
compendium of #Bookofwhy, and includes beautiful examples, non trivial exercises and software support. For a
quick glance click on Pearce and Lawlor 2016 review # epibookclub #causalinference

8.23.18 @5:57pm - Normally the choice of taxonomy is arbitrary. But the more I think about it, the more I
suspect the difference between Miguel's trichotomy and the Ladder of Causation may account for the hurdles
we encountered while reading chapters 7-8 of #Bookofwhy. Any #Epibookclub thoughts.

8.23.18 @4:23am - (Replying to @HolgerSteinmetz @PHuenermund @Noahpinion) Tell us more about psychology.
Any major article unveiling how far behind the field is? Any embarrassing statement by the leaders? I know that
Psychometrika just rejected an article on graphical methods, but this is not unusual; editors are paid to impede
progress. What's new?

8.23.18 @2:14am - (Replying to @wgeary @MSFTResearch @akelleh) This is great news, which should make
causal inference a household item. It reminds me of Max Planck's quote: "A new scientific truth does not
triumph by convincing opponents to see the light, but because a new generation grows up that is familiar with
it." (1949) #Bookofwhy

8.23.18 @1:49am - (Replying to @swanderingf) This is another good question: doing discovery in potential
outcomes (PO). The answer is OF COURSE, the two are equivalent, but why would anyone want to do arithmetic
with addition only? The SCM framework is PO + DAGs, the former emerging from the latter. So SCM is PO.
#Bookofwhy

8.23.18 @1:02am - (Replying to @yudapearl @_MiguelHernan @rdpeng) Note that Miguel's trichotomy is different
from our Ladder of Causation, which behooves readers to ask: 1) Where is "diagnosis" (or "abduction") situated?
and 2) Where is the barrier between intervention and counterfactuals situated?#Bookofwhy #causalinference

8.23.18 @12:45am - (Replying to @rjmori @PHuenermund @Noahpinion) This looks familiar!!! Its the definition
of (nonparametric) identifiability from Causality (2000)!- Thanks for sharing! Sadly, even the most recent survey
paper on econometric identification that I reviewed does not provide a definition, not to mention a solution.
#Bookofwhy
8.22.18 @10:22pm - (Replying to @ShlomoArgamon) My deep-learning problem is much more modest. Forget
solving, I am concerned merely with representing the sentence: "the barometer fall does not cause rain". I do
not know how to encode it in any deep-learning program. Once encoded we can try solving, but not before.
#Bookofwhy

8.22.18 @10:07pm - (Replying to @swanderingf) Great question! The answer is YES, under selection. Examine
the graph: X---->Z<---U--->Y. If we select patients having Z=1 (say pain from treatment X) then the path
between X and Y will be active, producing correlation w/o causation ie confounding. #Bookofwhy
#causalinference

8.22.18 @9:19pm - (Replying to @PHuenermund @eliasbareinboim) I'd never give up hope; see #Bookofwhy.
Several economists indeed Tweeted that things are waking up, and the paper by Abadie-Cataneo affirms it.
Still, the editors of Econometrica do not see urgency in teaching readers what identification and endogeneity
are from a causal lens.

8.22.18 @7:28pm - (Replying to @_MiguelHernan @rdpeng) I specially love: "Enough with platitudes like "gaining
insights" & "extracting meaning". These were favorites of pre-causal statisticians. Many still fail to realize that
"meaning" and "insights" have been transformed into formal questions and algorithmic answers #Bookofwhy

8.22.18 @3:59pm - (Replying to @PHuenermund @Noahpinion) Economists had to change the definition of
"identification" because 95% of them do not know what it means (especially the PhD's). It is the only field that
prides itself on embarrassment. The gurus admit it, but will do nothing about it #Bookofwhy

8.22.18 @2:04pm - (Replying to @djvanness @rdpeng) Hybrid process is indeed what you find today in
#causalinference and #Bookofwhy but the Bayes option is a dangerous Siren song. As I explain here

https://ucla.in/2mhxKdO , why I am only Half-Bayesian. Spraying priors and waiting for posteriors to peak will
never flip --> into <--

8.22.18 @6:58am - (Replying to @djvanness) The point I make in https://ucla.in/2umzd65 is that, even if we let
a deep-learning algorithm digest the data for three years it still has no language to postulate a scientific
hypothesis, such as "the barometer fall does not cause rain" #Bookofwhy @rdpeng no C in, no C out.

8.22.18 @6:26am - (Replying to @PHuenermund @eliasbareinboim) It is easier to convince the Pope to preach
Voodoo than to teach an economist to speak cause-effect. But its not entirely their fault -- it has been so long
since they last communicated with the outside world, and their journals would not publish well-spoken papers.
#econometrics

8.22.18 @5:55am - Very well put! But I would go even further: "To be a science, data-science should start where
science does." I tried to add it here https://ucla.in/2umzd65 but one reviewer got offended. #Bookofwhy
#ecobookclub #Epibookclub

8.21.18 @6:15pm - (Replying to @lisabodnar @ja_labrecque_ and 5 others) The correct link to the revised "On
non-manipulable causes" is https://ucla.in/2wpolWr Sorry.

8.21.18 @6:05pm - (Replying to @lisabodnar @ja_labrecque_ and 5 others) For passengers on board of the
"Obesity" cruiser, I have just revised the paper "On non-manipulable causes" https://ucla.in/2MSj4xw . Thanks
for all your comments Key message: Time to separate epidemiological research from its potential outcome
for all your comments. Key message: Time to separate epidemiological research from its potential outcome
roots #Bookofwhy #causalinference

8.21.18 @3:13pm - (Replying to @Undercoverhist @causalinf and 9 others) Frisch "thoughts experiments" may be
imaginative and insightful (like "Maxwell's demon"). But Haavelmo's priority shines through the credo: "A formula
is a baked idea. Words are ideas in the oven" #Bookofwhy p 335 Note, Granger and Sims are regressionists - no C
in, no C out.

8.21.18 @2:40pm - (Replying to @ProfMattFox) No need to struggle. The "Probability of sufficiency"


(#Bookofwhy) stands for what it says, i.e., the prob. that an event X=1 would be sufficient to produce outcome
Y=1. It is informed by the data. What PIE stands for is still waiting for experts' interpretation in #Epibookclub

8.21.18 @3:31am - (1/3) Anyone interested in the historical origins of the "well-defined intervention" fixation
can watch Don Rubin's talk on youtube (nov 2016), describing where potential outcomes came from, and how
they can influence users choice of treatment. #Bookofwhy
8.21.18 @3:40am - (2/3) Rubin also stresses the importance of separating Science from what we do to learn
about it. I mention it hoping to convince Rubin's disciples to separate causes (the Science) from interventions
(what we do to learn about them) #Bookofwhy #causalinference #Epibookclub
8.21.18 @3:51am - (3/3) The link to Rubin's talk is here : https://www.youtube.com/watch?v=AH7lvy45k2U
Note also how the antecendant x in the potential outcome Y(x) is always a "treatment", never a state (e.g.,
temperature = 30) on event (e.g., earthquake). and this is 2017. #Bookofwhy #causalinference #Epibookclub

8.19.18 @11:23pm - (Replying to @thosjleeper @eliasbareinboim and 2 others) Our ability to define (race =
black) has little to do with manipulation. Suppose we make the term very very precise, say by requiring a
spectral content of skin pigments. Now what? Would the sentence "The cup would not have shot him if only he
was white" be more "well-define"?

