Psychological Review Copyright 1986 by the American Psychological Association, Inc.

1986, Vol. 93, No. 2, 136-153 0033-295X/86/$00.75

Norm Theory: Comparing Reality to Its Alternatives

Daniel Kahneman Dale T. Miller
University of British Columbia Simon Fraser University
Vancouver, British Columbia, Canada Burnaby, British Columbia, Canada

A theory of norms and normality is presented and applied to some phenomena of emotional responses,
social judgment, and conversations about causes. Norms are assumed to be constructed ad hoc by
recruiting specific representations. Category norms are derived by recruiting exemplars. Specific objects
or events generate their own norms by retrieval of similar experiences stored in memory or by con-
struction of counterfactual alternatives. The normality of a stimulus is evaluated by comparing it to
the norms that it evokes after the fact, rather than to precomputed expectations. Norm theory is
applied in analyses of the enhanced emotional response to events that have abnormal causes, of the
generation of predictions and inferences from observations of behavior, and of the role of norms in
causal questions and answers.

This article is concerned with category norms that represent produced by aggregating the set of recruited representations. The
knowledge of concepts and with stimulus norms that govern assumptions of distributed activation and rapid aggregation are
comparative judgments and designate experiences as surprising. not unique to this treatment. Related ideas have been advanced
In the tradition of adaptation level theory (Appley, 1971; Helson, in theories of adaptation level (Helson, 1964; Restle, 1978a,
1964), the concept of norm is applied to events that range in 1978b) and other theories of context effects in judgment (N. H.
complexity from single visual displays to social interactions. We Anderson, 1981; Birnbaum, 1982; Parducci, 1965, 1974); in
first propose a model of an activation process that produces connectionist models of distributed processing (Hinton & An-
norms, then explore the role of norms in social cognition. derson, 1981; McClelland, 1985; McClelland & Rumelhart,
The central idea of the present treatment is that norms are 1985); and in holographic models of memory (Eich, 1982; Met-
computed after the event rather than in advance. We sketch a calfe Eich, 1985; Murdock, 1982). The present analysis relates
supplement to the generally accepted idea that events in the most closely to exemplar models of concepts (Brooks, 1978, in
stream of experience are interpreted and evaluated by consulting press; Hintzman, in press; Hintzman & Ludlam, 1980; Jacoby
precomputed schemas and frames of reference. The view devel- & Brooks, 1984; Medin & Schaffer, 1978; Smith & Medin, 1981).
oped here is that each stimulus selectively recruits its own alter- We were drawn to exemplar models in large part because they
natives (Garner, 1962, 1970) and is interpreted in a rich context provide the only satisfactory account of the norms evoked by
of remembered and constructed representations of what it could questions about arbitrary categories, such as "Is this person
have been, might have been, or should have been. Thus, each friendlier than most other people on your block?"
event brings its own frame of reference into being. We also explore Exemplar models assume that several representations are
the idea that knowledge of categories (e.g., "encounters with Jim") evoked at once and that activation varies in degree. They do not
can be derived on-line by selectively evoking stored representa- require the representations of exemplars to be accessible to con-
tions of discrete episodes and exemplars. scious and explicit retrieval, and they allow representations to
The present model assumes that a number of representations be fragmentary. The present model of norms adopts all of these
can be recruited in parallel, by either a stimulus event or an assumptions. In addition, we propose that events are sometimes
abstract probe such as a category name, and that a norm is compared to counterfactual alternatives that are constructed ad
hoc rather than retrieved from past experience. These ideas ex-
This work was supported by the Office of Naval Research under Grant tend previous work on the availability and simulation heuristics
NR 197-058, and by grants from the National Science and Engineering (Kahneman & Tversky, 1982; Tversky & Kahneman, 1973).
Research Council of Canada and from the Social Sciences and Humanities A constructive process must be invoked to explain some cases
Research Council of Canada (410-68-0583). of surprise. Thus, an observer who knows Marty's affection for
Dale Griffin, Leslie McPherson, and Daniel Read provided valuable his aunt and his propensity for emotional displays may be sur-
assistance. Many friends and colleagues commented helpfully on earlier prised if Marty does not cry at her funeral—even if Marty rarely
versions. The comments of Anne Treisman and Amos Tversky were cries and if no one else cries at that funeral. Surprise is produced
especially influential. The preparation of the manuscript benefited greatly in such cases by the contrast between a stimulus and a counter-
from a workshop entitled "The Priority of the Specific," organized by factual alternative that is constructed, not retrieved. Constructed
Lee Brooks and Larry Jacoby at Elora, Ontario, in June of 1983.
Correspondence concerning this article should be addressed to Daniel elements also play a crucial role in counterfactual emotions such
Kahneman, who is now at the Department of Psychology, University of as frustration or regret, in which reality is compared to an imag-
California, Berkeley, California 94720, or Dale Miller, Department of ined view of what might have been (Kahneman & Tversky, 1982).
Psychology, Simon Fraser University, Burnaby, British Columbia V5A At the core of the present analysis are the rules and constraints
1S6, Canada. that govern the spontaneous retrieval or construction of alter-

136
 NORM THEORY 137
natives to experience. A question of particular interest concerns explores the function of norms as representations of stored
the mutability of different attributes of an evoking stimulus: knowledge of persons and the role of norms in causal reasoning.
Which of its features will be retained in the norm elements that
it evokes? Which attributes will have highly variable norms? The A Model of Norms
differential mutability of attributes restricts both the retrieval
and the construction of norm elements. A model of norms is sketched in Figure 1. A probe (which
An abnormal event is one that has highly available alternatives, may be the experience of an object or event, or a reference to a
whether retrieved or constructed; a normal event mainly evokes concept) recruits an evoked set that consists of elements (A, B,
representations that resemble it. The treatment of normality in and C in the example). The elements of the evoked set are ac-
this article is guided by the phenomenology of surprise rather tivated (made available) to different degrees, as indicated in the
than by formal or informal conceptions of probability. The main figure by the thickness of the arrows. The elements are repre-
difference between the two notions is that probability is always sentations of objects, episodes, or classes of elements. Represen-
construed as an aspect of anticipation, whereas surprise is the tations of the neighbor's dog Fido or of the category "poodle"
outcome of what we shall call backward processing—evaluation could be recruited as elements of the set evoked by the probe
after the fact. Probability reflects expectations. Surprise (or its "dog." Each element is internally described by features, which
absence) reflects the failure or success of an attempt to make are specific values of attributes (X and Y in the example). The
sense of an experience, rather than an evaluation of the validity evoked set is characterized by norms for each of the attributes
of prior beliefs. In his critique of standard notions of probability, that describe its elements. Norms for the attributes X and Y are
the statistician Shafer (1976) developed the related idea that events shown in the bottom panels.
do not merely alter the strength of belief in existing possibilities; Elements are internally described in terms of physical attri-
they also evoke and shape the set of relevant possibilities. butes (e.g., size), more abstract ones (e.g., friendliness), and some
Of course, specific anticipations that exist in advance of an conjunctions of elementary attributes (e.g., size and strength).
event will be included in the norm to which it is compared. The For simplicity, the attributes X and Y are presented in Figure 1
event will then appear normal if it confirms expectations, ab- as ordered dimensions, but the treatment extends readily to at-
normal or surprising if it violates them. However, an unantici- tributes that have other similarity structures. As indicated in the
pated event will also be judged normal if it simply fails to evoke figure, each feature of an element is described by a profile or
strong alternatives. This formulation distinguishes two ways in distribution of activation over a range of attribute values. When
which an occurrence may affect the normality of subsequent the element represents an individual object or event, the shape
events: (a) by eliciting hypotheses and expectations, which later of the profile can be interpreted as a gradient of generalization.
events confirm or disconfirm, or (b) by laying down a trace that The profile is also flattened by any uncertainty or imprecision
is activated when a subsequent event provides an appropriate in the assignment of a feature to an element. When the element
reminder (Schank, 1982). stands for a class, the profile represents the internal variability
Consider an observer, casually watching the patrons at a of the class. The degree of activation of an element determines
neighboring table of a fashionable restaurant, who notices that the size of the profiles for its attributes.
the first guest to taste the soup winces, as if in pain. The normality The entire evoked set is described by summing, for each at-
of a multitude of events will be altered by this incident. For tribute, the profiles associated with all activated elements. We
example, it is now unsurprising for the guest who first tasted the shall say that the aggregate profile assigns a measure of availability
soup to startle violently when touched by a waiter; it is also to values of the attribute. A measure of normality is obtained
unsurprising for another guest to stifle a cry when tasting soup by rescaling the availability profile, assigning a normality of 1.0
from the same tureen. These events and many others appear to the most available value. The normality measure therefore
more normal than they would have done otherwise, but not nec- ranges between 0 and 1, and the normality of any attribute value
essarily because they confirm advance expectations. Rather, they is the ratio of the availability of that value to the modal (maximal)
appear normal because they recruit the original episode and are availability. A norm is a function that assigns a normality measure
interpreted in conjunction with it. In general, selective retrieval to values of an attribute.
of pertinent episodes tends to reduce surprise and to favor hind- In summary, a probe recruits an evoked set of individual ele-
sight about both the recent event and its predecessor (Fischhoff, ments, each of which is described by several features. The evoked
1975, 1982). set is described by an aggregate of individual features. The norm
Reasoning flows not only forward, from anticipation and hy- for an attribute is the envelope of this aggregate profile, scaled
pothesis to confirmation or revision, but also backward, from to assign unit normality to the modal value. In the text that
the experience to what it reminds us of or makes us think about. follows, "norm element" is often used as shorthand for "element
This article is largely dedicated to the power of backward think- of the evoked set."
ing. Its aim is not to deny the existence of anticipation and ex- As is evident in Figure 1, the principal independent variables
pectation but to encourage the consideration of alternative ac- of the model are the factors that control the activation of potential
counts for some of the observations that are routinely explained elements of an evoked set. Direct and indirect manifestations of
in terms of forward processing. normality are the dependent variables. The direct measures in-
We first introduce a model of norms and illustrate some of clude expressions of surprise and judgments of normality or typ-
its applications. The rules that govern the generation of norm icality. Indirect measures include intuitive predictions, emotional
elements are then introduced, as well as some consequences of responses to abnormal events, and various aspects of causal rea-
these rules in the domain of emotion. The remainder of the article soning. The schema for applying the theory is the following:
138 DANIEL KAHNEMAN AND DALE T. MILLER

