r351

Causality in the Sciences

Phyllis McKay Illari

University of Kent

Federica Russo
Université catholique de Louvain, University of Kent

Jon Williamson
University of Kent

CLARENDON PRESS . OXFORD

2009
 1
THE STRUCTURAL THEORY OF CAUSATION
JUDEA PEARL

Abstract
This paper presents a general theory of causation based on the Structural
Causal Model (SCM) described in (Pearl, 2000a). The theory subsumes
and unifies current approaches to causation, including graphical, poten-
tial outcome, probabilistic, decision analytical, and structural equation
models, and provides both a mathematical foundation and a friendly cal-
culus for the analysis of causes and counterfactuals. In particular, the
paper demonstrates how the theory engenders a coherent methodology
for inferring (from a combination of data and assumptions) answers to
three types of causal queries: (1) queries about the effects of potential
interventions, (2) queries about probabilities of counterfactuals, and (3)
queries about direct and indirect effects.
Keywords: Structural equation models, confounding, graphical methods, coun-
terfactuals, causal effects, potential-outcome, probabilistic causation.

1.1 Introduction
Twentieth-century science has witnessed a lingering tension between the ques-
tions that researchers wish to ask and the language in which they were trained
– statistics.
The research questions that motivate most studies in the health, social and
behavioral sciences are not statistical but causal in nature. For example, what
is the efficacy of a given drug in a given population? Whether data can prove
an employer guilty of hiring discrimination? What fraction of past crimes could
have been avoided by a given policy? What was the cause of death of a given
individual, in a specific incident? These are causal questions because they require
some knowledge of the data-generating process; they cannot be computed from
distributions alone.
Any conception of causation worthy of the title “theory” must be able to (1)
represent these questions in a formal language, (2) provide a precise language
for communicating assumptions under which the questions need to be answered,
(3) provide a systematic way of answering at least some of these questions and
labeling others “unanswerable,” and (4) provide a method of determining what
assumptions or new measurements would be needed to answer the “unanswer-
able” questions.1
1 For example, a theory may conclude that the information at hand is not sufficient for

determining the efficacy of a drug unless certain assumptions are deemed plausible, or unless
2 The Structural Theory of Causation Judea Pearl

A “general theory” of causation should do more. In addition to embracing

all questions judged to have causal character, a general theory must also sub-
sume any other theory or method that scientists have found useful in exploring
the various aspects of causation, be they epistemic, methodological or practical.
In other words, any alternative theory need to evolve as a special case of the
“general theory” when restrictions are imposed on either the model, the type of
assumptions admitted, or the language in which those assumptions are cast.
This paper presents a theory of causation that satisfies the criteria above.
It is based on the Structural Causal Model (SCM) developed in (Pearl, 1995,
2000a) which combines features of the structural equation models (SEM) used
in economics (Haavelmo, 1943) and social science (Duncan, 1975), the potential-
outcome notation of Neyman (1923) and Rubin (1974), and the graphical models
developed for probabilistic reasoning (Pearl, 1988; Lauritzen, 1996) and causal
analysis (Spirtes et al., 2000; Pearl, 2000a). The theory presented forms a coher-
ent whole that supercedes the sum of its parts.
Although the basic elements of SCM were introduced in the mid 1990’s (Pearl,
1995), and have been adapted warmly by epidemiologists (Greenland et al., 1999;
Glymour and Greenland, 2008), statisticians (Cox and Wermuth, 2004; Lau-
ritzen, 2001), and social scientists (Morgan and Winship, 2007), its potentials
as a comprehensive theory of causation are yet to be fully utilized. Some have
congratulated the SCM for generalizing econometric models from linear to non-
parametric analysis (Heckman, 2008), some have marveled at the clarity and
transparency of the graphical representation (Greenland and Brumback, 2002),
others praised the flexibility of the do(x) operator (Hitchcock, 2001; Lindley,
2002; Woodward, 2003) and, naturally, many have used the SCM to weed out
myths and misconceptions from outdated traditions (Meek and Glymour, 1994;
Greenland et al., 1999; Cole and Hernán, 2002; Arah, 2008; Shrier, 2009; Pearl,
2009b) Still, the more profound contributions of SCM, those stemming from its
role as a comprehensive theory of causation, have not been fully explicated. These
include:
1. The unification of the graphical, potential outcome, structural equations,
decision analytical (Dawid, 2002), interventional (Woodward, 2003), suffi-
cient component (Rothman, 1976) and probabilistic approaches to causa-
tion; with each approach viewed as a restricted special aspect of the SCM.
2. The axiomatization and algorithmization of counterfactual expressions.
3. Defining and identifying joint probabilities of counterfactual statements.
4. Reducing the evaluation of actions and policies to algorithmic level of anal-
ysis.
5. Solidifying the mathematical foundations of the potential-outcome model,
and formulating the counterfactual foundations of structural equation mod-
els.

data from a specific experimental study were made available. Such conclusion constitutes a
valid “solution,” provided no better solution exists.
 From Statistical to Causal Analysis: Distinctions and Barriers 3

6. Demystifying enigmatic notions such as “confounding,” “ignorability,” “ex-

changeability,” “superexogeneity” and others, which have emerged from
“black-box” approaches to causation.
7. Providing a transparent language for communicating causal assumptions
and defining causal problems.
This paper presents the main features of the structural theory by, first, con-
trasting causal analysis with standard statistical analysis (Section 1.2), second,
presenting a friendly formalism for counterfactual analysis, within which most
(if not all) causal questions can be formulated and resolved (Section 1.3 and
1.4) and, finally, contrasting the structural theory with two other frameworks:
probabilistic causation (Section 1.5) and the Neyman-Rubin potential-outcome
model (Section 1.6). The analysis will be demonstrated by attending to three
types of queries: (1) queries about the effect of potential interventions, (Section
1.3.1 and 1.3.2), (2) queries about counterfactuals (Section 1.3.3) and (3) queries
about direct and indirect effects (Section 1.4).

1.2 From Statistical to Causal Analysis: Distinctions and Barriers

1.2.1 The Basic Distinction: Coping With Change
The aim of standard statistical analysis, typified by regression, estimation, and
hypothesis testing techniques, is to assess parameters of a distribution from sam-
ples drawn of that distribution. With the help of such parameters, one can infer
associations among variables, estimate the likelihood of past and future events,
as well as update the likelihood of events in light of new evidence or new mea-
surements. These tasks are managed well by standard statistical analysis so long
as experimental conditions remain the same. Causal analysis goes one step fur-
ther; its aim is to infer not only the likelihood of events under static conditions,
but also the dynamics of events under changing conditions, for example, changes
induced by treatments or external interventions.
This distinction implies that causal and statistical concepts do not mix. There
is nothing in the joint distribution of symptoms and diseases to tell us that
curing the former would or would not cure the latter. More generally, there is
nothing in a distribution function to tell us how that distribution would differ
if external conditions were to change—say from observational to experimental
setup—because the laws of probability theory do not dictate how one property
of a distribution ought to change when another property is modified. This in-
formation must be provided by causal assumptions which identify relationships
that remain invariant when external conditions change.
These considerations imply that the slogan “correlation does not imply cau-
sation” can be translated into a useful principle: one cannot substantiate causal
claims from associations alone, even at the population level—behind every causal
conclusion there must lie some causal assumption that is not testable in obser-
4 The Structural Theory of Causation Judea Pearl

vational studies.2

1.2.2 Formulating the Basic Distinction

A useful demarcation line that makes the distinction between associational and
causal concepts crisp and easy to apply, can be formulated as follows. An as-
sociational concept is any relationship that can be defined in terms of a joint
distribution of observed variables, and a causal concept is any relationship that
cannot be defined from the distribution alone. Examples of associational con-
cepts are: correlation, regression, dependence, conditional independence, likeli-
hood, collapsibility, propensity score, risk ratio, odd ratio, marginalization, con-
ditionalization, “controlling for,” and so on. Examples of causal concepts are:
randomization, influence, effect, confounding, “holding constant,” disturbance,
spurious correlation, faithfulness/stability, instrumental variables, intervention,
explanation, attribution, and so on. The former can, while the latter cannot be
defined in term of distribution functions.
This demarcation line is extremely useful in causal analysis for it helps in-
vestigators to trace the assumptions that are needed for substantiating various
types of scientific claims. Every claim invoking causal concepts must rely on some
premises that invoke such concepts; it cannot be inferred from, or even defined
in terms statistical associations alone.

1.2.3 Ramifications of the Basic Distinction

This principle has far reaching consequences that are not generally recognized in
the standard statistical literature. Many researchers, for example, are still con-
vinced that confounding is solidly founded in standard, frequentist statistics, and
that it can be given an associational definition saying (roughly): “U is a potential
confounder for examining the effect of treatment X on outcome Y when both U
and X and U and Y are not independent.” That this definition and all its many
variants must fail (Pearl, 2000a, Section 6.2)3 is obvious from the demarcation
line above; if confounding were definable in terms of statistical associations, we
would have been able to identify confounders from features of nonexperimen-
tal data, adjust for those confounders and obtain unbiased estimates of causal
effects. This would have violated our golden rule: behind any causal conclu-
sion there must be some causal assumption, untested in observational studies.
Hence the definition must be false. Therefore, to the bitter disappointment of
generations of epidemiologist and social science researchers, confounding bias
cannot be detected or corrected by statistical methods alone; one must make
some judgmental assumptions regarding causal relationships in the problem be-

2 Theassumption of “faithfulness” or “stability” as defined in the “causal discovery” liter-

ature (Spirtes et al. 2000; Pearl 2000a, Chapter 2) is likewise a causal assumption, albeit a
genetic one, for it restricts any causal model from generating data that hide the structure of
the model (e.g., by cancellation).
3 For example, any intermediate variable U on a causal path from X to Y satisfies this

definition, without confounding the effect of X on Y .

