Sei sulla pagina 1di 9

Downloaded from jech.bmj.

com on 21 November 2008

Causation in epidemiology
M Parascandola and D L Weed

J Epidemiol Community Health 2001;55;905-912


Updated information and services can be found at:

These include:
References This article cites 36 articles, 18 of which can be accessed free at:

9 online articles that cite this article can be accessed at:

Rapid responses You can respond to this article at:

Email alerting Receive free email alerts when new articles cite this article - sign up in the box at the
service top right corner of the article


To order reprints of this article go to:

To subscribe to Journal of Epidemiology and Community Health go to:
Downloaded from on 21 November 2008

J Epidemiol Community Health 2001;55:905–912 905

Causation in epidemiology
M Parascandola, D L Weed

Abstract But despite much discussion of causes, it is not

Causation is an essential concept in clear that epidemiologists are referring to a sin-
epidemiology, yet there is no single, gle shared concept.
clearly articulated definition for the disci- Multiple definitions of cause have been
pline. From a systematic review of the lit- oVered in epidemiology. In 1970 MacMahon
erature, five categories can be delineated: and Pugh wrote that “the word cause is an
production, necessary and suYcient, abstract noun and, like beauty, will have diVer-
suYcient-component, counterfactual, and ent meanings in diVerent contexts.”2 Yet cause
probabilistic. Strengths and weaknesses of is also a scientific term, and it is important that
these categories are examined in terms of there is some common thinking among epide-
proposed characteristics of a useful scien- miologists about what is meant in saying “X
tific definition of causation: it must be causes Y.” In practice many causal statements
specific enough to distinguish causation are ambiguously stated; for example, “smoking
from mere correlation, but not so narrow is a cause of cancer” may mean “every smoker
as to eliminate apparent causal phenom- will develop cancer” or it can be construed as
ena from consideration. Two categories— “at least one smoker will develop cancer,”3
production and counterfactual—are depending upon the underlying concept of
present in any definition of causation but causation the speaker has in mind.
are not themselves suYcient as defini- There have been calls in the recent epide-
tions. The necessary and suYcient cause miological literature for more attention to the
definition assumes that all causes are theoretical and conceptual basis of epidemiol-
deterministic. The suYcient-component ogy.4 5 In this journal, Nijhuis and Van der
cause definition attempts to explain Maesen urged readers to investigate the philo-
probabilistic phenomena via unknown sophical foundations of public health, with
component causes. Thus, on both of these special attention to ontological concerns (for
example, the nature of basic concepts such as
views, heavy smoking can be cited as a
“public” and “health”).6 In this paper, we focus
cause of lung cancer only when the
on causation.
existence of unknown deterministic vari-
This paper reviews the epidemiological
ables is assumed. The probabilistic defini-
literature, seeking patterns and dominant defi-
tion, however, avoids these assumptions
nitions. The strengths and weaknesses of
and appears to best fit the characteristics diVerent approaches to defining causation are
of a useful definition of causation. It is also examined, drawing also on the rich philosophi-
concluded that the probabilistic definition cal literature on this topic. From this, a key dis-
is consistent with scientific and public tinction is drawn between deterministic and
health goals of epidemiology. In debates in probabilistic concepts of causation. The choice
the literature over these goals, proponents between these two approaches is in part driven
of epidemiology as pure science tend to by researchers’ views about the role of
favour a narrower deterministic notion of epidemiology in relation to other sciences and
causation models while proponents of epi- to public health. We explain this relation and
demiology as public health tend to favour make a recommendation about what type of
a probabilistic view. The authors argue causal definition best meets the goals of the
that a single definition of causation for the discipline of epidemiology.
discipline should be and is consistent with The focus of attention in this paper will be
both of these aims. It is concluded that a on chronic disease epidemiology. It is in this
counterfactually-based probabilistic defi- area that the concept of causation has proved
nition is more amenable to the quantita- most elusive and has provoked the most confu-
tive tools of epidemiology, is consistent sion and debate. The focus will also be on the
with both deterministic and probabilistic ontological nature of causation rather than on
Cancer Prevention phenomena, and serves equally well for
Fellowship Program, causal inference, a more epistemological con-
OYce of Preventive
the acquisition and the application of sci- cern; it is this conceptual understanding that
Oncology, National entific knowledge. those who call for more theory have noted is
Cancer Institute, (J Epidemiol Community Health 2001;55:905–912) deficient.
Executive Plaza South,
Suite T-41, 6120
Executive Blvd, MSC “The view [of causation] we adopt has Literature review
7105, Bethesda, MD consequences which reach far beyond informal We searched for definitions of causation in the
20892–7105, USA
discussion during coVee breaks.”1 epidemiological literature to determine how
Correspondence to: Causation is an essential concept in the epidemiologists characterise causes. Heuristic
Dr Parascandola practice of epidemiology. Causal claims like devices that are not strictly definitions, such as
( “smoking causes cancer” or “human papilloma the “web of causation”, were excluded.7 The
Accepted for publication virus causes cervical cancer” have long been a web is not intended to be a theory or to provide
22 May 2001 standard part of the epidemiology literature. causal explanations, but merely to act as a
Downloaded from on 21 November 2008