ATTRIBUTE ATTRIBUTE
X ! j Y
WEIGHTS
(AVAILABILITY)

ELEMENT A

ELEMENT B

ELEMENT C

NORMS

Figure 1. Model of norms. (A stimulus or category label, and its context, activates elements A, B, and C to
different degrees. The values of each element on two attributes, X and Y, are represented by profiles, which
are summed to establish a norm for each attribute. The normality of a value is denned as the ratio of its
aggregate availability [the height of the norm at that value] to the modal [maximum] availability in the
norm.)

Keeping all else constant, a manipulation that increases the bulk for a camping implement. Furthermore, these comparative
probability of a given element being recruited by a probe will judgments appear neither especially difficult nor especially slow.
also increase the weight of that element in the norms evoked by A dismissal can be judged outrageous or a sleeping bag bulky
that probe. without conscious examination of many (or any?) instances of
In the next sections we elaborate on three central aspects of the category. Multiple representations appear to be activated in
the present analysis: the recruitment of norm elements, the role parallel without entering consciousness or working memory.
of norms in the representation of knowledge, and the concept Summary statistics for the activated elements are used for com-
of normality. parative judgments, and perhaps for other purposes as well.
The recruitment of norm elements has many of the charac-
Modes of Recruitment teristics attributed to spreading associative activation (J. R. An-
derson, 1983; Ratcliff& McKoon, 1981). However, not all of the
The model of Figure 1 applies to two types of norms that are representations that are associatively connected to the probe will
distinguished by their evoking probe: (a) stimulus norms, which be included in its norm. Inclusion in a category norm, in par-
are evoked by experiences of objects and events, and (b) category ticular, must be restricted to members of the designated category.
norms, which are evoked by references to categories. Two modes The norm for horses should not include carriages, and the cat-
of recruitment are distinguished: (a) retrieval of memory rep- egory label "married graduate students" should not strongly ac-
resentations of individual objects and events or of subordinate tivate representations of married nonstudents or unmarried stu-
categories and (b) construction of counterfactual alternatives to dents (Osherson & Smith, 1982). The required selectivity of re-
experience. As these concepts are used here, the scope of "re- cruitment is achieved more economically by precisely controlled
cruitment" is broader than "retrieval," and "element" is broader activation than by inhibition of irrelevant activated elements. A
than "memory trace."1
A process that retrieves or generates specific exemplars appears
necessary to account for people's ability to deal with arbitrary 1
An analysis of perceptual norms requires a third process of recruit-
collections or functionally denned ad hoc categories, such as ment, in which a focal stimulus selectively interacts with representations
"reasons for firing an employee" or "things to take on a camping of some of the objects present in the current perceptual field. Context
trip" (Barsalou, 1983). Instances of ad hoc categories can be effects in perception are often analyzed in such terms (N. H. Anderson,
evaluated on such attributes as reasonableness for a dismissal or 1981;Coren&Girgus, 1978; Coren& Miller, 1974;Resue, 1978a, 1978b).
 NORM THEORY 139

plausible hypothesis is that the activating effects of distinct con- dence in the use of knowledge has prompted several theorists to
stituents of the probe, or of the probe and its context, are mutually search for a mode of representation that is more molecular and
reinforcing (Medin & Schaffer, 1978). The specification of a cat- more flexible than a schema (Alba & Hasher, 1983; Barsalou,
egory by a conjunction of features serves to restrict the spread 1985; Johnson, 1983; Lakoff, in press; Schank, 1982). The pos-
of activation to elements that possess most or all of them. There sibility considered here is an assemblage of exemplars (Brooks,
may also be a limit on the number of elements that can be si- 1978, in press; Medin & Schaffer, 1978). Memory is assumed to
multaneously activated by a single probe—perhaps half a dozen store content-addressable files, each containing data about a spe-
or less (J. R. Anderson, 1983; Mandler, 1967, 1975). cific episode of experience. When information is required some
The same probe can elicit both retrieval and construction. For of the files are selectively retrieved, computations are performed
example, consider a person who is involved in an accident. The as needed—and perhaps a book about the topic is printed and
occurrence is a memory probe, which acts as a reminder of sim- bound for conscious inspection, all within 200 milliseconds or
ilar experiences in the past. The current occasion will appear so. This mode of representation by exemplars appears necessary
more normal if traces of similar experiences are activated than for stimulus norms. It is also very plausible for ad hoc categories
otherwise (Schank, 1982). Any serious accident will also provoke such as "encounters with Jim in the elevator," which in turn
an examination of the sequence of events that led to it, and this shade into the domain of familiar categories and natural kinds.
examination in turn involves the generation of counterfactual It is important to stress that no exemplar model can account
alternatives. The occurrence will appear especially abnormal if for all variants of category knowledge (see also Murphy & Medin,
some scenarios that yield a different outcome are highly available. 1985). The inheritance of properties, in particular, cannot be
The outcome will appear inevitable if no such alternatives come explained by such a model. Thus, it is surely normal for birds
readily to mind. to have stomachs, but this norm is derived deductively from
The generation of alternatives to reality appears to be quite knowledge about animals, not induced from observed exemplars.
disciplined. Inclusion in a stimulus norm is restricted to objects The only reasonable claim for exemplar models is that some
and events that share the immutable features of the evoking knowledge of categories can be represented by norms that are
stimulus. For example, the alternative scenarios that are produced computed on the fly, not that all category knowledge is achieved
in mentally "undoing" an accident tend to alter some features in this manner. Furthermore, it appears likely that norms that
of the real sequence of events, leaving others constant (Kahneman are evoked repeatedly are eventually stored as summary statistics
& Tversky, 1982). We return later to a discussion of mutable rather than as raw data (Fried & Holyoak, 1984). In spite of
and immutable features, which also confronts an obvious diffi- these qualifications, the assumption that category knowledge is
culty for the present model: If stimuli only evoke norm elements often represented by exemplars appears necessary. It also appears
that are similar to them, how can a stimulus ever appear abnor- especially fruitful in the representation of knowledge about other
mal? In responding to this question we shall propose that re- people.
cruited elements are only constrained to share the immutable The properties of on-line computation and context sensitivity
features of the evoking stimulus. Norm elements are allowed to are common to exemplar models and to connectionist and holo-
differ from the evoking stimulus and from each other in other graphic representations (e.g., Eich, 1982; McClelland & Ru-
attributes, and any mutable feature of the evoking stimulus can melhart, 1985; Murdock, 1982). However, the current versions
therefore appear abnormal. of these models represent a category by a composite pattern of
memory activation, which obliterates information about both
Norms as Representations of Knowledge individual elements and higher order statistics. The easy access
of observers to variances and covariances of attributes in collec-
The model of Figure 1 invokes the same exemplar represen- tions of instances is perhaps the strongest reason to reserve a
tation for category norms and for stimulus norms and the same role for individual exemplars in a model of category knowledge
diffuse description for individuals and for sets. We next sketch (Medin, 1983; Medin, Altom, Edelson, & Freko, 1982; Medin
the reasoning that led to the apparent neglect of these distinctions. & Schaffer, 1978).
The decision to describe both category norms and stimulus The model of Figure 1 assigns a diffuse description to the
norms as temporary patterns of memory activation is a response features of individual objects and events. As in the classic com-
to a failure. Although our concern was mainly with stimulus posite photograph, the representation of the evoked set is an
norms, which are temporary by definition, we were unable to aggregate, but in this case each of the pictures in the composite
deal with these norms without invoking category knowledge, is fuzzy on its own. The view of norms as aggregates of diffuse
and equally unable to draw a clean boundary between temporary descriptions accommodates the basic finding that an unpresented
and durable representations of such knowledge. At issue is the prototypical member of a category is likely to be erroneously
representation of knowledge that is not at the moment in use. recognized as familiar (Posner & Keele, 1968). Because activation
One common view is that knowledge of categories (e.g., "dogs" spreads to the vicinity of each presented value, a feature that
or "encounters with Jim") is contained in a single compact rep- has not been presented can be highly normal (typical) if neigh-
resentation in semantic memory, like a book in a library, which boring values of the attribute have frequently appeared.
is consulted when information is needed about dogs or about An important implication of diffuse description is that the
Jim. This notion of a schema is most useful in dealing with model does not distinguish individuals from sets. As a conse-
information that is needed often and that can be applied time quence, the elements of a norm can be either tokens (episodes
after time with little variation (Hastie, 1981; Mandler, 1984; Ru- of experience) or types (concepts of lower order categories). For
melhart, 1980). However, the need to represent context depen- example, it is standard to view the category "bird" as consisting
140 DANIEL KAHNEMAN AND DALE T. MILLER

