Out11 PDF

This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
Psychological Bulletin
1986, VoL 99, No. 1,3-19
Copyright 1986 by the American Psychological Association, Inc.

<XU3-2909/86/$00.75
Judging Probable Cause

Hillel J. Einhorn and Robin M. Hogarth
Graduate School of Business
Center for Decision Research
University of Chicago
Despite the debate and lack of agreement as to a formal definition of causality, we argue that people
use systematic rules for assessing cause, both in science and everyday inference. By explicating the
processes that underlie the judgment of causation, we review and integrate various theories of causality
proposed by psychologists, philosophers, statisticians, and others. Because causal judgment involves
inference and uncertainty, the literature on judgment-under-uncertainty is also considered. Our review
is organized around four concepts, (a) The idea of a "causal field" is central for determining causal
relevance, differentiating causes from conditions, determining the salience of alternative explanations,
and affecting molar versus molecular explanations, (b) Various "cues-to-causality," such as covariation,
temporal order, contiguity in time and space, and similarity of cause and effect, are discussed. In
doing so, we show how these cues can conflict with probabilistic ideas, (c) A model for combining the
cues and the causal field allows us to discuss a wide range of studies on causal judgments and explicates
methodological issues such as spurious correlation, "causalation," and causal inference in case studies,
(d) The discounting of an explanation by specific alternatives is discussed as a special case of the
sequential updating of beliefs. Finally, we extend our approach to consider conjunctive explanations
in multiple causation.
The topic of causality is central to many areas of psychology,
the scientific method more generally, is at present in a productive
such as developmental psychology (Piaget, 1974; Sedlak & Kurtz,
state of near chaos" (p. 10).
1981), perception and learning (Michotte, 1946), social psy-
Psychology is not alone, however, in presenting a fragmented
chology (Harvey & Weary, 1984; Heider, 1958; Kelley, 1967),
view of causality. In philosophy, the definition and meaning of
methodology (Cook & Campbell, 1979), text comprehension
cause have been debated for centuries (for a review, see Bunge,
(Trabasso, Secco, & van den Broek, 1984), and so on. However,
1979). Indeed, Bertrand Russell (1913) even suggested that the
on reading these various literatures, one is struck by their relative
concept of cause does not play a useful role in science. Thus,
isolation from each other. Indeed, there seems to be little contact
lacking a clear definition, psychologists within different research
or even awareness of mutual problems, issues, and concepts. For
traditions have focused on specific aspects of causality that have
example, discussions of how people perceive physical causation
been emphasized by different philosophers and made operational
are rarely linked to the processes of causal attribution in the
tests of these notions. For example, workers in attribution theory
social domain, or vice versa. Moreover, even within areas, there
have tended to follow Kelley (1967) in emphasizing Mill's (1872)
is considerable dissatisfaction concerning the incomplete nature
criteria of concomitant variation and the method of differences;
of our understanding of the psychology of cause. For example,
Michotte's (1946) classic demonstrations of how people perceive
in a review of attribution theory, Kelley and Michela (1980)
causes relies heavily on ideas advanced by Hume (1886/1964);
wrote, ". . . the theories are piecemeal and greatly in need of
Shultz's (1982) work has been influenced by Kant's (1781 /1982)
synthesis" (p. 494). In a similar vein, Cook and Campbell (1979)
notions that causal relations are characterized by forces of "gen-
have pointed out that, "The epistemology of causation, and of
erative transmission" between cause and effect; and Campbell

and colleagues (see, e.g., Cook & Campbell, 1979) have selected
from various philosophical positions to reach guidelines concerning when assertions of causation are justified. Similarly, the
This research was supported by a contract from the Office of Naval

Research. The Lady Davis Foundation also provided support for the first
author while visiting the Hebrew University of Jerusalem.
We would like to thank the following people for their comments on
earlier versions of this manuscript: Robert Abelson, Beradt Brehmer,
Colin Camerer, Norman Dalkey, Donald Fiske, William Goldstein, Kenneth Hammond, Reid Hastie, Betty House, Joshua Klayman, Howard
Kunreuther, Marlys Lipe, Lola Lopes, Albert Madansky, Haim Mano,
Ann McGill, John Payne, Nancy Pennington, Jay Russo, Paul Schoemaker, Thomas Trabasso, Arnold Zellner, and four anonymous referees.
Correspondence concerning this article should be addressed to either
Hillel J. Einhorn or Robin M. Hogarth, Center for Decision Research,
Graduate School of Business, University of Chicago, 1101 East 58th Street,
Chicago, Illinois 60637.
meaning of cause and its operational definition have also attracted

controversy and debate in other areas, such as economics (Granger, 1969; Zellner, 1984), law (Fincham & Jaspars, 1980; Hart
& Honore, 1959), medicine (Shapiro, 1960; Susser, 1973), statistics (Cohen, 1977; Suppes, 1970; Toda, 1977), and sociology
(Blalock, 1971;Costner&Leik, 1964).
Despite the lack of both formal definition and agreement as
to what does or does not constitute a cause, it is significant that
people do have strong intuitions concerning the presence or absence of causes in particular instances. From our perspective,
this implies that people use systematic rules and strategies for
assessing cause, both in science and everyday inference. Thus,
HILLEL J. EINHORN AND ROBIN M. HOGARTH

the assessment of cause can be seen as the result of a judgment
process. Moreover, it is by explicating this process that we shall
discuss and integrate the various theories of causation in this
article.
In addition to providing a general framework, a judgment approach has the distinct advantage of highlighting the inferential
nature of causal reasoning that typically goes "beyond the information given" (Bruner, 1957), and thus involves uncertainty;
hence, our emphasis on judging probable cause. Indeed, as
pointed out by Cook and Campbell (1979),". . . molar causal
laws, because they are contingent on many other conditions and
causal laws, are fallible and hence probabilistic" (p. 33). The
explicit recognition that uncertainty plays a major role in assessing cause also allows us to consider the burgeoning literature
on judgment-under-uncertainty in the assessment of probable
cause (for reviews see Einhorn & Hogarth, 1981; Kahneman,
Slovic, & Tversky, 1982; Pitz & Sachs, 1984). Although there
have been few attempts to link probabilistic and causal notions
(cf. Suppes, 1970), we believe that such links are crucial for
understanding both causal and probabilistic ideas.
Purpose and Plan of Article

The purpose of this article is to review different theoretical
approaches to the psychology of cause within a framework that
recognizes such attributions as resulting from processes of judgment-under-uncertainty (Brunswik, 1952; Einhorn & Hogarth,
1981; Hammond, 1955; Kahneman et al., 1982).
To encompass the wide diversity of ideas concerning causation,
our framework consists of four components. These are: (a) the
causal field or context in which probable cause judgments are
made (Mackie, 1965; van Fraassen, 1980); (b) the use of various
probabilistic indicators of causal relations, called "cues-to-causality" (Hume, 1886/1964); (c) judgmental strategies employed
for combining the causal field with the cues-to-causality; and (d)
the discounting of causal strength by alternative explanations
(Cook & Campbell, 1979; Kelley, 1973; Mill, 1872). We stress
that whereas aspects of each of these components have been
treated in detail elsewhere, our contribution lies in their consolidation. To do so, we occasionally "fill in the gaps" where
there is no existing theory and/or data. However, even in these
cases, the framework extrapolates from either theoretical or empirical results that have been discussed in one field or another.
The advantages of organizing our review in this manner are
to: (a) show how various theoretical perspectives do or do not
address different aspects of causal judgment; (b) illustrate real
and apparent conflicts between viewpoints; (c) illuminate issues
in the literature on judgment-under-uncertainty, which has typically adopted an acausal perspective; (d) assess existing experimental literature; and (e) indicate specific areas where more research is needed.
The four components of our framework can be briefly illustrated by the following example. Imagine that a watch face has
been hit by a hammer and the glass breaks. How likely was the
force of the hammer the cause of the breakage? Because no explicit context is given, an implicitly assumed neutral context is
invoked in which the cues-to-causality point strongly to a causal
relation; that is, the force of the hammer precedes the breakage
in time, there is high covariation between glass breaking (or not)
with the force of solid objects, contiguity in time and space is

high, and there is congruity (similarity) between the length and
strength of cause and effect. Moreover, it is difficult to discount
the causal link because there are few alternative explanations to
consider. Now imagine that the same event occurred during a
testing procedure in a watch factory. In this context, the cause
of the breakage is more often judged to be a defect in the glass
(for specific experimental evidence, see Einhorn & Hogarth,
1983). We consider this in detail below but note that although
the cues have not changed, the shift in the causal field or context
affects their interpretation as well as the strengths of competing
explanations.
More generally, consider an effect, Y, and a possible cause, X.
In our framework, we distinguish between what we term the
gross and net strengths of evidence that A" causes Y. Gross strength
is determined by the causal field and the cues-to-causality. Net
strength results when the gross strength of the X, Y relation is
discounted by the gross strengths of alternative causes or explanations. In other words, the net strength of a causal explanation
is increasing in its gross strength but decreasing in the gross
strengths of its alternatives.
To illustrate, reconsider the initial watch-hammer scenario
and contrast the net strength of the "force of the hammer" explanation with the net strength of any single explanation for the
following questions:
1. Why are the outer rings of Saturn braided?
2. Why was Ronald Reagan elected President in 1980?
For the first question, even if one could generate an explanation,
its gross strength would be low because there is little in the way
of either context or causal cues to explain this anomalous finding.
Moreover, net strength is low because competing explanations
would only discount the original explanation further. For the
second question, there are many strong explanations (e.g., the
situation of the economy, the rise of the moral majority, the
unresolved Iranian hostage problem, etc.). Therefore, although
the gross strength of each of these explanations is high, the net
strength of any single one is low precisely because the others
provide plausible alternatives. On the other hand, the watchhammer question leads to high net strength because the explanation is strong and there are few plausible alternatives. In short,
we maintain that like good patterns, good explanations have few
alternatives (Garner, 1970); or, to be more precise, whereas good
explanations imply few strong alternatives, the lack of strong
alternatives does not imply good explanations.
Our discussion of judging probable cause reviews, in turn,
each of the four components of the general framework illustrated
above.
Causal Field
Philosophical discussions of causality have frequently been
plagued by the inability to specify criteria for determining which
variables are causally relevant in particular circumstances. One
reason for this quandary is nicely illustrated by Hanson (1958):
There are as many causes of x as there are explanations of x. Consider
how the cause of death might have been set out by the physician as
"multiple haemorrhage," by the barrister as "negligence on the part of
the driver," by a carriage-builder as "a defect in the brakeblock construction," by a civic planner as "the presence of tall shrubbery at that turning."
(P. 54)
PROBABLE CAUSE
In other words, the determination of causal relevance depends

critically on contextual factors (see also van Fraassen, 1980,
Ch. 5).
The philosopher Mackie (1965, 1974) has made an important
theoretical contribution to our understanding of the role of context in judgments of probable cause. Mackie has argued that
judgments of causal relevance are generally related to the degree
that a variable is a diflerence-in-a-background. This means that
factors that are part of some presumed background or "causal
field" (in Mackie's terminology) are judged to be of little or no
causal relevance. For example, does birth cause death? Although
the former is both necessary and sufficient for the latter (and
thus covaries perfectly with it), it seems odd to consider one the
cause of the other. The reason is that death presupposes that one
has been born. Therefore, birth is part of the causal field (it is
not a difference-in-a-background) and its causal relevance is low.
Mackie's conceptualization has three important implications
for understanding judgments of probable cause. These concern:
(1) the characteristics of events that trigger causal reasoning, (2)
the role of the background in distinguishing causes from conditions, and (3) how shifts in the causal field make alternative
causes more or less salient.
1. There is much evidence in cognitive psychology that people
engage in causal reasoning in order to "make sense of " the world
(Abelson, 1981; Hastie, 1984; Schank & Abelson, 1977). Thus,
because differences-in-a-background involve events that are unusual, abnormal, or unlikely, an important test of Mackie's notion
of the causal field is that these kinds of events should characterize
the arousal of causal interest. Experimental evidence shows this
to be the case. In a recent review of both laboratory and field
studies investigating the issue of when people engage in spontaneous causal reasoning (as opposed to simply answering questions posed by experimenters), Weiner (1985) stated:
There is reasonable consensus in the reviewed research (6 of the 8 pertinent
publications) that search is elicited by an unexpected eventa win by
an underdog, a loss by a favored team, more or less profits than anticipated,
unexpected academic success or failure, unusual willingness or unwillingness to help, and inconsistent behavior, (p. 81)
Thus, one rarely seeks the cause of why one feels "average," why
traffic flowed normally, or why some accident is typical. To be
sure, the need for explanations can be aroused vis-a-vis normal
events. However, this is most likely to happen when those events
violate expectations and are therefore seen as unusual after all.
For example, we might want to know why traffic flowed normally
if major highway improvements were just completed, or why we
feel "average" after hearing about a death in the family.
2. It is important to note that not alt differences-in-a-background are seen as causally relevant. In particular, one must
distinguish between causes and conditions. To illustrate the distinction, consider diseases that are confined to particular segments of the population. A classic example, discussed by van
Fraassen (1980), is paresis, a disease that is restricted to people
with latent, untreated syphilis but that only occurs in a low percentage of such persons. It seems odd to say that untreated syphilis causes paresis. Rather, having the former suggests various
enabling factors that, when conjoined with some other event(s)
(e.g., virus, bacteria, etc.) produce the disease. Indeed, having
untreated syphilis may be a necessary condition for the disease,
but not its cause. As we argue below, the distinction between