8.19.18 @6:48pm - (Replying to @robertwplatt @ProfMattFox) Hard to measure, or no way to measure are
different from "not well-defined" . The structural framework makes this distinction crisp, the potential-outcome

framework conflates the two. Why not choose clarity? If something has no effect then that effect must first be
well-defined.

8.19.18 @6:19pm - (Replying to @jamessseattle) I am not familiar with Eliasmith's work. If you have the source,
please try to summarize it within Twitter's constraints. Just what the input is and what the output we can
expect. Thanks, JP

8.19.18 @5:22pm - (1/2) (Replying to @EpiEllie) Ellie, Congratulations on finishing the @Bookofwhy. It has been
truly edicational for me to work with the #Epibookclub and learn all the obstacles that readers may have and
which we assumed away. Please pass my hugs to all the survivors. Hey, how was the paradox party? 8.19.18
@5:36pm - (2/2) (Replying to @yudapearl @EpiEllie) Speaking of learning more #causalinference and going a
step beyond the non-technical style of #Bookofwhy, I must suppress modesty and recommend
https://ucla.in/2n3VkuU Uniquely liberated from "potential outcomes" handcuffs, yet smiling in coherence and
examples.#Datascience

8.19.18 @4:21pm - (Replying to @ProfMattFox) http://ucla.in/2MQF9MY Sec 5: "The basis for rejecting the new
drug is precisely your understanding that "obesity has no effect on outcome," the very quantity that some
epidemiologists now wish to purge from science, all in the name of caring only about "what to do"#Bookofwhy

8.19.18 @3:09pm - (Replying to @thosjleeper @eliasbareinboim and 2 others) You hit it on the nail: "phenomena
for which effects cannot be defined because the cause is poorly defined." Isn't it possible that such phenomena
p y p p
are fictional? or cultural habit?, a relic of Rubin's RCT roots which epidemiologists now should, but find hard to
shake off?

8.19.18 @6:24am - (1/6) (Replying to @mendel_random) #Bookofwhy has two missions: 1. To lay before readers
what can and cannot be done given a set of "defensible anchors" (DAs) that they may possess. 2. To make those
DAs advertise their own defense by making them meaningful to and scrutizable by as wide a pool of peers
8.19.18 @6:28am - (2/6) (Replying to @yudapearl @mendel_random) a pool of peers and adversaries as possible.
The fact that you are able today to find flaws in some of those DAs, articulate by your peers, proves that
mission (2) has been accomplished. Had these DAs been made under the rug of ignorability or, worst yet,
8.19.18 @6:33am - (3/6) (Replying to @yudapearl @mendel_random) under the methodology that ruled epi prior
to 1990, you would be unable to judge their plausibility -- they would be buried in hand waving. As to mission
(1) it is purely mathematical, hence unassailable. Still, showing what is impossible saves many hours of futile
research
8.19.18 @6:37am - (4/6) (Replying to @yudapearl @mendel_random) and showing what is possible tickles the
imagination to find defensible opportunities for the unveiled possibilities. There remains the problem of what
to do when a researcher finds himself with no DAs. You addressed it in Ref 1, and proposed a "broaden scope"
approach.
8.19.18 @6:44am - (5/6) (Replying to @yudapearl @mendel_random) I then showed step by step in
https://ucla.in/2MZJRgf or https://doi.org/10.1093/ije/dyy068 that what your "broadened scope" proposal
aspires to do is already done in DAGs. Including "triangulation" "reasoning to the best explanation" and other
aspirations. #Bookofwhy
8.19.18 @6:54am - (6/6) (Replying to @yudapearl @mendel_random) The argument that powerful tools are
dangerous because they invite misuse was made against the telescope in 17th century Italy. I dont think it
applies to 21st-Century Epi. The priests are scientists, epidemiologists are smarter, and their tools are
transparent. #Bookofwhy

8.18.18 @6:38pm - (Replying to @EpiEllie @eliasbareinboim) Happy that you pointed to the source paper. Note
though that the paper starts quite friendly and provides ample examples. Importantly, the paper aims to be
"anti-standardization", showing that standard "standardization" methods (ie, re-calibtation) must be revamped.
#Bookofwhy

8.18.18 @6:17pm - I am sharing a new paper on non-manipulable causes https://ucla.in/2MQF9MY which


summarizes my position in the semi-heated discussion we have had here regarding Obesity and its
consequences. May all discussions come to a good cause. #Bookofwhy #causalinference #Epibookclub

8.18.18 @5:39pm - (Replying to @eliasbareinboim @EpiEllie @leskocar) The PO perspective has advanced
considerably in the past decade. It started with "DAGs are not helpful", and advanced to "DAGs are helpful but
only to verify ignorability". I now see it heading towards: "Ignorability is not really needed" #Bookofwy
#causalinference

8.18.18 @4:00pm - (Replying to @bobgrossman) There is also free access to Chapter 2, https://ucla.in/2MZn7bi
for those who care about the genesis of causal analysis, how statistics squandered it, and how my hero, Sewal
Wright, single-handedly stood in: "And yet it moves!" Where did he get this Chuzpa? #Bookofwhy

8.18.18 @4:03am - (Replying to @fadinberg) It is not only a relic of the past. I am asking myself what big-data
and deep learning would recommend to patients after analyzing megatons of medical records. #bookofwhy
#epibookclub

8.18.18 @1:03am - (Replying to @BreskinEpi @eliasbareinboim and 5 others) Eager to hear what the restrictions
are: Number of pies? content of pies? 2-SAT? Horn restriction? Moreover, given a restricted PIE, does it let you
"visualize" things like synergism or interactions, or send you back to sweat over the various pies in the system?
#B k f h
#Bookofwhy

8.17.18 @8:40pm - (Replying to @BreskinEpi @eliasbareinboim and 5 others) One reason computer scientists are
baffled by PIEs is that, recognizing that "A and B only cause Y when both are present" (given a bunch of N PIEs)
amounts to solving an N-variable SAT problem, which is hard. Can PIE do what math says we can't? #Bookofwhy
#Epibookclub

8.16.18 @10:04pm - (Replying to @EconTalker @SpenceKjell) Russell, Rumor has it that you are trying to reach
me. I am alive and reachable on judea@cs.ucla.edu. Somewhat tired, but alive. Judea

8.16.18 @8:06pm - (Replying to @JaimieGradus @EpiEllie and 2 others) As far as #bookofwhy is concerned, the
purpose of Fig. 7.5 is to show: (1) that even toy problems can present a formidable challenge to the unaided
mind (and PO). (2) Mt. Intervention is conquered, (3) science can be fun. Readers who internalized these 3
points need not solve 7.5

8.16.18 @6:55pm - (Replying to @malco_barrett @EpiEllie and 3 others) Back to PIEs. The fact that so many
smart people are in love with PIEs gives me 99% assurance that there is something good there. The fact that
none was able to explain how to look at a bunch of pies and get the insight they swear by tells me that someone
should do it. #Bookofwhy

8.16.18 @3:15am - (Replying to @statsepi @tmorris_mrc and 5 others) What most people do not realize is that
MAR can now be tested, see https://ucla.in/2LdEjZW So, I imagine 2 years from now your abstract will read::
"We used MI and passed our data through the famous MAR test." No editor will resist ! No reviewer will squeak!
#Bookofwhy

8.16.18 @1:09am - (Replying to @julesgreig) Fascinating article, thanks for sharing. No, my math cannot prove
causation, it is after all only math, namely going from assumptions to conclusions. But even this dry exercise is
fun, because you find things you did not expect. For example, that Reichenbach forgot colliders.