of kinds of birds, such as robin and chicken (Smith & Medin,

1981). In applying an exemplar model to such classes, the sub-
ordinate categories contribute their norms to the norm for the
inclusive class. Here again, some elements will contribute more STIMULUS CONTEXT STIMULUS

than others: a reference to "bird" is likely to activate the rep-

resentations of "robin" and "sparrow" more strongly than those
of "ostrich" or "chicken" (Rosch & Mervis, 1975; Rosch, Mervis, CATEGORY CATEGORY
LABEL LABEL
Gray, Johnson, & Boyes-Braem, 1976). The present treatment
offers the mode of the category norm as a counterpart to the
notion of prototypical value. It also incorporates a notion of a b
differential activation of elements, akin to the idea of graded Figure 2. Two models of recruitment of norm elements.
membership in a category (Barsalou, in press; Rosch, 1978; Rosch
& Mervis, 1975; Zadeh, 1965, 1976). The name of a category will surely evoke (make available) its
representative exemplars more strongly than others (Barsalou,
Normality and Probability in press; Rosch & Mervis, 1975; Rosch et al., 1976). The char-
acteristics of the most available instances, in turn, are likely to
A norm and a probability distribution can both be displayed be judged most representative of the category. The present anal-
as a function of attribute values, but normality and probability ysis focuses on availability as the immediate determinant of
differ in their formal characteristics. Normality values are rep- norms.
resented by heights (scaled to a maximum of 1.0), whereas prob-
abilities are represented by areas (scaled so that the total area
Comparative Judgments
under the profile is 1.0). One immediate consequence is that,
unlike probability, the normality of a value of an attribute can We have distinguished two types of probes that evoke norms:
increase without a corresponding reduction of the normality of experiences of objects and events and references to categories.
any other. It can be seen in Figure 1, for example, that the addition An application of this distinction to comparative judgment is
of element C to elements A and B made high values of X more illustrated in Figure 2. In category-centered comparisons, the
normal without reducing the normality of low values of that object of judgment is compared to the norm for a specified cat-
attribute. egory. In stimulus-centered comparisons, the elements of the
Another contrast concerns disjunctions and conjunctions of norm tend to be recruited directly by the stimulus itself.
values. The probability of a disjunction of mutually exclusive "Jane owns a small dog" is an example of a category-centered
events is the sum of their probabilities, but the most reasonable judgment. To make and to interpret such judgments, a norm of
rule for the normality of a disjunction is that it equals the max- size for a particular category must be invoked. However, the set
imal normality of its constituents (Zadeh, 1965). In particular, of recruited exemplars can be biased by context or background.
extending the range of a category does not necessarily increase The same statement will yield different norms of size and different
normality. It is no more normal for a man to be 5 ft 10 in. tall ideas of the size of Jane's dog if she is known to live in a New
(to the nearest inch) than to be 175 cm tall (to the nearest cm), York apartment or on a farm in Maine (Barsalou, in press; Roth
although the former event is substantially more probable. & Shoben, 1983).
A message is compared to alternatives at the same level of Stimulus-centered judgments are more elusive. Consider the
detail. Thus, the information that the favorite lost the first set following information: "Ms. Z is 26 years old, with a Master of
of a tennis match will only evoke the alternative possibility of Science degree in geography. She earns $33,000 a year." Most
the favorite winning that set. A more detailed message, stating readers will probably recognize that, although not instructed to
that the favorite lost the first set but won the match will evoke do so, they have already evaluated Ms. Z's salary as high or low.
alternative conjunctions of outcomes for the first set and the The norm for this judgment is not precomputed: Few people
match. Because normality is relative, it is quite possible for a will have access to stored statistics for the income of 26-year-old
single event (e.g., that the favorite lost the first set) to be more women with a master's degree in geography. Furthermore, we
surprising than the conjunction of that event with another (e.g., suspect that the norm that yields the spontaneous judgment of
that the favorite lost the set and won the match). The rule that Ms. Z's salary is not quite the same as would be elicited by the
the probability of a conjunction cannot exceed that of any of its category-centered instruction "Compare Ms. Z's income to 26-
constituents extends neither to normality nor to surprise. Con- year-old.. . ."In particular, some friends have joined us in con-
fusion of probability with normality or surprise may be a con- fessing that thoughts of their own past and present income were
tributing factor in some—not all—violations of the conjunction not irrelevant to their evaluation of Ms. Z's salary. Stimulus-
rule in naive judgments of probability (Tversky & Kahneman, centered norms are not restricted to members of a particular
1983). category and are likely to be biased toward highly available ex-
The concept of normality is closely related to both typicality amples.
(Rosch, 1973) and representativeness (Tversky & Kahneman, The distinction between stimulus-centered and category-cen-
1983). A significant theoretical difference is that representative- tered judgments identifies a continuum rather than a dichotomy.
ness is modeled as a similarity relation, whereas normality has In the pure case of category-centered judgment, only instances
been traced to the availability of exemplars. However, availability of the category will be evoked, and the stimulus and context
and representativeness are intertwined in the generation of norms. further narrow the selection of category members. At the other
 NORM THEORY 141