particular causes and conditions is highly dependent on some
assumed context. However, as Mackie argued (1974), perceptions
of causes arc typically distinguished from conditions in that: (a)
Events are more causal than standing conditions (viruses are
more causal than, say, being old); (b) intrusive events are more
causal than events that generally occur; (c) something abnormal
or wrong is more causal than what is normal and right (e.g., the
car accident was caused by the person veering to the left, not by
the other person who drove straight ahead, cf. Kahneman &
Tversky, 1982).
Now consider how a shift in the causal field can change a
probable cause into a condition, with a corresponding reduction
in its causal strength. Recall our earlier example of a watch face
being hit by a hammer. In the original context, a hammer that
hits glass is an intrusive event that does not generally occur
whereas a defect in the glass is a standing state. However, in a
watch factory, the hitting of glass with hammers generally occurs
during testing whereas a defect is something abnormal or wrong.
Hence, in this context, the hitting of glass becomes more of a
condition or standing state and the explanation loses much of
its causal strength. Similarly, the defect explanation gains in causal
strength due to its abnormality in these circumstances. However,
it is possible to decrease the relative causal strength of the defect
explanation by increasing the abnormality of the testing procedure. For example, consider that all of the watch faces were being
broken in the testing procedure. This suggests that the force of
the hammer was too great (it being inappropriate for a testing
procedure), and its causal strength is thereby increased.
The importance of shifts in context on causes versus conditions
is not limited to physical causation. Indeed, consider the work
on actor versus observer differences in attribution theory (Jones
& Nisbett, 1972; see Watson, 1982, for a review). The basic finding is that whereas observers tend to attribute the causes of an
actor's behavior to the dispositional traits of the actor, actors
tend to attribute the causes of their own behavior to situational
variables. Note that from the actor's perspective, his or her own
traits are standing states, whereas situations are more intrusive
and hence more causal. From the observer's perspective, situations are seen as standing states and the dispositions of actors
are more intrusive. In fact, this asymmetry was noted by Jones
and Nisbett (1972) who stated, "For the observer behavior is
figural against the ground of the situation. For the actor it is the
situational cues that are figural and are seen to elicit behavior"
(p. 93). In our analysis, therefore, much should depend on the
familiarity of observers with particular actors (e.g., if person A
is well known to the observer, A's dispositions may be seen as
standing states), as well as the actor's familiarity with particular
situations. Indeed, although we know of no studies that have
addressed the latter prediction, some evidence exists supporting
the former (Goldberg, 1981; but see Watson, 1982). Moreover,
a shift in the causal field does not mean that previous causes
lose all their strength or that prior conditions are now seen as
strong causal candidates. Indeed, both absolute and relative causal
strength can vary considerably. In his review, Watson (1982) noted
that dispositions generally received more weight than situations
for both actors and observers, but the relative weight for dispositions was greater for observers than for actors.
The influence of shifts in the causal field also clarifies a peren-
HILLEL J. E1NHORN AND ROBIN M. HOGARTH
nial issue in scientific inquiryreductionism in causal explanation (causes at a molar level are different from those at a molecular level). That is, meaningful statements can be made about
causal links at different levels of understanding. As noted by
Cook and Campbell (1979):
Causal assertions are meaningful at the molar level even when the ultimate
micromediation is not known. . , . It seems meaningful to assert that,
when activated, a particular light switch A causes a lightbulb B to burn.
. . . We also accept as meaningful and open to criticism the causal assertions that the electric current flowing into the bulb caused the light
and that the heat of the filament in the lightbulb caused the light, (p. 32)
A useful analogy is to think of the causal field as the scope of

vision under a microscope such that shifts in the magnification
of the lens define different fields. Moreover, because relevance is
at least a difference-in-the-field, a cause at one level will not
necessarily be relevant at another. The microscope analogy makes
clear that the "appropriate" level of magnification depends on
one's purposes and the extent of one's knowledge of the phenomenon in question. Thus, a biochemist may see the causal
link between smoking and lung cancer as due to chemical effects
of tar, nicotine, and the like, on cell structure, whereas an immunologist might see the causal link as due to the suppression
of the immune system in controlling diseases in general.
3. In addition to defining differences/deviations and distinguishing causes from conditions, the causal field plays an important role in affecting the number and salience of alternative
explanations. That is, if one considers the causal field as containing the set of alternatives against which a particular causal
candidate is evaluated, then shifts in the field can highlight, or
de-emphasize, classes of alternatives. For example, imagine the
following scenario (Einhorn & Hogarth, 1982): Joe is a chemical
worker who contracts lung cancer and sues the company for
causing his disease. His lawyer argues that the cancer rate of
workers in this factory is nine times the national average for
workers in comparable industries. Note that the field or background in this argument is industries of a certain type and the
causal argument rests on there being a difference (i.e., higher
cancer rates) in this field. However, the lawyers for the chemical
company can shift the field by arguing that Joe has smoked cigarettes for years, comes from a family with a history of cancer,
and so on. Note that here Joe is contrasted against a different
field (involving people with certain personal habits and family
characteristics) in which cancer is not unusual and hence not a
difference in the field. Thus, Joe's case of cancer is now judged
as being less related to working in a chemical factory. Note that
the shift in the field changes the outcome (cancer) from an unusual event that requires a special causal explanation (e.g., working in a particular chemical plant), to a less unusual event that
does not require such an explanation.
The breadth of the field can also determine whether specific
alternatives are ruled in or out. In the above scenario, for example,
note how Joe's case would be strengthened if it could be shown
that the cancer rate in his factory was nine times the rate of
other chemical factories making exactly the same product. The
reason is that by narrowing the field to chemical plants making
the same product, the number of alternative explanations is reduced, thereby making the difference in the narrowed field more
causally relevant.
Finally, although both theoretical arguments and empirical

evidence attest to the role of the causal field in judging probable
cause, we note the importance of context and background effects
to other areas of cognitive psychology. Two immediate parallels
are the work of Simon and Hayes (1976) on "problem isomorphs," and that of Tversky and Kahneman (1981) on "decision
frames." In common with the effects of the causal field, these
phenomena illustrate how different representations of structurally
identical stimuli can lead to different responses (for a more general
review, see Hogarth, 1982). What makes these issues problematic
for study is that the meaning people give to stimuli is dependent
on prior knowledge and experience. Thus, whereas judgmental
processes are sensitive to differences and deviations (see e.g.,
Helson, 1964; Kahneman & Tversky, 1979), such deviations are
often defined relative to expectations, which are difficult to know
and control.
Cues-to-Causality
Many philosophers (including Hume, 1886/1964; and Mill,
1872) have noted that certain variables, such as temporal order
and contiguity, tend to indicate causal relations. However, the
mere presence of such variables does not guarantee that relations
will be perceived as causal. In our framework, we maintain that
once a causal field or context has been evoked, the gross strength
of a causal link between X and Y depends on various cues-tocausality. Moreover, we attach to the term cue the specific meaning accorded in Brunswik's psychology (1952; see also Campbell,
1966; Hammond, 1955). Thus, (a) the relation between each cue
and causality is probabilistic. That is, each cue is only a fallible
sign of a causal relation, (b) People learn to use multiple cues in
making inferences in order to mitigate against the potential errors
arising from the use of single cues, (c) The use of multiple cues
is often facilitated by the intercorrelation (redundancy) between
them in the environment. This both reduces the negative effects
of omitting cues, and aids in directing attention to the presence
of others, (d) Although multiple cues can reduce uncertainty in
causal inference, conflicting cues can increase it. As we shall see,
conflicting cues often require that distinctions be made with respect to the type of causal relation present (e.g., precipitating vs.
underlying cause).
In discussing the cues, we consider each at some molar level
although we are fully aware that they can be decomposed into
more molecular components. For example, contiguity can be
decomposed into temporal and spatial cues and temporal contiguity can be further divided into the time interval between
cause and effect and the regularity of the interval. The importance
of considering the elements of each cue will become apparent
as we proceed.
The cue of covariation is consistent with traditional notions
of cause advanced by both Mill (1872) and Hume (1886/1964)
and has also received considerable attention from numerous researchers (Alloy & Tabachnik, 1984; Crocker, 1981;Kelley, 1967,
1973; Nisbett & Ross, 1980). An important question centers on
why covariation need not be perfect for people to infer cause.
We consider this issue by building on ideas proposed by Mackie
(1965, 1974). This involves (1) asking what people mean when
they say that ^causes (caused or will cause) Y, and (2) showing
PROBABLE CAUSE
how such inferences are consistent with differing levels of covariation.

1. We begin by noting that when people judge that X caused
Y, they rarely mean that X is either necessary or sufficient for
Y. For example, if a fire inspector said that a short circuit (X)
caused a house fire (Y), it is clear that A"is neither necessary nor
sufficient for Y. Rather, it means that X, when conjoined with a
particular set of conditions, leads to Y (e.g., flammable material
near the short circuit, no sprinkler system, etc.). If we denote W
as representing flammable material near the short circuit and Z
as the absence of a sprinkler system, then define as the scenario
in which X, W, and Z occurred; that is, a = (ATPl WT\Z\B\ where
B - the causal field. Following Mackie (1965), we call a minimally sufficient condition for Y if
p(Y\a)= 1, but
(1)
Thus, given , Y always follows; however, given Y, a is neither

certain nor impossible. Moreover, a minimally sufficient scenario
is defined as one that would no longer be sufficient if any of its
conjuncts were deleted. It is also important to note that the sufficiency of a always presupposes some causal field, otherwise the
number of conjuncts in a would be very large. For example, it
is assumed that oxygen is part of the causal field because there
can be no fires without its presence. Indeed, because oxygen is
a standing state, it can be seen as a necessary condition and thus
part of the causal field in this context.
If we now consider the short circuit as the probable cause of
Y, note that X is necessary for the scenario a but that it is not
sufficient; that is,
p(X\a)= 1, but
p(a\X)*Q,\.
(2)
We can now say that X is an insufficient but necessary part of a

complex scenario (a), which is itself wnnecessary but sufficient
for y (INUS for short). Mackie notes that the statement, "The
short-circuit caused the house fire," implies that: (a) X is at least
an INUS for y; (b) X occurred; (c) the other conjuncts occurred;
and (d) all minimally sufficient conditions for Y not having X in
them were absent on the occasion in question.
2. To show how judgments of probable cause are consistent
with differing levels of covariation, consider a 2 X 2 table in
which X and X are crossed with Y and Y. An example is provided
by Figure 1, where the upper left-hand cell (Cell A) contains the
number (and frequency) of scenarios (m*) in which X conjoins
with various conditions to cause Y. The upper right-hand cell
(Cell B) contains scenarios in which A'occurs but the remaining
conditions are insufficient for y (am*)- This can result if there
are too few enabling conditions to allow the occurrence of Y; or
conditions exist that cancel or prevent X from causing y. For
example, imagine that X was a particular virus, Y was some
disease, and J^was an antibody that existed in some proportion
of the population. If W counteracts X, then the conjunction
(AT) W) does not produce the disease. We denote the frequency
of ot fnx scenarios as indicating the "conditionality of causation."
That is, the larger this cell, the more X is either sensitive to cancellation or the more it has to conjoin with other conditions to
produce Y.
Now consider the lower left-hand cell (Cell C), which represents
the scenarios ayns- This cell contains the minimally sufficient
Effect
Suspected
Cause
Y nX
X
(A)
Q
Y n >(
multiplicity
(C)
Y nX
conditionally
(B)
a
7n"x
(D)
Figure 1. Contingency table representing causal relations. (Y is an effect