906 Parascandola, Weed

metaphor for the idea that causal pathways are the definition is vague about what “produc-
complex and interconnected.8 tion” or “creation” means. Thus, the elusive
The MEDLINE database was searched concept of causation is defined only in terms of
using PUBMED. The English language litera- another equally elusive concept. Philosophers
ture from 1990 to August 1999 was searched have long been aware of the weaknesses inher-
for MeSH terms “causality” and either “epide- ent in this definition of causation; David Hume
miology” or “epidemiologic methods.” The and Bertrand Russell both reject the notion of
terms were used unexploded so that subhead- causation because it retains this mysterious
ings of these terms would be excluded. element.10 11 The shortcomings of this defini-
Altogether 148 records were retrieved. “Cau- tion have also been noted in the epidemiologi-
sality” was only introduced as a MeSH heading cal literature.12 13 Thus, a more robust defini-
in 1990. The literature from 1966 through tion is needed.
1989 was searched for “logic” and “epidemiol-
ogy” as keywords. A search for “causation” and NECESSARY CAUSES
“epidemiology” during this period also re- Many discussions of causation in epidemiology
vealed some further references. Articles also incorporate the concepts of necessary and suf-
were identified through the authors’ collections ficient causes. Succinctly, a necessary cause is a
of papers on the topic of causation in epidemi- condition without which the eVect cannot
ology. The reference lists of review articles on occur, and a suYcient cause is a condition with
causation in epidemiology were perused for which the eVect must occur.(page 191–2)14
mention of additional relevant articles. Major (page 4–5)15 (page 326–7)16 (page 21)17 (page
epidemiology textbooks were also reviewed for 261)18 (page 27)19 (page 45–47)20 Four diVerent
discussions of causation. From among these types of causal relations can be derived from
sources, at least five distinct definitions can be these two definitions: necessary and suYcient,
identified. necessary but not suYcient, suYcient but not
necessary, and neither necessary nor suYcient.
A small minority of epidemiologists main-
Making sense of multiple definitions tain that the term “cause” should be limited to
After reviewing the literature, we have found highly specific necessary conditions.21 22 The
that cause is defined in the following ways: pro- view that all causes must be necessary for their
duction, necessary causes, suYcient- eVects is traditionally associated with the germ
component causes, probabilistic causes, and theory of disease, wherein each disease—for
counterfactuals (table 1). Interestingly, these example, tuberculosis—is caused by a specific
five types of definitions correspond with major infectious agent—for example, tubercle bacil-
approaches to causation found in the philo- lus.23 Stehbens applies this model to all
sophical literature. However, discussion of diseases, claiming that no particular disease has
these definitions in the epidemiological litera- been proven to have more than one cause.
ture fails to recognise some important distinc- Instead, says Stehbens, multi-causal models
tions and criticisms from the philosophical lit- merely indicate gaps in scientific understand-
erature. In particular, it is important to ing, as scientists have not yet uncovered a sole,
distinguish between deterministic and probabi- specific cause or adequately defined a disease
listic causes and between ontological defini- outcome. Thus, for example, a high cholesterol
tions of causation and causal models used in diet cannot be called a cause of heart disease.21
scientific inference. While these five definitions Charlton similarly claims that the basic sci-
are not all mutually exclusive, there are signifi- ences are built on the concept of necessary
cant consequences to choosing one definition causes and that epidemiology, in order to be
over another. scientific, should follow that model.22
Support for this narrow definition of causa-
PRODUCTION tion is ultimately a product of the lingering his-
Under this definition, a cause is something torical influence of scientific determinism.
that, in short, creates or produces an eVect. A Since Galileo and Newton, classical physics has
similar notion is that causes aVect outcomes been rooted in a world view in which complex
(page 637)9 or that causes alter outcomes.2 In phenomena can always be reduced to simple,
contrast, a non-causal association does not deterministic mechanisms. Strict determinism
involve production; A coincides with B, but A requires a one to one correspondence between
does not produce B. But while the notion of cause and eVect; the same cause invariably
production draws an ontological distinction leads to the same eVect, with no role for chance
between causal and non-causal associations, or stochastic variation. In the biomedical

Table 1 Definitions of causation from the epidemiological literature

Production Causes are conditions that play essential parts in producing the occurrence of disease.2 9
Necessary causes A necessary cause is a condition without which the eVect cannot occur. For example, HIV infection is a
necessary cause of AIDS.14–22
SuYcient-component A suYcient cause guarantees that its eVect will occur; when the cause is present, the eVect must occur.
causes A suYcient-component cause is made up of a number of components, no one of which is suYcient on
its own but which taken together make up a suYcient cause.26 29
Probabilistic cause A probabilistic cause increases the probability of its eVect occuring.1 12 15 19 35–37 Such a cause need not be
either necessary or suYcient.
Counterfactual causes A counterfactual cause makes a diVerence in the outcome (or the probability of the outcome) when it is
present, compared with when it is absent, while all else is held constant.46–48 The counterfactual
approach also does not specifically require that causes must be necessary or suYcient for their eVects.
Downloaded from on 21 November 2008