extreme, a stimulus-centered norm can be evoked without any the context of adaptation level theory (Appley, 1971; Kelson,
explicit categorization of the evoking stimulus. As indicated by 1964). The instructions given to subjects in adaptation level
the broken arrow in Figure 2, classification is optional in stimulus- studies do not usually specify a reference category. The judgments
centered judgments. If an explicit category label is activated, are therefore stimulus centered rather than category centered.
elements attached to it will be evoked, but the norm will not be As a consequence, stimuli encountered outside the immediate
restricted to these elements. task context often have substantial effects on observed adaptation
Probes and task demands are an important aspect of the con- levels. A 5-g weight is rarely judged heavy, regardless of the ex-
text that affects the recruitment of elements in both stimulus- perimental context. Stimuli presented for judgment also vary in
centered and category-centered judgments. In particular, the their influence on the adaptation level. Two sets of factors de-
designation of an attribute as focal tends to increase the muta- termine the weight of previous stimuli in a norm: (a) factors that
bility of that attribute, that is, its variability among members of affect the salience or memorability of specific experiences and
the evoked set. For example, questions about Ms. Z's income or (b) the strength of the connection between norm elements and
educational level will increase the respective variability of each the evoking stimulus, which is mainly determined by the simi-
of these attributes in the norms to which Ms. Z is compared. larity between them.
Avant and Kelson (1973, p. 440) offered the following summary
Control of Recruitment list of the factors that control the impact of stimuli on adaptation
Explicit reference to a category permits a high degree of control level: "recency, frequency, intensity, area, duration, and higher-
over the selection of norm elements. In particular, the use of a order attributes such as meaningfulness, familiarity and ego-
category label is very effective in restricting the evoked set to involvement." This list could serve equally well as a list of de-
category members (Rips & Turnbull, 1980). Consider the sen- terminants of availability for retrieval. There is evidence that the
tence "The large fly climbed up the trunk of the small elephant." weight of stimuli in the norm is higher for recent stimuli than
The sentence is understood without difficulty, although its in- for those presented earlier (Lockhead& King, 1983; Ward, 1979).
terpretation almost simultaneously invokes two different norms The weight of a stimulus is also increased by making it distinctive
of size. The work of Barsalou (1983, in press) shows that norms or salient on some irrelevant attribute such as duration of ex-
may be generated even for ad hoc categories, and also that people posure (Kelson & Kozaki, 1968). On the other hand, an anchor
can take the point of view of others in considering exemplars of stimulus that is redundantly presented before each trial has much
a category. The ability to simulate another person's norms is less weight in the norm than the aggregate of other stimuli, al-
often essential to successful communication. The task is pre- though the total frequency of presentation is the same for the
sumably carried out by selecting a subset of one's own relevant anchor and for all other stimuli combined (Kelson, 1964).
experiences. Parents seem to have no difficulty in determining The notion of postcomputed norms implies that recruitment
that a particular animal is the largest that their 2-year-old has favors elements that resemble the evoking stimulus. As a con-
ever seen. sequence, different objects of judgment may be compared to
Voluntary control of invoked categories is of course not perfect. somewhat different norms even in the context of a single task.
The idea that recruitment is controlled by selective activation In support of this idea, Restle (1978a) highlighted the finding by
rather than by inhibition suggests that it might be difficult to Sarris (1976) that a repeated anchor does not have the same
exclude designated instances from a category norm—just as it effect on the judgments of all stimuli. The anchor is assigned the
is difficult to obey the instruction not to think of elephants. This largest weight in judgments of its nearest neighbors in the test
reasoning entails an interesting asymmetry: An observer might series.
be able to include selected elements in a norm by deliberately The first trial in a judgment experiment provides the best ex-
thinking about them but fail to exclude specified elements from ample of the process in which a stimulus evokes its own context.
a norm if they have been associatively activated. When Parducci The present model suggests that some features of the evoking
(1956) informed subjects of a change in the composition of a stimulus are treated as immutable in that process: The recruit-
series that they were judging, he observed rapid adjustment to ment of the evoked set tends to be restricted to elements that
the news that additional stimuli would extend the range previously share these features. A plausible hypothesis is that the essential
included, but he observed almost no adjustment in response to features that define the identity of the stimulus are most likely
an announced restriction of the range. This finding suggests that to be maintained as immutable. This hypothesis has surprising
size judgments assigned to common mammals will change consequences: It entails that judgments of a stimulus evaluated
abruptly when respondents are informed that whales and ele- in isolation will tend to be dominated by features that are not
phants may occur in the list, but the judgments will not adjust the most central.
with equal ease to the information that the larger mammals will A striking result observed by Slovic (1985) provides an illus-
not, after all, be included. In a weight judgment task, Brown and tration. Slovic asked a group of subjects to evaluate on a 20-
Reich (1971) found that the instruction to make the judgment point scale the attractiveness of the following bet: "a 7/36 chance
of each weight relative only to weights of the same color was to win $9." The mean evaluation was 9.4. A different group of
largely ineffective if it was given after a common frame of ref- subjects evaluated the bet "a 7/36 chance to win $9 and a 29/
erence was established. 36 chance to lose 5 cents." Although the second bet is strictly
inferior to the first, its average rating on the attractiveness scale
Effects of Availability and Mutability was 14.4. The two bets appear to be compared to different norms.
The factors that govern the weighting of norm elements in The second bet appears very favorable among bets that involve
comparative judgments have been studied most extensively in a risk of loss, but a modest chance to win $9 is mediocre in a
142 DANIEL KAHNEMAN AND DALE T. MILLER

context of purely positive prospects. The attributes of certain the raped virgin is more discrepant in the norm that she evokes
objects, such as risky prospects, are hierarchically related. Dom- than the raped divorcee is in hers. The effect is strongly reversed
inant attributes—in this case the presence or absence of a risk when the same subjects judge both cases. Birnbaum (1982) com-
of loss—are relatively immutable and therefore control the re- mented that in a between-subjects design, the stimulus is com-
cruitment of the norm. The attributes that are allowed to vary pletely confounded with its context—which is another way of
in the norm account for most of the variance in judgments of saying that the stimulus brings its context into being.
such objects. The role of the immutable features of a stimulus in recruiting
A similar interpretation applies to another curious phenom- its norms is the same as the role of the category label in recruiting
enon in the evaluation of positive bets: The probability of winning category norms. A category label—whether it is a single name
and the amount of the prize have different weights in choices or a complex specification—constrains the retrieval process to
and in judgments of attractiveness (Goldstein, 1982). Consider, category members. Other factors, such as typicality or recency,
for example, the two bets "a 31/36 chance to win $3" and "a 7/ control further selection among representations of exemplars and
36 chance to win $ 13." In an experiment conducted at the Uni- ultimately determine how strongly individual instances are ac-
versity of British Columbia, the two bets were preferred about tivated. We propose that the immutable features of a stimulus
equally often in a direct choice. When they were rated for at- similarly guide and constrain the spontaneous recruitment of
tractiveness, however, the bet with the higher probability of win- alternatives to it.
ning was rated 12.1 on a 20-point scale, and the bet with the The processes that have been sketched in this discussion of
lower probability of winning was rated only 6.5. As this example comparative judgment yield norms for the various attributes of
illustrates, the probability attribute accounts for most of the a stimulus. The availability profile (see Figure 1) defines a range
variance in ratings of attractiveness. This is the pattern of judg- of possible values and provides a proxy for a frequency distri-
ments that would be expected if a bet that offers a probability P bution. In particular, the rank of the stimulus in its norm is
to win $X is compared mainly to other bets in which the same readily computable from this information. Thus, a norm has
amount Xcan be won with different probabilities. The dominant the characteristics necessary to support the scaling operations
role of probability in these judgments fits easily within the present envisaged in range-frequency theory (Parducci, 1965, 1974) and
framework, with the added assumption that the amount to be in related analyses of comparative judgment (Birnbaum, 1982;
won is the central attribute of a positive bet—and therefore the Mellers & Birnbaum, 1983). More complex judgments involve
attribute most likely to be adopted as immutable in the recruit- matching values of different attributes by their position in their
ment of a norm. Indeed, most people state that they would rather respective norms. For example, nonregressive predictions are
know the prize than the probability of winning in assessing the made by choosing a criterion value that matches the position of
attractiveness of a bet. the predictor feature in its norm (Kahneman & Tversky, 1973).
Evaluations of the attractiveness of bets illustrate what appears Similarly, judgments of equity appear to be based on a compar-
to be a more general effect. Consider two factors that may de- ison of the relative positions of an individual in norms of salary
termine the impression of the career success of a civil servant: and merit (Mellers, 1982).
rank and performance ratings. The hierarchical ordering of the
two attributes is clear: Rank and the tasks associated with it are
commonly presupposed in evaluating performance, and the very Mutability and the Availability of Counterfactuals
meaning of successful performance is altered by varying rank.
One theme of this article is that the experienced facts of reality
Now imagine two civil servants of different ranks who are judged
evoke counterfactual alternatives and are compared to these
equally successful in direct comparison, because the one at the
alternatives. The development of this theme takes us to regions
lower rank has a higher performance rating. The present analysis
more often traveled by philosophers than by psychologists. Phil-
entails that when the success of the two individuals is evaluated
osophical treatments of counterfactuals and possible worlds have
in isolation, the judgments will be mainly determined by the
explored the compelling intuition that some alternatives are closer
performance ratings. The general rule is that, other things being
to reality than others and that some changes of reality are smaller
equal, the more mutable and less important of two attributes
than others (see Lewis, 1973, for a particularly engaging treat-
will have a disproportionately large effect in single-stimulus
judgments. ment). As Hofstadter (1985) noted, the word "almost" provides
The application of this analysis of mutability is straightforward a key to some of these intuitions. For example, the statement "I
when a single object is presented for judgment, with no prior almost caught the flight" is appropriate for an individual who
experimental context. The situation is ambiguous in a within- reached the departure gate when the plane had just left but not
subject design, where several objects are evaluated in immediate for a traveler who arrived half an hour late. The world in which
succession. The early items in the series are likely to play a sig- the passenger arrives five minutes earlier than she did is closer
nificant role in the evaluation of later ones, and some of the to reality than a world in which she arrives half an hour earlier.
paradoxical effects of the hierarchy of attributes are weakened The present analysis links these intuitions to mutability: A coun-
or even reversed in such cases. When positive and mixed bets terfactual possibility should appear "close" if it can be reached
are evaluated in the same series, for example, the positive ones by altering some mutable features of reality.
tend to be rated higher. Birnbaum (1982) described another par- Our notion of mutability is similar to the concept of slippability
adoxical result that is eliminated in a within-subject design. When introduced by Hofstadter (1979, 1985). The shared ideas are
a single case is evaluated, judgments of a rape victim's respon- that the mental representation of a state of affairs can always be
sibility are higher for a virgin than for a divorcee, perhaps because modified in many ways, that some modifications are much more
 NORM THEORY 143