and X is a possible cause. Overbars indicate that X or Y did not occur.)
conditions for Y in which X is absent and thus indicates the

existence of alternative causes of y. To contrast this source of
uncertainty with the "conditionality of causation," we consider
the frequency of rrvf scenarios as representing the "multiplicity
of causation." Indeed, as Mackie (1974) noted, in assessing
whether X caused Y, one implicitly asks the counter-factual
question, "Would y have occurred if X had not?" Note how this
question focuses attention on the multiplicity of causation by
asking one to imagine minimally sufficient conditions for Y that
do not contain X (i.e., ayfvf). We refer to this process as backward
causal inference in contrast to forward causal inference, where
attention is primarily directed toward the conditionality of causation (i.e., Cell B), as in "Will X lead to y?" Experimental
evidence concerning the effects of question type on causal judgments is discussed below.
Finally, the lower right-hand cell contains those scenarios in
which neither X nor Y occurs, a-fnx This cell is problematic
because it implicitly defines the context in which X and Y are
being considered. For example, consider the relation between
smoking (X) and lung cancer (y). Cell A contains the minimally
sufficient scenarios for y in which X occurs; Cell B contains
insufficient scenarios for y in which X occurs; Cell C contains
sufficient scenarios for y that do not contain X; and Cell D
contains the scenarios in which no smoking conjoins with other
conditions to lead to no lung cancer. In this latter cell, questions
arise as to what objects to include as having X and y. For example,
should we allow children to be included (after all, they neither
smoke nor have lung cancer)? If not, at what age are we to define
the relevant population? Should, for example, the analysis be
different for men versus women or for people with a history of
cancer in the family versus those without? What is important to
note is that covariation, by its very nature, implies a reference
population and thus some causal field against which the relation
should be assessed.
The above conceptualization has several general implications.
First, the traditional notions of necessity and sufficiency of causation can be represented by covariation in that a necessary cause
is one that is in all minimally sufficient scenarios for y (implying
that Cell C is empty); a sufficient cause is one that always conjoins
with conditions to produce Y (implying that Cell B is empty).
Second, the idea of considering scenarios of events as comprising
the "raw data" for judging covariation is consistent with the use
of this cue in singular cases; that is, where the data consist of a
single instance. For example, imagine that you feel ill today and
HILLEL J. E1NHORN AND ROBIN M. HOGARTH
recall that you ate fish last night. Because there is only one observation, covariation might be thought irrelevant. However, one
can imagine (or remember) scenarios involving illness without
eating fish; one can also imagine (or remember) scenarios in
which eating fish produces (or produced) no ill effects. Furthermore, even if one's imagination or experience is limited, one is
aware of the uncertainties associated with both the multiplicity
and eonditionality of causation. That is, one is aware of one's
incomplete knowledge regarding: (a) all the conditions that can
conjoin with X to produce Y; and (b) the alternatives that can
cause Y in the absence of X. Put succinctly, people know that
they do not have complete causal knowledge. Therefore, the
awareness of incomplete knowledge provides the basis for why
causal judgments are generally judgments of probable cause.
Many researchers have examined issues concerning judgments
of covariation (see e.g., Crocker, 198 l;Shaklee& Tucker, 1980).
A convenient way of summarizing the generally contradictory
experimental results in this literature is to represent covariation
judgments (denoted as Qc) as resulting from a weighted linear
combination of the cell frequencies from a 2 X 2 table. Thus,
Qc- 2 ft<?/
where 0, = arm, <h = frtr, 13 = yrw, It
weight given to <?,.
(3)
vf, and ft = the
Some early and influential studies (Jenkins & Ward, 1965;

Smedslund, 1963) showed that subjects' judgments were based
almost exclusively on am* (i.e., ft > 0, ft = ft = ft = 0). Ward
and Jenkins (1965), however, changed the way information was
presented to subjects (from sequential to intact displays), and
found different patterns of use (many subjects ignored disconfirming evidence, i.e., 02 ~ & = 0; many other subjects weighted
all cells). Einhorn and Hogarth (1978) noted that information
is frequently absent from real-world tasks, such that ft , ft > 0
but ft = ft = 0. Because much of the research on covariation
judgments compares intuitive responses with some statistical
measure of covariation (usually the correlation coefficient in
which the four cells are equally weighted), ignoring certain cells
is often taken as evidence that people have a poor intuitive concept
of correlation. However, numerous studies have also shown that
people can and do make use of all of the q/s that are available
(Alloy & Abramson, 1 979; Crocker, 1 98 1 ). A recent meta-analysis
speaks directly to this issue. Lipe (1982) has shown that when
subjects' responses were regressed onto the four data cells over
a number of studies, the weights for all four g$ were significant
and in the expected directions (ft , ft > 0; ft , ft < 0). Therefore,
the question is not whether people sometimes ignore relevant
information but rather, under what conditions is particular information ignored or not? For example, Crocker (1982) and Seggie and Endersby (1972) have demonstrated that the type of
question asked can greatly affect attention, and thus the weight
given to each cell (also see Arkes & Harkness, 1983).
It is important to note that many of the above studies are
confounded by whether subjects actually interpreted the experimental tasks as requiring judgments of covariation or causation.
This is particularly the case when the context of judgment in
covariation studies implicitly invokes causal reasoning, such as
problems involving the control of one variable over another, or
when relations are assessed between, say, symptoms and diseases.

(Incidentally, we note that warnings issued in introductory statistics courses that correlation does not imply causation also attest
strongly to confusion on this issue in untutored minds). Equation
3, however, suggests an important distinction between judgments
of covariation and causation. Specifically, because covariation
data convey information concerning the multiplicity and eonditionality of cause, the weights accorded the off-diagonal cells
(i.e., ft and ft) should differ in causal judgments depending on
whether multiplicity or eonditionality is of greater concern. For
covariation judgments, on the other hand, there should be more
equal allocation of attention to all cells.
Several studies are relevant to these issues. In an investigation
of correlational reasoning, Adi, Karplus, Lawson, and Pulos
(1978) examined and categorized responses made by 9th- and
12th-grade students to problems in which relations between
variables were expressed in both causal and coincidental contexts.
For the causal task, the modal class of subjects' responses involved
comparisons between data from Cells A and C, whereas the modal
class of responses for coincidental data involved all four cells.
The former result is particularly significant to the framework
developed here in that the causal question was framed in terms
of backward inference (i.e., ft > ft). In contrast, Schustack and
Sternberg (1981) asked subjects to make forward causal inferences
based on covariation data. By regressing subjects' judgments onto
the covariation data in the manner of Equation 3, they found
that the regression weight for Cell B was larger than for Cells C
and D (Cell A had the largest weight; all weights were significantly
different from zero). This result is also consistent with our
framework (i.e., in this case ft > ft). In another study, Waller
and Felix (1982) asked subjects to judge the same information
by answering both a causal and a correlational question. In accord
with the view that covariation is a fallible cue to causality, they
found a moderate but significant correlation between the two
types of judgment (r = .57). Finally, Lipe (1985) also compared
judgments of the same stimuli using causal (specifically backward
inference) and correlational questions in two conditions. In the
first, subjects made judgments on the basis of covariation data
embedded in content-rich scenarios; in the second, subjects were
given the same covariation data without any context. For the
content-rich condition, no differences in the use of data from
Cell C between causal and correlational questions were found.
However, there were significant differences in the use of data
from Cell C when scenarios were presented in abstract form.
In addition to the direction of inference, another factor is likely
to affect causal judgments via the covariation cue. Specifically,
recall that Cell D (ajrtf) is often problematic in that the definition
of what constitutes the absence of X and Y depends on some
assumed background or reference population. Consistent with
this is the fact that Cell D is rarely weighted as heavily as the
other three cells. However, there is a sense in which this cell is
important in that it raises the issue of the robustness of the X,
Y relation over other backgrounds or populations. For example,
imagine a high correlation between a particular diet and heart
disease in the U.S., but a lack of any such correlation in Western
Europe. We would expect that in judging the causal strength of
A" and Y, robustness will be seen as an important cue-to-causality.
Indeed, an interesting and unresearched question concerns
whether, or to what degree, people are willing to trade off co-
PROBABLE CAUSE
variation (in one setting) with lower covariation that is robust

over several backgrounds.
P(Y I X)
Temporal Order, Contiguity, and Similarity

We now consider the cues of'temporal order, contiguity in time
and space, and similarity of cause and effect. Temporal order
(Qr) is discussed first. The importance of temporal order in causal
judgments seems obvious as it is essential for determining which
of two variables in a relation is cause and which is effect. However,
whereas temporal order greatly affects causal judgments, it has
no role in formal probability theory. This suggests that there is
likely to be much conflict and confusion in probability judgments
when temporal order is salient. We illustrate this conflict by considering three issues: the confusion of joint and conditional
probabilities, causal versus diagnostic probabilities, and the confusion of inverse probabilities (Dawes, undated).
An important feature of temporal order is the speed and facility
with which it is usedoften without explicit awareness. This is
particularly the case in the interpretation of language; consider
the conjunction "and," which frequently implies temporal order
in everyday English (Strawson, 1952). However, the word "and"
is used in formal probability theory to denote a conjunction of
events, irrespective of temporal order. Therefore, how should
one understand a question like, "What is the probability of going
into the supermarket and buying some coffee?" If S is "going
into the supermarket," and K is "buying coffee," the probability
of S and K could be interpreted in several ways. Whereas a statistician would represent the question asp(SC\K) and ignore the
temporal meaning of "and," others may perceive the question
as requiring the conditional probability, /KA^IS). That is, one
could consider the sequence of events as being important and
condition the uncertainty of the later event (K) on the earlier
one (S). Indeed, to direct attention to the conjunction of the
events, it might be helpful to reverse S and K in order to break
the implied time order, that is, "What is the probability of buying
some coffee (K) and going into the supermarket (S)?" An experiment to test this assertion was performed using graduate
students who had taken at least one statistics course. One group
(i = 24) was asked to choose how they would represent "S and
K" probabilistically, while a second group (2 = 24) was asked
to represent "K and S" probabilistically. Subjects chose from
either p(ST\K), p(K\S\ p(S\K), or "none of the above." The
results showed an increase for p(ST\K) when the time order was
reversed (58% vs. 75%). Of further interest was the finding that
38% of the subjects in the first group chose p(K\S) (in accord
with the natural order of the events), whereas no subjects in the
second group chose p(K\S). Clearly, temporal order is an important cue that is difficult to ignore, even when it may be appropriate to do so.
A further example of the importance of temporal order in
probability judgments has been discussed by Tversky and
Kahneman (1980). They argued that people judge the probability
of data that have causal significance as being greater than data
that have diagnostic value, even when the information is equally
informative from a statistical viewpoint. For example, compare
the probabilities that: a girl has blue eyes (Y) if her mother has
blue eyes (X) versus a mother has blue eyes if her daughter has
blue eyes. Although p(Y\X) - p(X\Y), making X and Y equally
Cause
X
E f f e c t (Symptoms)
Y
P(X I Y)
Predict
back in time
Figure 2. Loop-like structure for confusion of inverse probabilities.
informative, many subjects judge the causal/forward probability

as greater than the diagnostic/backward one (i.e., p[Ir|Ar] >
Pl-Yl^])- Tversky and Kahneman also demonstrated that when
temporal cues are ambiguous, the causal interpretation of data
tends to receive disproportionate weight such that elementary
rules of probability can be violated.
The above discussion suggests that temporal order has a
straightforward effect on probability judgments. However, the
situation is more complex; in particular, people often have difficulty in distinguishing between p(Y\X) and p(X\Y) when X
and Y are not equally informative. This "confusion of inverse
probabilities" (Dawes, undated) seems to contradict the importance of temporal order as a cue that unequivocally leads to
causal probabilities being judged larger than diagnostic ones.
However, we now discuss when the confusion is most likely to
occur and argue that it is due to the ambiguity of temporal cue.?
in the task. To illustrate, consider the case of mammography
and breast cancer (Eddy, 1982). For women with breast cancer (C), the probability of a positive mammogram (M) is .79
(p[M\C) = .79). Eddy reported that most physicians misinterpret
the statements about the test and estimate the probability of
breast cancer given a positive test to be approximately .75
(p[C\M] = .75).
When asked about this, the erring physicians usually report that they
assumed that the probability of cancer given that the patient has a positive
X-ray . . . was approximately equal to the probability of a positive Xray in a patient with cancer.. . .The latter probability is the one measured
in clinical research programs and is very familiar, but it is the former
that is needed for clinical decision making. It seems that many if not
most physicians confuse the two. (1982, p. 254)
Under what conditions is the confusion most likely to occur?