8.15.18 @11:32pm - (Replying to @MaartenvSmeden) Even after data collection it is difficult to infer causes of
unexpected missing data, or even expected missing data. But we found https://ucla.in/2LdEjZW that the easier
problem too is in total confusion: Suppose we postulate those causes. Can we estimate causal effects?

8.15.18 @10:19pm - (Replying to @causalinf @jakewertz and 9 others) Philosophy aside, to me, the causal
revolution in econ begins with Haavelmo, and I justify it step by step in https://ucla.in/2mhxKdO . The ET
Editors asked me to comment on the "credibility hype." My conclusions (Section 4): they did not pass the causal
litmus test. #Bookofwhy

8.15.18 @9:29pm - (Replying to @thosjleeper @TheBrettGall) In #Bookofwhy I praise Morgan and Winship for
saving social science from the fate of economics. Indeed they cover both sides, missing one tiny step -- showing
that the two sides are ONE. Namely, results obtained in one are obtainable in the other. A DAG generates all
PO's.

8.15.18 @8:57pm - (Replying to @AndersHuitfeldt @MaartenvSmeden) The most crucial thing to learn from data
is which variables are contaminated by missingness and which are not. We usually know this in advance. What
this paper tells us https://ucla.in/2LdEjZW is, assuming that we know which is which, can we estimate what we
need? #Bookofwhy

8.15.18 @8:42pm - (Replying to @thosjleeper @TheBrettGall) You dont need to drop potential outcomes (PO).
8. 5. 8 @8: p ( eply g to @t osjleepe @ e ettGall) ou do t eed to d op pote t al outco es ( O).
All you need to do is to see how PO's are derived from DAGs, then join the PO camp pretending you are one of
them, unsure of where PO's come from -- you will perform miracles. #Bookofwhy #causalinference #Epibookclub

8.15.18 @6:37am - (Replying to @MaartenvSmeden) My faithful oracle #Bookofwhy says that "missing data
mechanism" is a causal notion, hence it cannot be "inferred" by data collection. It sometimes has testable
implications, and this paper tells us when/how https://ucla.in/2LdEjZW . It even tests MAR, and some MNAR.
New and fun

8.14.18 @6:45pm - Intrigued by the "Paradox of inevitable regret", https://ucla.in/2OV2mPp ,some readers
asked how Forney and Bareinboim managed (2017) to combine information from observational studies to
improve on RCT performance. Here it is: https://ucla.in/2MImBye #Bookofwhy

8.14.18 @3:29pm - (Replying to @AndersHuitfeldt @_MiguelHernan @analisereal) Whatever helps detect model
misspecification is useful. But philosophical differences make a huge difference in practice. They lure whole
communities to dismiss the need for model specification, or even for models per se. Look at all the model-blind
armies around us. #Bookofwhy

8.14.18 @5:54am - (1/2) (Replying to @AndersHuitfeldt @_MiguelHernan @analisereal) First, "obesity" is not a
"node", it is a vector of 17 factors that captures the word "obesity". Second, a model is not a 'rug", it is our only
window to reality. Third, the 17 factors may be "misspecified", fine, but their effects remain intervention-
neutral. #Bookofwhy.
8.14.18 @5:05am - (2/2) (Replying to @AndersHuitfeldt @_MiguelHernan @analisereal) Surely there is ambiguity
in defining "obesity" and in handling side-effects of interventions. But the whole point of this discussion is to
treat these sources of ambiguity as instances of "model misspecification" not of "undefined-ness" or "non-
scientific-ness" #BookofwhyReplying to @AndersHuitfeldt @_MiguelHernan

8.14.18 @2:56am - (Replying to @_MiguelHernan @analisereal) This text makes me cringe: "different
interventions often result in different effect estimates... " Are we are all doomed? No! In SCM, different
interventions NEVER result in different effect estimates. X in P(y|do(x)) is not the Soda, it is Obesity. Breath
safely #Bookofwhy

8.13.18 @8:14am - (Replying to @EpiEllie @l__ds and 3 others) Correction: These two sets are logically
equivalent:
set1 (A=1,C=0) (B=1,C=1)
set 2 (A=1,B=0, C=0) (A=1, B=1) (A=0, B=1, C=1)
A and B interact in set 2, but not in set 1 Can you share "the rules of causal pies"? #Bookofwhy #causalinference
#Epibookclub

8.13.18 @7:11am - (Replying to @EpiEllie @l__ds and 3 others) I used to think the way you do. But consider
these three pies:
(A=1,C=0) (B=1,C=1) (A=0, C=1)

A and B do not interact. Now consider:


(A=1,B=0, C=0) (A=1, B=1) (A=0, B=0, C=1)
Here A and B do interact, right? Alas, the two are logically equivalent. #Bookofwhy #causalinference

8.13.18 @6:02am - (Replying to @EpiEllie @l__ds and 3 others) If a PIE helps "visualize how interaction could
arise" then I expect someone to be able to look at the PIE and tell whether "interaction could arise". This
amount to "detecting" interaction. So, how do we tell interactions (or some other relation) by looking at a set of
PIEs ?
8.13.18 @2:27am - (Replying to @AndersHuitfeldt @eliasbareinboim) Put yourself in the shoes of a 1925 farmer:
"The RCT is interesting because and only because it corresponds to what will happen if I use-Fertilizer1. If those
two ever came apart (and Fisher did not proved they wont), the RCT is no longer interesting." Thus RCT comes
2nd to do(x)

8.13.18 @1:07am - (Replying to @evpatora) Thanks for sharing the arithmetics with other followers of
#Bookofwhy. The three steps needed in computing potential outcomes are fundamental to understanding
counterfactuals and, honestlyy, I dont believe anyone can understand counterfactuals without doing these steps
ONCE. Thkx

8.13.18 @12:55am - (Replying to @AndersHuitfeldt) The decision makers that I know could not care less about
"what would have happened if we ran RCT". They care about "what will happen if we use fertilizer F1 (vs F2) on
the entire field" And this is what the do-operator simulates. #Bookofwhy #causalinference

8.12.18 @8:44pm - (Replying to @l__ds @EpiEllie and 3 others) I also thought that the relationships between
SCC and PO are well formalized. Today I begin to suspect that they were left hanging. I dont know what
"synergism" is, but formal definition of "Sufficient Causation" is on page 289, line 4. #bookofwhy Can Ellie read it
from PIEs ???