 From Statistical to Causal Analysis: Distinctions and Barriers 5

fore an adjustment (e.g., by stratification) can safely correct for confounding

bias.
Another ramification of the sharp distinction between associational and causal
concepts is that any mathematical approach to causal analysis must acquire new
notation for expressing causal relations – probability calculus is insufficient. To
illustrate, the syntax of probability calculus does not permit us to express the
simple fact that “symptoms do not cause diseases”, let alone draw mathematical
conclusions from such facts. All we can say is that two events are dependent—
meaning that if we find one, we can expect to encounter the other, but we can-
not distinguish statistical dependence, quantified by the conditional probability
P (disease|symptom) from causal dependence, for which we have no expression in
standard probability calculus. Scientists seeking to express causal relationships
must therefore supplement the language of probability with a vocabulary for
causality, one in which the symbolic representation for the relation “symptoms
cause disease” is distinct from the symbolic representation of “symptoms are
associated with disease.”

1.2.4 Two Mental Barriers: Untested Assumptions and New Notation

The preceding two requirements: (1) to commence causal analysis with untested,4
theoretically or judgmentally based assumptions, and (2) to extend the syntax of
probability calculus, constitute the two main obstacles to the acceptance of causal
analysis among statisticians and among professionals with traditional training
in statistics.
Associational assumptions, even untested, are testable in principle, given suf-
ficiently large sample and sufficiently fine measurements. Causal assumptions, in
contrast, cannot be verified even in principle, unless one resorts to experimental
control. This difference stands out in Bayesian analysis. Though the priors that
Bayesians commonly assign to statistical parameters are untested quantities, the
sensitivity to these priors tends to diminish with increasing sample size. In con-
trast, sensitivity to prior causal assumptions, say that treatment does not change
gender, remains substantial regardless of sample size.
This makes it doubly important that the notation we use for expressing causal
assumptions be meaningful and unambiguous so that one can clearly judge the
plausibility or inevitability of the assumptions articulated. Statisticians can no
longer ignore the mental representation in which scientists store experiential
knowledge, since it is this representation, and the language used to access it that
determine the reliability of the judgments upon which the analysis so crucially
depends.
How does one recognize causal expressions in the statistical literature? Those
versed in the potential-outcome notation (Neyman, 1923; Rubin, 1974; Holland,
1988), can recognize such expressions through the subscripts that are attached
to counterfactual events and variables, e.g. Yx (u) or Zxy . (Some authors use

4 By “untested” I mean untested using frequency data in nonexperimental studies.

6 The Structural Theory of Causation Judea Pearl

parenthetical expressions, e.g. Y (0), Y (1), Y (x, u) or Z(x, y).) The expression
Yx (u), for example, stands for the value that outcome Y would take in indi-
vidual u, had treatment X been at level x. If u is chosen at random, Yx is a
random variable, and one can talk about the probability that Yx would attain
a value y in the population, written P (Yx = y) (see Section 1.6 for semantics).
Alternatively, (Pearl, 1995) used expressions of the form P (Y = y|set(X = x))
or P (Y = y|do(X = x)) to denote the probability (or frequency) that event
(Y = y) would occur if treatment condition X = x were enforced uniformly over
the population.5 Still a third notation that distinguishes causal expressions is
provided by graphical models, where the arrows convey causal directionality.6
However, few have taken seriously the textbook requirement that any intro-
duction of new notation must entail a systematic definition of the syntax and
semantics that governs the notation. Moreover, in the bulk of the statistical liter-
ature before 2000, causal claims rarely appear in the mathematics. They surface
only in the verbal interpretation that investigators occasionally attach to cer-
tain associations, and in the verbal description with which investigators justify
assumptions. For example, the assumption that a covariate not be affected by a
treatment, a necessary assumption for the control of confounding (Cox, 1958, p.
48), is expressed in plain English, not in a mathematical expression.
Remarkably, though the necessity of explicit causal notation is now recognized
by many academic scholars, the use of such notation has remained enigmatic to
most rank and file researchers, and its potentials still lay grossly underutilized in
the statistics based sciences. The reason for this, can be traced to the unfriendly
semi-formal way in which causal analysis has been presented to the research com-
munity, resting primarily on the restricted paradigm of controlled randomized
trials advanced by (Rubin, 1974).
The next section provides a conceptualization that overcomes these mental
barriers; it offers both a friendly mathematical machinery for cause-effect analysis
and a formal foundation for counterfactual analysis.

1.3 Structural Causal Models (SCM) and The Language of

Diagrams
1.3.1 Semantics: Causal Effects and Counterfactuals
How can one express mathematically the common understanding that symptoms
do not cause diseases? The earliest attempt to formulate such relationship math-
ematically was made in the 1920’s by the geneticist Sewall Wright (1921), who
5 Clearly, P (Y = y|do(X = x)) is equivalent to P (Y = y), This is what we normally assess
x
in a controlled experiment, with X randomized, in which the distribution of Y is estimated for
each level x of X. Still, the former can be defined without resorting to counterfactual notation
(Pearl, 2000a, pp. 23-4) to the delight of those who prefer to deny mathematical notation to
any assertion that is not experimentally testable in isolation (Dawid, 2002).
6 These notational clues should be useful for detecting inadequate definitions of causal con-

cepts; any definition of confounding, randomization or instrumental variables that is cast in

standard probability expressions, void of graphs, counterfactual subscripts or do(∗) operators,
can safely be discarded as inadequate.
 Structural Causal Models (SCM) and The Language of Diagrams 7

used a combination of equations and graphs. For example, if X stands for a dis-
ease variable and Y stands for a certain symptom of the disease, Wright would
write a linear equation:
y = βx + uY (1.1)
where x stands for the level (or severity) of the disease, y stands for the level (or
severity) of the symptom, and uY stands for all factors, other than the disease in
question, that could possibly affect Y . In interpreting this equation one should
think of a physical process whereby Nature examines the values of x and uY and,
accordingly, assigns variable Y the value y = βx + uY . Similarly, to “explain”
the occurrence of disease X, one could write x = uX , where UX stands for all
factors affecting X.
To express the directionality inherent in this process, Wright augmented the
equation with a diagram, later called “path diagram,” in which arrows are drawn
from (perceived) causes to their (perceived) effects and, more importantly, the
absence of an arrow makes the empirical claim that Nature assigns values to
one variable while ignoring the other. In Figure 1.1, for example, the absence
of arrow from Y to X represent the claim that symptom Y is not among the
factors UX which affect disease X.
The variables UX and UY are called “exogenous”; they represent observed or
unobserved background factors that the modeler decides to keep unexplained,
that is, factors that influence but are not influenced by the other variables (called
“endogenous”) in the model.
If correlation is judged possible between two exogenous variables, UY and
UX , it is customary to connect them by a dashed double arrow, as shown in
Figure 1.1(b).

UX UY UX UY
x = uX
y = βx + uY
X β Y X β Y

(a) (b)

Fig. 1.1. A simple structural equation model, and its associated diagrams. Un-
observed exogenous variables are connected by dashed arrows.

To summarize, path diagrams encode causal assumptions via missing arrows,

representing claims of zero influence, and missing double arrows (e.g., between
UX and UY ), representing the assumption Cov(UY , UX )=0. Note that, despite its
innocent appearance in associational vocabulary, the latter assumption is causal,
not statistical, for it cannot be confirmed or denied from the joint distribution
of observed variables, in case the U ’s are unobservable.
The generalization to nonlinear systems of equations is straightforward. For
example, the non-parametric interpretation of the diagram of Figure 1.2(a) cor-
8 The Structural Theory of Causation Judea Pearl

U U U U U U
Z X Y Z X Y
x0

Z X Y Z X Y

(a) (b)

Fig. 1.2. (a) The diagram associated with the structural model of equation
(1.2). (b) The diagram associated with the modified model, Mx0 , of equation
(1.3), representing the intervention do(X = x0 ).

responds to a set of three functions, each corresponding to one of the observed

variables:

z = fZ (uZ )
x = fX (z, uX ) (1.2)
y = fY (x, uY )

where UZ , UX and UY are assumed to be jointly independent but, otherwise,

arbitrarily distributed.
Remarkably, unknown to most economists and pre-2000 philosophers,7 struc-
tural equation models provide a formal interpretation and symbolic machinery
for analyzing counterfactual relationships of the type: “Y would be y had X been
x in situation U = u,” denoted Yx (u) = y. Here U represents the vector of all
exogenous variables.8
The key idea is to interpret the phrase “had X been x0 ” as an instruction
to modify the original model and replace the equation for X by a constant x0 ,
yielding the sub-model.

z = fZ (uZ )
x = x0 (1.3)
y = fY (x, uY )

the graphical description of which is shown in Figure 1.2(b).