From our perspective, first note that cancer causes the positive
test result and not vice-versa. Therefore, denote cancer as the
cause (X) that temporally precedes the test as an effect (Y). However, the test result is known first in time and it is used to predict
the cause, X. The peculiarity here is that the test predicts a prior
state (cancer) that presumably existed before the test and that
caused the test result itself. Hence, this situation has a looplike
structure that makes the temporal order of events particularly
confusing. This looplike structure is shown in Figure 2.
10
Note that the causal probability, p(Y\X), is not predictive in

the usual sense; that is, whereas having cancer predicts the test,
this hardly seems relevant to the practical matter. On the other
hand, the test (which is known first in time), is used to predict
the disease that preceded it (because the disease causes the test
result). Therefore, there are two temporal cues in problems of
this type: causes precede effects, but effects are seen first and are
used to predict their prior causes. Under such conditions, confusion of inverse probabilities is quite understandable. To make
matters worse, in statistical jargon, the "diagnosticity" of a datum
(D) to some hypothesis (//) refers top(D\H). However, in ordinary
discourse, the diagnostic value of D for H means how well D
predicts H (i.e., p[H\D]). Therefore, to avoid such confusion, it
would be useful when discussing p(D\H) to say that D is symptomatic rather than diagnostic ofH,
The next cue to be considered is contiguity (Qc). The extent
to which events are contiguous in time and space is an important
cue to causality, as evidenced in the work of Michotte (1946).
Siegler has also shown that for young children (5-6 years old),
temporal contiguity is a very strong cue for inferring causality
(Siegler, 1976; Siegler & Liebert, 1974). Moreover, these studies
show that older children are less dependent on contiguity alone,
a belief was erroneous, and many cases exist in which physical

resemblance has been misleading. For example, Nisbett and Ross
(1980) pointed out that physical resemblance was the cornerstone
of a medical theory called the "doctrine of signatures" whereby
cures for diseases were thought to be marked by their resemblance
to the symptoms of the disease. Thus, the curing of jaundice was
attributed to a substance that had a brilliant yellow color (see
also Shapiro, 1960, and Shweder, 1977). However, whereas physical resemblance may be a cue of low validity, it does not mean
it has no validity. For example, consider how the physical resemblance between offspring and adults implicates particular adults
as the causal agents for those offspring.
At a second level, one can consider similarity based on the
congruity of the length and strength of cause and effect. That is,
if the effect of interest is large (i.e., is of substantial duration
and/or magnitude), people will expect the cause(s) to be of comparable size. For example, the germ theory of disease advanced
by Pasteur must have seemed incredible to his contemporaries
in that people were asked to believe that invisible creatures caused
death, plagues, and so on. In the same way, it is equally difficult
for many to believe that billions of dollars spent on social programs in the 60s and 70s could have had little or no effect, or
being able to make use of multiple cues. Nevertheless, contiguity

remains an important cue for directing attention to contingencies
between variables, and such contingencies may then be considered
as to their causal significance. When temporal and/or spatial
contiguity is low (or temporal contiguity is erratic), inferring
causality becomes more difficult. That is, in the absence of contiguity, relations are hard to develop unless one uses intermediate
causal models to link the events (see the next section). For instance, the temporal gap between intercourse and birth requires
some knowledge of human biology and chemistry to maintain
the links between those events.
Whereas the above seems obvious, it should first be noted that
the sign and form of the relation between contiguity and judgments of probable cause depend on one's knowledge of a particular situation. For example, imagine that one has advertised
a product and sales go up dramatically the next day. If one believes
that advertising works by a gradual diffusion process, the sales
increase may not be attributed to the ad precisely because the
events occurred too close in time. On the other hand, contiguity
could be seen as monotonic with causal strength by others with
different theories, or by the same person in another context. Second, contiguity is particularly prone to conflict with other cues,
which can then require compromises of several sorts. For example, if increased contiguity in time leads to greater causal
strength, this cue can conflict with temporal order when cause
and effect occur simultaneously. If temporal order has priority,
the relation between contiguity and judgments of probable cause
is discontinuous.
The cue of similarity between cause and effect (Qs) is of importance to judgments of probable cause. To discuss this, we
need to consider the meaning of "similarity" at several levels (cf.
Tversky, 1977). First, there is the long-standing notion that cause
that long-term and complex effects like poverty can have shortterm and simple causes.
and effect should exhibit some degree of physical resemblance.

Mill noted that this is a deeply rooted belief that, "not only
reigned supreme in the ancient world, but still possesses almost
undisputed dominion over many of the most cultivated minds"
(cited in Nisbett & Ross, 1980, p. 115). Mill thought that such
Determinants of Gross Strength

Causal chains. Although the causal field and cues-to-causality
are important in determining the gross strength of an explanation,
are other factors involved? Several authors have argued that attributions of physical causation depend on the perception of a
"generative" force that links causes to effects (e.g., Shultz, 1982).
This implies a strong mechanistic conception of causation in
which effects result from a physical transmission of causal "energy" from A" to Y. The importance of this idea is that it implies
that events need to be linked via a causal chain so that the force
can be transmitted from one link to the next. Conversely, if no
connecting links can be made between X and Y, the strength of
a causal explanation will be zero. To illustrate, how likely is it
that sunspots (X) cause price changes on the stock market (Y)1
(Similar concerns, having to do with the relation between sunspots and business cycles, were discussed by Jevons, 1884, and
more recently by Sheehan & Grieves, 1982). In order to link
these events, one needs a causal chain that bridges the spatial
and temporal gap (assume that temporal order and covariation
both point to a causal relation and that prices are affected 6
months after sunspot activity). If someone cannot construct a
causal chain to link X and Y, we argue that this is sufficient for
zero gross strength. Put differently, the existence of a causal chain
is necessary for nonzero gross strength regardless of the levels of
the cues. Therefore, causal relations must be "achieved" in the
sense that prior knowledge and imagination are needed to construct a schema, scenario, or chain, to link cause and effect. On
the other hand, a scenario that does link X and Y may be so
weak as to result in zero or near-zero gross strength. For example,
imagine the following causal chain: sunspots affect weather conditions, which affect agricultural production, which affects economic conditions, which affect profits, which affect stock prices.
However, before explicating the implausible nature of this chain,
PROBABLE CAUSE
Congruity
High
High
Contiguity
Low
Low
Few links
needed (if
any) .
(1)
Links needed
to amplify or
dampen p r o c e s s .
(2)
Links needed
to bridge
contiguity gap.
(3)
Links needed to
bridge contiguity
gap & amplify or
dampen p r o c e s s .
(4)
Figure 3. Effects of contiguity and congruity on constraining the causal

chain.
we need to consider the role of the cues and background in the

construction of such chains.
If a causal chain is necessary for linking cause and effect, what
rules govern its construction? Of particular importance is the
fact that the cues-to-causality constrain the possible causal chains
that can be imagined. This is most obvious in considering the
cue of temporal order. That is, if A7 does not temporally precede
7, it cannot be a cause of Y. However, the cues of contiguity in
time and space, and congruity (i.e., similarity of length and
strength of cause and effect) also constrain the linkage process.
First, consider contiguity in time and space and recall the sunspots
and stock prices example. In order to link these events at all, it
was necessary to bridge the spatial and temporal gap by positing
a change in weather, and so on. However, imagine that price
changes occur immediately after sunspot activity (rather than 6
months later). Note that the high contiguity in time precludes
the weather-economic conditions scenario because that scenario
requires a time delay between X and Y. Hence, in order to link
X and y, another scenario is necessary that is consistent with
high contiguity in time (and low contiguity in space). Similarly,
imagine taking up smoking and getting lung cancer the next
week. It seems highly unlikely that smoking is the cause because
a causal chain with high contiguity in time is not easy to construct. Now consider incongruity of cause and effect; that is,
small causes and big effects or, big causes and small effects. To
account for incongruity, the linkage process must involve some
type of "amplification" in the first instance, and "dampening"
in the second. For example, reconsider the incongruity between
germs and illness or death, as exemplified by the germ theory of
disease. Without knowledge of how germs enter the body, multiply, disseminate throughout the system, and so on, it would be
difficult (and was difficult at the time), to accept such a theory.
Figure 3 summarizes the effects of the cues on constraining
the type of causal chain necessary to link X and y, where the
cues of contiguity, and congruity are shown at each of two levels.
Constrast, for example, Cells 1 and 4. In Cell 1, the two cues
have high values such that one would expect a model requiring
few links, if any. Recall, for instance, the example of a hammer
hitting a watch face and causing the glass to break. In Cell 4,
both cues are low and therefore place strong demands on the
type of causal chain that can meet specifications.
There is an important connection between the construction
of a chain that meets the constraints of the cues and the strength
11
of that chain. To explicate the nature of this relation, we first

denote QL as the strength of the causal chain that links X with
y. If one cannot construct a linkage between X and Y, QL = 0.
If a causal chain can be constructed, then QL measures the
strength of that chain. In order to discuss the factors that determine <2u> recall the sunspot-stock price causal chain given above:
sunspots weather > agriculture
economy > profits > price changes
Note that whereas this scenario meets the constraints in Figure
3, it requires a long chain of events to bridge the temporal and
spatial gap (we would argue that congruity is not a problem
here). Let / be defined as the number of links in the constructed
chain, where / is equal to the number of intermediate variables
in the chain (jcs) plus one. Thus, in the above chain there are
five links. What determines the strength of each link, and, how
do the individual link strengths combine to determine the strength
of the whole chain (i.e., d.)?
Before individual link strength can be determined, the constraints of temporal order, contiguity, and congruity must have
been met. Therefore, we argue that it is the covariation between
variables at each link that determines link strength. For example,
the link between sunspots and weather is due to the covariation
of X with X i , the weather-agricultural production link is due to
the covariation of Xi with xi, and so on. Now consider how the
individual links combine to determine the strength of the causal
chain, QL (cf. Costner & Leik, 1964). To examine this, let Cj be
the absolute size of the covariation of variables at they'th link in
the chain 0 = 1 , 2 , . . . , /) and assume that the relation between
the CjS and QL is as follows:
QL = II Cj.
j-i
(4)
By assuming a multiplicative function, Equation 4 implies the

following: (a) The strength of a chain is at best equal to its weakest
link. Thus, if c, has the lowest link strength in the chain, QL is
at best equal to c{ (if all other links are I). Indeed, note that if
any c, = 0, QL = 0, regardless of the strength of the other links,
(b) Note that because Cj is between 0 and 1, the longer the chain
(/), the lower the strength of the whole chain (as one is multiplying
fractions).1 This makes clear why the sunspot-stock price scenario is implausible on the basis of QL. Because / = 5 in that
scenario, even if, say, C, = .6 for all five links, QL is only .08.
Therefore, long chains are generally weaker than short ones.
However, Equation 4 also implies that if all the cjs are high, the
length of the chain is less important in affecting QL. Indeed, if
Cj = 1 for ally, the length of the chain is irrelevant. For example,
consider the following scenario:
1
We can allow for the possibility that each ct is raised to a power such
that the addition of elements to a chain increases, rather than decreases,
judgments of gross strength (cf. Einhorn, 1985). Indeed, some evidence
exists (Slovic, FischhofT, & Lichtenstein, 1976) that longer scenarios are
sometimes judged as more likely than shorter ones. Also see our discussion
of conjunction fallacies in the section on Multiple Causation and Conjunction Fallacies.
12
Table 1
Variables and Assumed Relations in Judgments
of Gross Strength
Variable
Definition
QB
Degree to which A" is a difference-in-the-background