Causation in epidemiology 907

context, 19th century physiologist Claude Ber- “these risks are either one or zero, according to
nard argued that for medicine to be scientific, whether the individual will or will not get lung
like physics or chemistry, it must recognise only cancer.”(page 9)29 Thus, all events are wholly
causes that operate through specific and deter- “deterministic” because their occurrence or
ministic biological mechanisms (of course, non-occurrence is completely determined by
Bernard was writing before the 20th century the existing circumstances. While few epidemi-
probabilistic revolution in physics).24 ologists explicitly subscribe to a strict deter-
But epidemiologists and pathologists have, at minism, the fact that this principle underlies
least since the advent of the chronic disease era, Rothman’s widely cited account is noteworthy.
acknowledged that not all diseases can be What is the problem with a principle of
attributed to a single necessary cause.25 Ciga- determinism? Few causes that epidemiologists
rette smoke, for example, is not necessary for (or other biomedical scientists) identify dem-
development of lung cancer. In fact, some epi- onstrate a pattern of one to one correspond-
demiology texts and commentaries have stated ence with their eVects. Smokers develop more
that causes of complex chronic diseases, like lung cancers, but smoking is not by itself
cancer and heart disease, tend to fit into the necessary or suYcient for developing lung can-
“neither necessary nor suYcient” category. cer. The suYcient-component cause definition
(page 175)14 (page 5)15 (page 46)20 Medicine is postulates that smoking is one element in a
full of such relations. Moreover, the arguments suYcient cause and that the other elements
of Charlton and Stehbens are circular; they simply have not been identified yet. But this is
argue that science has never identified non- a strong belief to hold in the absence of empiri-
specific, unecessary causes because they ex- cal evidence. In short, the suYcient-
clude such causes by definition. In sum, while component cause definition requires that we
some causes may be necessary for their eVects, assume the existence of countless hidden eVect
this formulation is inadequate as a definition of modifiers to turn every less than perfect corre-
causation. lation into pure determinism. Moreover, be-
cause many causes identified via epidemiology
SUFFICIENT-COMPONENT CAUSES are weak, we would be required to assume that
The suYcient-component cause definition, many of these hidden eVect modifiers exert
articulated by Rothman,26 improves upon the eVects far stronger than those of known,
necessary cause view presented above by observable causes. No doubt numerous un-
admitting causes that are neither specific nor known eVect modifiers do exist, but to assume
strictly necessary for their eVects. A suYcient- (absent evidence) that they must exist in every
component cause is made up of a number of corner of the natural world is not a minor pre-
components, no one of which is suYcient for sumption. Is such a global assumption biologi-
the disease on its own. When all the compo- cally plausible? Scientists have long noted that
nents are present, however, a suYcient cause is biological processes frequently fail to behave
formed. Because more than one set of compo- deterministically.30 31
nents may be suYcient for the same eVect, a The suYcient-component cause definition
disease may have multiple causes. This defini- also has diYculty in generating models that
tion is essentially identical to an influential capture the dynamics of cause and eVect
account in the philosophical literature intro- relations. Because a suYcient cause is fully suf-
duced by John Stuart Mill27 and more recently ficient for its eVect, it cannot readily explain
promoted by Mackie.28 While in practice scien- how changes in the quantity of a component
tists and public health practitioners may focus cause (say, dose of an antibiotic) could lead to
their interest on a particular component of a a corresponding change in the eVect (say, suc-
suYcient cause (such as tobacco advertising), cess in breaking up an infection). Rothman26
the true cause of a disease, according to this responds to this concern by positing a set of
view, is the total assemblage of conditions that suYcient causes, each including a diVerent
are together suYcient for the disease. dose of the varying component cause (the anti-
But, like the necessary cause definition, the biotic). The problem with that response is that
suYcient-component cause definition retains it still fails to capture dose-response relations as
an assumption of scientific determinism that a continuum rather than a series of discrete
often goes unacknowledged. There are, in fact, steps. It is also ontologically unwieldy, intro-
two ways in which a cause can be necessary for ducing unnecessary complexity to a phenom-
some eVect: (1) it can be necessary in any set of enon more easily explained by other means (as
circumstances (the tubercle bacillus is neces- we discuss in the following section) solely to
sary for any case of tuberculosis) or (2) it can preserve the suYcient-component cause view.
be necessary only in a particular set of circum- By extension, as noted in the literature, similar
stances in which no other suYcient causes are worries apply to the view’s treatment of
present (uranium exposure is not a necessary interaction.32–34 In sum, the suYcient-
precursor for lung cancer, but perhaps for a component cause view has substantial draw-
particular non-smoking uranium miner his backs as a definition of causation for epidemi-
radiation exposure was necessary for his lung ology.
cancer to develop when it did). The suYcient-
component cause definition includes causes PROBABILISTIC CAUSATION
that are not necessary in the former sense, but Some commentators and textbooks have pro-
requires that, within a given set of circum- vided a “probabilistic” or “statistical” defini-
stances, all causes must be necessary and suY- tion of causation,1 12 15 19 35–37 where a cause
cient. For the individual, fatalism prevails: increases that probability (or chance) that its
Downloaded from on 21 November 2008