8.12.18 @8:15pm - (Replying to @l__ds @EpiEllie and 3 others) In the SCM model, the function Child=f(Parents)
is part of a complete specification of the model, usually kept implicit. The SCC model explicates this function. I
am still waiting to hear from Ellie how she detects notions eg. "interaction" from a collection of
PIEs.#bookofwhy

8.12.18 @7:18pm - Errata: A diligent reader of #Bookofwhy, anxious to verify the fallability of MATCHING in
potential-outcome analysis, has found a typo on page 282. Line 22 should read: -$9,500, instead of $5,000. No
change in conclusions. Please mark your copy. #Epibookclub

8.12.18 @6:19pm - Congratulations go to co-follower @eliasbareinboim, + his student and collaborator, who
were just awarded the UAI best student paper award! The paper provides an interesting way of identifying
causal effects from equivalence classes of models. http://causalai.net/r35.pdf #Bookofwhy

8.12.18 @4:48pm - (1/n) Heavens, No! If there is ONE thing I hope readers will take from #Bookofwhy it is that
causal quantities (eg, causal effects) should not be left to the mercy of English or Potuguese, but be defined as
a property of a formal model. Not back to Alchemy! #causalinference 1/n
8.12.18 @5:10pm - (2/n) I hope that #Bookofwhy convinces readers that defining "causal effect" in English won't
let us verify if any proposed estimator yields the quantity we wish estimated. If doubts remain, then a more
convincing book must be written dedicated entirely to this crucial p oint. 2/n

8.12.18 @5:48am - Before quitting #Bookofwhy chap 9 with counterfactual heart-breaking headache, recall, this
is precisely the headache that causal models prevent. Modelers think only ONCE, when constructing the
diagram. The rest is algorithmic, including mediation. Headache is for the birds.

8.12.18 @4:38am - RCT-ists beware: The important problem of "surrogate endpoint" has been lingering in
decades of quasi-formal confusion. A new paper by Tikka and Karavanen https://arxiv.org/pdf/1806.07172.pdf
... now submits the problem to full causal analysis. Worthy of attention! #Bookofwhy

8.12.18 @3:40am - (1/2) (Replying to @swanderingf) Before we ask "how do we measure the goodness of the
assumptions" we should ask "why should not the assumptions themselves advertise their goodness"? After all
assumptions we should ask why should not the assumptions themselves advertise their goodness ? After all
they came from the roots of OUR knowledge, so we should be able to judge their plausibility. Next we can assist
our 1/2
8.12.18 @3:52am - (2/2) (Replying to @swanderingf) Next we can assist our plausibility-judging software by
doing sensitivity analysis, which is a way of focusing this software on the most crucial assumptions first. Finally,
we submit a PROVISIONAL policy recommendation coupled with the assumptions. Why numbers? #Bookofwhy 2/2

8.11.18 @3:32am - (Replying to @cygarde) Off hand, "predictive models that change causal variables" sounds
like a violation of Cartwright's dictum. I would love to see and toy example. Perhaps the example will also help
me understand your proposal of "conjoint modeling". #Bookofwhy

8.11.18 @1:31am - (Replying to @FriedrichHayek) Pragmatically, I hope you agree that econometric education
and methodology need foundational reforms, that such reforms will not come from its current leadership, and
that the democratization of causal inference e.g., #bookofwhy is a step in the right direction. @PHuenermund

8.11.18 @12:55am - (Replying to @LauraBBalzer @Ron_Wasserstein @AmstatNews) In reading this interview on


@AnstatNews, please note my super emphasis on the "three bullets" at the end. They are meant not only for
hard core PO analysts, but also for those tormented by split loyalty between PO habits and new structural
thinking. #Causalinference #bookofwhy

8.10.18 @2:06am - (Replying to @l__ds @EpiEllie and 3 others) Indeed, that PIEs were re-discovered in
philosophy, epi and law says a lot for their naturalness. That's why I am so eager to explore what makes them
"natural" and what function they serve in causal modeling, as well as in teaching epi. Help me explore, What is
it? #Bookofwhy

8.9.18 @8:03pm - (Replying to @EpiEllie @malco_barrett and 3 others) No miscommunication, just systematic
explication. If one's goal is quickly verify absence of interaction, three questions arise: (1) What for? (2) Is PIE
the best representation for doing it? (3) What's our input? i.e., where is the knowledge needed to construct a
PIE? @Bookofwhy

8.9.18 @6:48pm - (Replying to @EpiEllie @malco_barrett and 3 others) Gladly. In DAGs, interactions are
presumed to exist a priori, with no bells or chimes. Why? Because the theory of Boolean functions (1848) tells us
that the percentage of non-interactions in the space of functions Child=f(Parents) is super infinitesimal.
#Bookofwhy cont. 1/n
8.9.18 @6:57pm - (Replying to @EpiEllie @malco_barrett and 3 others) In DAGs, if you want to quantify a
specific interaction, all you need to do is to estimate P(Y|do(x),do(z)) and find out to what degree the effect of
X on Y is modified by Z. This is a do-expression, so it is estimable by do-calculus. Done. No bells, no fancy
articles @Bookofwhy
8.9.18 @7:20pm - (Replying to @EpiEllie @malco_barrett and 3 others) Continuing with interactions. Note that I
wrote a do-expression P(Y|do(x),do(z)) and NOT a conditional expression as habits and PIEs so seductively
tempt us to do. Note also that PIEs tacitly insist on components being Parents(Y), not so X and Z above.
#Bookofwhy 3/n
8.9.18 @7:35pm - (Replying to @EpiEllie @malco_barrett and 3 others) This is not to say that PIEs have no role
in SCM. But as analysts we need to explicate this role carefully if we want robots to use PIEs, or even just to
teach Epi students, and "help them think". What can they do with a PIE at hand, compared with PIE in the sky?
#Bookofwhy n/n

8.9.18 @5:57am - (Replying to @yudapearl @malco_barrett and 4 others) Given that Causal PIE Models do not
tell us whether Oxygen or Match was a sufficient cause for Fire, we conclude that additional information is
needed to make this distinction. Where does it reside? The #Bookofwhy says: it resides in the pre-fire
distribution, and given by PS.
8.9.18 @5:50am - (Replying to @yudapearl @malco_barrett and 4 others Thinking of PIEs. Suppose Joe is naive
causal analyst who never heard about PIEs, but manages nevertheless to express Child=f(Parents) in some other
convenient way (say conjunctive normal form). What would Joe not be able to do that a PIE lover can.
#Bookofwhy #Epibookclub

8.9.18 @5:45am - (Replying to @yudapearl @malco_barrett and 4 others) Thinking about Causal PIEs. Suppose
Joe is a naive causal analyst who never heard about causal PIEs, and is using a DAG in which ACE is identified
(say by backdoor.) What would Joe gain by learning about PIE. #Bookofwhy #Epibookclub

8.9.18 @2:59am - (Replying to @OstlundOllie @AndersHuitfeldt) To rule out interpretations based on selection
bias, let us assume that the 200 students were randomly selected; not "volunteers". Let us also assume that
students did not know whether they are in the randomized group or in the free choice group. Let's keep it
simple. #Bookofwhy

8.9.18 @12:08am - Rumours say that a party is planned in a bookclub to celebrate the successful crossing of
Chapter 6 - Paradoxes Galore! I would love to spice up the party with a humble a paradox of my own,
https://ucla.in/2OV2mPp , discarded from the #Bookofwhy by space limitations. Enjoy!

8.7.18 @1:12pm - This is a great article, by Chris Hitchcock, both profound and comprehensive. Thanks for
bringing it to our attention here, at the Tweeting trenches. It even mentions #Bookofwhy, despite of what the
book says about the philosophy of "probability raising" in chapter one.