This replacement permits the constant x0 to differ from the actual value of
X (namely fX (z, uX )) without rendering the system of equations inconsistent,
thus yielding a formal interpretation of counterfactuals in multi-stage models,
7 Connections between structural equations and a restricted class of counterfactuals were

recognized by (Simon and Rescher, 1966). These were generalized by (Balke and Pearl, 1995)
who used modified models to permit counterfactual conditioning on dependent variables. This
development seems to have escaped Collins et al. (2004).
8 Because U = u may contain detailed information about a situation or an individual, Y (u)
x
is related to what philosophers called “token causation,” while P (Yx = y|Z = z) characterizes
“Type causation,” that is, the tendency of X to influence Y in a sub-population characterized
by Z = z.
 Structural Causal Models (SCM) and The Language of Diagrams 9

where the dependent variable in one equation may be an independent variable in

another (Balke and Pearl, 1994ab; Pearl 2000b). For example, to compute E(Yx0 ),
the expected effect of setting X to x0 , (also called the average causal effect of
X on Y , denoted E(Y |do(x0 )) or, generically, E(Y |do(x))), we solve equation
(1.3) for Y in terms of the exogenous variables, yielding Yx0 = fY (x0 , uY ), and
average over UY . It is easy to show that in this simple system, the answer can be
obtained without knowing the form of the function fY (x, uY ) or the distribution
P (uY ). The answer is given by:

E(Yx0 ) = E(Y |do(X = x0 ) = E(Y |x0 )

which is computable from the distribution P (x, y, z), hence estimable from ob-
served samples of P (x, y, z). This result hinges on the assumption that UZ , UX ,
and UY are mutually independent and on the topology of the graph (e.g., that
there is no direct arrow from Z to Y .)
In general, it can be shown (Pearl, 2000a, Chapter 3) that, whenever the
graph is Markovian (i.e., acyclic with independent exogenous variables) the post-
interventional distribution P (Y = y|do(X = x)) is given by the following expres-
sion:
X
P (Y = y|do(X = x)) = P (y|t, x)P (t) (1.4)
t

where T is the set of direct causes of X (also called “parents”) in the graph.
Again, we see that all factors on the right hand side are estimable from the
distribution P of observed variables and, hence, the counterfactual probability
P (Yx = y) is estimable with mere partial knowledge of the generating process
– the topology of the graph and independence of the exogenous variables is all
that is needed.
When some variables in the graph (e.g., the parents of X) are unobserved,
we may not be able to learn (or “identify” as it is called) the post-intervention
distribution P (y|do(x)) by simple conditioning, and more sophisticated meth-
ods would be required. Likewise, when the query of interest involves several
hypothetical worlds simultaneously, e.g., P (Yx = y, Yx′ = y′ )9 , the Markovian
assumption may not suffice for identification and additional assumptions, touch-
ing on the form of the data-generating functions (e.g., monotonicity) may need
to be invoked. These issues will be discussed in Sections 1.3.3 and 1.6.
This interpretation of counterfactuals, cast as solutions to modified systems of
equations, provides the conceptual and formal link between structural equation
models, used in economics and social science and the Neyman-Rubin potential-
outcome framework to be discussed in Section 1.6. But first we discuss two
long-standing problems that have been completely resolved in purely graphical
terms, without delving into algebraic techniques.

9 Read: The probability that Y would be y if X were x and y′ if X were x′ .

10 The Structural Theory of Causation Judea Pearl

1.3.2 Confounding and Causal Effect Estimation

The central target of most studies in the social and health sciences is the elu-
cidation of cause-effect relationships among variables of interests, for example,
treatments, policies, preconditions and outcomes. While good statisticians have
always known that the elucidation of causal relationships from observational
studies must be shaped by assumptions about how the data were generated,
the relative roles of assumptions and data, and ways of using those assumptions
to eliminate confounding bias have been a subject of much controversy. The
structural framework of Section 1.3.1 puts these controversies to rest.

Covariate Selection: The back-door criterion Consider an observational study

where we wish to find the effect of X on Y , for example, treatment on response,
and assume that the factors deemed relevant to the problem are structured as
in Figure 1.3; some are affecting the response, some are affecting the treatment

Z1
Z2
W1
Z3 W2

X
W3
Y

Fig. 1.3. Graphical model illustrating the back-door criterion for identifying
the causal effect of X on Y . Error terms are not shown explicitly.

and some are affecting both treatment and response. Some of these factors may
be unmeasurable, such as genetic trait or life style, others are measurable, such
as gender, age, and salary level. Our problem is to select a subset of these fac-
tors for measurement and adjustment, namely, that if we compare treated vs.
untreated subjects having the same values of the selected factors, we get the
correct treatment effect in that subpopulation of subjects. Such a set of factors
is called a “sufficient set” or a set “appropriate for adjustment”. The problem
of defining a sufficient set, let alone finding one, has baffled epidemiologists and
social science for decades (see (Greenland et al., 1999; Pearl, 1998; Pearl, 2003b)
for review).
The following criterion, named “back-door” in (Pearl, 1993a), settles this
problem by providing a graphical method of selecting a sufficient set of factors for
adjustment. It states that a set S is appropriate for adjustment if two conditions
hold:

1. No element of S is a descendant of X
 Structural Causal Models (SCM) and The Language of Diagrams 11

2. The elements of S “block” all “back-door” paths from X to Y , namely all

paths that end with an arrow pointing to X.10
Based on this criterion we see, for example, that the sets {Z1 , Z2 , Z3 }, {Z1 , Z3 },
and {W2 , Z3 }, each is sufficient for adjustment, because each blocks all back-door
paths between X and Y . The set {Z3 }, however, is not sufficient for adjustment
because, as explained above, it does not block the path X ← W1 ← Z1 → Z3 ←
Z2 → W2 → Y .
The implication of finding a sufficient set S is that, stratifying on S is guar-
anteed to remove all confounding bias relative the causal effect of X on Y . In
other words, it renders the causal effect of X on Y estimable, via

P (Y = y|do(X = x))
X
= P (Y = y|X = x, S = s)P (S = s) (1.5)
s

Since all factors on the right hand side of the equation are estimable (e.g., by
regression) from the pre-interventional data, the causal effect can likewise be
estimated from such data without bias.
The back-door criterion allows us to write equation (1.5) directly, after select-
ing a sufficient set S from the diagram, without resorting to any algebraic manip-
ulation. The selection criterion can be applied systematically to diagrams of any
size and shape, thus freeing analysts from judging whether “X is conditionally
ignorable given S,” a formidable mental task required in the potential-outcome
framework (Rosenbaum and Rubin, 1983). The criterion also enables the ana-
lyst to search for an optimal set of covariate—namely, a set S that minimizes
measurement cost or sampling variability (Tian et al., 1998).

General control of confounding Adjusting for covariates is only one of many

methods that permits us to estimate causal effects in nonexperimental studies. A
much more general identification criterion is provided by the following theorem:
Theorem 1.1 (Tian and Pearl 2002)
A sufficient condition for identifying the causal effect P (y|do(x)) is that every
path between X and any of its children traces at least one arrow emanating from
a measured variable.11
For example, if W3 is the only observed covariate in the model of Figure 1.3,
then there exists no sufficient set for adjustment (because no set of observed
covariates can block the paths from X to Y through Z3 ), yet P (y|do(x)) can
nevertheless be estimated since every path from X to W3 (the only child of X)
10 A set S of nodes is said to block a path p if either (i) p contains at least one arrow-emitting
node that is in S, or (ii) p contains at least one collision node that is outside S and has no
descendant in S. See (Pearl, 2000a, pp. 16-7). If S blocks all paths from X to Y it is said to
“d-separate X and Y .”
11 Before applying this criterion, one may delete from the causal graph all nodes that are not

ancestors of Y .
12 The Structural Theory of Causation Judea Pearl

traces either the arrow X → W3 , or the arrow W3 → Y , both emanating from a

measured variable (W3 ). In this example, the variable W3 acts as a “mediating
instrumental variable” (Pearl, 1993b; Chalak and White, 2006) and yields the
estimand:

P (Y = y|do(X = x))
X
= P (W3 = w3 |do(X = x))P (Y = y|do(W3 = w3 ))
w3
X X
= P (w3 |x) P (y|w3 , x′ )P (x′ ) (1.6)
w3 x′

More recent results extend this theorem by (1) presenting a necessary and suf-
ficient condition for identification (Shpitser and Pearl, 2006a), and (2) extending
the condition from causal effects to any counterfactual expression (Shpitser and
Pearl, 2007). The corresponding unbiased estimands for these causal quantities
are readable directly from the diagram.
The mathematical derivation of causal effect estimands, like equations (1.5)
and (1.6) is merely a first step toward computing quantitative estimates of those
effects from finite samples, using the rich traditions of statistical estimation and
machine learning. Although the estimands derived in (1.5) and (1.6) are non-
parametric, this does not mean that one should refrain from using paramet-
ric forms in the estimation phase of the study. For example, if the assump-
tions of Gaussian, zero-mean disturbances and additive interactions are deemed
reasonable, then the estimand given in (1.6) can be converted to the product
E(Y |do(x)) = rW3 X rY W3 ·X x, where rY Z·X is the (standardized) coefficient of Z
in the regression of Y on Z and X. More sophisticated estimation techniques can
be found in (Rosenbaum and Rubin, 1983), and (Robins, 1999). For example,
the “propensity score” method of (Rosenbaum and Rubin, 1983) was found to
be quite useful when the dimensionality of the adjusted covariates is high and
the data is sparse (see Pearl (2009a, pp. 348-42)).
It should be emphasized, however, that contrary to conventional wisdom (e.g.,
(Rubin, 2009)), propensity score methods are merely efficient estimators of the
right hand side of (1.5); they cannot be expected to reduce bias in case the set
S does not satisfy the back-door criterion (Pearl, 2009abc).