908 Parascandola, Weed

eVect will occur. For example, one source

defines “a cause of cancer as a factor that
x Multiple definitions of causation exist in
increases the probability that cancer will
the epidemiologic literature, but some
develop in an individual.”37 Under this defini-
prominent accounts rely on unfounded
tion, the occurrence of cancer in an individual
biological assumptions.
may be in part a matter of chance (that is, it is x A probabilistic definition of causation,
a “stochastic” or “indeterministic” process). along with the notion of counterfactuals,
Thus, a probabilistic cause may be neither nec- avoids these assumptions and oVers prac-
essary nor suYcient for disease. However, the tical advantages for epidemiology.
definition also does not exclude necessary and x Exclusive attention to events at the
suYcient causes; a suYcient cause is simply molecular level has been associated with
one that raises the probability of its eVect an overly restrictive definition of causa-
occurring to 1, and a necessary cause raises tion.
that probability from 0. x Requirements for an eVective definition
Probabilistic causation oVers an alternative of causation should be distinguished from
to the determinism inherent in necessary and those for constructing a useful causal
suYcient causation. In recent years, philoso- model in a particular investigation.
phers of science have developed sophisticated
theories of probabilistic causation to explain
apparently indeterministic processes.38–44 These answered by noting that a probabilistic defini-
theories provide the means to construct models tion allows for the possibility that other undis-
of interaction and dose-response relations in covered causes may also be at work. A newly
quantitative terms through a continuum of discovered genetic factor might reveal that
probability values.42 For example, the relation smoking dramatically increases lung cancer
between dose and, say, the probability of the risk for some individuals but has only a moder-
eVect occurring can be described mathemati- ate eVect on others. There is no reason to
cally. Thus, these probabilistic theories provide assume that a cause increases each individual’s
a better fit to existing epidemiologic tools for risk by the same amount. In fact, because of the
eVect measurement. mathematical continuum of probability, the
A probabilistic definition of causation is probabilistic model allows for a greater range of
more inclusive than the suYcient-component possible eVects. Risks for individuals might dif-
cause definition; while necessary and suYcient fer in more ways than simply being equal to a
causes can be described in probabilistic terms, probability of either 1 or 0.
probabilistic causes cannot be described in Another concern is that the definition
deterministic language. Additionally, the remains unclear about what it means to say
probabilistic definition makes fewer biological smoking raises one’s probability of developing
assumptions, as it does not require believing in lung cancer. For example, statisticians Cox and
countless hidden eVect modifiers for every less Holland45 46 both object to a prominent philo-
than perfect correlation. Science aims to make sophical account of probabilistic causation38 on
as few assumptions as possible, and it is these grounds. They argue that a definition of
contrary to that spirit to assume (by definition) causation based on statistical inequalities (that
that there can be no probabilistic causes. is, the probability of the eVect is diVerent when
A simple fix allows Rothman’s “pie” charts the cause is present than when it is absent) is
for suYcient-component causes26 to be used in inadequate. Olsen voices a similar criticism.
representing complex probabilistic causes as (page 3)35 In such cases, how is it possible to
draw a distinction between causal relations and
well. Simply think of the components as
non-causal associations? In order to meet this
contributing together to the probability of the
concern a further element must be added to
eVect, rather than being suYcient for it. If one
the definition—a counterfactual.
component is missing, the probability is
decreased. Rothman and Greenland29 them-
selves suggest a fix to account for stochastic
A few statisticians and epidemiologists have
phenomena. They suggest that the presence or advocated a counterfactual definition of
absence of one component might be deter- causation.46–48 A counterfactual statement
mined by a random process; yet on their draws a contrast between one outcome given
approach the relation between cause and eVect certain conditions and another outcome given
remains deterministic, but simply more diY- alternative conditions. For example, “if an hour
cult to predict. ago I had taken two aspirins instead of just a
While probabilistic definitions have ap- glass of water, my headache would now be
peared in the epidemiological literature, there gone.” Thus, Rubin defines the causal eVect of
has been little discussion of their strengths and treatment T, compared to absence of treat-
weaknesses or how they relate to other ment, as the outcome given T minus the
definitions. One concern likely to be raised outcome given the absence of T.47 The
against a probabilistic definition is that it fails definition requires a ceteris paribus condition
to explain why some smokers develop cancer that everything else is held constant.
and others do not, and that is what the “whole Counterfactuals can be either deterministic
discipline of epidemiology” is after.(page 205)1 or probabilistic. The aspirin counterfactual can
Probability is therefore a “euphemism for be revised to read: “If an hour ago I had taken
ignorance.” However, this concern can be two aspirins instead of just a glass of water, I
Downloaded from on 21 November 2008