8.8.18 @4:50am - (Replying to @causalinf) Thanks for the paper. I am totally speechless, and stand in awe to
celebrate the publication of the FIRST econometric paper since Herbert Simon to contain a graphical model. As
we say in the trenches: there is hope to econometric after all. #Bookofwhy Correction: forgot Hal White

8.8.18 @1:42am - (Replying to @malco_barrett @NeuroStats and 3 others) Agree. The role that Rothman's PIE
plays in causal modelling is simple: You zoom in on any family in the DAG and, if the function Child = f(Parents)
is Boolean, you can give it visual description, called Disjunctive Normal Form (DNF). But are the slices
"sufficient causes" ? 1/3 8.8.18 @1:46am - (Replying to @malco_barrett @NeuroStats and
3 others) Is it justified to call the slices of the PIE "sufficient causes" ? Consider the famous oxygen-match-fire
problem: Was the Presence of Oxygen a sufficient cause for the fire?. More trouble: The DNF is not unique. If A
is a sufficient cause, so is any event B=1. Why? see 2/3
8.8.18 @2:03am - (Replying to @malco_barrett @NeuroStats and 3 others) If A is a sufficient cause, so is any
event B=1. Why? Because A is logically equivalent to (A & B=1) OR (A & B=0). [See Causality Section 10.1.4. for
discussion]. The notion of "sufficiency" described in #Bookofwhy is invariant to syntax and, surely, prefers Match
to Oxygen. 3/3

8.7.18 @11:43pm - (Replying to @DKedmey) My all time favorite economist is Joseph. In econometrics I vote for
T. Haavelmo and H. Simon. Modern heretics? None! None of the priests can solve any of the toy problems posed
to them in https://ucla.in/2mhxKdO . But I see sparks of heresy among the young, even here on Twitter.

8.7.18 @10:53pm - (Replying to @MariaGlymour) Beautifully said. Now I understand why I find more common
language with epidemiologists than economists. #Bookofwhy

8.7.18 @12:22pm - (Replying to @malco_barrett @NeuroStats and 3 others) Great. But since every potential
outcome model is a derivative of a SCM (structural causal model) we can ask: What connects SSC and SCM? To
see this, lets ponder (1) What this pie means to you? (2) What the pie would look like when Y=1 if and only if
A=B #Epibookclub #Bookofwhy
A=B. #Epibookclub. #Bookofwhy

8.7.18 @6:07pm - (Replying to @erikbryn) Sorry, I overlooked your Aug 5 tweet. No reason for Kleinberg etal to
remind us that prediction is important for policy making; else why would politicians would be hiring expensive
pollsters. We are now learning the surprising converse; even sworn pollsters should study causation

8.7.18 @5:52pm - (Replying to @malco_barrett @NeuroStats and 3 others) At this juncture, my college professor
would say: Now that we understand what a model is and how it helps us answer hard questions on interventions

and counterfactuals, lets go back to our favorite PIE and ask where it fits in the #Bookofwhy scheme of things.
Do I see any hands?

8.7.18 @12:22pm - (Replying to @malco_barrett @NeuroStats and 3 others) Great. But since every potential
outcome model is a derivative of a SCM (structural causal model) we can ask: What connects SSC and SCM? To
see this, lets ponder (1) What this pie means to you? (2) What the pie would look like when Y=1 if and only if
A=B. #Epibookclub. #Bookofwhy

8.7.18 @2:00am - (Replying to @malco_barrett @ken_rothman) I am sure there is a connection between the pie
and the notion of "necessary causation" mentioned in #Bookofwhy but to explicate it one needs to know: 1.
what this pie means to you? and 2. what the pie would look like when Y=1 if and only if A=B. #Epibookclub.
@sharlagelfand @epi

8.6.18 @6:56pm - (Replying to @DataPuzzler) As the next technical introductionI beyond #Bookofwhy I strongly
recommend the Primer http://bayes.cs.ucla.edu/PRIMER/index.html It contains references to software
(Daggity) and has entertaining homework problems that even the gurus will find hard to solve.

8.6.18 @12:31pm - (Replying to @eddelbuettel) Thanks for noting this oversight; it will be corrected. My
skepticism of econometrics is not about its scientific foundations but about its current cultish and insular
leadership (see https://ucla.in/2L8OCyl ) #Bookofwhy. @ndoogan

8.6.18 @12:14pm - (Replying to @malco_barrett @ken_rothman) My difficulty w Macksey's INNUS condition (


Causality chap 8) stems from a simple principle: You can't express counterfactual concepts (rung-3) in
propositional logic (rung-1), the "necessary" (and "sufficient") idea requires counterfactual specification.
#Bookofwhy @EpiEllie

8.6.18 @2:55am - (Replying to @stephensenn) I hope you agree there is a difference between "not throwing
information away" and "use baseline as covariate". If we blindly stratify on (or "control for") every measured
baseline we will get into the M-bias fallacy (e.g., page 161 @Bookofwhy). Any qualification? @f2harrell
8.6.18 @11:59am - (Replying to @yudapearl @stephensenn @f2harrell) The M-Bias fallacy does not afflict RCT's,
so my request for qualification is answered. Thanks. #Bookofwhy #Epibookclub

8.6.18 @2:12am - (Replying to @EpiEllie) Let us not conflate specific conditions needed for estimation with
relationships that hold universally in every problem. The first law of causal inference is universal, and
consistency follows from this law, regardless of debatable concerns over "well-defineded-ness" #Bookofwhy

8.4.18 @1:57pm - This is what we used to believe, until problems of transportability, robustness and
explainability taught us that even purely predictive tasks should be model-guided. See https://ucla.in/2umzd65
#Bookofwhy Gush! It rains whenever I forget my umbrella (selection bias in big data).
8.4.18 @10:45pm - (Replying to @erikbryn) I think you are saying: Yes to commonsense, Yes to the Ladder of
Causation. Yes to modern #causalinference. Join us at #Bookofwhy.

8.4.18 @1:15pm - Well put, with one twist: replace "a good empirical strategy" with "a provable strategy",
because most big data guys think they "have a good empirical strategy for identifying causal effects", citing
leading statisticians who play down the importance of identification. #Bookofwhy

8.2.18 @1:48pm - To all captains and sailors on this ship, The errata sheet for #Bookofwhy has been updated.
See https://ucla.in/2L8jUFg . Thank you all for your comments and suggestions. Sail safely and enjoy the rest
of your (smooth) voyage. #epibookclub #causality #causalinference

8.2.18 @2:01pm - Sorry, the correct link to the errata sheet for #Bookofwhy is https://ucla.in/2NRof0D . Sail
safely . #epibookclub #causality #causalinference

8.1.18 @12;59PM - My interview on #Bookofwhy has just come out on Amstat News https://ucla.in/2v72QK5 .
Please note my "three bullets" summary of what potential-outcome enthusiasts can learn from causal modeling.
#epibookclub #causalinference #Datascience

7.29.18 @8:38pm - For those who are hooked on Y(x) go ahead and replace P(Y|do(x)) with P(Y(x)), no rule will
fail. Just recall: P(Y|do(x), do(z), W=w) goes into P(Y(x,z)|W(x,z)=w). Also note: the #bookofwhy does not go
into the do-calculus, except conceptually. #epibookclub

7.29.18 @8:16pm - Got it, thanks. It is ready to be corrected

7.29.18 @6:54pm - Econ should have striven to be the "crown" because it had such a huge head-start over other
disciplines. We need to ask: "What happened!" if we are serious about removing the obstacles that resist
working together in true openness.

7.29.18 @6:39pm - Where exactly is the "land grab"? That I classify maximum likelihood in "traditional

statistics"? or that I harness modern machine learning techniques to compute maximum likelihood in much
larger models?