1.3.3 Counterfactual Analysis in Structural Models

Not all questions of causal character can be encoded in P (y|do(x)) type expres-
sions, in much the same way that not all causal questions can be answered from
experimental studies. For example, questions of attribution (also called “causes
of effects” (Dawid, 2000), e.g., I took an aspirin and my headache is gone, was it
due to the aspirin?) or of susceptibility (e.g., I am a healthy non-smoker, would
I be as healthy had I been a smoker?) cannot be answered from experimental
studies, and naturally, this kind of questions cannot be expressed in P (y|do(x))
 Structural Causal Models (SCM) and The Language of Diagrams 13

notation.12 To answer such questions, a probabilistic analysis of counterfactu-

als is required, one dedicated to the relation “Y would be y had X been x in
situation U = u,” denoted Yx (u) = y.
As noted in Section 1.3.1, the structural definition of counterfactuals involves
modified models, like Mx0 of equation (1.3), formed by the intervention do(X =
x0 ) (Figure 1.2(b)). Call the solution of Y in model Mx the potential response of
Y to x, and denote it by the symbol Yx (u). In general, then, the formal definition
of the counterfactual Yx (u) in SCM is given by (Pearl, 2000a, p. 98):
Yx (u) = YMx (u). (1.7)
The quantity Yx (u) can be given experimental interpretation; it stands for the
way an individual with characteristics (u) would respond, had the treatment been
x, rather than the treatment x = fX (u) actually received by that individual. In
our example, since Y does not depend on v and w, we can write:
Yx0 (uY , uX , uZ ) = Yx0 (uY ) = fY (x0 , uY ).
Clearly, the distribution P (uY , uX , uZ ) induces a well defined probability on the
counterfactual event Yx0 = y, as well as on joint counterfactual events, such as
‘Yx0 = y AND Yx1 = y′ ,’ which are, in principle, unobservable if x0 6= x1 . Thus,
to answer attributional questions, such as whether Y would be y1 if X were
x1 , given that in fact Y is y0 and X is x0 , we need to compute the conditional
probability P (Yx1 = y1 |Y = y0 , X = x0 ) which is well defined once we know the
forms of the structural equations and the distribution of the exogenous variables
in the model. For example, assuming linear equations (as in Figure 1.1),
x = uX , y = βx + uY ,
the conditions Y = y0 and X = x0 yield uX = x0 and uY = y0 − βx0 , and
we can conclude that, with probability one, Yx1 must take on the value: Yx1 =
βx1 + uY = β(x1 − x0 ) + y0 . In other words, if X were x1 instead of x0 , Y would
increase by β times the difference (x1 −x0 ). In nonlinear systems, the result would
also depend on the distribution of U and, for that reason, attributional queries
are generally not identifiable in nonparametric models (Pearl, 2000a, Chapter
9).
In general, if x and x′ are incompatible then Yx and Yx′ cannot be measured
simultaneously, and it may seem meaningless to attribute probability to the
joint statement “Y would be y if X = x and Y would be y′ if X = x′ .”13
12 The reason for this fundamental limitation is that no death case can be tested twice,

with and without treatment. For example, if we measure equal proportions of deaths in the
treatment and control groups, we cannot tell how many death cases are actually attributable
to the treatment itself; it is quite possible that many of those who died under treatment would
be alive if untreated and, simultaneously, many of those who survived with treatment would
have died if not treated.
13 For example, “The probability is 80% that Joe belongs to the class of patients who will be

cured if they take the drug and die otherwise.”

14 The Structural Theory of Causation Judea Pearl

Such concerns have been a source of objections to treating counterfactuals as

jointly distributed random variables (Dawid, 2000). The definition of Yx and Yx′
in terms of two distinct submodels neutralizes these objections (Pearl, 2000b),
since the contradictory joint statement is mapped into an ordinary event, one
where the background variables satisfy both statements simultaneously, each in
its own distinct submodel; such events have well defined probabilities.
The structural interpretation of counterfactuals also provides the concep-
tual and formal basis for the Neyman-Rubin potential-outcome framework, an
approach to causation that takes a controlled randomized trial (CRT) as its
starting paradigm, assuming that nothing is known to the experimenter about
the science behind the data. This “black-box” approach, which has thus far
been denied the benefits of graphical or structural analyses, was developed by
statisticians who found it difficult to cross the two mental barriers discussed in
Section 1.2.4. Section 1.6 establishes the precise relationship between the struc-
tural and potential-outcome paradigms, and outlines how the latter can benefit
from the richer representational power of the former.

1.4 Mediation: Direct and Indirect Effects

1.4.1 Direct versus Total Effects:
The causal effect we have analyzed so far, P (y|do(x)), measures the total effect of
a variable (or a set of variables) X on a response variable Y . In many cases, this
quantity does not adequately represent the target of investigation and attention
is focused instead on the direct effect of X on Y . The term “direct effect” is
meant to quantify an effect that is not mediated by other variables in the model
or, more accurately, the sensitivity of Y to changes in X while all other factors
in the analysis are held fixed. Naturally, holding those factors fixed would sever
all causal paths from X to Y with the exception of the direct link X → Y , which
is not intercepted by any intermediaries.
A classical example of the ubiquity of direct effects involves legal disputes
over race or sex discrimination in hiring. Here, neither the effect of sex or race
on applicants’ qualification nor the effect of qualification on hiring are targets
of litigation. Rather, defendants must prove that sex and race do not directly
influence hiring decisions, whatever indirect effects they might have on hiring by
way of applicant qualification.
Another example concerns the identification of neural pathways in the brain
or the structural features of protein-signaling networks in molecular biology
(Brent and Lok, 2005). Here, the decomposition of effects into their direct and
indirect components carries theoretical scientific importance, for it predicts be-
havior under a rich variety of hypothetical interventions.
In all such examples, the requirement of holding the mediating variables
fixed must be interpreted as (hypothetically) setting the intermediate variables
to constants by physical intervention, not by analytical means such as selection,
conditioning, or adjustment. For example, it will not be sufficient to measure the
association between gender (X) and hiring (Y ) for a given level of qualification Z,
 Mediation: Direct and Indirect Effects 15

because, by conditioning on the mediator Z, we may create spurious associations

between X and Y even when there is no direct effect of X on Y . This can easily
be illustrated in the model X → Z ← U → Y , where X has no direct effect
on Y . Physically holding Z constant should eliminate the association between
X and Y , as can be seen by deleting all arrows entering Z. But if we were to
condition on Z, a spurious association would be created through U (unobserved)
that might be construed as a direct effect of X on Y .
Using the do(x) notation, and focusing on differences of expectations, this
leads to a simple definition of controlled direct effect:
∆
CDE = E(Y |do(x′), do(z)) − E(Y |do(x), do(z))

or, equivalently, using counterfactual notation:

∆
CDE = E(Yx′ z ) − E(Yxz )

where Z is any set of mediating variables that intercept all indirect paths be-
tween X and Y . Graphical identification conditions for expressions of the type
E(Y |do(x), do(z1), do(z2 ), . . . , do(zk )) were derived by (Pearl and Robins, 1995)
(see Pearl (2000a, Chapter 4)) and invoke sequential application of the back-door
conditions discussed in Section 1.3.2.

1.4.2 Natural Direct Effects

In linear systems, the direct effect is fully specified by the path coefficient at-
tached to the link from X to Y ; therefore, the direct effect is independent of the
values at which we hold Z. In nonlinear systems, those values would, in general,
modify the effect of X on Y and thus should be chosen carefully to represent
the target policy under analysis. For example, it is not uncommon to find em-
ployers who prefer males for the high-paying jobs (i.e., high z) and females for
low-paying jobs (low z).
When the direct effect is sensitive to the levels at which we hold Z, it is often
meaningful to define the direct effect relative to some “natural” base-line level
that may vary from individual to individual and represents the level of Z just
before the change in X. Conceptually, we can define the average direct effect
DEx,x′ (Y ) as the expected change in Y induced by changing X from x to x′
while keeping all mediating factors constant at whatever value they would have
obtained under do(x). This hypothetical change, which (Robins and Greenland,
1991) called “pure” and (Pearl, 2001) called “natural,” mirrors what lawmakers
instruct us to consider in race or sex discrimination cases: “The central question
in any employment-discrimination case is whether the employer would have taken
the same action had the employee been of a different race (age, sex, religion,
national origin etc.) and everything else had been the same.” (In Carson versus
Bethlehem Steel Corp., 70 FEP Cases 921, 7th Cir. (1996)).
Extending the subscript notation to express nested counterfactuals (Pearl,
2001) gave the following definition for the “natural direct effect”:
16 The Structural Theory of Causation Judea Pearl

DEx,x′ (Y ) = E(Yx′ ,Zx ) − E(Yx ). (1.8)

Here, Yx′ ,Zx represents the value that Y would attain under the operation of
setting X to x′ and, simultaneously, setting Z to whatever value it would have
obtained under the original setting X = x. We see that DEx,x′ (Y ), the natu-
ral direct effect of the transition from x to x′ , involves probabilities of nested
counterfactuals and cannot be written in terms of the do(x) operator. Therefore,
the natural direct effect cannot in general be identified, even with the help of
ideal, controlled experiments, for we cannot rerun history and re-condition on
an action actually taken (see footnote 12). Pearl (2001) has nevertheless shown
that, if certain assumptions of “no confounding” are deemed valid,14 the natural
direct effect can be reduced to
X
DEx,x′ (Y ) = [E(Y |do(x′ , z)) − E(Y |do(x, z))]P (z|do(x)). (1.9)
z

The intuition is simple; the natural direct effect is the weighted average of the
controlled direct effect, using the causal effect P (z|do(x)) as a weighing function.
In particular, expression (1.9) is both valid and identifiable in Markovian
models, where each term on the right can be reduced to a “do-free” expression
using equation (1.4).