(0<<2Bl)
Covariation of X with Y (0 < Qc < 1)
Temporal order of X and Y (0, 1)
Contiguity in time and space
Similarity of cause and effect
Congruity of duration and magnitude
Physical similarity (>s)
Causal chain strength" (0 < QL < 1)
Cc
Qr
Qc
Qs
QL
1
QL = 0 if no construction, H Cj otherwise. / = the number of links in
y-i
the causal chain due to contiguity and congruity of cause and effect;
Cj = covariation at the/th link of the chain.
Mrs. Jones, a middle-aged lady, was looking for her seat at a football
game. While trying 10 ease past some other spectators on a steep staircase,
she lost her balance and fell over. She hit another spectator, who was also
off balance. This person, in turn, fell on someone else who unwittingly
pushed Mr. Smith, a 70-year-old man, against some iron railings. Mr.
Smith broke a leg in the incident. How likely did Mrs. Jones' fall cause
Mr. Smith to break his leg?
We designate such scenarios as "domino-chains," for obvious

reasons. The basic characteristics of these chains are that all QS
are high, / is high, and QL is strong. On the other hand, when
any cj is moderate or low, even small / leads to a low value of
QL. (c) Equation 4 suggests that the level of the causal background
can also affect QL by changing the number of links in the causal
chain. To see this, recall the child's game of repeatedly asking
"Why?" after answering a question. This game highlights the
fact that the links in a causal chain are subject to further analysis
at a more reductionistic level. As pointed out earlier, the causal
field implies a particular level of explanation, and as such, it
constrains the type and length of the causal chain used to connect
X with 7. In particular, as each link is decomposed into further
components, the number of total links {/) increases. However,
unless the c/s also increase (which is possible, as the phenomenon
may be better understood at lower levels), QL could be seen to
decrease. Clearly, the rates at which /and Cj increase (if they do)
determine whether QL is higher or lower at different levels of the
causal background.
The above analysis deals with chains underlying judgments of
probable cause in which the links between causes and effects are
direct (see also Brickman, Ryan, & Wortman, 1975). Causal
knowledge about a substantive domain, however, is often represented in memory by chains involving notions of indirect as
well as direct causal links (see, e.g., Axelrod, 1976 and Trabasso
& Sperry, 1985). We believe that integrating the framework developed here with problems of how causal knowledge comes to
be represented in memory is a fruitful path for research.
A combining rule for gross strength. As should be clear from
the preceding sections, judgments of gross strength are composed
of numerous factors. To recapitulate, Table 1 lists these factors,
including some of the assumed relations between them. We have
also specified the causal field or background by a variable QB

for reasons we explain below.
Evidence concerning the manner in which the variables listed
in Table 1 combine in affecting gross strength can be found in
many areas of psychology. Moreover, this evidence can be conceptualized as illuminating two key issues: (1) What are the necessary conditions for perceiving causal relations? (2) How are
conflicts between the cues resolved?
1. Three variables can be considered necessary to inferring
cause in that if any is zero, so is gross strength. These are: QB,
Qr, and QL. As discussed previously, if a variable is not perceived
as a difference-in-a-field, it will not be deemed causally relevant.
Similarly, temporal order requires causes to precede effects. Now
consider the evidence regarding the necessity of QL. Several researchers have demonstrated that despite the presence of strong
cues-to-causal ity, people will still fail to make causal attributions
if they cannot perceive a causal link (i.e., QL = 0). For example,
whereas Michotte (1946) showed that subjects perceived causal
relations when the movement of objects after contact was congruent with prior trajectories and/or positions, this was not the
case when one object touched the other and the latter changed
color, got larger, or rose. Consider further the literature demonstrating the limits of classical conditioning. For instance, although Watson and his colleagues were able to condition little
Albert to fear rabbits by pairing the appearance of a rabbit with
that of a large noise, they could not produce the same effect
when the rabbit was replaced by a block of wood or a cloth
curtain (Nisbett & Ross, 1980, p. 104).
We note that the classical conditioning paradigm implicitly
makes use of the cues-to-causality, especially contiguity (Qo),
temporal order (Qr), and covariation (Qc), in establishing "cognitive maps in rats and men" (Tolman, 1948). Thus, it is significant that Garcia and his colleagues (Garcia, 1981; Garcia,
McGowan, Ervin, & Koelling, 1968; Garcia. McGowan, &
Green, 1972) have demonstrated that rats can learn to associate,
after one trial, distinctive tasting food and a gastrointestinal illness
(induced by X-rays) several hours later, that is, despite a large
temporal gap. However, the rats do not learn to associate a differential shape of food with the illness, when the taste is familiar.
It seems odd to think of rats judging probable cause. However,
this evidence is relevant to the evolutionary interpretation of
Kant's (1781/1982) notions of a priori causal categories that has
been espoused by Campbell and others (see, e.g., Cook & Campbell, 1979). Moreover, in a provocative review of the literature
on classical conditioning, Seligman (1970) suggested that organisms are differentially prepared to learn different types of relations. Specifically, stimuli can be arranged along a continuum
reflecting the ease with which organisms can learn to perceive
contingencies. Thus, animals may be "prepared" to learn some
relations but "contra-prepared" to learn others. In essence,
QL ~ 0 for contra-prepared relations. However, the extent to
which the source of this phenomenon is biological and could be
overcome by relevant environmental contingencies is controversial. (For an extensive review of animal studies of covariation
assessment, see Alloy & Tabachnik, 1984, pp. 130-139.)
2. If QB, Or. and Q\. are all necessary for the perception of
a causal relation, they should combine multiplicatively in affecting gross strength because if any variable is 0, so is gross
strength. On the other hand, because the other variables in Table
PROBABLE CAUSE
13
I are not necessary, their effect on gross strength should be compensatory. Indeed, our review of the literature (see below), suggests
the following functional form for relating gross strength to the
cues,
ture of our model helps to integrate the views of those who have
argued both for and against the priority of cues-to-causality. On
the one hand, researchers such as Shultz (1982) have presented
convincing evidence that children will ignore certain cues-to-
s(X, Y) = QjQBQL(\cQc + Xs<2s),
causality when these conflict with a perceived link between cause

and effect (QL > 0). Shultz interpreted his results as indicating
that the rules governing attributions of causality give precedence
to notions of "generative transmission" over the (Humean) cuesto-causality. On the other hand, we believe that posing these
issues as a controversy between the Humean or "regularity" approach to causality and the "generative" theory is misplaced.
First, we stress the notion that the cues-to-causality are not rules;
they are imperfect, probabilistic indicators (Brunswik, 1952;
Campbell, 1966). Thus, whereas the presence of the cues can
strengthen judgments of probable cause, they do not establish it
as such. Indeed, the discovery of new causal mechanisms often
involves thinking in ways that counteract some of the cues. Second, because the cues play a crucial role in delimiting the particular models of "generative transmission" that people use in
causal reasoning, it cannot be said that the cues per se are disregarded in favor of the latter. Third, although our approach
puts considerable emphasis on generative transmission via QL,
it also allows for trade-offs if QL is above a threshold level. Therefore, if the Humean principles are given a probabilistic interpretation and gross strength is comprised of both compensatory
and noncompensatory aspects, we see considerable complementarity between the two approaches.
Although generative links are represented in Equation 5 by
QL, QL can also be taken to represent the strength of any theory
(whether acquired via experience or innate) that people bring to
bear in causal reasoning. Thus, QL can be likened to a schema
that the person brings to bear in a particular situation (Abelson,
1981; Schank & Abelson, 1977). Such schemas are usually considered to be based on past experience and world knowledge,
and have been studied from a variety of viewpoints. (See also
the growing literature on "story grammars," Stein & Glenn, 1979;
Trabasso et al., 1984; Trabasso & Sperry, 1985.) As the reader
will note, in addition to acting as a necessary condition in Equation 5, QL is represented in a form that allows trade-offs with
both covariation and physical similarity. This form of equation
was chosen to reflect the considerable literature summarized by
Nisbett and Ross (1980) under the heading of "theory driven vs.
data driven" judgments of covariation. In short, Nisbett and Ross
show that there is much evidence to suggest that people will often
allow their prior conceptions of relations (i.e., QL) to outweigh
considerations of more data-driven estimates of covariation (i.e.,
(5)
where the As are attentional weights. Note that whereas contiguity

and congruity are not explicitly shown in the equation, they are
included via their effects on QL. We now consider further evidence
supporting the form of Equation 5.2
Several studies have examined how the cues trade off. In developmental psychology, for example, Sedlak and Kurtz (1981)
have reviewed research on three cues: temporal order, covariation,
and contiguity. In summarizing the results of studies in which
these cues conflict, they stated:
When a temporal delay is introduced, only children 8 years of age or
older offer reliable evidence of covariation. Specifically, the covariation
use of younger children deteriorates when the covarying cause is temporally non-contiguous with the effect and they encounter problems in
(1) rationalizing the delay, or (2) recognizing the constancy of the delay.
Where these tasks are made difficult, a temporally contiguous (albeit
inconsistent) cause is preferred. Thus we can tentatively rank order the
three Humean principles. It appears that temporal antecedence supercedes
spatial contiguity (Bullock & Gelman, 1979), and that temporal (and
spatial) contiguity may sometimes outweigh the covariation factor in
children's judgments (Mendelson & Shultz, 1976; Siegler, 1975; Siegler
&Liebert, 1974). (1981, p. 763)
Furthermore, in studies that have contrasted covariation with

similarity, Shultz and Ravinsky (1977) found that 6-year-olds
were unwilling to label dissimilar factors as causes, even in the
presence of systematic covariation. On the other hand, older
children (10-12-years old) favored covariation over similarity.
They also found that the relative weights given to similarity versus
temporal contiguity varied according to age; similarity outweighed contiguity for 6-year-olds, but older children favored
contiguity over similarity. In studies involving adult subjects,
Einhorn and Hogarth (1983) varied covariation, contiguity, and
similarity in a factorial design. In one experiment, covariation
traded off with similarity but there was no effect for contiguity.
However, when contiguity was made more salient in the experimental stimuli, it did trade off with covariation and similarity,
thereby indicating the effects of attention as captured by the Xs
in Equation 5. Brown (1984, 1985) also investigated the role of
causal cues in accounting tasks. He found that whereas temporal
order (Qt) was not compensatory, covariation (Qc) was. Furthermore, when he varied covariation and similarity, they traded
off in affecting causal judgments. Finally, it is interesting to note
that conflicts between similarity (congruity) and contiguity can
sometimes be resolved by distinguishing between "precipitating"
and "underlying" causes. The former is generally some action
or event that is high in temporal and spatial contiguity but low
in similarity of length or strength with the event. The latter is
generally based on high similarity of length and strength, with
contiguity being less important. For example, the precipitating
cause of World War I was an assassination in Sarajevo, but the
underlying cause(s) were economic upheaval, German nationalism, and so on.
An important aspect of Equation 5 is that it includes both the
necessity of certain cues and trade-offs amongst others. This fea-
Qc)- Thus, even if \cQc - 0, strong causal relations can still be

perceived. In analogous fashion, Shweder (1977) has summarized
much literature that goes under the heading of "illusory correlation." Equation 5 allows for this phenomenon as judgments of
* If one assumes that AC + ^s = I, this implies that gross strength is

made up of two major parts: a multiplicative component (Qr-Qa-QL)
and a weighted average (M2c + ^sQ's). Furthermore, if 0 ^ Q's <. 1, the
gross strength of A' for Y will be appropriately bounded between 0 and
1; that is, O^sfJf, Y)& 1.It should be stressed that the value ofEquation
5 lies in integrating various theoretical and empirical findings. We are
well aware of other specific forms for expressing gross strength as a function
of the cues and the causal background.
14
HFLLEL J. EINHORN AND ROBIN M. HOGARTH
Pregnancy
Yes
Intercourse
Yes
No
No
20
80
100
95
100
25
175
200
Figure 4 Data matrix for hypothetical intercourse-pregnancy experiment.
gross strength can be high if the lack of covariation is compensated