Causation in epidemiology 909

would be much less likely to still have a a limiting or extreme case. Thus, it also makes
headache.” Some recent philosophical ac- fewer assumptions about unobserved natural
counts of probabilistic causation incorporate phenomena, removing the need to always posit
counterfactuals.39 40 43 44 For example, under hidden deterministic component causes. While
Cartwright’s definition43 C causes E if the a causal model of a specific phenomenon may
probability of E given C is greater than the posit hidden mechanisms, a definition requires
probability of E given the absence of C, while generalisability and should make few assump-
all else is held constant. tions. So why do some commentators continue
A counterfactual definition is not inconsist- to resist the probabilistic definition? DiVerent
ent with the necessary cause, suYcient- views about the role of epidemiology relative to
component cause, or probabilistic cause defini- other sciences and to the practice of public
tions. Rather, counterfactuals articulate an health influence choices about causal models.
additional attribute that we suspect will In the following two sections, we explain how
enhance any definition of causation by challenges in defining causation are related to
strengthening the distinction between causa- two current debates in epidemiology.
tion and mere correlation. While some com-
mentators have emphasised the role of counter- Causation at many levels
factuals more than others, none have argued Is there a hierarchy of causation at diVerent
that causes do not behave according to levels of observation? According to the discus-
counterfactuals. Nevertheless, the counterfac- sion so far, there is no reason to assert that
tual definition is insuYcient by itself as a defi- causes at one level, such as molecules, are any
nition of causation. While the counterfactual more real or significant than causes at another
definition essentially states that the presence or level, such as social factors. Yet the priority of
absence of the cause “makes a diVerence”, the diVerent levels of observation has been a topic
necessary and suYcient, suYcient-component, of debate in the epidemiological literature. In
and probabilistic definitions clarify what kind 1973 Susser20 described the importance of rec-
of diVerence it must make; on the probabilistic ognising diVerent levels of organisation in
view a cause makes a diVerence in the chance evaluating causal hypotheses. During the past
of the eVect. 15 years, much ink has been spilled over the
The ceteris paribus condition must be inter- value of “black box” strategies for investigating
preted carefully, however. While the ceteris causal relations.52–54 Critics of “black box” epi-
paribus condition is key to distinguishing cau- demiology argued that scientific understanding
sation from non-causal associations in theory, was only advanced by looking inside the black
it is not intended to be met in practice. Indeed, box to understand the underlying mechanisms
the condition is rarely met in actual scientific connecting cause and eVect at the biological
research; that is, the same individual cannot be level.55 56 More recently, advocates of the
observed in the exact same circumstances as importance of social factors in disease have
both as a smoker and a non-smoker.46 argued that causal phenomena at the social
Karhausen erroneously objects that a probabi- level cannot be fully reduced to biology or to
listic definition as outlined by philosophers of individual behaviours, like smoking.57–60 For
science43 is too strict for epidemiological prac- example, characteristics of the quality of a
tice, because it is defined through a ceteris neighbourhood may influence health outcomes
paribus condition.12 at the individual level, such as cases of
This confusion between criteria for defining gonorrhea61; these neighbourhood characteris-
a cause (ontology) and criteria for identifying a tics cannot be captured by looking solely at
cause empirically (epistemology) should be characteristics of individual residents (such as
avoided. Some epidemiologists have failed to income). This debate has substantial import
make this distinction, including intervention as for the allocation of resources in research and
part of a definition of what a cause is46 49 or intervention.
referring to observed relative frequencies in What does this have to do with the diVerent
stating a probabilistic definition of causation. notions of causation discussed? In short, the
(page 263)18 35 But a definition of cause should positions in the black box debate are polarised
not depend on what diVerences have been along similar lines as those in discussion of dif-
measured or whether a particular intervention ferent definitions of causation (particularly
has been attempted. Rather, the ceteris paribus deterministic versus probabilistic). Advocates
condition is hypothetical (that is, the contami- of the strict necessary cause view insist that
nated Broad Street well is a cause of the chol- causes should be identified at the level of inter-
era epidemic when it is the case that if the nal biological mechanisms (rather than in
Broad Street pump was shut down, then behavioural or social factors).21 22 24 And those
cholera incidence would decrease). Extra care who advocate a prominent role for social and
is required in the context of counterfactuals group level eVects8 57 62 are also critical of the
because they are often used not only in defining suYcient-component cause and necessary
causation but also in constructing causal mod- cause definitions, urging instead a broader cat-
els for scientific inference.50 51 egory of causation.
We argue that the probabilistic definition These diVerences correspond, in turn, to
combined with a counterfactual condition, like diVerent types of scientific explanations. In
that oVered by Cartwright, provides the great- developing a causal explanation of an observed
est promise for epidemiology. It is consistent association, there is a popular tendency among
with both deterministic and probabilistic both lay persons and scientists to prioritise
causal models, treating deterministic models as knowledge at the molecular level.63 In the
Downloaded from on 21 November 2008