1.4.3 Natural Indirect Effects

Remarkably, the definition of the natural direct effect (1.8) can easily be turned
around and provide an operational definition for the indirect effect – a concept
shrouded in mystery and controversy, because it is impossible, using the do(x)
operator, to disable the direct link from X to Y so as to let X influence Y solely
via indirect paths.
The natural indirect effect, IE, of the transition from x to x′ is defined as the
expected change in Y affected by holding X constant, at X = x, and changing
Z to whatever value it would have attained had X been set to X = x′ . Formally,
this reads (Pearl, 2001):
∆
IEx,x′ (Y ) = E[(Yx,Zx′ ) − E(Yx )], (1.10)

which is almost identical to the direct effect (equation (1.8)) save for exchanging
x and x′ .
Indeed, it can be shown that, in general, the total effect T E of a transition
is equal to the difference between the direct effect of that transition and the
indirect effect of the reverse transition. Formally,
∆
T E x,x′ (Y ) = E(Yx′ − Yx ) = DEx,x′ (Y ) − IEx′ ,x (Y ). (1.11)

14 One sufficient condition is that Z ⊥

x ⊥Yx′ ,z |W holds for some set W of measured covariates.
See details and graphical criteria in (Pearl, 2001; Pearl, 2005) and in (Petersen et al., 2006).
 Structural versus Probabilistic Causality 17

In linear systems, where reversal of transitions amounts to negating the signs

of their effects, we have the standard additive formula

T Ex,x′ (Y ) = DEx,x′ (Y ) + IEx,x′ (Y ). (1.12)

Since each term above is based on an independent operational definition, this

equality constitutes a formal justification for the additive formula used routinely
in linear systems.
Note that, although it cannot be expressed in do-notation, the indirect effect
has clear policy-making implications. For example: in the hiring discrimination
context, a policy maker may be interested in predicting the gender mix in the
work force if gender bias is eliminated and all applicants are treated equally—
say, the same way that males are currently treated. This quantity will be given
by the indirect effect of gender on hiring, mediated by factors such as education
and aptitude, which may be gender-dependent.
More generally, a policy maker may be interested in the effect of issuing a
directive to a select set of subordinate employees, or in carefully controlling the
routing of messages in a network of interacting agents. Such applications motivate
the analysis of path-specific effects, that is, the effect of X on Y through a selected
set of paths (Avin et al., 2005).
Note that in all these cases, the policy intervention invokes the selection of
signals to be sensed, rather than variables to be fixed. (Pearl, 2001) has suggested
therefore that signal sensing is more fundamental to the notion of causation
than manipulation; the latter being but a crude way of stimulating the former
in experimental setup. The mantra “No causation without manipulation” must
be rejected. (See Pearl (2009a, Section 11.4.5.),
It is remarkable that counterfactual quantities like DE and ID that could not
be expressed in terms of do(x) operators, and appear therefore void of empirical
content, can, under certain conditions be estimated from empirical studies. A
general characterization of those conditions is given in (Shpitser and Pearl, 2007).
Additional examples of this “marvel of formal analysis” are given in (Pearl
2000a, Chapters 7 and 9; Pearl 2009a, Chapter 11). It constitutes an unassailable
argument in defense of counterfactual analysis, as expressed in (Pearl, 2000b)
against the stance of (Dawid, 2000).

1.5 Structural versus Probabilistic Causality

Probabilistic causality (PC) is a branch of philosophy that has attempted, for the
past several decades, to characterize the relationship between cause and effect
using the tools of probability theory (Hitchcock, 2003; Williamson, ming). Our
discussion of Section 1.2 rules out any such characterization and, not surprisingly,
the PC program is known mainly for the difficulties it has encountered, rather
than its achievements. This section explains the main obstacle that has kept PC
at bay for over half a century, and demonstrates how the structural theory of
causation clarifies relationships between probabilities and causes.
18 The Structural Theory of Causation Judea Pearl

1.5.1 The “Probability Raising” Trap

The idea that causes raise the probability of their effects has been the engine
behind most of PC explorations. It is a healthy idea, solidly ensconced in intu-
ition. We say, for example, “reckless driving causes accidents” or “you will fail
the course because of your laziness” (Suppes, 1970), knowing quite well that the
antecedents merely tend to make the consequences more likely, not absolutely
certain. One would expect, therefore, that probability raising should become the
defining characteristic of the relationship between a cause (C) and its effect (E).
Alas, though perfectly valid, this intuition cannot be expressed using the tools
of probabilities; the relationship “raises the probability of” is counterfactual (or
manipulative) in nature, and cannot, therefore, be captured in the language of
probability theory.
The way philosophers tried to capture this relationship, using inequalities
such as15
P (E|C) > P (E) (1.13)
was misguided from the start – counterfactual “raising” cannot be reduced to ev-
idential “raising,” or “raising by conditioning.” The correct inequality, according
to the structural theory of Section 1.3, should read:

P (E|do(C)) > P (E) (1.14)

where do(C) stands for an external intervention that compels the truth of C.
The conditional probability P (E|C), as we know from Section1.3 represents a
probability resulting from a passive observation of C, and rarely coincides with
P (E|do(C)). Indeed, observing the barometer falling increases the probability of
a storm coming, but does not “cause” the storm; if the act of manipulating the
barometer were to change the probability of storms, the falling barometer would
qualify as a cause of storms.
Reformulating the notion of “probability raising” within the calculus of do-
operators resolves the difficulties that PC has encountered in the past half-
century.16 Two such difficulties are worth noting here, for they can be resolved
by the analysis of Section 1.3.

1.5.2 The mystery of “background context”

Recognizing that the basic inequality P (E|C) > P (E) may yield paradoxical
results in the presence of confounding factors (e.g., the atmospheric pressure in
the example above), philosophers have modified the inequality by conditioning
on a background factor K, yielding the criterion: P (E|C, K = k) > P (E|K = k)
where K consists on a set of variables capable of creating spurious dependencies
15 Some authors write P (E|C) > P (E|¬C), which is equivalent to (1.13); the latter is easier

to generalize to the non-binary case.

16 This paper focuses on “type causation” namely, the tendency of the cause to bring about

the effect. Token causation, also known as “actual causation” (Pearl, 2000a, Chapter 10) re-
quires heavier counterfactual machinery.
 Structural versus Probabilistic Causality 19

between the cause and the effect. However, the question of what variables should
enter K led to speculations, controversies and fallacies.17
Cartwright (1983), for example, states that a factor F should enter into K
if and only if F is causally relevant to the effect, that is, F tends to either
promote or prevent E. Eells (1991) on the other hand dropped the “only if” part
and insisted on the “if.” The correct answer, as we know from our analysis of
Section 1.3, is neither Cartwright’s nor Eell’s; K should merely satisfy the back-
door criterion of Section 1.3.2, which may or may not include variables that are
causally relevant to the effect E.
The background-context debate is symptomatic of the fundamental flaw of
the probabilistic causality program; the program first misrepresented the causal
relation P (E|do(C)) by a conditional probability surrogate P (E|C), and then,
to escape the wrath of spurious associations, attempted to patch-up the dis-
tortion by adding remedial conditionalizations, only to end up with a contested
P (E|C, K). The correct strategy should have been to define “probability raising”
directly in terms of the do(x) operator (or counterfactual variables Yx ), which
would have yielded general and coherent results with no need for remedies.18
1.5.3 The epistemology of causal relevance and probability raising
The introduction of a “causal relevance” relation into the definition of “cause”
is of course circular, for it compromises the original goal of reducing causality to
purely probabilistic relations. It gave rise however to an interesting epistemolog-
ical problem whose aim is not reductive but interpretative: Given that humans
store experience in the form of qualitative “causal relevance” relationships, (with
variable X being “causally relevant” to Y whenever it can influence Y in some
way), we ask whether this knowledge, together with a probability function P is
sufficient for determining whether event X = x is a cause of event Y = y in the
“probability raising” sense.19
The problem is interesting because it connects judgments of three different
types: judgments about “causal relevance” (R), about probabilities (P ), and
about cause-effect relations (CE). There is little doubt that causal-relevance
relationships form part of an agent epistemic state; such relationships are im-
plied by people’s understanding of mechanisms, and how mechanisms are put
together in the world around them. It is also reasonable to assume that an
agent’s epistemic state contains some representation of a probability function
P that summarizes facts, observations, and associations acquired by the agent,
either directly or indirectly (say through hearsay, or reading scientific reports).
17 Conditioning on all factors F preceding C (Good, 1961; Suppes, 1970) would lead to

counter intuitive conclusions (Pearl, 2000a, p. 297).

18 Lewis (1986) proposed indeed to treat probability raising in the context of his counter-

factual theory. However, lacking structural semantics, PC advocates viewed Lewis’s counter-
factuals as resting on shaky formal foundation “for which we have only the beginnings of a
semantics (via the device of measures over possible worlds)” (Cartwright, 1983, p. 34).
19 This is my interpretation of Eell’s (1991) epistemic consistency problem (Pearl, 2000a, p.