for by prior theory (i.e., QL) and/or physical similarity between
the variables (i.e., Q's),
The tension between theory and data (i.e., QL and Qc) that is
captured in Equation 5 is also helpful for understanding a perennial methodological issue: spurious correlation. The existence
of this concept suggests that some correlations are more (or less)
causally related than others, and thereby raises the issue of how
to tell the difference (cf. Simon, 1954). For example, consider
the correlation between the number of pigs and the amount of
pig iron (Ehrenberg, 1975). Such a correlation seems spurious
when the common causal factor, "economic activity," is considered. On the other hand, consider the correlation between amount
of rain and number of traffic accidents in a city over the course
of a year. Such a correlation does not seem spurious (or at least,
less spurious). What is the difference between these two cases?
If one uses Equation 5, the spuriousness of the correlation
between pigs and pig iron becomes apparent. That is, although
the covariation cue points to a causal relation, temporal order
cannot be used to specify which variable is cause or effect
(Q-f = 0); in addition, most people cannot construct a causal
chain linking pigs to pig iron (QL = 0). Indeed, the judgment
that the relation is spurious is made easily. Now consider the
second case: Assuming that there is a statistical relation (Qc >
0), note how the other cues reinforce that link. The temporal
order of rain and accidents is clear; it is relatively easy to construct
a chain linking the controllability of cars on slippery roads (as
long as there is high contiguity in time and space between rain
and accidents); in addition, rain is not a standing state or condition (QB= 0). There seems less doubt that the correlation is
real.
When cues-to-causality conflict, spurious correlation is not
the only outcome; for example, alow or zero statistical correlation
could mask a true causal relation (Einhorn & Hogarth, 1982).
To illustrate, imagine that we were ignorant of the cause of birth.
However, it has been suggested that sexual intercourse is related
to pregnancy, and the following experiment was designed to test
this hypothesis: 100 couples were allocated at random to an intercourse condition, and 100 to a nonintercourse condition. As
indicated in Figure 4, 25 women became pregnant, and 175 did
not. In light of our current knowledge (but unknown to our
hypothetical selves), we can state that the 5 people in the nointercourse/yes-pregnancy cell represent "measurement error,"
that is, faulty memory in reporting, lying, and so on. Because
the statistical correlation is small (r = 0.34), we might
question whether the hypothesis is worth pursuing. Indeed, if
the sample size were smaller, the correlation might not even be
significant. Moreover, even with a significant correlation, r2 =
0.12, which is hardly a compelling percentage of the Y variance
accounted for by X.
There are two important implications of this example. First,
whereas statistics texts correctly remind us that correlation does
not necessarily imply causation, the imperfect nature of this cueto-causality is also reflected in the statement: Causation does
not necessarily imply correlation. We have somewhat facetiously
labeled examples of the latter as "causalations," giving them equal
standing with the better-known and opposing concept of spurious
correlation. Second, causalation demonstrates that sole reliance
on a single cue, such as covariation, is inadequate for understanding causal relations (see also Sheehan & Grieves, 1982).
Indeed, as Equation 5 shows, gross strength can be nonzero even
if Qc is low. Therefore, the use of multiple cues highlights the
role of judgment in interpreting data (see also Simon, 1954) and
Equation 5 provides a basis for understanding how this is done.
Most of the examples in this article have involved judging
probable cause in the physical as opposed to the social domain.
Our framework, however, applies to both. The main difference
between physical and social causation lies in the fact that events
in the social domain are assumed to result from goal-oriented
factors of some type. Thus, outcomes are frequently explained
by unobservable constructs such as motivations, intentions, or
goals of actors. Moreover, the presumed existence of such causes
affects the cues-to-causality such that gross strength is increased.
For example, note that a motive or intention exists prior to the
taking of action and is thus consistent with the temporal order
cue. Furthermore, high contiguity in time and space, as well as
high congruity between motives and actions, results in a short
and direct causal chain linking outcomes to goals, intentions, or
motives. Thus, in distinguishing between causal candidates,
studies of probable cause judgments in the social domain typically
depend on two cues, covariation and the salience of X as represented by the degree to which it is a difference-in-the-background (i.e., ?B)- Indeed, both covariation and salience have been
the subject of intense investigation in the literature on social
attribution (Harvey & Weary, 1984;Hastie, 1983). Kelley's( 1967,
1973) influential analysis of variance (ANOVA) model, for example, was constructed solely on the basis of covariation patterns.
Studies using examples of physical causation, on the other hand,
typically involve varying the other cues (e.g., contiguity, similarity,
etc.). It is important, however, not to make too much of the
distinction between physical and social causation because experience of the real world typically involves both. This is evident,
for example, in studies of text comprehension (e.g., Trabasso &
Sperry, 1985), in which to understand stories, people must be
able to reason causally about both social and physical events
(e.g., in inferring both a protagonist's motives and the physical
means used to achieve them).
Role of Specific Alternatives
The importance of specific alternative explanations in assessing
the strength of causal candidates has been stressed by many researchers. Of particular importance is the work of Campbell and
colleagues on threats to "internal validity" (Campbell, 1969;
Campbell & Stanley, 1963; Cook & Campbell, 1979). Indeed,
PROBABLE CAUSE
the assessment of internal validity, whereby one asks what factors

other than X could have produced Y, is a basic component of
causal inference (see also Mackie, 1974). Similarly, attribution
theorists (e.g., Kelley, 1973) have discussed the discounting of a
causal explanation on the basis of plausible alternatives. From
our perspective, specific alternatives are important for two reasons
beyond their obvious role in lowering the necessity of X for Y
(by increasing the frequency of Cell C). First, specific alternatives
allow one to use the cues-to-causality to assess their plausibility
(i.e., determine their gross strengths). Hence, the content knowledge associated with a specific alternative affects its gross strength
(via Equation 5), which in turn can be used to discount the
original explanation. Second, the power of specific alternatives
lies in the possibility of replacing the current explanation rather
than simply discounting it. That is, because the goal of causal
inference is to find some explanation for the observed effects,
the discounting of an explanation by specific alternatives still
leaves one with the question, "If X did not cause Y, what did?"
Therefore, although the testing of hypotheses via comparison
with alternatives is part of the causal inference process, one is
still left with finding a plausible explanation. In fact, the distinction between testing hypotheses and searching for better ones
can be likened to the difference between a "disconfirmation"
versus "replacement" mode of inference. The replacement view
is consistent with the Kuhnian notion that theories in science
are not discarded, despite evidence to the contrary, if they are
not replaced by better alternatives (Kuhn, 1962). Indeed, the
replacement view is equally strong in everyday inference. A useful
analogy might be the following: How many people would read
detective stories if the author only revealed who did not do it?
The notion of a discounting process provides an important
connection between research on causal reasoning and probabilistic inference since much of the latter concerns how people
update their present beliefs on the basis of new information (see,
e.g., Anderson, 1981;Edwards, 1968; Einhorn& Hogarth, 1985;
Fischhoff & Beyth-Marom, 1983; Robinson & Hastie, 1985;
Shanteau, 1970). Therefore, the discounting of a causal explanation by specific alternatives involves a process of updating one's
(causal) beliefs on the basis of new, and in this case, negative
evidence (i.e., alternative explanations). Moreover, the idea that
judgments of probable cause result from an updating process
focuses attention on the dynamic processes underlying judgment
and choice (Hogarth, 1981). We now discuss several issues that
follow from considering judgments of probable cause as resulting
from an updating process.
To begin, consider the following model of the updating process:
= s(X,Y)
2 w(z*,n
(6)
k- 1
where S^X,Y) is the net strength of explanation X for Y after k

alternatives are discounted (0 <. Sk[X, Y] < 1); s(X, Y) is the gross
strength ofX for Y; s(Zk,Y) is the gross strength of alternative
Zt for Y; and wk is the weight given to the Ztth alternative.
Equation 6 can be interpreted in a variety of ways; for example,
an anchoring-and-adjustment model in which the gross strength
of an explanation serves as the anchor and adjustments are made
on the basis of the weighted strengths of the alternatives; a
"Bayesian-like" update model in which gross strength is analo-
15
gous to one's prior opinion and the diagnosticity of new evidence

(i.e., the gross strength of alternatives) is incorporated to yield a
posterior judgment (net strength); and so on.
There are several important aspects of Equation 6 that link it
to our earlier discussion and to the literature on how people
update their beliefs on the basis of new evidence.
1. Much theoretical and empirical work suggests that the updating process can be conceptualized via a sequential anchoringand-adjustment process (e.g., probabilistic inference, Edwards,
1968; attitude change, Fishbein & Ajzen, 1975; impression formation, Hovland, 1957). That is, one's current belief serves as
an anchor and adjustments to the anchor are made on the basis
of new evidence. Indeed, Jones (1979) has suggested such an
anchoring-and-adjustment process in the formation of causal
attributions and evidence for this has been found when people
were asked to re-assess the strength of their original causal hypothesis after seeing alternatives of varying gross strengths (Einhorn & Hogarth, 1983). From our perspective, the general updating model shown in Equation 6 can accommodate such a
process because it can be written in a sequential form, as follows:
Sk(X,Y) = S^,
wks(Zk,Y).
(1)
Hence, the updating model in Equation 6 is general enough to

capture both the idea of net strength as a function of gross strength
and the discounting of alternatives, as well as a more processoriented sequential model that has been found to fit various
types of inference judgments. Finally, the process captured by
Equations 6 and 7 is also consistent with much evidence showing
that people tend to focus on single rather than multiple hypotheses
when presented with new information (see, e.g., Doherty, Mynatt,
Tweney, & Schiavo, 1979, Fischhoff & Beyth-Marom, 1983, and
Schum & Martin, 1982).
2. An important aspect of Equation 6 is that it is a subtractive
rather than a ratio model. This is important because a ratio
model of the form
wks(Zk,Y)
s(X,Y)Sk(X,Y)
s(X,Y)
(8)
has the unacceptable property that S^X, Y) = 1 if there are no

specific alternatives. Indeed, an important implication of the
subtractive form is that the net strength of an explanation can
be low even though specific alternatives do not exist (when gross
strength is low). On the other hand, if gross strength is high and
few specific alternatives exist, the explanation will be seen as a
highly probable cause even if the inference is based on a oneshot case study (see Campbell, 1975, for an illuminating discussion of this issue). As a case in point, consider the occurrence
of a huge explosion near Los Alamos, New Mexico, in July 1945.
No one doubts this to be the effect of detonating an atomic bomb.
Clearly, inferring causality in this poorly designed experiment
was not difficult whereas assessing causality in the most meticulously designed experiments in social science is often problematic at best. When one considers why the causal inference is so
definite in the bomb example, first note that there is a strong
causal chain for connecting the events (atomic physics) and the
cues-to-causality strongly point to a causal relation. Second, ask
yourself, "Would an explosion of such magnitude have occurred
16
if an atomic bomb had not gone off?" Although it is possible to
The notion that the judged probability of a causal explanation
think of alternative explanations for the explosion, they are so
is "single-peaked" (Coombs & Avrunin, 1977) over the number
unlikely as to be virtually nonexistent. In short, strong expla-
of reasons in that explanation, is an intriguing idea that we con-
nations have high gross strength and few plausible alternatives.
sider below. In the second article, Locksley and Stangor (1984)
3. The weight, wk, given to the fcth alternative explanation,
presented subjects with either a rare event (suicide) or a common
can be used to capture a variety of psychological factors involved
event (getting married) and asked them to rank 10 explanations
in the updating process. For example, weights can reflect the
in terms of the probability that they caused the event. The 10
amount of attention given to the &th alternative (for a review of
explanations consisted of both single and conjunctive reasons.
both theoretical and empirical evidence, see Anderson, 1981),
The results showed substantial conjunction fallacies when the
redundancy between alternatives, and contrast/surprise effects
outcome was rare but many fewer fallacies when the outcome
(Einhorn & Hogarth, 1985). A particularly important issue that
was common.
has received little attention is the role of redundancy in the up-
To discuss conjunction fallacies within our framework, we use
dating of causal beliefs. By analogy with multiple regression, Wk
our earlier discussion of the role of necessity and sufficiency in
can be thought of as a weight that reflects the independent con-
judging probable cause. To begin, designate a, as a scenario that
tribution of a particular alternative explanation. Thus, alternative
contains causal explanation X\, as well as various conditions;
explanations that are either more specific or more general in-
that is, a! = (Xif~\
stances of X should not discount X. Similarly, alternatives that
outcome Y, a second causal reason, X2, is added to the expla-
are highly redundant should not be treated as if they were in-
nation and denote the augmented explanation as a 2 ; that is,
|B). Now imagine that in explaining
dependent explanations. However, although there is evidence that
a2 = (JfirUT2n |B). Let us assume that the addition of A"2
people have difficulty ignoring redundant testimony in experi-
to the explanation makes a2 more sufficient for Y than ai. This
ments simulating legal proceedings (Schum & Martin, 1982),
can be expressed as
p(Y\a2)>p(Y\at).
the issue merits much more study.