7.29.18 @6:22pm - The counterfactual Y_x (e.g, "Wet had it rained") is a unit property; it refers to identical
circumstances, same season, same day, same grass etc. Do-expressions are population properties. Typical grass
on a typical day will get wet if we do(rain). #bookofwhy #epibookclub

7.29.18 @3:02pm - The do(x) operator simulates what will happen if you actually apply the atomic action
do(X=x). If you cant implement such action (eg. do(rain)) it still tells you what will happen if you could, and
sometimes it does not matter, (eg, the grass will still get wet). #bookofwhy

7.30.18 @12:50am (Replying to @EpiEllie @eliasbareinboim @PHuenermund) Everything we do with do-calculus


and counterfactuals, in the entire #bookofwhy is WELL-DEFINED, because "well-defineness" is a mathematical
notion that we adhere to religiously by defining things unambiguously on the model. Estimation and
implementation are orthogonal questions

7.29.18 @1:14pm - Absolutely, as long as it encode causal assumptions. And, BTW, I do not know of an economic
model that is not a causal model, though many economists shake when you ask them.
7.29.18 @1:08pm - P(Y|do(x)) tells you that, to get ATE, you do not need to know counterfactuals (rung-3,
Chapter 8), nor worry about "ignorability" magic before you adjust. Just make sure you adjust for variables
satisfying the backdoor condition. #Bookofwhy #epibookclub #causalinference

7.29.18 @12:14am - #Bookofwhy #epibookclub. Chapter 6 should actually be most fun and rewarding Even the
first section (the back-door) already walks you through territories that economists will find unchartered and
leading statisticians will deem heretical, if not "strongly ignorable." Happy voyage

7.28.18 @8:38pm For those who are hooked on Y(x) go ahead and replace P(Y|do(x)) with P(Y(x)), no rule will
fail. Just recall: P(Y|do(x), do(z), W=w) goes into P(Y(x,z)|W(x,z)=w). Also note: the #bookofwhy does not go
into the do-calculus, except conceptually. #epibookclub

7.29.18 @11:36pm 1/2 (Replying to @_MiguelHernan @EpiEllie) This debate surfaces at least twice a year since
Paul Holland's announced his model-blind dictum (1986) "No causation without manipulation". I have defended a
model-based position in Causality (2009) Sections 11.4.3, 11.4.5, 11.4.6, and many more places. To summarize:
2/2

7.28.18 @1:45pm - Your review is excellent and comprehensive. Now that we dig the limitations of classic IV's,
attention should perhaps be shifted to "mediating IV's" (frontdoor) which provide nonparametric point
identification of ATE and ATT (#bookofwhy, p. 235-7), albeit under different setting.

7.27.18 @6:24pm - Lauren has discovered a typo in #bookofwhy. It should be P(no tar|do(smoking)). A few typos
have already been corrected here: https://ucla.in/2NRof0D . A few more will be corrected soon. I suggest to
mark your copy accordingly. Please let me know if any more. #Epibookclub

7.28.18 @1:28pm (Replying to @zongel @EpiEllie) This is another correction on our plate. Andrew Forbes alerted
us to Elizabeth's major contribution and we will reflect that in the next revised printing. #Bookofwhy

7.30.18 @12:04pm (Replying to @yudapearl @_MiguelHernan @EpiEllie) Deja vu !!! I recognized this debate from
the ancient past. Here it is again: "Does obesity shorten life?"https://ucla.in/2NSDZk3 . I hope it makes the
reading of #Bookofwhy more enjoyable and less dusty: If you suspect side-effects, you must put them in the
model - No cheating!

7.26.18 @11:12pm (replying to @IronEconomist) "... most important and underrated statistician". Hmm. The
former is overrated and the latter makes me cry. But given that #Bookofwhy is #1 on Amazon best seller lists, in
both stat. and biostat., I think statistics shows promising signs of awakening to modernity.

7.26.18 @9:24pm (Replying to @Jsevillamol) I just posted the answer on


http://math.stackechange.com/questions/2857 and invited you to view the full solution, which was posted

here #bookofwhy yesterday. Glad mathematically inspired students see the challenge and beauty of graphical
modeling.

7.26.18 @6:16am (Replying to @AndersHuitfeldt) It is not notational convenience but a semantical leap. To
define Pr(Y_x) I need to define Y_x, not so in defining Pr(Y|do(x)); it is what we measure in a randomized
experiment (rung-2). The fact that the two collapse under Pr does not make them equally meaningful (eg to
computer
7.26.18 @4:13am (1/2) (Replying to @_MiguelHernan) I am over-flattered by your kind words about #Bookofwhy
and thanks for posting our interview. We disagree on the role of intervention in the definition of "effect" and I
explained my reasons for defining "effects" prior to intervention. (eg the moon affecting tides). 7.26.18
@4:42am (2/2) (Replying to @_MiguelHernan) We also differ on whether #Bookofwhy (p. 9) description of the
do-operator is overrated in light of previous attempts to distinguish doing vs. seeing. I consider it pivotal, being
the first to offer a knowledge-based counterfactual-free semantics for the distinction. (rung-2)

7.25.18 @5:37am (Replying to @aynumazi @Harvard) In the 1960-1980's, Simon, Lewis, Rubin and Robins devised
various counterfactual theories (Chapters 8-9) which could distinguish actions from observations but did not
become operational until the do-operator (1993) told us when the effects of actions can be estimated from
data.

7.25.18 @1:18pm (Replying to @cdsamii @analisereal and 3 others) I like potential outcomes too, and use them
as much as graphs. You know why? Because, given a graphical model, I know what Y_x means, where it is in the
model, how it is derived from the model, etc. It is a good language to define the quantities we wish estimated.
#bookofwhy

7.25.18 @7:38pm (Replying to @aynumazi @Harvard) Let me explain why your marked sentence is valid in its
context. Prior to 1993 there was no formal way of taking what people understand about storms and barometers
and proving them right on their intuitive predictions. Counterfactuals would not have gotten us very far.

7.25.18 @1:49pm (Replying to @_MiguelHernan @aynumazi @Harvard) Under certain conditions, the do-operator
bears syntactic resemblance to g-formula ie, both are products of conditional probabilities. The resemblance
ends here. The do-operator is defined on my knowledge -- a causal diagram, while the g-formula on an object I
do not recognize.

7.25.18 @6:31pm (Replying to @AndersHuitfeldt @_MiguelHernan and 2 others) Are you familiar with Stalnaker-
Lewis definition of "effect" in terms of "closest worlds"? It is also "not as easy to interpret" and formally
isomorphic to DAGs and FFR.... Yet it is metaphysical, for it is not defined in terms of the way knowledge is
represented by researchers.

7.25.18 @2:50pm (Replying to @eliasbareinboim @AndersHuitfeldt and 4 others) I would not insist on "solving",
let's focus on "encoding," to see how the quantity "the effect of X on Y" is represented in "FFRCISTG" which is
proposed as alternative representation of causal knowledge. We just want to be sure that the answer is not pre-
encoded in the FF...

7.25.18 @6:19pm (Replying to @AndersHuitfeldt @analisereal and 5 others) It is not a matter of "user friendly".
It is a matter of defining "effect" in terms of what you know (diagram) or defining it in terms of metaphysical
quantities, eg counterfactuals, and later show formal isomorphism. In SCM, counterfactuals are defined BY THE
Model not converse.