natural than others, and that some attributes are particularly time or route altered the exceptional value and made it normal.
resistant to change. The results support two related propositions about the availability
Another cognate of mutability is the distinction between the of counterfactual alternatives: (a) Exceptions tend to evoke con-
information presupposed and the information asserted in a verbal trasting normal alternatives, but not vice versa, and (b) an event
message (Clark & Haviland, 1977). The cleft sentence "It was is more likely to be undone by altering exceptional than routine
Tom who set fire to the hotel" designates an immutable aspect aspects of the causal chain that led to it.
of the situation (the hotel was set on fire) and a mutable one (the
identity of the arsonist). Note that either aspect of the sentence Ideals and Violations
could be presupposed: "It was a hotel that Tom set on fire" has Barsalou (1985) and Lakoff (in press) have emphasized the
the same basic structure, with the two components interchanging role of distance from an ideal or paragon as a determinant of
their roles. The cleft sentence invites the listener to consider al- typicality. For example, zero-calorie foods are judged to be highly
ternatives to the asserted content, even as it denies these alter- typical members of the category "things to eat on a diet," although
natives. The presupposition is shared by all the alternatives. they are neither the most common nor the most similar to the
As this example illustrates, presuppositions are highly flexible, prototypical diet food (Barsalou, 1985). In the terms of the pres-
and the relative mutability of attributes can be controlled almost ent model, elements that have ideal values on significant attributes
at will. In the absence of deliberate intent or conversational guid- appear to be highly available. A hypothesis about differential
ance, however, differences in the mutability of attributes will mutability follows: When an alternative to an event could be
affect the spontaneous recruitment of norm elements. produced either by introducing an improvement in some ante-
In the following sections we examine several hypotheses about cedent or by introducing a deterioration, the former will be more
factors that determine the relative mutability of different aspects available.
of an event. We also illustrate some of the ways in which the Evidence for this proposition was obtained in unpublished
elusive concept of "availability of counterfactual alternatives" research by D, Read (1985). Subjects were taught the rules of a
can be operationalized. simple two-person card game. They were then shown pictures
of the players' hands and were asked to complete the blanks in
Exception and Routine the following statement by changing one card: "The outcome
would have been different if the had been a " The
A complex situation may combine some routine and some question of interest was whether the subjects would choose to
exceptional features. Kahneman and Tversky (1982) tested the weaken the winning hand or to strengthen the losing one. The
hypothesis that exceptional features are more mutable than rou- rule discussed in the preceding section suggests that the winning
tine ones by eliciting alternatives to a stipulated reality. Subjects hand might be more readily altered, since the strongest combi-
were given a story describing a fatal road accident, in which a nations (e.g., four of a kind) are more exceptional than weaker
truck driven by a drug-crazed teenager ran a red light and crashed ones (e.g., three of a kind). However, the tendency to eliminate
into a passing car, killing Mr. Jones, its occupant. The following exceptions was overcome in the data by a tendency to approach
instructions were given: an ideal value. In a significant majority of cases, subjects chose
to modify the outcome by strengthening the losing hand rather
As commonly happens in such situations, the Jones family and their than by weakening the stronger one. Informal observations of
friends often thought and often said "If only . . ." during the days spectators at sports events suggest that the outcome of a contest
that followed the accident. How did they continue that thought?
Please write one or more likely completions. is more commonly undone by improving the losing performance
(e.g., imagining the successful completion of a long pass in the
Two versions of the story were constructed, labeled route and last seconds) than by imagining a poorer performance of the
time, which were identical except for one paragraph. In the route winning team.
version the critical paragraph read as follows: Differential availablility of changes that improve or degrade a
performance could be one of the factors that explain the answers
On the day of the accident, Mr. Jones left his office at the regular to the following question:
time. He sometimes left early to take care of home chores at his
wife's request, but this was not necessary on that day. Mr. Jones did Tom and Jim both were eliminated from a tennis tournament, both
not drive home by his regular route. The day was exceptionally clear on a tie-breaker. Tom lost when his opponent served an ace. Jim lost
and Mr. Jones told his friends at the office that he would drive along on his own unforced error. Who will spend more time thinking about
the shore to enjoy the view. the match that night?
The time version of this paragraph was as follows: Jim 85% Tom 15% (N=92)

On the day of the accident, Mr. Jones left the office earlier than Note that Tom and Jim could both imagine themselves winning
usual, to attend to some household chores at his wife's request. He the game, but the judgment of our subjects is that these thoughts
drove home along his regular route. Mr. Jones occasionally chose to are likely to be more available when they involve an imagined
drive along the shore, to enjoy the view on exceptionally clear days,
but that day was just average.
improvement of one's own performance than an imagined de-
terioration of the opponent's game.
Both versions suggest route and time as possible attributes that
might be changed to undo the accident, but the change introduces Reliable and Unreliable Knowledge
an exception in one case, and restores the routine in the other. Tversky and Kahneman (1982) noted an asymmetry in the
As predicted, over 80% of the responses that mentioned either confidence with which people made inferences and predictions
144 DANIEL KAHNEMAN AND DALE T. MILLER

