Sei sulla pagina 1di 43

MICRO-ECONOMETRICS

ECO 6175

ABEL BRODEUR

Week 6

1/ 43
Regression Discontinuity Design

Outline:
I (1) Regression Discontinuity Design

I (2) Dell (2010)

I (3) Stata

2/ 43
Objective:

How can we evaluate government policies?


I Understanding RDD

Reference:
I Dell, 2010. “The Persistent Effects of Peru’s Mining
Mita”
I Beland, 2015. “Political Parties and Labor Market
Outcomes: Evidence from U.S. States” AEJ: Applied
Economics

3/ 43
Regression Discontinuity Design
Assignment to treatment is based on a clearly defined
index or parameter with a known cut-off for eligibility
I Here the control and treatment groups are different
ex ante (only ex post for RCTs)
I Possible when units can be ordered along a
quantifiable dimension which is systematically related
to the assignment of treatment
I Compare individuals just above and below the cut-off
for eligibility

Note that the effect is measured at the discontinuity.


Estimated impact around the cut-off may not generalize to
entire population

4/ 43
Regression Discontinuity Design

Rules with cut-offs are common in social programs


I Anti-poverty programs targeted to households below
a given poverty index
I Pension programs targeted to population above a
certain age
I Scholarships targeted to students with high scores on
standardized test

For example, beneficiaries and non-beneficiaries are


systematically different. BUT: for people near the
threshold, differences are slight

5/ 43
Regression Discontinuity Design

Are these thresholds applied?


I Easily seen with a graph

I Sharp discontinuity that determines the treatment

For example, people who earn $50,000 or less (or more)


are (not) eligible
I Treatment is a tax credit

I Dependent variable is happiness

6/ 43
Baseline (Not Always a Baseline)

7/ 43
Follow-Up

8/ 43
Regression Discontinuity Design

Another Example:
I Score between 0 (healthy) and 100 (dead) of health
status
I A treatment is given to individual with a score over 50

I Give treatment to sickest patients

I We get causal effect only at the discontinuity


(score = 50)

9/ 43
Regression Discontinuity Design

Post = β0 + β1 pre + β2 treatment + µ

where β2 is the LATE!

10/ 43
Regression Discontinuity Design

E(postscore|pre = 50− ) = β0 + β1 50− + E(µ|pre = 50− ) (1)

E(postscore|pre = 50+ ) = β0 + β1 50+ + β2 + E(µ|pre = 50+ )


(2)
We get from equations (1) and (2)

β2 = E(postscore|pre = 50+ ) − E(postscore|pre = 50− )


− [E(µ|pre = 50+ ) − E(µ|pre = 50− )]

11/ 43
Regression Discontinuity Design

If [E(µ|pre = 50+ ) − E(µ|pre = 50− )] = 0, then β2 is the


RDD coefficient: causal effect of the treatment for
individuals with a prescore around 50

Assumption:
I Unobservable variables are continuous around 50
(vary smoothly)
I Cut-offs are often not arbitrary!
I And manipulations of pre-score!

12/ 43
Manipulation of the Assignment Variable
Hard to test!

McCrary Test (2008)


I Assess whether there is discontinuity in the density of
observations at the cut-off
I But not sufficient to prove that there is no
manipulation...

Also test that there are no jump at the discontinuity for


other covariates
I Imbens and Lemieux (2008) look at one side of the
discontinuity and test for discontinuity in that part

13/ 43
Distribution of the Running Variable

14/ 43
Distribution of the Running Variable

15/ 43
Nonlinearities in Functional Form

So far we measured the LATE as the size of the


discontinuity at the cut-off

We thus assumed the relationship between assignment


and dependent variable was linear
I Beware of the functional form, could be nonlinear!
And heterogeneous effect!
I Causal effect is obtained only if the functional form of
the relationship between the assignment variable and
the outcome variable is correctly modeled

16/ 43
Linear

17/ 43
Nonlinear

18/ 43
Nonlinearity Mistaken for Discontinuity

19/ 43
Sharp RD
I Conditional independence assumption
E[Y0i |Xi , Di ] = E[Y0i |Xi ]. Once we control for a
confounder Xi , treatment assignment is as good as
random.
I Need deep understanding of the mechanism which
underlies the assignment of treatment Di .
Assignment usually depends on a single variable Xi
(i.e. the running variable). In the sharp design, this
variable fully determines the treatment according to
the cut-off rule.

1 if Xi ≥ X0
Di = (3)
0 if Xi < X0

20/ 43
Sharp RD

21/ 43
Sharp RD

E[Y0i |Xi ] is a function of Xi , so write


I Y0i = f (Xi ) + εi

I Y1i = Y0i + β

I Yi = f (Xi ) + βDi + εi

I Yi = f (Xi ) + β1(Xi ≥ X0 ) + εi

The function f (·) must be continuous at X0 . In practice,


need to assume some flexible functional form for f (·), for
example a polynomial.

22/ 43
Functional Form

How to choose the form:


I Flexible enough to get the functional form of E[Y0i |Xi ]
right away from X0
I Make the window of your data around X0 smaller and
stick with linear
I RD provides local information on treatment at X0
I Need lot of data near X0
Optimal bandwidth choice (Imbens and Kalyanaraman,
RESTUD 2012)

23/ 43
Nonlinearities: Heterogeneous Effect?