7.25.18 @2pm - Notation is only part of counterfactuals, semantics is no less important. The diagram is more
than a visual aid; it DEFINES the do-operator. Without it I would not know what to cut and what to leave uncut.
Without it I can't explain why tweaking the barometer will not stop rain.

7.27.18 @10:31pm (Replying to @aynumazi) Glad you liked the answers to your question. The main point is that
the do-opeartor cannot be separated from the diagram on which it operates, and from the features of the
diagram that legitimize its predictions. (e.g., which factors to wipe out). #Bookofwhy #causalinference
7.24.18 @10:27pm - You might be one of the enlightened few, but I have spoken to literally hundreds of
microeconometrics graduates, across the country. Infinitesimally few can solve the toy problems presented in
my ET paper https://ucla.in/2mhxKd0 and many many resent me for counting -- I stopped.

7.25.18 @2:09am (Replying to @fabinger) - Gee, Michal, I hope I do not sound "complaining". No way! I am
describing the sad state of econometric education to entice students and researchers to study causal inference
on their own, eg.through #bookofwhy rather than stale textbooks and conservative professors. It is working

7.26.18 @3:27am (Replying to @juli_schuess @fabinger) The Chen and Pearl survey has been updated recently
by Chris Auld, https://chrisaulddotcom.wordpress.com/2013/10/08/remarks-on-chen-and-pearl-on-causality-in-
econometrics-textbooks/ . Moreover, some authors of econ. texts are calling us to review their upcoming
revised editions. There is hope, but it is very hard to uproot a culture baked in confusion. #Bookofwhy

7.24.18 @10:54pm- (Replying to @rationalexpec) The correct link to the ET paper is https://ucla.in/2mhxKdO .
The title is "TRYGVE HAAVELMO AND THE EMERGENCE OF CAUSAL CALCULUS" and the toy problems are in
Section 3.2. I have mentioned this section before to #bookofwhy --it is worth a second reading.

7.24.18 @3:46pm - Economics has had all the potentials for becoming the queen of social science. It blew it by
fostering an orthodox and insular culture. Do you know that to post a comment on NBER Webserver you need to
be an "approved NBER family member"? #bookofwhy is not very eco-charitable.

7.24.18 @3:46pm - And do you know the percentage of economists who can tell (in their own model) which
parameter can be identified by OLS? Or which variable would qualify as a (conditional) instrumental variable? Or
whether a model has testable implications? Or,...Or.. It could have been the queen

7.24.18 @10:10pm (Replying to @stewarthu) You made my day, Stewart, thanks. And, as our sages said: "One
convert weighs more than ten born-righteous" (free translation from the Hebrew). Welcome to the global march
on commonsense. #bookofwhy

7.24.18 @6:25pm (Replying to @intensivemargin) I am not so sure. Readers of Morgan and Winship (2014) and a
good chunk of political science (eg #bookofwhy p.228-230) can answer these elementary questions. The

enlightened camp of econometrics is still dormant/suppressed. And it shows in the literature, in paper after
paper.

7.24.18 @11:47pm - (Replying to @PHuenermund @vhranger) Thanks, Paul. But, to escape charges of bias, I
would not state a problem in graphical terms. The problem is Eco. 101. You take ANY economic model and ANY
parameter BETA in the model, and you ask an innocent question: Can BETA be estimated using OLS with ANY
control whatsoever?

7.25.18 @ 6:07am - (Replying to @vhranger @PHuenermund) I am not speaking graphical models. I am speaking
econ 101 where you teach simultaneous equation models and you explain why OLS does not always work. A

curious student asks: when can we estimate a parameter, say beta, using OLS. This is not a provocative
question. Why fight it?

7.24.18 @2:59pm - My opinion on @aaronecarroll article http://nyti.ms/2ND756J is echoed here:


https://ucla.in/2JfdQcv . Trialists and writers need to know that mathematical tools for managing and
interpreting "pragmatic trials" are now available, hence they can safely enter the folds of science.
7.24.18 @1:47am (to Paul Hunermund @PHuenermund) Well put, Paul. I often say it in less poetic terms: "The
assumptions embedded in the model act like a catalyst to steer the data to support the conclusion that we seek
to obtain". This clumsy sentence may be the formal definition of "Extract insight". Well put. #Thebookofwhy

7.23.18 @9:14pm - I can think of (at least) two ways of achieving model-free behavior: (1) rely on data only,
and (2) match any new situation to one of your prestored models, then use the latter. Which one do you
advocate? Which one is currently practiced in ML? #Bookofwhy #DataScience

7.23.18 @8:19pm - Caution: "garbage in garbage out" is sometimes used against causal inference, saying: "your
conclusions are not more informative than your assumptions". Wrong! Our conclusions are quantitative,
informed by data, our assumptions are qualitative, reflecting prior knowledge.

7.23.18 @12:42am - Hector Geffner just gave an interesting talk at IJCAI titled "Model-free, Model-based, and
General Intelligence" https://ucla.in/2O8rDoQ . He likens these two modes of reasoning to Kahneman's System-
1 and System-2. #bookofwhy #Epibookclub #DataScience #causalinference

7.23.18 @12:57am (Replying to @corbyjerez @AlexVasilescu) I hope this work: https://ucla.in/2O8rDoQ

7.15.18 @11:32am - This is tyranny galore!! Just because some path analysts disavow the causal meaning of
their models you were forced to do the same? By AJE? We truly need a Hall of Shame and we truly need to
educate those "satisfied referees" that this is the 21st century! #bookofwhy should help.

7.15.18 @11:19am - I agree with your conclusion, but not with your first sentence. Do say "X causes Y", with the
understanding that it is an abbreviation for the rest. Science has known many abbreviations before (eg,
density=weight/volume) why not this one? It is not fair to ask us to repeat.

7.15.18 4:24am - "Causal inference" is benign, it goes from causal assumptions to causal conclusions, never
claiming to "infer causality" from naked data. Making this explicit, and justifying some of the assumptions,
should cure editors allergy. Mentioning the "causal revolution" might also help.

7.15.18 - I am surprised that, in the 21st century, submissions to medical journals still require special
justification for causal inference. I propose starting a new blog, "Hall of Shame", to post reviews and editorial
letters from journals that do require such justifications. #bookofwhy.

7.14.18 @7:08pm - For the curious, the second part of Cinelli's tutorial has been posted on:
https://listserv.ua.edu/cgi-bin/wa?A2=ind1807&L=SEMNET&P=142106 It demonstrates the logical equivalence of
Rubin's potential-outcome framework and the structural causal model used in #bookofwhy. A conclusion in one,
is a conclusion in the other

7.15.18 @12:50pm (Replying to @robinberjon) Sorry, this is all I have. My Login was free and easy.

7.14.18 @4:01pm - It is not "forbidden" physically. If we do not observe a person more than once we escape the
criticism: "Once treated, it is not the same person " or, "Once operated on, you cannot undo the surgery."
#bookofway #epibookclub It makes the results applicable to undoable treatments.