from one attribute to another. Inferences from a reliable measure process (Suls & Miller, 1977). People should prefer to compare
to an unreliable one are made with greater confidence than in- themselves to others who resemble them on causal factors rather
ferences in the opposite direction, although correlation is actually than to others who resemble them on outcome variables. For
symmetric. For example, people were more confident in pre- example, a student should be more interested in discovering how
dicting the score on a short IQ test from a long form of the test well a peer of comparable industry did on an exam than in dis-
than vice versa. They also believed, erroneously, that the state- covering the industriousness of a student who achieved a com-
ment "The individual who won the decathlon won the first event" parable grade.
is more probable than "The individual who won the first event
won the decathlon." Focal and Background Actors
Similar asymmetries are observed when subjects are presented
with a statement that includes an apparent discrepancy and are We propose that the mutability of any aspect of a situation
asked to choose how to eliminate the discrepancy. Most subjects increases when attention is directed to it and that unattended
do so by altering the less reliable item (e.g., the performance on aspects tend to become part of the presupposed background.
the short test or in the first event of the decathlon) to fit the more The hypothesis that the attributes of a focal object or agent are
reliable one. Thus, attributes about which little is known appear more mutable than those of nonfocal ones was explored by D.
to be relatively mutable. Read (1985). In the card-game study described earlier, he showed
subjects the hands of two players, A and B, and asked them to
Causes and Effects "complete stems such as "A would have won if . . ." or "A
We propose the hypothesis that, when people consider a cause- would have lost if ... ." The large majority of completions
effect pair, alternatives to the effect will be more available than involved changes in the hand held by A, although the same out-
alternatives to the cause. The tendency to presuppose causes is come could have been generated just as well by altering B's hand.
reflected in everyday conversation. When an observation departs Other tests of the hypothesis used vignettes such as the following:
from the normal covariation of cause and effect, the discrepancy Helen was driving to work along a three-lane road, where the middle
is usually attributed to the effect rather than to the cause. Thus, lane is used for passing by traffic from both directions. She changed
a child may be described as "big for her age" but not as "young lanes to pass a slow-moving truck, and quickly realized that she was
headed directly for another car coming in the opposite direction.
for her size," and students may be described as overachievers, For a moment it looked as if a collision was inevitable. However,
not as undertalented. this did not occur. Please indicate in one sentence how you think
When a particular conjunction of effect and causal attribute the accident was avoided.
is observed, the alternatives that are recruited should mainly
consist of cases in which the same cause is followed by variable The situation of the two cars is symmetric, or perhaps biased
effects. A spectator at a weight lifting event, for example, will against Helen's being able to do much to prevent the accident,
find it easier to imagine the same athlete lifting a different weight given that the circumstances of the other car are not described.
than to keep the achievement constant and vary the athlete's Nevertheless, a substantial majority of subjects completed the
physique. We turned this hunch into a small experiment. The story by ascribing the critical action to Helen.
participants were given information that was described as a form The idea that the actions of a focal individual are mutable
sheet for the members of a club of weight lifters. The data were may help explain the well-documented tendency for victims of
presented in two columns, stating the body weight of each athlete violence to be assigned an unreasonable degree of responsibility
and his best achievement (the order of columns was varied in for their fate (Lerner & Miller, 1978). Information about a
alternate forms). The numbers in both columns were in strict harmful act often presents the actions of the perpetrator in a
ascending order. Data for 10 athletes were given, with the 10th way that makes them part of the presupposed background of the
observation deviating markedly from the trend established in the story, and therefore relatively immutable. Alternatives to the vic-
first 9 cases. Half of the participants received a form in which tim's actions are likely to be more mutable, and counterfactual
the 10th athlete was only heavier than the 9th by 3 kg but lifted scenarios in which the harm is avoided are therefore likely to be
30 kg more. In the other forms the 10th athlete was 30 kg heavier ones that change the victim's actions but keep the aggressor's
than the 9th and lifted only 3 kg more. All forms included the behavior essentially constant. The high availability of such coun-
following question: terfactual scenarios can induce an impression that the victim is
responsible for her fate—at least in the sense that she could easily
Do you find the relationship between body weight and lifted weight have altered it. Any factor that increases the attention focused
in the last entry surprising? (Yes/No) If you do, please change it to on the victim increases the availability of alternatives to the vic-
make it conform better to what you would expect. Mark your change
on the sheet and return it. tim's reactions and the blame attached to the victim. The finding
that emotional involvement with victims can increase the blame
Most subjects found the entry to be surprising and changed attributed to them (Lerner, 1980) is consistent with this specu-
only one of the two items of information on the critical line. As lation.
predicted, a substantial majority (86%) of those who changed Our analysis is based on the idea that features of a situation
one item altered the weight lifted by the athlete rather than his that have highly available alternatives are attributed greater causal
body weight. The order of the two columns on the sheet did not effectiveness than equally potent but less mutable factors. This
matter. analysis does not imply, of course, that causal responsibility is
The differential mutability of effects and causes suggests an never assigned where it belongs. Probes that draw attention to
apparently untested hypothesis concerning the social comparison the perpetrator, such as those concerning the punishment to be
 NORM THEORY 145

given, can be expected to evoke appropriate degrees of blame. to arise from the availability of a counterfactual construction.
The point made here is simply that the actions of the perpetrator, Mr. C. and Mr. D. differ in the ease with which they can imagine
unless made salient, are likely to be presupposed, with the result themselves—contrary to fact—catching up with their flights. This
that constructed scenarios of how the victimization might have is easier for Mr. C., who needs only to imagine making up 5
been avoided will focus on the actions of the victim. Although minutes, than for Mr. D., who must construct a scenario in which
seemingly paradoxical, it is not a new discovery that apparent he makes up half an hour. A preferred alternative is thus more
immutability reduces attributions of responsibility. This fact has available (normal) for Mr. C. than for Mr. D., which makes his
been known to bullies since time immemorial. experience more upsetting.
The hypotheses that we have discussed illustrate, but do not The next demonstration tests the prediction that outcomes
exhaust, the factors that control the differential mutability of that follow exceptional actions—and therefore seem abnormal—
attributes and the differential availability of alternatives to reality. will elicit stronger affective reactions than outcomes of routine
In other exploratory work we have found suggestive support for actions.
several additional hypotheses. One of these hypotheses is that
temporal order affects mutability. Consider the pair of consonants Mr. Adams was involved in an accident when driving home after
work on his regular route. Mr. White was involved in a similar ac-
XF. Now quickly change it by replacing one of the letters by cident when driving on a route that he only takes when he wants a
another consonant. Which letter did you change? A robust ma- change of scenery. Who is more upset over the accident?
jority of respondents replace the second letter rather than the Mr. Adams 18% Mr. White 82% (N=92)
first. This example illustrates a general rule: The second member
of an ordered pair of events is likely to be more mutable than As predicted, the same undesirable outcome is judged to be
the first. Another hypothesis is that a change that can be visualized more upsetting when the action that led to it was exceptional
(e.g., undoing an accident) is more available than a change that than when it was routine.
cannot be visualized (e.g., undoing a heart attack). The next example tests the prediction that people are most
In concluding this section, we note that the differential avail- apt to regret actions that are out of character:
ability of counterfactual alternatives defines the grain of expe-
Mr. Jones almost never takes hitch-hikers in his car. Yesterday he
rience—the fault lines along which reality is likely to be undone gave a man a ride and was robbed. Mr. Smith frequently takes hitch-
in mental transformations (Hofstadter, 1985). In particular, dif- hikers in his car. Yesterday he gave a man a ride and was robbed.
ferences of mutability determine which among the many alter- Who do you expect to experience greater regret over the episode?
natives to a particular occurrence will be viewed as normal. Mu- Mr. Jones 88% Mr. Smith 12%
tability also determines which aspects of reality are likely to be Who will be criticized most severely by others?
accepted with relatively little questioning. In the next section we
Mr. Jones 23% Mr. Smith 77% (N = 138)
consider some effects of the differential availability of counter-
factuals on the intensity of emotional responses. The results confirm the hypothesis and incidentally indicate that
regret cannot be identified with an internalization of others'
Affective Role of Counterfactuals blame.
The final example (Kahneman & Tversky, 1982) tests the hy-
This section develops some implications of the analysis of
pothesis that consequences of actions evoke stronger emotional
counterfactual thought for the domain of affect. We examine
responses than consequences of failures to act. The intuition to
this question in conjunction with a hypothesis of emotional am-
which this example appeals is that it is usually easier to imagine
plification, which states that the affective response to an event is
oneself abstaining from actions that one has carried out than
enhanced if its causes are abnormal. In each of the following
carrying out actions that were not in fact performed.
examples, the same misfortune is produced by two sequences
of events, which differ in normality. The respondents assess the Mr. Paul owns shares in company A. During the past year he con-
intensity of the affective responses that are likely to arise in the sidered switching to stock in company B, but he decided against it.
two situations. The first demonstration (Kahneman & Tversky, He now finds out that he would have been better offby $1,200 if he
1982) tested the prediction that outcomes that are easily undone had switched to the stock of company B. Mr. George owned shares
in company B. During the past year he switched to stock in company
by constructing an alternative scenario tend to elicit strong af- A. He now finds that he would have been better off by $1,200 if he
fective reactions. had kept his stock in company B. Who feels greater regret?
Mr. C and Mr. D were scheduled to leave the airport on different Mr. Paul 8% Mr. George 92% (N = 138)
flights, at the same time. They traveled from town in the same lim-
ousine, were caught in a traffic jam, and arrived at the airport 30 The finding that acts of commission produce greater regret than
minutes after the scheduled departure time of their flights. Mr. D is acts of omission was replicated by Landman (1984) and is in
told that his flight left on time. Mr. C is told that his flight was
delayed, and only left 5 minutes ago. Who is more upset? accord with formulations that distinguish omission from com-
mission in attributions of causality and responsibility (Hart &
Mr. D 4% Mr. C 96% (N= 138)
Honore, 1959; Heider, 1958).
There is essentially unanimous agreement that Mr. C. is more Miller and McFarland (1986) conducted a series of studies to
upset than Mr. D., although their objective situations are iden- test the hypothesis that the abnormality of a victim's fate affects
tical—both have missed their planes. Furthermore, their expec- the sympathy that the victim receives from others. Subjects were
tations were also identical, since both had expected to miss their told that the purpose of the studies was to provide victim com-
planes. The difference in the affective state of the two men appears pensation boards with information about the public's view of
146 DANIEL KAHNEMAN AND DALE T. MILLER