910 Parascandola, Weed

1950s, for instance, sceptics doubted whether of science, is frequently cited in this context for
cigarette smoking patterns could provide a suf- his claim that there are more forms of determi-
ficient explanation of lung cancer occurrence. nation than strict determinism.69 Even some
Instead, they sought a highly specific necessary advocates of determinism admit that while
cause (a molecule within cigarette smoke) that probabilistic causal models are not wholly
would provide a one to one correlation with accurate, they can provide quick and dirty
disease outcome, thus providing a seemingly descriptions of phenomena that may be
more complete explanation.64 Lurking behind useful.19 Thus, Rothman suggests that while an
this belief are the assumptions that lung cancer individual’s risk of disease at time T must be
occurrence can be fully reduced to and either 1 or 0, probabilistic risk statements can
explained by causes at the molecular level and be viewed as statements about the likelihood of
that those causes behave deterministically. a suYcient cause being present.(page 589)26
Among biologists, however, this simple reduc- Yet this practical notion of causation tends to
tionism has been widely criticised.65 In con- be viewed as a lesser compromise to the more
trast, Susser and Susser have recently called for robust “scientific” notion of cause because of a
epidemiological theory that unifies disparate lingering assumption of determinism.70 The
phenomena, from causal pathways at the soci- unfortunate consequence is that the claims of
etal level to pathogenesis at the molecular level; epidemiology seem to be practical but not truly
these are multi-level explanations.66 scientific; sceptics of epidemiological methods
Probabilistic causes are often charged with might wonder, if epidemiologists use an unsci-
oVering only incomplete explanations because entific definition of causation, how can they be
they retain an irreducible element of chance. expected to identify genuine causes? Drawing a
Mid-20th century philosophers of science distinction between a “scientific” and a “prac-
argued that unlikely events could not be tical” definition of causation further contrib-
explained.67 But more recent scholars have utes to a longstanding tension within the disci-
rejected that view, urging that probabilistic pline of epidemiology between allegiance to
causes oVer partial explanations for particular science and allegiance to public health. Some
outcomes.40 41 68 This view more closely mirrors recent commentators have responded to this
epidemiological thinking. However, in this tension by restraining the boundaries of epide-
context it is important, again, to distinguish miology and maintaining that the discipline
between a satisfactory causal model and a best contributes to public health by maintain-
definition of causation. In developing causal ing scientific rigour and separating science
models scientists sometimes do seek to reduce from public health policy.71
the influence of chance as much as possible, as However, we maintain that a strong and
it may be inversely related to the model’s eVective definition of causation for epidemiol-
explanatory power. While the claim that smok- ogy must be consistent with the aims of both
ing causes lung cancer explains diVerences in science and public health. Although epidemi-
lung cancer rates between groups, it does not ologists may debate their proper role in public
explain why some smokers develop lung cancer health, it is clear that epidemiology cannot be
and other do not. There is still hope that the divorced from the application of its findings.
latter can eventually be fully explained by While we argued earlier that a probabilistic
deterministic molecular mechanisms. We have view of causation is consistent with modern
no objection to the use of deterministic models theories of biological science, here we note that
where appropriate. But a definition of causa- it also has some distinct advantages for the
tion must allow for the possibility that chance is application of causal knowledge in practical
inherent in some natural processes. The proba- public health eVorts.
bilistic definition of causation allows for For public health it is often not enough to
construction of both deterministic and proba- know that “x is part of a constellation of condi-
bilistic models, which are essential to both the tions suYcient for y.” Public health eVorts tar-
biological and social sciences. get particular causes chosen for practical
reasons,72 rather than targeting Mill’s total
Practical versus scientific goals of constellation of causes. Moreover, quantifying
epidemiology the relative contribution of diVerent types of
A number of commentators have drawn a dis- causes is important, but suYcient-component
tinction between a strict “scientific” or “logi- causes and necessary causes oVer little help
cal” definition of causation and a more flexible here. Rothman and Greenland note that it can
practical definition.9 12 37 49 They do so because be said for many diseases that they are 100%
they believe that the aims of science and public caused by genetics and 100% by environment,
health are diVerent. The primary aim of public because the appropriate suYcient causes con-
health is “to intervene to reduce morbidity and tain both genetic and environmental compo-
mortality from the disease.”(page 3)2 In nents.29 The influence of a necessary or
contrast, the primary aim of science is, as com- suYcient cause cannot be measured in degrees;
monly stated, to explain the world; such inves- its eVect is all or nothing. A probabilistic view
tigations may or may not lead to eVective pub- of causation, however, explains how diVerent
lic health strategies. causes can exert diVerent degrees of influence
Thus, some have recommended that epide- on an eVect by referring to the amount by
miologists abandon the traditional scientific which each contributing cause increases the
concept of causes as necessary and suYcient probability of the eVect.
conditions in favour of a broader concept with Moreover, a probabilistic definition is al-
more practical value.9 57 Bunge, the philosopher ready implicit in practical reasoning about
Downloaded from on 21 November 2008

Causation in epidemiology 911

causes in epidemiology and public health. nature of causation. Unfortunately, philosophi-