252).
20 The Structural Theory of Causation Judea Pearl

Finally, people usually reach consensus judging whether a given event X = x

“causes” event Y = y, and generally agree with the “probability raising” maxim.
The epistemic question above amounts to asking whether the three types of
judgments, R, P, and CE, are compatible with each other. Put differently, the
question we may ask is whether CE judgments are compatible with the pair
< R, P > and the probability raising maxim given in (1.14). To answer such
questions we must first determine whether the pair < R, P > is sufficient for
deriving inequalities of the type given in (1.14).
The structural theory of causation gives a definitive solution to this problem
which reads as follows:
Given: A graph G on a set V of variables, such that there is a directed path from
X to Y in V iff X is judged to be “causally relevant” to Y .
Also given: a probability measure P (v) that is compatible with G.
Problem: Decide, for a given X and Y in V , whether the probability raising
inequality (1.14) holds for C : X = x and E : Y = y, namely whether the causal
effect
CE = P (y|do(x)) − P (y) (1.15)

is greater than zero, given G and P .

The solution follows immediately from the identification of causal effects in
Markovian models, which permits the derivation of CE from G and P , for ex-
ample, by the causal effect formula of equation (1.4).
The solution is less obvious when P is defined over a proper subset W of
V , where {V − W } represents the set of unmeasured variables. The problem
then reduces to that of identifying CE in semi-Markovian models such as those
addressed in Theorem 1. Fortunately, the completeness results of Tian and Pearl
(2002) and Shpitser and Pearl (2006b) reduce this problem to algorithmic routine
on the graph G and, furthermore, they provide a guarantee that, if the algorithm
fails, then any algorithm would fail, namely the causal effect of x on y does not
have a unique value, given R and P .
I venture to conjecture that every epistemic problem concerned with the
relationship between causes and probabilities is now amenable to algorithmic
solution, provided that one explicates formally what is assumed and what needs
to be decided.

1.5.4 Is Probabilistic Causality Subsumed by the Structural Theory?

In view of the difficulties described above, it is fair to ask whether PC should be
regarded as a special case of the structural theory, or, for that matter, whether it
should qualify as a theory of causation by the four criteria set forth in Section 1.1.
The answer is that, although PC fails to satisfy these criteria, its aspirations were
to provide a formal language for causal assertions of the “probability raising”
variety. While the notation chosen for the task was inadequate, the reasoning
behind most PC investigations was clearly guided by structural considerations.
The introduction of a “causal relevance” relation into the theory attests to the
 Comparison to the Potential-Outcomes Framework 21

structural nature of that reasoning. The structural theory now permits PC in-
vestigators to re-articulate philosophical and epistemological problems in an un-
ambiguous formal language and derive, using the notational machinery provided
by the SCM, answers to pending questions in this area of inquiry. Section 1.5.3
demonstrates the benefits of this machinery; similar benefits were demonstrated
in problems posed by Woodward (Pearl, 2003a) and Cartwright (Pearl, 2009a,
pp. 362-5)

1.6 Comparison to the Potential-Outcomes Framework

The primitive object of analysis in the potential-outcome framework is the unit-
based response variable, denoted Yx (u), read: “the value that outcome Y would
obtain in experimental unit u, had treatment X been x” (Neyman, 1923; Rubin,
1974). Here, unit may stand for an individual patient, an experimental subject,
or an agricultural plot. In Section 1.3.3 we saw that this counterfactual entity has
the natural interpretation as representing the solution for Y in a modified system
of equations, where unit is interpreted a vector u of background factors that
characterize an experimental unit. Each structural equation model thus carries
a collection of assumptions about the behavior of hypothetical units, and these
assumptions permit us to derive the counterfactual quantities of interest. In the
potential-outcome framework, however, no equations are available for guidance
and Yx (u) is taken as primitive, that is, an undefined quantity in terms of which
other quantities are defined; not a quantity that can be derived from some model.
In this sense the structural interpretation of Yx (u) given in (1.7) provides the
formal basis for the potential-outcome approach; the formation of the submodel
Mx explicates mathematically how the hypothetical condition “had X been x”
could be realized, and what the logical consequence are of such a condition.

1.6.1 The “Black-Box” or “Missing-data” Paradigm

The distinct characteristic of the potential-outcome approach is that, although
investigators must think and communicate in terms of undefined, hypothetical
quantities such as Yx (u), the analysis itself is conducted almost entirely within
the axiomatic framework of probability theory. This is accomplished, by postu-
lating a “super” probability function on both hypothetical and real events. If
U is treated as a random variable then the value of the counterfactual Yx (u)
becomes a random variable as well, denoted as Yx . The potential-outcome anal-
ysis proceeds by treating the observed distribution P (x1 , . . . , xn ) as the marginal
distribution of an augmented probability function P ∗ defined over both observed
and counterfactual variables. Queries about causal effects (written P (y|do(x)) in
the structural analysis) are phrased as queries about the marginal distribution
of the counterfactual variable of interest, written P ∗ (Yx = y). The new hypo-
thetical entities Yx are treated as ordinary random variables; for example, they
are assumed to obey the axioms of probability calculus, the laws of conditioning,
and the axioms of conditional independence.
22 The Structural Theory of Causation Judea Pearl

Naturally, these hypothetical entities are not entirely whimsy. They are as-
sumed to be connected to observed variables via consistency constraints (Robins,
1986) such as
X = x =⇒ Yx = Y, (1.16)
which states that, for every u, if the actual value of X turns out to be x, then
the value that Y would take on if ‘X were x’ is equal to the actual value of Y .
For example, a person who chose treatment x and recovered, would also have
recovered if given treatment x by design. Whether additional constraints should
tie the observables to the unobservables is not a question that can be answered
in the potential-outcome framework, which lacks an underlying model.
The main conceptual difference between the two approaches is that, whereas
the structural approach views the intervention do(x) as an operation that changes
the distribution but keeps the variables the same, the potential-outcome ap-
proach views the variable Y under do(x) to be a different variable, Yx , loosely
connected to Y through relations such as (1.16), but remaining unobserved when-
ever X 6= x. The problem of inferring probabilistic properties of Yx , then becomes
one of “missing-data” for which estimation techniques have been developed in
the statistical literature.
Pearl (2000a, Chapter 7) shows, using the structural interpretation of Yx (u),
that it is indeed legitimate to treat counterfactuals as jointly distributed random
variables in all respects, that consistency constraints like (1.16) are automatically
satisfied in the structural interpretation and, moreover, that investigators need
not be concerned about any additional constraints except the following two:

Yyz = y for all y, subsets Z, and values z for Z (1.17)

Xz = x ⇒ Yxz = Yz for all x, subsets Z, and values z for Z (1.18)

Equation (1.17) ensures that the interventions do(Y = y) results in the condition
Y = y, regardless of concurrent interventions, say do(Z = z), that may be applied
to variables other than Y . Equation (1.18) generalizes (1.16) to cases where Z is
held fixed, at z.

1.6.2 Problem Formulation and the Demystification of “Ignorability”

The main drawback of this black-box approach surfaces in problem formulation,
namely, the phase where a researcher begins to articulate the “science” or “causal
assumptions” behind the problem at hand. Such knowledge, as we have seen in
Section 1.1, must be articulated at the onset of every problem in causal analysis
– causal conclusions are only as valid as the causal assumptions upon which they
rest.
To communicate scientific knowledge, the potential-outcome analyst must
express assumptions as constraints on P ∗, usually in the form of conditional in-
dependence assertions involving counterfactual variables. For instance, in our ex-
ample of Figure 1.2(a), to communicate the understanding that the Z is random-
ized (hence independent of UX and UY ), the potential-outcome analyst would use
 Comparison to the Potential-Outcomes Framework 23

the independence constraint Z⊥ ⊥{Yz1 , Yz2 , . . . , Yzk }.20 To further formulate the
understanding that Z does not affect Y directly, except through X, the analyst
would write a, so called, “exclusion restriction”: Yxz = Yx .
A collection of constraints of this type might sometimes be sufficient to per-
mit a unique solution to the query of interest. For example, if one can plausibly
assume that, in Fig. 1.3, a set Z of covariates satisfies the conditional indepen-
dence
Yx ⊥⊥X|Z (1.19)

(an assumption termed “conditional ignorability” by Rosenbaum and Rubin

(1983),) then the causal effect P (y|do(x)) = P ∗(Yx = y) can readily be eval-
uated to yield
X
P ∗ (Yx = y) = P ∗(Yx = y|z)P (z)
z
X
= P ∗(Yx = y|x, z)P (z) (using (1.19))
z
X
= P ∗(Y = y|x, z)P (z) (using (1.16))
z
X
= P (y|x, z)P (z). (1.20)
z