7 14 18 @3:35pm This paper would not have been published were it not for Phil Dawid's editorial leadership to
7.14.18 @3:35pm - This paper would not have been published were it not for Phil Dawid s editorial leadership to
overrule all the authoritative reviewers and move the field forward. Today,such editors are an endangered
specie. And that is why, Kuhn explains, paradigm shifts take forever. #bookofwhy

7.14.18 @11:25am - So what says the Jury? (1) The heterogeneity tests proposed have been considered before
and found to be useless, or (2) they havn't been considered and the jury is still out, or (3) they haven't been
considered but are judged useless apriori? Where are we? #bookofwhy #epibookclub

7.14.18 @10:28am - No radical steps allowed. Every person is observed only ONCE. The observational study is
conducted separately, in the general population, not involving subjects chosen for the experimental

study.#Bookofwhy #epibookclub @stephensenn @AlexJohnLondon @f2harrell

7.14.18 @9:46am - This uncertainty in response holds for an isolated RCT, true. But if you also run a separate
observational study, then the two studies give us an inequality that, if violated, rules out the second option,
namely, the population cannot be homogeneous. #BookOfWhy #EpiBookClub

7.14.18 @9:08am - Using the inequalities in https://ucla.in/2uqDjdp we can only tell that there are substantial
differences in the population, that is, some patients benefit much more than others, but we cannot tell who
those are. #bookofwhy #epibookclub

7.13.18 @3:13pm - I had to pacify a reviewer's demand to remove the reference to Babylon, possibly to protect
the honor of ML. I truly recommend Toulmin's "Foresight and Understanding" (1961). Published papers must
compromise to the lowest denominator. Hoping #BookOfWhy will see an uprising.

7.14.18 @4:17pm (Replying to @aesopesque) It is not unusual. After all, reviewing is a boring job unless you see
it as part of a mission to defend your ideology, or your profession. That is why the professional literature is so
slow to catch up with progress (see econometrics). That is why I wrote #Bookofwhy #epibookclub

7.13.18 - The heterogeneity tests derived in https://ucla.in/2uqDjdp only require one observation per
individual. Did I derive the impossible? I doubt it! But I used counterfactual logic and causal models, which may
have not been available to more senior analysts. #BookOfWhy #EpiBookClub

7.13.18 - A reader asked if there is a test for heterogeneity, e.g., whether a drug is uniformly beneficial or kills
some and saves more. Remarkably, three such tests have been uncovered in https://ucla.in/2ugpzTg . Can
someone in #BookOfWhy or #EpiBookClub check if the FDA is aware?'

7.13.18 - Speaking about Reichenbach, the #BookOfWhy (p. 47-51) is highly critical of "probabilistic causality" -
a branch of philosophy that emerged from Reichenbach's work, and which was misguided from day one. See
Section 33.5 in https://ucla.in/2ugpzTg . Philosophers can make mistakes.

7.13.18 - Readers ask if the #BookOfWhy describes the work of Hans Reichenbach, the first to attempt a
connection between causation and probabilities. Briefly. Noting that his failure to account for colliders

represents our mind's tendency to assume:"no correlation without causation."

7.11.18 - Perhaps "Let's fake it" is too harsh. Anyone who tells us: "Let's assume ignorability" is really saying:
"Let's move to where I feel more comfortable - statistical estimation - and leave causal inference to others.
#BookOfWhy
7.11.18 - Why I can't trust "ignorability" assumptions? Because I am yet to find an ordinary mortal who can
discern ignorability claims in his/her head, even in toy problems. The mental task is formidable so, anyone who
tells us: "Let's assume ignorability" is saying: "Let's fake it". #Boo

7.11.18 - Fashion pushes researchers to say they are combining causality with ML. So, watch for imposters.
How? They hide assumptions. Some start: "Let's assume ignorability"-- suspect them. Others will not even say
that -- shun them. I trust those who climb the 3-step ladder. #BookOfWhy

7.10.18 - In the spirit of Twitting, I recently completed a paper that summarizes the #BookOfWhy in seven
words, each standing for a principle or a tool: http://ftp.cs.ucla.edu/pub/stat_ser/r481.pdf Once you acquire
these seven tools you would qualify as top Causal Inference Expert - a champion of commonsense.

7.8.18 (1/4) - 10 days ago, when I decided to go on Twitter, I did not know what I am getting into. Today, with
over 5,000 followers, I feel a sense of obligation to give you a progress report, and to reflect on this
experience. This will probably take 4 Tweets - stay tuned #bookofwhy
(2/4) - First, I find the intellectual exercise of compacting pages of technical analysis into 3-4 sentences
challenging, educational and empowering. Second, the genuine quest for understanding I see among you gives
me the hope that my primary goal of writing the Book is realizable.
(3/4) - My goal was and is: the democratization of causal inference . By that I mean empowering students and
researchers to understand causal inference on their own, without waiting for professors or journal editors to
catch up. Toward that end, I will continue to bring up ...
(4/4) - for discussion topics that have been ignored or suppressed by mainstream literature and, simultaneously,
try to be as flexible, honest and open-minded as I can in responding to your questions and comments.
Unortunately, I can only Tweet once or twice a day. #bookofwhy

7.7.18 - The logical equivalence between the structural vs. potential outcome frameworks is one of those issues
that researchers are taking for granted but rarely discuss. For those interested, Carlos Cinneli is giving a tutorial
on this topic here: https://listserv.ua.edu/cgi-bin/wa?A2=ind1807&L=SEMNET&P=42219 #bookofwhy

7.7.18 - My dream is "No Reader Left Behind". What's the problem? Conceptual or mathematical? #Bookofwhy

7.6.18 - Cartwright"s dictum is so simple and so universal, like the conservation of mass and energy. Yet it is so
often ignored, especially by Bayesians. Bayesianism today is a license to spray priors on whatever you can't
estimate, then wait for the posteriors to peak, in case they do.Judea Pearl added,

7.6.18 - Kudos to Jonathan Kirsch for a most insightful review of #BookOfwhy, which makes even humble me
understand what the Book is all about. Who said you have to be a rocket scientist to understand cause and
effect?

7.3.18 - Someone asked: Why a revolution and not a dialogue? T. Kuhn lamented the same, and concluded that,
given scientists ego and investment, no dialogue can avert a revolution. I tried it, for the past 25 years - and
wrote the book becasues I failed. #BookOfWhy #Epibookclub

7.1.18 - Someone asked (can't find who): Doesn't correlation always imply some form of causation? Answer:
Almost! This was Reichenbach's dictum (1956), But he forgot about colliders (See The Book of Why, page 199,
and the Monty Hall paradox). #BookOfWhy

7.1.18 - Where do we fit? For a physicist, the equations of physics are sufficient. Not so for a robot who would
b il t t dt t k th b t t t i St t l ti ti li f th
be easily tempted to tweak the barometer to prevent rain. Structural equations operationalize some of the
implicit information a physicist must use before applying the eqs.

7.1.18 - "Only focused on robots?" God forbids! I am focused on science, and thinking "robots" helps us uncover
all the implicit information that a scientist must bring to bear on a problem, before applying the explicit
information encoded in the equations. It's a good thinking device.

6.28.18 - An interviewer asked me today what I believe will be the most important practical implication of the
Causal Revolution. My answer: Serious erosion of the hegemony of RCT's. See
http://ftp.cs.ucla.edu/pub/stat_ser/r477-online-reprint.pdf @deatonangus

6.30.18 - I am not sure what is meant by: "response schedule/counterfactual etc controversy in statistics". From
what I recall, "response schedule" was David Freedman's round-about way of saying "function", and
"counterfactuals" were made controvertials by Dawid (2000), but no longer are.

6.27.18 - Hi everybody, the intense discussion over The Book of Why drove me to add my two cents. I will not
be able to comment on every tweet, but I will try to squeak where it makes a difference.

5.15.17 - I exist therefore I tweet!

Potrebbero piacerti anche