various types of victims. Subjects were presented with a brief Codes and Category Norms in Person Perception
description of an incident and then were asked to indicate on an
11-point scale how much compensation they thought the victim The present approach, like several others (N. H. Anderson,
should receive. Normality was manipulated in one study by 1981; Hastie & Kumar, 1979; Higgins & Lurie, 1983; Wyer &
varying the mutability of an action that led to a bad outcome. Gordon, 1984; Wyer & Srull, 1981), assumes the possibility of
The victim was a man who had been severely injured during a dual memory representations—raw memories of episodes and
robbery. In one condition, the robbery took place in the store at stored codes. Many features of a person can be stored (a) as
which the victim shopped most frequently. In a second condition, comparative trait labels, which assign the individual to a position
the robbery took place in a store to which the victim had decided in an interpersonal norm, (b) as a set of episodes that define an
to go only after finding his regular store closed for renovations. intrapersonal norm of behavior for that individual, or (c) in both
It was predicted that subjects would view the fate that befell the forms at once. In this section we pursue the implications of norm
victim in the "unusual" store to be more abnormal, and hence theory for two aspects of knowledge about persons: the ambiguity
more unjust, than the fate that befell the victim in the "usual" of codes and the formation and retrieval of intrapersonal norms.
store. Consistent with this hypothesis, subjects recommended We discuss these issues in turn.
significantly more compensation (over $100,000 more) for the
same injury in the exceptional context than they did in the routine Norms and Ambiguous Codes
context.
This study demonstrates that even morally charged judgments As was seen earlier in the discussion of comparative judgment,
such as those involving compensation can be influenced by the the interpretation of a comparative code is necessarily dependent
normality of the outcome. It is as though a negative fate for on the norm to which the object of judgment is related. For
which a more positive contrast is highly available is worse or example, hearing that "Jim has been given a long jail term" will
more unfair than one for which there is no highly available pos- suggest different jail terms to listeners, if they differ in the norm
itive alternative. It is important to note that the different reaction for jail terms that they attribute to the speaker. Communication
to the two victims is not due to the perceived probability of their will fail if speakers and listeners do not share, or at least coor-
fate. The probabilities of being shot in the two stores were both dinate, their norms. Coordination of norms is also involved when
judged very low and indistinguishable from one another. The an individual uses a remembered code to reconstruct the literal
subjects apparently presupposed both the robbery and the store detail of an experience—for example, the length of a jail sentence
at which it took place. Given this presupposition, it is relatively that is only remembered as long. Accurate performance depends
more normal for an individual to be shot where it is normal for on the match between the norm that is applied when the code
that individual to be—in the store that is regularly frequented. is interpreted and the norm that supported the initial judgment.
A second compensation study, paralleling the missed-flight As Higgins and Lurie (1983) demonstrated in an impressive ex-
script, varied the distance between the negative outcome and a periment, the reconstruction of the initial episode will be sys-
more positive alternative. The victim in this study had died from tematically biased if the norm changes in the interval. This effect,
exposure after surviving a plane crash in a remote area. He had which Higgins and Lurie termed change of standard, can yield
made it to within 75 miles of safety in one condition and to a range of cognitive and affective responses, including the dis-
within 'A mile in the second condition. Assuming that it is easier appointment that people often experience when they meet a for-
to imagine an individual continuing another '/4 mile than another mer teacher whom they had always remembered as brilliant
75 miles, it was predicted that the fate of the "close" victim (Higgins & King, 1981).
would be perceived to be more abnormal, and hence more unfair, Trait labels and expressions fall into two categories: (a) relative
than the fate of the "distant" victim. The results supported the predicates, which specify the individual's position on an inter-
prediction, inasmuch as subjects once again recommended sig- personal norm, and (b) absolute predicates, which summarize
nificantly more compensation for the family of the victim whose an intrapersonal norm of actions or feelings on relevant occasions.
fate was more easily undone. The same trait name can sometimes serve in both functions. The
These results confirm the correlation between the perception statement that "Jane is assertive" can be understood as saying
of abnormality of an event and the intensity of the affective re- either that she is more assertive than most people or that her
action to it, whether the affective reaction be one of regret, horror, behavior is assertive on most occasions. In the latter interpre-
or outrage. This correlation can have consequences that violate tation, the word "assertive" is a category label, which evokes
other rules of justice. An example that attracted international exemplars of assertive behaviors.
attention a few years ago was the bombing of a synagogue in Trait labels that have both an absolute and a relative sense are
Paris, in which some people who happened to be walking their potentially ambiguous. This ambiguity appears to underlie the
dogs near the building were killed in the blast. Condemning the tendency of people to accept general descriptions as uniquely
incident, a government official singled out the tragedy of the relevant to them, also known as the Barnum effect (Snyder,
"innocent passers-by." The official's embarrassing comment, with Shenkel, & Lowery, 1977). Barnum statements typically evoke
its apparent (surely unintended) implication that the other vic- both absolute and relative interpretations. For example, the
tims were not innocent, merely reflects a general intuition: The statement "You are shy in the presence of strangers" can be
death of a person who was not an intended target is more poignant recognized by most people as a valid description of themselves—
than the death of a target. Unfortunately, there is only a small if shyness indicates that one is less comfortable with strangers
step from this intuition to the sense that the persons who are than with familiar others. In its relative sense, of course, the
chosen as targets thereby lose some of their innocence. description is applicable only to the minority of people who are
 NORM THEORY 147