The last expression contains no counterfactual quantities (thus permitting us to

drop the asterisk from P ∗) and coincides precisely with the standard covariate-
adjustment formula of equation (1.5).
We see that the assumption of conditional ignorability (1.19) qualifies Z as
a sufficient covariate for adjustment; it is entailed indeed by the “back-door”
criterion of Section 1.3.2, which qualifies such covariates by tracing paths in the
causal diagram.
The derivation above may explain why the potential-outcome approach ap-
peals to mathematical statisticians; instead of constructing new vocabulary (e.g.,
arrows), new operators (do(x)) and new logic for causal analysis, almost all math-
ematical operations in this framework are conducted within the safe confines of
probability calculus. Save for an occasional application of rule (1.18) or (1.16)),
the analyst may forget that Yx stands for a counterfactual quantity—it is treated
as any other random variable, and the entire derivation follows the course of rou-
tine probability exercises.
However, this mathematical orthodoxy exacts a very high cost: all back-
ground knowledge pertaining to a given problem must first be translated into
the language of counterfactuals (e.g., ignorability conditions) before analysis can
commence. This translation may in fact be the hardest part of the problem. The

20 The notation Y ⊥
⊥X|Z stands for the conditional independence relationship P (Y = y, X =
x|Z = z) = P (Y = y|Z = z)P (X = x|Z = z) (Dawid, 1979).
24 The Structural Theory of Causation Judea Pearl

reader may appreciate this aspect by attempting to judge whether the assump-
tion of conditional ignorability (1.19), the key to the derivation of (1.20), holds
in any familiar situation, say in the experimental setup of Figure 1.2(a). This
assumption reads: “the value that Y would obtain had X been x, is independent
of X, given Z”. Even the most experienced potential-outcome expert would be
unable to discern whether any subset Z of covariates in Figure 1.3 would satisfy
this conditional independence condition.21 Likewise, to derive equation (1.6) in
the language of potential-outcome (see Pearl (2000a, p. 223)), one would need
to convey the structure of the chain X → W3 → Y using the cryptic expression:
W3x ⊥ ⊥{Yw3 , X}, read: “the value that W3 would obtain had X been x is inde-
pendent of the value that Y would obtain had W3 been w3 jointly with the value
of X”. Such assumptions are cast in a language so far removed from ordinary
understanding of scientific theories that, for all practical purposes, they cannot
be comprehended or ascertained by ordinary mortals. As a result, researchers
in the graph-less potential-outcome camp rarely use “conditional ignorability”
(1.19) to guide the choice of covariates; they view this condition as a hoped-for
miracle of nature rather than a target to be achieved by reasoned design.22
Replacing “ignorability” with a simple condition (i.e., back-door) in a graph-
ical model permits researchers to understand what conditions covariates must
fulfill before they eliminate bias, what to watch for and what to think about
when covariates are selected, and what experiments we can do to test, at least
partially, if we have the knowledge needed for covariate selection.
Aside from offering no guidance in covariate selection, formulating a problem
in the potential-outcome language encounters three additional hurdles. When
counterfactual variables are not viewed as byproducts of a deeper, process-based
model, it is hard to ascertain whether all relevant counterfactual independence
judgments have been articulated, whether the judgments articulated are redun-
dant, or whether those judgments are self-consistent. The need to express, defend,
and manage formidable counterfactual relationships of this type explain the slow
acceptance of causal analysis among health scientists and statisticians, and why
economists and social scientists continue to use structural equation models in-
stead of the potential-outcome alternatives advocated in (Angrist et al., 1996;
Holland, 1988; Sobel, 1998).
On the other hand, the algebraic machinery offered by the counterfactual
notation, Yx (u), once a problem is properly formalized, can be extremely power-
ful in refining assumptions (Angrist et al., 1996), deriving consistent estimands
(Robins, 1986), bounding probabilities of necessary and sufficient causation (Tian
21 Inquisitive
readers are invited to guess whether Xz ⊥⊥Z|Y holds in Figure 1.2(a).
22 The opaqueness of counterfactual independencies explains why many researchers within the
potential-outcome camp are unaware of the fact that adding a covariate to the analysis (e.g., Z3
in Figure 1.3) may actually increase confounding bias. Paul Rosenbaum, for example, writes:
“there is no reason to avoid adjustment for a variable describing subjects before treatment”
(Rosenbaum, 2002, p. 76). Rubin (2009) goes as far as stating that refraining from conditioning
on an available measurement is “nonscientific ad hockery” for it goes against the tenets of
Bayesian philosophy (see (Pearl, 2009bc) for a discussion of this fallacy).
 Conclusions 25

and Pearl, 2000), and combining data from experimental and nonexperimental
studies (Pearl, 2000a). Pearl (2000a, p. 232) presents a way of combining the
best features of the two approaches. It is based on encoding causal assumptions
in the language of diagrams, translating these assumptions into counterfactual
notation, performing the mathematics in the algebraic language of counterfactu-
als (using (1.16), (1.17), and (1.18)) and, finally, interpreting the result in plain
causal language. The mediation problem of Section 1.4 illustrates such symbiosis.
In comparison, when the mediation problem is approached from an orthodox
potential-outcome viewpoint, void of the structural guidance of equation (1.7),
pardoxical results ensue (Rubin, 2004). For example, the direct effect is definable
only in units absent of indirect effects. This means that a grandfather would be
deemed to have no direct effect on his grandson’s behavior in families where
he has had some effect on the father. This leaves us mostly with odd families,
absent of grandfathers or fathers. In linear systems, to take a sharper example,
the direct effect would be undefined whenever indirect paths exist from the cause
to its effect. Such paradoxical conclusions underscore the wisdom, if not necessity
of a symbiotic analysis, in which the counterfactual notation Yx (u) is governed
by the structural semantics of the SCM.

1.7 Conclusions
Theories of causation require two ingredients that are absent from probabilistic
or logical theories; a science-friendly language for articulating causal knowledge,
and a mathematical machinery for processing that knowledge, combining it with
data and drawing new causal conclusions about a phenomenon. This paper intro-
duces a general theory of causation, based on nonparametric structural equations
models, that supplements statistical methods with the needed ingredients. The
algebraic component of the theory coincides with the potential-outcome frame-
work, and its graphical component embraces Wright’s method of path diagrams
(in its nonparametric version). When unified and synthesized, the two compo-
nents offer empirical investigators a powerful and comprehensive methodology
for causal inference. and a general framework for viewing other, less general ap-
proaches to causation, including probabilistic causation (Section 1.5) and the
potential-outcome model (1.6).

Acknowledgments
Portions of this paper are based on my book Causality (Pearl, 2000, 2nd edition
forthcoming 2009), and have benefited appreciably from conversations with Chris
Hitchcock. This research was supported in parts by an ONR grant #N000-14-
09-1-0665.
 REFERENCES

Angrist, J.D., Imbens, G.W., and Rubin, D.B. (1996). Identification of causal
effects using instrumental variables (with comments). Journal of the American
Statistical Association, 91(434), 444–472.
Arah, O.A. (2008). The role of causal reasoning in understanding Simp-
son’s paradox, Lord’s paradox, and the suppression effect: Covariate se-
lection in the analysis of observational studies. Emerging Themes in
Epidemiology, 4, doi:10.1186/1742–7622–5–5. Online at <http://www.ete-
online.com/content/5/1/5>.
Avin, C., Shpitser, I., and Pearl, J. (2005). Identifiability of path-specific effects.
In Proceedings of the Nineteenth International Joint Conference on Artificial
Intelligence IJCAI-05, Edinburgh, UK, pp. 357–363. Morgan-Kaufmann Pub-
lishers.
Balke, A. and Pearl, J. (1994a). Counterfactual probabilities: Computational
methods, bounds, and applications. In Uncertainty in Artificial Intelligence
10 (ed. R. L. de Mantaras and D. Poole), pp. 46–54. Morgan Kaufmann, San
Mateo, CA.
Balke, A. and Pearl, J. (1994b). Probabilistic evaluation of counterfactual
queries. In Proceedings of the Twelfth National Conference on Artificial Intel-
ligence, Volume I, pp. 230–237. MIT Press, Menlo Park, CA.
Balke, A. and Pearl, J. (1995). Counterfactuals and policy analysis in struc-
tural models. In Uncertainty in Artificial Intelligence 11 (ed. P. Besnard and
S. Hanks), pp. 11–18. Morgan Kaufmann, San Francisco.
Brent, R. and Lok, L. (2005). A fishing buddy for hypothesis generators. Sci-
ence, 308(5721), 523–529.
Cartwright, N. (1983). How the Laws of Physics Lie. Clarendon Press, Oxford.
Chalak, K. and White, H. (2006, July). An extended class of instrumental
variables for the estimation of causal effects. Technical Report Discussion
Paper, UCSD, Department of Economics.
Cole, S.R. and Hernán, M.A. (2002). Fallibility in estimating direct effects.
International Journal of Epidemiology, 31(1), 163–165.
Collins, J., Hall, N., and Paul, L.A. (eds.) (2004). Causation and Counterfac-
tuals. MIT Press, Cambridge, MA.
Cox, D.R. (1958). The Planning of Experiments. John Wiley and Sons, NY.
Cox, D.R. and Wermuth, N. (2004). Causality: A statistical view. International
Statistical Review , 72(3), 285–305.
Dawid, A.P. (1979). Conditional independence in statistical theory. Journal of
the Royal Statistical Society, Series B, 41(1), 1–31.
Dawid, A.P. (2000). Causal inference without counterfactuals (with comments
and rejoinder). Journal of the American Statistical Association, 95(450), 407–
 References 27