What does a physician mean when she tells her cal thinking about causation has been largely
patient that he can reduce his risk of developing driven by the physical sciences, focusing on
lung cancer by giving up smoking? Does she simple chains of events rather than the complex
mean that he might be one of those individuals multi-level relations that make up biology.
for whom smoking might be the final compo- Thus, this is an area that needs further
nent in a suYcient cause for lung cancer? The research. How are explanations at diVerent lev-
implication here is that if he only knew the sta- els, from the molecular to the social, related?
tus of the other component causes (that is, does And further inquiry is needed into the relation
he carry the right form of gene x?) he might not between concepts of causation and causal
have to give up smoking after all. But this view inference. Guidelines for drawing causal infer-
trivialises the nature of public health advice. In ences may diVer according to how causation is
practice, the clinician means that giving up defined. Because causal inference is at the heart
smoking will actually lower that individual’s of what epidemiologists do, the theoretical
probability of developing lung cancer—not study of causation has important consequences
lower it from 1 to 0, but lower it significantly. for the everyday practice of epidemiology.
The deterministic assumption of hidden eVect
modifiers is thus more likely to thwart than to The authors are grateful to Ellen Velie, Steve Goodman,
help public health eVorts. The underlying defi- Mervyn Susser, Michael S Kramer, Karen Goodman, Fred
nition of causation used is important for other Suppe, and two anonymous referees for discussions and
comments on earlier versions of this manuscript.
applications of epidemiological findings as
well, such as in toxic tort cases; in these cases,
Funding: none.
an implicit assumption of deterministic causa-
Conflicts of interest: none.
tion has led some courts to apply unreasonable
criteria for evaluating epidemiologic evi-
dence.73 74 1 Olsen J. Some consequences of adopting a conditional
deterministic causal model in epidemiology. Eur J Public
Health 1993;3:204–9.
2 MacMahon B, Pugh T. Epidemiology: principles and methods.
Conclusion Boston: Little, Brown, 1970.
3 Kramer MS, Lane DA. Causal propositions in clinical
We have drawn on both epidemiological and research and practice. J Clin Epidemiol 1992;45:639–49.
philosophical literature on causation, but phi- 4 Krieger N, Zierler S. The need for epidemiologic theory.
Epidemiology 1997;8:212–14.
losophers and epidemiologists, of course, pursue 5 Susser M, Susser E. Choosing a future for epidemiology: I.
very diVerent objects of study. While philoso- Eras and paradigms. Am J Public Health 1996;86:668–73.
phers are interested in general principles of cau- 6 Nijhuis HGJ, Van Der Maesen LJG. The philosophical
foundations of public health: an invitation to debate. J Epi-
sation, epidemiologists tend to be interested in demiol Community Health 1994;48:1–3.
particular examples of causal relations. Thus, 7 MacMahon B, Pugh TF, Ipsen J. Epidemiologic methods.
Boston: Little, Brown, 1960.
philosophers seek definitions while epidemiolo- 8 Krieger N. Epidemiology and the web of causation: has any-
gists construct causal models. However, we have one seen the spider? Soc Sci Med 1994;39:887–903.
9 Susser M. What is a cause and how do we know one? Am J
argued that an epidemiologist’s beliefs and Epidemiol 1991;133:635–48.
assumptions about cause carry practical conse- 10 Hume D. An enquiry concerning human understanding.
Indianapolis: Hackett, 1977.
quences, because causal models are constructed 11 Russell B. On the notion of cause. In: Mysticism and logic.
within limits set by that definition the epidemi- New York: W W Norton, 1929.
12 Karhausen LR. The logic of causation in epidemiology.
ologist has in mind (whether explicit or im- Scand J Soc Med 1996;24:8–13.
plicit). Thus, if a deterministic definition of cau- 13 Renton A. Epidemiology and causation: a realist view. J Epi-
demiol Community Health 1994;48:79–85.
sation is assumed, only deterministic models will 14 Gordis L. Epidemiology. 2nd ed. Philadelphia: W B
be recognised as causal. Because causation is so Saunders, 2000.
15 Elwood M. Causal relationships in medicine: a practical system
central to what epidemiologists do, we argue for critical appraisal. Oxford: Oxford University Press, 1988.
that it is crucial that epidemiologists understand 16 Timmreck TC. An introduction to epidemiology. Boston: Jones
and Bartlett, 1994.
the implications of applying a particular defini- 17 Last JMA. Dictionary of epidemiology, 2nd ed. New York:
tion of causation. Oxford University Press, 1988.
18 Lilienfeld DE, Stolley PD. Foundations of epidemiology. 3rd
We argue for using a probabilistic definition ed. New York: Oxford University Press, 1994.
of causation because it is more inclusive than 19 Kleinbaum DG, Kupper LL, Morgenstern HL. Epidemio-
logic research: principles and quantitative methods. Belmont,
the necessary and suYcient and suYcient- CA: Lifetime Learning Publications, 1982.
component cause definitions. Deterministic 20 Susser M. Causal thinking in the health sciences: concepts and
strategies of epidemiology. New York: Oxford University
models can be treated as extreme cases of the Press, 1973.
probabilistic definition; thus, the probabilistic 21 Stehbens WE. The concept of cause in disease. J Chronic Dis
definition incorporates the concept of neces- 22 Charlton BG. Attribution of causation in epidemiology:
sary and suYcient causes. In addition, there are chain or mosaic? J Clin Epidemiol 1996;49:105–7.
23 Carter KC. Koch’s postulates in relation to the work of
some concepts, while insuYcient for defining Jacob Henle and Edwin Klebs. Med Hist 1985;29:353–74.
causation themselves, that cannot be dispensed 24 Bernard C. An introduction to the study of experimental medi-
cine. (Translated by Henry Copley Greene). New York:
with in any causal definition—counterfactuals, Dover, 1957.
production, and temporality. Additionally, the 25 Evans AS. Causation and disease: a chronological journey. New
York: Plenum, 1993.
probabilistic definition does not require as- 26 Rothman KJ. Causes. Am J Epidemiol 1976;104:587–92.
sumptions about hidden biological mecha- 27 Mill JS. Philosophy of scientific method. New York: Hafner,
1950 [1843].
nisms (that observed probabilistic phenomena 28 Mackie JL. The cement of the universe: a study of causation.
always reduce to some underlying deterministic Oxford: Clarendon Press, 1974.
29 Rothman K, Greenland S. Modern epidemiology. 2nd ed.
mechanism). Philadelphia: Lippincott Williams and Wilkins, 1998.
Nevertheless, a definition is not itself a 30 Fisher RA. Indeterminism and natural selection. Philosophy
of Science 1934;1:99–117.
theory of causation; there is more conceptual 31 Mayr E. Cause and eVect in biology. Science 1961;134:
work to be done towards further specifying the 1501–6.
Downloaded from on 21 November 2008