From our perspective, the various interpretations of wk suggest
many ways of incorporating different cognitive processes in the
sequential updating of beliefs. Indeed, the importance of viewing
discounting within an updating framework lies in linking basic
cognitive processes to causal inferences in general and judgments
of probable cause in particular.
(9)
The important point to note is that although a2 is more sufficient

for ythan &{, it is also less necessary. That is, Equation 9, taken
together with the fact that a2 is less probable than a,, implies
that
p(a 2 |r)<p(a,|r).
(10)
Hence, adding reasons that increase the sufficiency of an expla-
Multiple Causation and Conjunction Fallacies
nation also decreases its necessity. Thus, if people are influenced

by both sufficiency and necessity (however weighted), there will
The framework presented thus far represents what might be
be a point at which the "cost" of lowered necessity will overcome
called the "univariate" case; that is, a single causal candidate,
increased sufficiency, resulting in more reasons being judged as
X, and a single effect, Y. The difficulties of delineating the "mul-
worse than less. Therefore, the single-peakedness of the judged
tivariate" case are many; however, because multiple causation is

crucial to many real-world inferences, we focus on one particular
probability of a conjunctive explanation as a function of the

number of reasons in that explanation results from the conflicting
aspect because it involves an important conflict between causal
forces of necessity and sufficiency.
and probabilistic thinking. Specifically, Tversky and Kahneman
The same idea also accounts for the Locksley and Stangor
(1983) have shown that people often judge the probability of a
(1984) results. In general, rare events require more reasons for
conjunctive event as being greater than the probability of one
a sufficient explanation than common events. Hence, adding
(or more) of its constituents. From a statistical viewpoint, this
multiple reasons to explain a rare event increases the sufficiency
is an error because a conjunction cannot be more probable than
of an explanation at a fast rate. On the other hand, the loss of
any of its components. Accordingly, Tversky and Kahneman
necessity is likely to be small because rare events, by their very
labeled this response a "conjunction fallacy," and gave several
nature, have a small multiplicity of causation. However, even for
reasons for its occurrence. Two recent articles, however, raise
rare events, we would expect a limit on the number of reasons
questions about the generality of conjunction fallacies in causal
one can add to an explanation beyond which it will be judged
explanations.
as less probable (for a mathematical model of the conjunction
In the first article (Leddo, Abelson, & Gross, 1984), subjects
fallacy, see Einhorn, 1985).
were presented with vignettes about individuals engaged in various activities and were asked to judge the probability of both
Conclusions
single and conjunctive reasons for why the behavior occurred.

Although conjunction fallacies were generally found (i.e., expla-
Bertrand Russell (1948) stressed the uncertain nature of in-
nations with more reasons were judged as being more probable
ferences in science and common sense by noting that they "differ
than those with fewer reasons), they were not found when the
from those of deductive logic and mathematics in a very impor-
story involved an actor not doing something (e.g., why did Fred
tant sense, namely, that when the premises are true and the rea-
not stop at the Italian restaurant while he was driving down the
soning correct, the conclusion is only probable" (1948, p. 335).
highway?). Leddo et al. argued that in some situations, the judged
On the other hand, the great Scottish philosopher David Hume
probability of an explanation may increase with the number of
w r o t e , " . . . and as these [resemblance, contiguity, and causation]
reasons up to a point, after which the judged probability declines.
are the only ties of our thoughts, they are really to us the cement
17
PROBABLE CAUSE
of the universe, and all the operations of the mind must, in a
great measure, depend on them" (1740/1938, p. 32). Because
causal inference is an essential and ubiquitous cognitive activity,
we have focused our review on how people make judgments of
probable cause. In providing both a conceptual framework and
review of the relevant literature, we have provided links between
the extensive literature on causation and the burgeoning literature
Brunswik, E. (1952). The conceptual framework of psychology. Chicago:

University of Chicago Press.
Bunge, M. (1979). Causality and modern science. New York: Dover.
Campbell, D. T. (1966). Pattern matching as an essential in distal knowing.
In K. R. Hammond (Ed.), The psychology ofEgon Brunswik (pp. 81106). New York: Holt, Rinehart & Winston.
Campbell, D. T. (1969). Reforms as experiments. American Psychologist,
24, 409-429.
on judgment-under-uncertainty. In doing so, we have argued that
Campbell, D. T. (1975). Degrees of freedom and the case study. Com-
judgments of probable cause are affected by a causal background,
parative Political Studies, 8, 178-193.

Campbell, D. T., & Stanley, J. C. (1963). Experimental and quasi-ex-
probabilistic cues-to-causality, and a discounting process for

dealing with specific alternatives. Moreover, these general ideas
can be summarized by a perceptual analogy in which causal
candidates are differences-in-a-background
(figures are seen
against ground), good explanations arise from internally consistent patterns of cues (good figures form a gestalt), and good explanations have few plausible alternatives (as do good figures).
Whereas our framework accounts for much theoretical and
empirical literature, it by no means explicates all aspects of causal
inference. In particular, inferences made on the basis of complex
scenarios, defining and measuring the coherence of causal explanations (Trabasso & Sperry, 1985), issues of multiple causation
(Downing, Sternberg, & Ross, 1985), and so on, present formi-
perimental designs for research. Chicago: Rand McNally.

Cohen, L. J. (1977). The probable and the provable. Oxford, England:
Clarendon Press.
Cook, T. D., & Campbell, D. T. (1979). Quasi-experimentalion: Design
and analysis for field settings. Chicago: Rand McNally.
Coombs, C. H., & Avrunin, G. S. (1977). Single-peaked functions and
the theory of preference. Psychological Review, 84, 216-230.
Costner, H. L., & Leik, R. K. (1964). Deductions from "axiomatic theory."
American Sociological Review, 29, 819-835.
Crocker, J. (1981). Judgment of covariation by social perceivers. Psychological Bulletin, 90, 272-292.
Crocker, J. (1982). Biased questions in judgment of covariation studies.
Personality and Social Psychology Bulletin, S, 214-220.
and challenges for psychological research.
Dawes, R. M. (undated). How to use your head and statistics at the same
However, given the complexity of these issues, it seems appro-
time or at least in rapid alternation. Unpublished manuscript, University

of Oregon working paper, Eugene.
dable difficulties
priate to have organized the literature on simple causation in

terms of alternatives, background, and cues; that is, the ABC of
probable cause.
Doherty, M. E., Mynatt, C. R., Tweney, R. D., & Schiavo, M. D. (1979).

Pseudodiagnosticity. Acta Psychologica, 43, 111-121.
Downing, C. J., Sternberg, R. J., & Ross, B. H. (1985). Multicausal inference: Evaluation of evidence in causally complex situations. Journal
References
Abelson, R. P. (1981). Psychological status of the script concept. American
Psychologist, 36, 715-729.
Adi, H., Karplus, R., Lawson, A., & Pulos, S. (1978). Intellectual development beyond elementary school VI: Correlational reasoning. School
Science and Mathematics, 78, 675-683.
Alloy, L. B., & Abramson, L. Y. (1979). Judgment of contingency in
depressed and nondepressed students: Sadder but wiser? Journal of
Experimenta( Psychology: General, 108, 441-485.
Alloy, L. B., & Tabachnik, N. (1984). Assessment of covariation by humans and animals: The joint influence of prior expectations and current
situational information. Psychological Review, 91, 112-149.
Anderson, N. H. (1981). Foundations of information integration theory.
New York: Academic Press.
Arkes, H. R., & Harkness, A. R. (1983). Estimates of contingency between
of Experimental Psychology: General, 114, 239-263.

Eddy, D. M. (1982). Probabilistic reasoning in clinical medicine: Problems
and opportunities. In D. Kahneman, P. Slovic, & A. Tversky (Eds.),
Judgment under uncertainty: Heuristics and biases (pp. 249-267). New
York: Cambridge University Press.
Edwards, W. (1968). Conservatism in human information processing. In
B. Kleinmuntz (Ed.), Formal representation of human judgment (pp.
17-52). New York: Wiley.
Ehrenberg, A. S. C. (1975). Data reduction: Analyzing and interpreting
statistical data. New York: Wiley.
Einhorn, H. J. (1985). A model of the conjunction fallacy. Unpublished
manuscript, University of Chicago, Center for Decision Research,
Graduate School of Business.
Einhorn, H. J., & Hogarth, R. M. (1978). Confidence in judgment: Persistence of the illusion of validity. Psychological Review. 85, 395-416.
Einhorn, H. J., & Hogarth, R. M. (1981). Behavioral decision theory:
two dichotomous variables. Journal of Experimental Psychology: Genera!, 112. 117-135.
Processes of judgment and choice. Annual Review of Psychology, 32,
Axelrod, R. (Ed.). (1976). The structure of decision: The cognitive maps

of political elites. Princeton, NJ: Princeton University Press.
Einhorn, H. J., & Hogarth, R. M. (1982). Prediction, diagnosis, and
Blalock, H. M., Jr. (Ed.). (1971). Causal models in the social sciences.
Chicago: Aldine"
Brickman, P., Ryan, K., & Wortman, C. B. (1975). Causal chains: Attribution of responsibility as a function ofimmediate and prior causes.
Journal of Personality and Social Psychology, 32, 1060-1067.
Brown, C. (1984). Diagnostic inference in performance evaluation: Effects
of cause and event covariation and similarity. Unpublished manuscript,
University of Illinois at Urbana-Champaign.
53-88.
causal thinking in forecasting. Journal of Forecasting, 1, 23-36.
Einhorn, H. J., & Hogarth, R. M. (1983). A theory of diagnostic inference:
Judging causality. Unpublished manuscript, University of Chicago,
Center for Decision Research, Graduate School of Business.
Einhorn, H. J., & Hogarth, R. M. (1985). A contrast/surprise model jor
updating beliefs.
Unpublished manuscript, University of Chicago,
Center for Decision Research, Graduate School of Business.

Fincham, F. D., & Jaspars, J. M. (1980). Attribution of responsibility:
of cause and effect temporal order and covariation. Accounting, Or-
From man the scientist to man as lawyer. Advances in Experimental

Social Psychology, 13, 82-138.
ganizations, and Society, 10, 255-266.

Bruner, J. S. (1957). Going beyond the information given. In H. E. Gruber,
Fischhoff, B., & Beyth-Marom, R. (1983). Hypothesis evaluation from a

Bayesian perspective. Psychological Review. 90, 239-260.
K. R. Hammond, & R. Jessor (Eds.), Contemporary approaches to
Fishbein, M., & Ajzen, I. (1975). Belief, attitude, intention and behavior:
Brown, C. (1985). Causal reasoning in performance assessment: Effects
cognition (pp. 41-69). Cambridge, MA: Harvard University Press.
An introduction to theory and research. Reading, MA: Addison-Wesley.
18
Garcia, J. (1981). Tilting at the paper mills of academe. American Psychologist, 36, 149-158.
Garcia, J., McGowan, B., Ervin, F. R., & Koelling, R. (1968). Cues:
Their relative effectiveness as reinforcers. Science, 160, 794-795.
Garcia, J., McGowan, B., & Green, K. (1972). Sensory quality and integration: Constraints on conditioning. In A. H. Black & W. F. Prokasy
(Eds.), Classical conditioning II: Current research and theory (pp. 327). New York: Appleton-Century-Crofts.
Garner, W. R. (1970). Good patterns have few alternatives. American
Scientist, 58, 34-42.
Goldberg, L. R. (1981). Unconfounding situational attributions from
uncertainty: Heuristics and biases. Cambridge, MA: Cambridge University Press.

Kant, I. (1982). Critique of pure reason. (W. Schwarz, Trans.). Aalen,
West Germany: Scientia Verlag. (Original work published 1781)
Kelley, H. H. (1967). Attribution theory in social psychology. In D. Levine
(Ed.), Nebraska Symposium on Motivation, 1967 (pp. 192-238). Lincoln: University of Nebraska Press.
Kelley, H. H. (1973). The processes of causal attribution. American Psychologist, 28, 107-128.
Kelley, H. H., & Michela, J. L. (1980). Attribution theory and research.
Annual Review of Psychology, 31, 457-501.
uncertain, neutral, and ambiguous ones: A psychometric analysis of
Kuhn, T. S. (1962). The structure of scientific revolutions. Chicago: Uni-
descriptions of oneself and various types of others. Journal of Personality

and Social Psychology, 41, 517-552.
Leddo, J., Abelson, R. P., & Gross, P. (1984). Conjunctive explanations:
Granger, C. W. J. (1969). Investigating causal relations by econometric
When two reasons are better than one. Journal of Personality and
models and cross-spectral methods. Econometrica, 37, 424-438.

Hammond, K. R. (1955). Probabilistic functionalism and the clinical
method. Psychological Review, 62, 255-262.
Hanson, N. R. (1958). Patterns of discovery: An inquiry into the conceptual
versity of Chicago Press.
Social Psychology, 47, 933-943.