448.
Dawid, A.P. (2002). Influence diagrams for causal modelling and inference.
International Statistical Review , 70, 161–189.
Duncan, O.D. (1975). Introduction to Structural Equation Models. Academic
Press, New York.
Eells, E. (1991). Probabilistic Causality. Cambridge University Press, Cam-
bridge, MA.
Glymour, M.M. and Greenland, S. (2008). Causal diagrams. In Modern Epi-
demiology (3rd edn) (ed. K. Rothman, S. Greenland, and T. Lash), pp. 183–
209. Lippincott Williams & Wilkins, Philadelphia, PA.
Good, I.J. (1961). A causal calculus (I). British Journal for the Philosophy of
Science, 11, 305–318.
Greenland, S. and Brumback, B. (2002). An overview of relations among causal
modelling methods. International Journal of Epidemiology, 31, 1030–1037.
Greenland, S., Pearl, J., and Robins, J.M (1999). Causal diagrams for epidemi-
ologic research. Epidemiology, 10(1), 37–48.
Haavelmo, T. (1943). The statistical implications of a system of simultaneous
equations. Econometrica, 11, 1–12. Reprinted in D.F. Hendry and M.S. Mor-
gan (Eds.), The Foundations of Econometric Analysis, Cambridge University
Press, 477–490, 1995.
Heckman, J.J. (2008). Econometric causality. International Statistical Re-
view , 76(1), 1–27.
Hitchcock, C. (2001). Book reviews: Causality: Models, Reasoning and Infer-
ence. The Philosophical Review , 110(4), 639–641.
Hitchcock, C.R. (2003). Probabilistic causation. In Stanford Ency-
clopedia of Philosophy (Winter 2003 Edition) (ed. E. Zalta). URL =
<http://plato.stanford.edu/entries/causation-probabilistic/>, Stanford, CA.
Holland, P.W. (1988). Causal inference, path analysis, and recursive structural
equations models. In Sociological Methodology (ed. C. Clogg), pp. 449–484.
American Sociological Association, Washington, D.C.
Lauritzen, S.L. (1996). Graphical Models. Clarendon Press, Oxford.
Lauritzen, S.L. (2001). Causal inference from graphical models. In Complex
Stochastic Systems (ed. D. Cox and C. Kluppelberg), pp. 63–107. Chapman
and Hall/CRC Press, Boca Raton, FL.
Lewis, D. (1986). Philosophical Papers, Volume II. Oxford University Press,
New York.
Lindley, D.V. (2002). Seeing and doing: The concept of causation. International
Statistical Review , 70, 191–214.
Meek, C. and Glymour, C.N. (1994). Conditioning and intervening. British
Journal of Philosophy Science, 45, 1001–1021.
Morgan, S.L. and Winship, C. (2007). Counterfactuals and Causal Inference:
Methods and Principles for Social Research (Analytical Methods for Social
Research). Cambridge University Press, New York, NY.
Neyman, J. (1923). On the application of probability theory to agricultural
28 References

experiments. Essay on principles. Section 9. Statistical Science, 5(4), 465–

480.
Pearl, J. (1988). Probabilistic Reasoning in Intelligent Systems. Morgan Kauf-
mann, San Mateo, CA.
Pearl, J. (1993a). Comment: Graphical models, causality, and intervention.
Statistical Science, 8(3), 266–269.
Pearl, J. (1993b). Mediating instrumental variables. Technical Report TR-
210, <http://ftp.cs.ucla.edu/pub/stat ser/R210.pdf>, Department of Com-
puter Science, University of California, Los Angeles.
Pearl, J. (1995). Causal diagrams for empirical research. Biometrika, 82(4),
669–710.
Pearl, J. (1998). Graphs, causality, and structural equation models. Sociological
Methods and Research, 27(2), 226–284.
Pearl, J. (2000a). Causality: Models, Reasoning, and Inference. Cambridge
University Press, New York. Second ed., 2009.
Pearl, J. (2000b). Comment on A.P. Dawid’s, Causal inference without counter-
factuals. Journal of the American Statistical Association, 95(450), 428–431.
Pearl, J. (2001). Direct and indirect effects. In Proceedings of the Seven-
teenth Conference on Uncertainty in Artificial Intelligence, pp. 411–420. Mor-
gan Kaufmann, San Francisco, CA.
Pearl, J. (2003a). Reply to Woodward. Economics and Philosophy, 19, 341–
344.
Pearl, J. (2003b, December). Statistics and causal inference: A review. Test
Journal , 12(2), 281–345.
Pearl, J. (2005). Direct and indirect effects. In Proceedings of the American Sta-
tistical Association, Joint Statistical Meetings, pp. 1572–1581. MIRA Digital
Publishing, Minn., MN.
Pearl, J. (2009a). Causality: Models, Reasoning, and Inference (Second edn).
Cambridge University Press, New York.
Pearl, J. (2009b). Letter to the editor: Remarks on the method
of propensity scores. Statistics in Medicine, 28, 1415–1416.
<http://ftp.cs.ucla.edu/pub/stat ser/r345-sim.pdf>.
Pearl, J. (2009c). Myth, confusion, and science in causal analysis.
Technical Report R-348, University of California, Los Angeles, CA.
<http://ftp.cs.ucla.edu/pub/stat ser/r348.pdf>.
Pearl, J. and Robins, J.M. (1995). Probabilistic evaluation of sequential plans
from causal models with hidden variables. In Uncertainty in Artificial Intelli-
gence 11 (ed. P. Besnard and S. Hanks), pp. 444–453. Morgan Kaufmann, San
Francisco.
Petersen, M.L., Sinisi, S.E., and van der Laan, M.J. (2006). Estimation of direct
causal effects. Epidemiology, 17(3), 276–284.
Robins, J.M. (1986). A new approach to causal inference in mortality stud-
ies with a sustained exposure period – applications to control of the healthy
workers survivor effect. Mathematical Modeling, 7, 1393–1512.
 References 29

Robins, J.M. (1999). Testing and estimation of directed effects by reparameter-

izing directed acyclic with structural nested models. In Computation, Causa-
tion, and Discovery (ed. C. Glymour and G. Cooper), pp. 349–405. AAAI/MIT
Press, Cambridge, MA.
Robins, J.M. and Greenland, S. (1991). Estimability and estimation of expected
years of life lost due to a hazardous exposure. Statistics in Medicine, 10, 79–93.
Rosenbaum, P.R. (2002). Observational Studies (Second edn). Springer-Verlag,
New York.
Rosenbaum, P. and Rubin, D. (1983). The central role of propensity score in
observational studies for causal effects. Biometrika, 70, 41–55.
Rothman, K.J. (1976). Causes. American Journal of Epidemiology, 104, 587–
592.
Rubin, D.B. (1974). Estimating causal effects of treatments in randomized and
nonrandomized studies. Journal of Educational Psychology, 66, 688–701.
Rubin, D.B. (2004). Direct and indirect causal effects via potential outcomes.
Scandinavian Journal of Statistics, 31, 161–170.
Rubin, D.B. (2009). Author’s reply: Should observational studies be designed
to allow lack of balance in covariate distributions across treatment group?
Statistics in Medicine, 28, 1420–1423.
Shpitser, I. and Pearl, J (2006a). Identification of conditional interventional
distributions. In Proceedings of the Twenty-Second Conference on Uncertainty
in Artificial Intelligence (ed. R. Dechter and T. Richardson), pp. 437–444.
AUAI Press, Corvallis, OR.
Shpitser, I. and Pearl, J. (2006b). Identification of joint interventional dis-
tributions in recursive semi-Markovian causal models. In Proceedings of the
Twenty-First National Conference on Artificial Intelligence, pp. 1219–1226.
AAAI Press, Menlo Park, CA.
Shpitser, I. and Pearl, J (2007). What counterfactuals can be tested. In Proceed-
ings of the Twenty-Third Conference on Uncertainty in Artificial Intelligence,
pp. 352–359. AUAI Press, Vancouver, BC, Canada. Also, Journal of Machine
Learning Research, 9:1941–1979, 2008.
Shrier, I. (2009). Letter to the editor: Propensity scores.
Statistics in Medicine, 28, 1317–1318. See also Pearl 2009
<http://ftp.cs.ucla.edu/pub/stat ser/r348.pdf>.
Simon, H.A. and Rescher, N. (1966). Cause and counterfactual. Philosophy and
Science, 33, 323–340.
Sobel, M.E. (1998). Causal inference in statistical models of the process of
socioeconomic achievement. Sociological Methods & Research, 27(2), 318–348.
Spirtes, P., Glymour, C.N., and Scheines, R. (2000). Causation, Prediction, and
Search (2nd edn). MIT Press, Cambridge, MA.
Suppes, P. (1970). A Probabilistic Theory of Causality. North-Holland Pub-
lishing Co., Amsterdam.
Tian, J., Paz, A., and Pearl, J. (1998). Finding minimal separating sets. Tech-
nical Report R-254, University of California, Los Angeles, CA.
30 References

Tian, J. and Pearl, J. (2000). Probabilities of causation: Bounds and identifi-

cation. Annals of Mathematics and Artificial Intelligence, 28, 287–313.
Tian, J. and Pearl, J. (2002). A general identification condition for causal
effects. In Proceedings of the Eighteenth National Conference on Artificial
Intelligence, pp. 567–573. AAAI Press/The MIT Press, Menlo Park, CA.
Williamson, J. (2010, Forthcoming). Probabilistic theories of causality. In The
Oxford Handbook of Causation (ed. H. Beebee, C. Hitchcock, and P. Peter).
Oxford University Press, New York.
Woodward, J. (2003). Making Things Happen. Oxford University Press, New
York, NY.
Wright, S. (1921). Correlation and causation. Journal of Agricultural Re-
search, 20, 557–585.