912 Parascandola, Weed

32 Koopman JS, Weed DL. Epigenesis theory: a mathematical 54 Peto R. The need for ignorance in cancer research. In: Dun-
model relating causal concepts of pathogenesis in individu- can R, Weston-Smith M, eds. The encyclopedia of medical
als to disease patterns in populations. Am J Epidemiol 1990; ignorance. Oxford: Pergamon Press, 1984:129–33.
132:366–90. 55 Vandenbroucke JP. Is “The causes of cancer” a miasma
33 Hogan MD, Kuper LL, Most BM, et al. Alternatives to theory for the end of the twentieth century? Int J Epidemiol
Rothman’s approach for assessing synergism (or antago- 1988;17:708–9.
nism) in cohort studies. Am J Epidemiol 1978;108:60–7. 56 Skrabanek P. The emptiness of the black box. Epidemiology
34 Koopman JS. Causal models and sources of interaction. Am 1994;5:553–5.
J Epidemiol 1977;106:439–44.
35 Olsen J. Causes and prevention. Scand J Soc Med 57 Diez-Roux AV. Bringing context back into epidemiology:
1991;19:1–6. variables and fallacies in multilevel analysis. Am J Public
36 Schlesselman JJ. “Proof ” of cause and eVect in epidemio- Health 1998;88:216–22.
logic studies: criteria for judgement. Prev Med 1987;16: 58 McMichael AJ. Prisoners of the proximate: loosening the
195–210. constraints on epidemiology in an age of change. Am J Epi-
37 International Agency for Research on Cancer. Cancer: demiol 1999;149:887–97.
causes, occurrences and control. IARC Scientific Publications 59 Fullilove RE, Edgoose JC, Fullilove MT. Chaos, criticality,
no. 100. Lyon, France: International Agency for Research and public health. J Natl Med Assoc 1997;89:311–16.
on Cancer, 1990. 60 McMichael AJ. Prisoners of the proximate: loosening the
38 Suppes P. A probabilistic theory of causality. Amsterdam: constraints on epidemiology in an age of change. Am J Epi-
North-Holland Publishing Company, 1970. demiol 1999;149:887–97.
39 Salmon WC. Scientific explanation and the causal structure of 61 Cohen D, Spear S, Scribner R, et al. “Broken windows” and
the world. Princeton: Princeton University Press, 1984. the risk of gonorrhea. Am J Public Health 2000;90:230–6.
40 Humphreys P. The chances of explanation: causal explanation 62 Koopman JS, Lynch JW. Individual causal models and
in the social, medical, and physical sciences. Princeton: Princ- population system models in epidemiology. Am J Public
eton University Press, 1989. Health 1999;89:1170–4.
41 Mellor DH. The facts of causation. London: Routledge, 1995.
42 Eells E. Probabilistic causality. Cambridge: Cambridge 63 Stout D. Direct link found between smoking and cancer.
University Press, 1991. The New York Times 1996 Oct 18; section A: 1 (col 3).
43 Cartwright N. Nature’s capacities and their measurement. 64 Yerushalmy J, Palmer CE. On the methodology of investiga-
Oxford: Clarendon Press, 1989. tions of etiologic factors in chronic diseases. J Chronic Dis
44 Lewis D. B. Chancy causation. Postscript to causation. In: 1959;10:27–40.
Lewis D. Philosophical papers. Vol II. New York: Oxford 65 Mayr E. The growth of biological thought: diversity, evolution,
University Press, 1986. and inheritance. Cambridge, MA: Belknap Press, 1982.
45 Cox DR. Causality: some statistical aspects. J R Stat Soc A 66 Susser M, Susser E. Choosing a future for epidemiology: II.
1992;155:291–301. From black box to Chinese boxes and eco-epidemiology.
46 Holland PW. Statistics and causal inference. J Am Stat Assoc Am J Public Health 1996;86:674–7.
1986;81:945–60. 67 Hempel CG. Aspects of scientific explanation and other essays
47 Rubin DB. Estimating causal eVects of treatments in rand- in the philosophy of science. New York: Free Press, 1965.
omized and nonrandomized studies. Journal of Educational 68 SchaVner KF. Discovery and explanation in biology and medi-
Psychology 1974;66:688–701. cine. Chicago: University of Chicago Press, 1994.
48 Maldonado G, Greenland, S. Estimating causal eVects. 69 Bunge M. Causality and modern science. 3rd ed. New York:
[abstract]. Presented at the Society for Epidemiologic Dover, 1959.
Research 31st Annual Meeting, Chicago, Illinois, June, 70 Parascandola M. Epidemiology: second rate science? Public
1998. Am J Epidemiol 1998;147:S80. Health Rep 1998;113:312–23.
49 Riegelman R. Contributory cause: unnecessary and insuY-
cient. Postgrad Med 1979;66:177–9. 71 Savitz DA, Poole C, Miller WC. Reassessing the role of epi-
50 Pearl J. Causality: models, reasoning, and inference. Cambridge: demiology in public health. Am J Public Health 1999;89:
Cambridge University Press, 2000. 1158–61.
51 Cook TD, Campbell DT. Quasi-experimentation: design and 72 Hart HLA, Honoré T. Causation in the law. 2nd ed. Oxford:
analysis issues for field settings. Boston: Houghton MiZin, Clarendon Press, 1985.
1979. 73 Parascandola M. What is wrong with the probability of cau-
52 Weed DL. Beyond black box epidemiology. Am J Public sation? Jurimetrics 1998;39:29–44.
Health 1998;88:12–14. 74 Beyea J, Greenland S. The importance of specifying the
53 Savitz DA. In defense of black box epidemiology. Epidemiol- underlying biologic model in estimating the probability of
ogy 1994;5:550–2. causation. Health Phys 1999;76:269–74.