Lipe, M. G. (1982). A cross-study analysis of covariation judgments.
Unpublished manuscript, University of Chicago, Center for Decision
Research, Graduate School of Business.
foundations of science. Cambridge, England: Cambridge University

Press.
Lipe, M. G. (1985). Attribution theory: A proposed model. Unpublished
Hart, H. L., & Honore, A. M. (1959). Causation in the law. Oxford,

England: Clarendon Press.
Locksley, A., & Stangor, C. (1984). Why vs. how often: Causal reasoning
Harvey, J. H., & Weary, G. (1984). Current issues in attribution theory

and research. Annual Review of Psychology, 35, 427-459.
Hastie, R. (1983). Social inference. Annual Review of Psychology, 35,
511-542.
Hastie, R. (1984). Causes and effects of causal attribution. Journal of
Personality and Social Psychology, 46, 44-56.
Heider, F. (1958). The psychology of interpersonal relations. New York:
Wiley.
Kelson, H. (1964). Adaptation-level theory. New York: Harper & Row.
Hogarth, R. M. (1981). Beyond discrete biases: Functional and dysfunctional aspects of judgmental heuristics. Psychological Bulletin, 90, 197217.
Hogarth, R. M. (Ed.). (1982). Question framing and response consistency.
San Francisco: Jossey-Bass.
Hovland, C. I. (Ed.). (1957). The order of presentation in persuasion.
New Haven, CT: Yale University Press.
Hume, D. (1938). An abstract of a treatise of human nature. London:
Cambridge University Press. (Original work published 1740)
Hume, D. (1964). A treatise of human nature. Aalen, West Germany:
Scientia Verlag. (Original work published 1739; reprint of the new
edition, London 1886)
Jenkins, H. M., & Ward, W. C. (1965). Judgment of contingency between
responses and outcomes. Psychological Monographs: General and Applied, 79(1, Whole No. 594), 1-17.
doctoral dissertation,4 University of Chicago.

and the incidence of judgmental biases. Journal of Experimental Social
Psychology, 20, 470-483.
Mackie, J. L. (1965). Causes and conditions. American Philosophical
Quarterly, 2, 245-264.
Mackie, J. L. (1974). The cement of the universe: A study of causation.
Oxford, England: Clarendon Press.
Michotte, A. (1946). La perception dela causalite. Paris: Vrin.
Mill, J. S. (1872). A system of logic raliocinalive and inductive (8th ed.).
London: Longmans, Green, Reader, & Dyer.
Nisbett, R. E., & Ross, L. D. (1980). Human inference: Strategies and
shortcomings of social judgment. Englewood Cliffs, NJ: Prentice-Hall.
Piaget, J. (1974). Understanding causality. New York: Norton.
PitE, G. F., & Sachs, N. J. (1984). Judgment and decision: Theory and
application. Annual Review of Psychology, 35, 139-163.
Robinson, L. B., & Hastie, R. (1985). Revision of beliefs: When a hypothesis is eliminated from consideration. Journal of Experimental
Psychology: Human Perception and Performance, 11, 443-456.
Russell, B. (1913). On the notion of cause. Proceedings of the Aristotelian
Society. 13, 1-26.
Russell, B. (1948). Human knowledge: Its scope and limit. New York:
Simon & Schuster.
Schank, R. C., & Abelson, R. P. (1977). Scripts, plans, goals and understanding: An inquiry into human knowledge structures. Hillsdale, NJ:
Erlbaum,
Schum, D. A., & Martin, A. W. (1982). Formal and empirical research
Jevons, W. S. (1884). Commercial crises and sun-spots. In H. S. Foxwell
on cascaded inference in jurisprudence. Law and Society Review, 17,
(Ed.), Investigations in currency and finance (pp. 221-243). London:
105-151.
Schustack, M. W., & Sternberg, R. J. (1981). Evaluation of evidence in
Macmillan.
Jones, E. E. (1979). The rocky road from acts to dispositions. American
Psychologist, 34, 107-117.
Jones, E. E., & Nisbett, R. E. (1972). The actor and the observer: Divergent
perceptions of the causes of behavior. In E. E. Jones, D. E. Kanouse,
causal inference. Journal of Experimental Psychology: General, 110,

101-120.
Sedlak, A. J., & Kurtz, S. T. (1981). A review of children's use of causal
inference principles. Child Development, 52, 759-784.
H. H. Kelley, R. E. Nisbett, S. Valins, & B. Weiner (Eds.), Attribution:
Seggie, J. L., & Endersby, H. (1972). The empirical implications of Piaget's
Perceiving the causes of behavior (pp. 79-94) Morristown, NJ: General

Learning Press.
concept of correlation. Australian Journal of Psychology, 24, 3-8.

Seligman, M E P. (1970). On the generality of the laws of learning.
Kahneman, D., & Tversky, A. (1979). Prospect theory: An analysis of

decision under risk. Econometrica, 47, 263-291.
Kahneman, D., & Tversky, A. (1982). The simulation heuristic. In D.
Kahneman, P. Slovic, & A. Tversky (Eds.), Judgment under uncertainty:
Heuristics and biases (pp. 201-208). Cambridge, MA: Cambridge
University Press.
Kahneman, D., Slovic, P., & Tversky, A. (Eds.). (1982). Judgment under
Psychological Review, 77, 406-418.

Shaklee, H., & Tucker, D. (1980). A rule analysis of judgments of covariation between events. Memory & Cognition. 8, 459-467.
Shanteau, J. C. (1970). An additive model for sequential decision making.
Journal of Experimental Psychology, 85, 181-191.
Shapiro, A. K. (I960). A contribution to a history of the placebo effect.
Behavioral Science, 5, 109-135.
19
PROBABLE CAUSE
Sheehan, R. G., & Grieves, R. (1982). Sunspots and cycles: A test of
causation. Southern Economic Journal, 48, 775-777.
Shultz, T. R. (1982). Rules of causal attribution. Monographs of the Society
for Research in Child Development. 47, 1-51.
Shultz, T. R., & Ravinsky, F. B. (1977). Similarity as a principle of causal
inference. Child Development, 48, 1552-1558.,
Tolman, E. C. (1948). Cognitive maps in rats and men. Psychological

Review. 55, 189-208.
Trabasso, T., Secco, T., & van den Broek, P. (1984). Causal cohesion and
story coherence. In H. Mandl, N. L. Stein, & T. Trabasso (Eds.),
Learning and comprehension of text (pp. 83-111). Hillsdale, NJ: Erlbaum.
Shweder, R. A. (1977). Likeness and likelihood in everyday thought:
Trabasso, T., & Sperry, L. L. (1985). Causal relatedness and importance
Magical thinking in judgments about personality. Current Anthropology,

Siegler, R. S. (1976). The effects of simple necessity and sufficiency re-
of story events. Journal of Memory and Language, 24, 595-611.

Tversky, A. (1977). Features of similarity. Psychological Review, 84, 327352.
lationships on children's causal inferences. Child Development, 47,
Tversky, A., & Kahneman, D. (1980). Causal schemas in judgments under
IS, 637-658.
1058-1063.
Siegler, R. S., & Liebert, R. M. (1974). Effects of contiguity, regularity,
and age on children's causal inferences. Developmental Psychology, 10,
574-579.
Simon, H. A. (1954). Spurious correlation: A causal interpretation. Journal of the American Statistical Association, 49, 467-479.
Simon, H. A., & Hayes, J. R. (1976). The understanding process: Problem
isomorphs. Cognitive Psychology, S, 165-190.
Slovic, P., Fischhoff, B., & Lichtenstein, S. (1976). Cognitive processes
and societal risk taking. In J. S. Carroll and J. W. Payne (Eds.), Cognition
and social behavior (pp. 165-184). Hillsdale, NJ: Erlbaum.
Smedslund, J. (1963). The concept of correlation in adults. Scandinavian
Journal of Psychology. 4, 165-173.
Stein, N. L., & Glenn, C. G. (1979). An analysis of story comprehension
in elementary school children. In R. O. Freedle (Ed.), Advances in
uncertainty. In M. Fishbein (Ed.), Progress in social psychology, Vol.

1 (pp. 49-72). Hillsdale, NJ: Erlbaum.
Tversky, A., & Kahneman, D. (1981). The framing of decisions and the
psychology of choice. Science, 211, 453-458.
Tversky, A., & Kahneman, D. (1983). Extensional versus intuitive reasoning: The conjunction fallacy in probability judgment. Psychological
Review. 90,293-315.
van Fraassen, B. C. (1980). The scientific image. Oxford, England: Clarendon Press.
Waller, W. S., & Felix, W. L. (1982). The auditor and learning from experience: Some empirical evidence. Unpublished manuscript, University
of Arizona, Department of Accounting, Tucson.
Ward, W. C., & Jenkins, H. M. (1965). The display of information and
the judgment of contingency. Canadian Journal of Psychology. 19,231241.
discourse processing: New directions in discourse processing, 2 (pp.

53-120). Norwood, NJ: Ablex.
Watson, D. (1982). The actor and the observer: How are their perceptions
Strawson, P. F. (1952). Introduction to logical theory. London: Methuen.
Weiner, B. (1985). "Spontaneous" causal thinking. Psychological Bulletin,

97, 74-84.
Suppes, P. (1970). A probabilistic theory of causality. Amsterdam: NorthHolland.

Susser, M. (1973). Causal thinking in the health sciences. New York:
Oxford University Press.
of causality divergent? Psychological Bulletin, 92. 682-700.
Zellner, A. (1984). Basic issues in econometrics. Chicago: University of

Chicago Press.
Toda, M. (1977). Causality, conditional probability and control. In A.

Aykac & C. Brumal (Eds.), New developments in the application of
Bayesian methods (pp. 109-124). Amsterdam: North-Holland.
Received August 13, 1984

Revision received July 10, 1985

Out11 PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Out11 PDF

Uploaded by

Copyright:

Available Formats

This document is copyrighted by the American Psychological Association or one of its allied publishers.

Copyright 1986 by the American Psychological Association, Inc.

Judging Probable Cause

The topic of causality is central to many areas of psychology,

the scientific method more generally, is at present in a productive

such as developmental psychology (Piaget, 1974; Sedlak & Kurtz,

state of near chaos" (p. 10).

1981), perception and learning (Michotte, 1946), social psy-

Psychology is not alone, however, in presenting a fragmented

chology (Harvey & Weary, 1984; Heider, 1958; Kelley, 1967),

view of causality. In philosophy, the definition and meaning of

methodology (Cook & Campbell, 1979), text comprehension

1979). Indeed, Bertrand Russell (1913) even suggested that the

on reading these various literatures, one is struck by their relative

concept of cause does not play a useful role in science. Thus,

isolation from each other. Indeed, there seems to be little contact

lacking a clear definition, psychologists within different research

or even awareness of mutual problems, issues, and concepts. For

traditions have focused on specific aspects of causality that have

example, discussions of how people perceive physical causation

been emphasized by different philosophers and made operational

are rarely linked to the processes of causal attribution in the

tests of these notions. For example, workers in attribution theory

social domain, or vice versa. Moreover, even within areas, there

have tended to follow Kelley (1967) in emphasizing Mill's (1872)

is considerable dissatisfaction concerning the incomplete nature

criteria of concomitant variation and the method of differences;

of our understanding of the psychology of cause. For example,

Michotte's (1946) classic demonstrations of how people perceive

in a review of attribution theory, Kelley and Michela (1980)

causes relies heavily on ideas advanced by Hume (1886/1964);

wrote, ". . . the theories are piecemeal and greatly in need of

Shultz's (1982) work has been influenced by Kant's (1781 /1982)

synthesis" (p. 494). In a similar vein, Cook and Campbell (1979)

notions that causal relations are characterized by forces of "gen-

have pointed out that, "The epistemology of causation, and of

erative transmission" between cause and effect; and Campbell

This research was supported by a contract from the Office of Naval

meaning of cause and its operational definition have also attracted

HILLEL J. EINHORN AND ROBIN M. HOGARTH

Purpose and Plan of Article

with the force of solid objects, contiguity in time and space is

In other words, the determination of causal relevance depends

but not its cause. As we argue below, the distinction between

HILLEL J. E1NHORN AND ROBIN M. HOGARTH

A useful analogy is to think of the causal field as the scope of

Finally, although both theoretical arguments and empirical

how such inferences are consistent with differing levels of covariation.

Thus, given , Y always follows; however, given Y, a is neither

We can now say that X is an insufficient but necessary part of a

Figure 1. Contingency table representing causal relations. (Y is an effect

conditions for Y in which X is absent and thus indicates the

HILLEL J. E1NHORN AND ROBIN M. HOGARTH

vf, and ft = the

Some early and influential studies (Jenkins & Ward, 1965;

when relations are assessed between, say, symptoms and diseases.

variation (in one setting) with lower covariation that is robust

Temporal Order, Contiguity, and Similarity

Figure 2. Loop-like structure for confusion of inverse probabilities.

informative, many subjects judge the causal/forward probability

Under what conditions is the confusion most likely to occur?

HILLEL J. EINHORN AND ROBIN M. HOGARTH

Note that the causal probability, p(Y\X), is not predictive in