24/ 43
Different Functions either Sides of the
Cut-Off

We can allow for different functions to the left and right of


the cut-off:

Yi = fl (Xi )1(Xi < X0 ) + fr (Xi )1(Xi ≥ X0 ) + βDi + εi (4)

Need to impose fl (X0 ) = fr (X0 ). For a linear function for


f (Xi ) this amounts to

Yi = α+γ1 1(Xi < X0 )(Xi −X0 )+γ2 1(Xi ≥ X0 )(Xi −X0 )+βDi +εi .
(5)

25/ 43
Fuzzy RD
What happens if the RDD is not sharp? Fuzzy RDD are
common!
I Some individuals “crossed-over” treatment or did not
“show-up” for the treatment
I Ex: Scholarship may not be wholly determined by
GPA, but threshold to become eligible
In those instances, we can get the LATE for the subset of
individuals who are induced into treatment at the cut-off
I In a regression framework, estimated through a 2SLS
approach
I Similar to non-compliance for RCTs: requires that the
instrument is good!

26/ 43
Fuzzy RD

27/ 43
Fuzzy RD

When there is not a deterministic assignment rule, then


we have a change in the probability of treatment at the
cut-off
I Probability of treatment p(Xi ) is also a continuous
function, except at X0 .
Sharp RD is the special case

1 if Xi ≥ X0
p(Xi ) = (6)
0 if Xi < X0

28/ 43
Fuzzy RD is IV

There are various regressions now


I Yi = f (Xi ) + βDi + εi structural equation

I Yi = f (Xi ) + π2 Xi ≥ X0 + ζ2i reduced form

I Di = g(Xi ) + π1 Xi ≥ X0 + ζ1i first stage

The cut-off induces a change in the probability of


treatment. If treatment matters, this induces a change in
the outcome. Since the treatment does not affect all units,
the jump at the cut-off in the outcome needs to be
rescaled by the jump at the cut-off in the probability of
treatment (β = ππ21 )

29/ 43
Fuzzy RD

30/ 43
Fuzzy RD

31/ 43
Estimation

1) Take means for the sample near the threshold


I Numerically equivalent to estimating IV

2) Kernel
I Discount weight of observations further away

I Better strategy if the sample size is small (may keep


all the observations)

32/ 43
Regression Discontinuity Design
Advantages:
I Many programs have eligibility rules

I Relatively weak assumptions (compared to other


non-experimental methods)
I Ethics

Disadvantages:
I Other programs might use the same (or related)
eligibility rules
I Households can respond to eligibility rules (test for
bunching)
I Thus there are threats to internal validity

I Only get local treatment effects

I Low power (fewer observations)

33/ 43
Local Average Treatment Effect

LATE: only get local treatment effects with RDD


I Not always generalizable

I Heterogeneous impacts (for the poorest for example)

I Only know the impact for those around the threshold


(not clear those are the most interesting ones)

34/ 43
Robustness Checks

Report results for both estimation types:


I Include also higher order polynomials in X and show
it does not affect the results
I Local linear regression

Show that the results are not affected by the window


around the cut-off

Check whether other variables that should not be affected


by the treatment jump at the discontinuity

35/ 43
Dell (2010)

Utilize RDD to examine the long-run impacts of the mita


I Mita started in 1573 and abolished in 1812

I Required over 200 indigenous communities to send


one-seventh of their adult male population to work in
the mines
I Discrete change at the boundary of the region:
I On one side, all communities affected
I On the other, all communities were exempt

36/ 43
Dell (2010)
Discrete change suggests a RDD approach for evaluating
the long-term effects of the mita
I Multidimensional discontinuity in longitude-latitude
space
I Validity of the RDD requires all relevant factors
besides treatment to vary smoothly at the mita
boundary
I Focus on a region in Peru where elevation and ethnic
distribution are identical across segment of the
boundary
No selective sorting across the treatment threshold (i.e.
out-migration)
I Hard to gain membership and land in different
indigenous communities
37/ 43
Summary Statistics

38/ 43
Summary Statistics

39/ 43
Dell (2010)

Basic regression framework

cidb = α + γmitad + Xid0 β + f (geographiclocationd ) + Φb + εidb ,

where cidb is the outcome variable for observation i in


district d along the segment b of the mita boundary
I Mita is an indicator equals to 1 if district d contributed
to the mita and 0 if it did not
I Φb is a set of boundary segment fixed effects that
denote which of four equal length segments of the
boundary is the closest to the observation’s district
capital

40/ 43
41/ 43
42/ 43
Stata

Beland, 2015. “Political Parties and Labor Market


Outcomes. Evidence from U.S. States”
I Estimates the causal impact of the party allegiance of
US governors on labor-market outcome
I RDD: 1977-2008

I Democratic governors cause an increase in the


annual hours worked by blacks relative to whites
I Stata files on the website of the journal (AEJ: Applied
Economics)

43/ 43

Potrebbero piacerti anche