sufficiently extreme to deserve special mention. It is the validity fact that the unhelpful behavior contradicted the observers' prior
of the Barnum statement in its absolute sense that makes it be- beliefs and expectancies.
lievable to the individual, but it is the unwarranted extension to Radical generalization from observed behavior to an intra-
the interpersonal comparison that makes it interesting. A similar personal norm is manifest in the nonregressiveness of behavioral
mixture of meanings was noted by Higgins and Winter (cited in predictions (Kahneman & Tversky, 1973; Ross & Anderson,
Kraut & Higgins, 1984) in their analysis of trait ambiguity. They 1982). A single observed instance of unusually generous tipping
asked subjects what percentage of people possessed various per- appears sufficient to set up a norn that will be consulted in pre-
sonality traits (e.g., friendliness, aggressiveness) and found many dictions of a person's future tips. The interpersonal base rate for
traits that were assumed to apply to an absolute majority of the behavior of other people is effectively excluded from consid-
people. It appears that the assignment of these traits involves a eration when thinking about this person's tips, just like one's
mixture of comparative and absolute criteria. knowledge of the size of insects in the previous example.
Trait descriptions vary in the degree to which they lend them- Radical generalizations can be made with high and unwar-
selves to interpersonal and intrapersonal interpretation. For ex- ranted confidence. Experimental results suggest that subjective
ample, the expression "He is not very intelligent" evokes an confidence depends mainly on the consistency of available evi-
interpersonal norm, but the statement "I like Coke" has an in- dence rather than on its quality or quantity (Einhorn & Hogarth,
trapersonal reference (more than other beverages, not necessarily 1982; Kahneman & Tversky, 1973; Tversky & Kahneman, 1971).
more than other people). The difference has a predictable effect Subjective confidence in a prediction is likely to reflect the
on the impact that information about others has on self-descrip- breadth, or variability, of the category norm on which the judg-
tion. Whether a person is more or less intelligent than her ref- ment is based (see Figure 1). Confidence will be high if all ele-
erence group will influence how intelligent she judges herself to ments cluster around a single value of the attribute, independently
be (Davis, 1966), but whether a person consumes more or less of the number of these elements. However, it would be incorrect
of a drink than her reference group will not influence her ex- to conclude that judgments are entirely insensitive to the quantity
pressed liking for the drink (Hansen & Donoghue, 1977; Nisbett, of evidence. The model of Figure 1 implies, in particular, that
Borgida, Crandall, & Reed, 1976). the susceptibility of a norm to change depends on the absolute
height of the availability profile, and thus on the number of ele-
Norms From Single Elements ments in the norm. A category norm that is based on one or two
The present model of category norms assumes both that a elements can support confident judgments, but it can also be
single instance suffices to set up a norm and that a reference to altered relatively easily under the impact of new evidence.
the category label serves to restrict consideration to members of Many inferences bridge across situations and across attributes.
that category. These assumptions entail nonregressive predictions In the absence of better evidence, people readily predict success
and a tendency to neglect relevant base rates. in graduate school from an IQ test score, research productivity
To illustrate, imagine that you were shown a single exemplar from performance in a colloquium, or the size of a mother's
of an unfamiliar species of insect, which was larger than any graduation gift to her daughter from the size of a tip that she
other insect you have ever seen. How large would you expect gave to a waiter. Such generalizations are involved both in delib-
another exemplar of that species to be? Would it occur to you erate predictions of future behaviors and in spontaneous infer-
that the single insect you saw is likely to be larger than most of ences about traits and about unobserved aspects of situations.
its conspecifics? The canons of inference prescribe that the single Inferences that bridge attributes are no more regressive than
insect you saw is likely to be larger than most others in its species. direct generalizations (Nisbett & Ross, 1980; Ross & Anderson,
Predictions of the size of the next member of the new species 1982). Nonregressive predictions can be generated in two ways.
should accordingly regress toward the general insect norm. In- First, the predicted value may be chosen so that its position
tuitive expectations, in contrast, are firmly centered on the single matches that of the known attribute in their respective inter-
observed value, which constitutes the norm for that category. In personal norms (Kahneman & Tversky, 1973). Thus, a predicted
the context of social judgment, the ease with which category graduation gift can be chosen that is as extreme in the norm for
norms are established leads to radical generalization from a single gifts as the observed tip is in the interpersonal norm for tips.
observation of behavior to an interpersonal norm for the behavior Second, nonregressive predictions can also be generated by
of other people in the same setting, or to a norm for a person's matching descriptive labels (e.g., if a tip is remembered as very
behavior on future occasions (Nisbett & Ross, 1980; S. Read, generous, what graduation gift would be considered equally gen-
1984). erous?). The intention to predict a behavior elicits a search for
Radical generalization to interpersonal norms was neatly relevant incidents or for pertinent descriptive labels. The search
demonstrated by Hamill, Wilson, and Nisbett (1980), who is guided by the similarity of potential elements to the target
showed that subjects centered their norms for an entire social attribute, and it is probably concentric: The nearest incidents or
group (prison guards, in one study) on an observation of a single labels that turn up in the search are used, nonregressively, to
member, even when he was described as atypical. In the same generate a prediction. In the absence of better data, people are
vein, Nisbett and Borgida (1975) exposed observers to brief and willing to make extreme predictions from evidence that is both
innocuous interviews with two students, allegedly participants flimsy and remote. The process of concentric search yields radical
in an experiment on helping behavior. From the information (and overconfident) predictions from observations of dubious
that these two individuals had not helped a stranger in distress, relevance. However, the same process also makes it likely that
the observers generalized that most people would also fail to help distant labels or incidents will be ignored when evidence that is
in the same situation. This generalization occurred despite the closer to the target attribute is available. Thus, generous tipping
148 DANIEL KAHNEMAN AND DALE T. MILLER

habits will not be given much weight in predicting a mother's concerning those opinions. As in the preceding example, this
graduation gift to her daughter if pertinent incidents of their norm will be biased even for an observer who "knows" that the
interaction can be retrieved. actor's behavior is constrained, and therefore uninformative.
Quite simply, pro-Castro behaviors are more normal for the actor
Unwanted Elements of Norms than for a random stranger. Belief perseverance, generalizations
The elicitation of a norm has been described here as a process from atypical examples, and failures of discounting all illustrate
of parallel activation of multiple representations that are recruited the same principle: Any observation of behavior—even if it is
by a stimulus or by a category. We have assumed that the process discounted or discredited—increases the normality of subsequent
of recruitment is rapid, automatic, and essentially immune to recurrences of compatible behaviors.
voluntary control after its initiation. In particular, it is not pos-
sible for the individual to sift through activated elements and Causal Questions and Answers
discard irrelevant or misleading ones. This limitation on the vol- In this section we consider the role of norms in causal judg-
untary control of norms helps explain two well-documented ments. This topic was chosen to illustrate the concepts of pre-
phenomena of social judgment: the perserverance of discredited supposition and norm coordination that were introduced earlier.
beliefs and the correspondence bias in person perception. We begin by examining a routine conversational exchange, which
Imagine a discussion of a Canadian athlete, in which someone provides a conceptual model for much attribution research. A
who is unfamiliar with metric measures reads from a sheet: questioner, whom we call Quentin (Q), asks a why question and
"Brian weighs 102 kg. That's 280 Ib, I think. No, it's actually receives an answer from Ann (A). An example might be:
about 220 Ib." Will the speaker's initial error affect listeners'
subsequent responses to questions about Brian's size and Q: "Why did Joan pass this math exam?"
strength? The literature on perseverance of discredited beliefs A: "She used the Brown textbook."
(C. A. Anderson, 1983; Fleming & Arrowood, 1979; Ross & An- We focus on two issues: (a) the inferences that A must make
derson, 1982; Ross & Lepper, 1980; Schul & Bernstein, 1985) about Q's norms to interpret a why question and (b) the con-
suggests that it will. The message of this literature is that traces straints that this interpretation of the question places on appro-
of an induced belief persist even when its evidential basis has priate answers.
been discredited. The discarded message is not erased from
memory, and the norm elicited by a subsequent question about Norms and Causal Questions
Brian's weight could therefore contain the original message as
well as its correction. Thus, a listener might "know" immediately Causal questions about particular events are generally raised
after the message that Brian's true weight is 220 Ib, and this only when these events are abnormal. The close connection be-
value would presumably retain an availability advantage, but the tween causal reasoning and norms is evident in the rules that
category norm associated with Brian's weight would still be biased govern the homely why question as it is used in conversations
toward the erroneous value of 280 Ib. Judgments that depend about particular events (Lehnert, 1978). The why question im-
indirectly on the activation of the norm would be biased as well. plies that a norm has been violated. Thus, the question "Why
Some aspects of the phenomenon known as the correspondence was John angry?" indicates Q's belief that it was normal for him
bias (Jones, 1979) or the fundamental attribution error (Ross, not to be, and the question "Why was John not angry?" indicates
1977) could be explained in similar terms. Many studies have the contradictory belief. Even the question "Why is John so nor-
shown that people make unwarranted inferences concerning mal?" implies that he is normal to an abnormal degree. A why
personal traits and attitudes from observations of behavior that question, then, presupposes that some state X is the case, and
is in fact entirely constrained by the situation. Subjects in one also implies an assertion that not-X was normal. The strongest
famous series of experiments in this tradition observed an in- indication of the implicit assertion of a norm is that the why
dividual who was explicitly instructed to write an essay or to question, unlike most others, can be denied. The denial can be
read aloud a speech advocating an unpopular position (Jones, expressed by a question, as in the familiar exchange:
1979; Jones & Harris, 1967; Jones & McGillis, 1976). In response
Q: "Why do you so often answer a question with a question?"
to subsequent questions, observers commonly attributed to target A: "Why not?"
persons an attitude consistent with the position that these persons
had been constrained to advocate. The why-not retort denies the assertion that not-X is normal. It
In these experiments, as in studies of discredited beliefs, a is a legitimate answer that, if accepted, leaves nothing to be ex-
behavioral observation is accompanied by information that plained.
challenges its validity. As in a theatrical performance, the actor We suggest that why questions (at least those of the deniable
in the experiments of Jones and his colleagues engages in a be- variety, for which "why not?" is a sensible answer) are not requests
havior (e.g., advocating the regime of Fidel Castro) that does not for the explanation of the occurrence or nonoccurrence of an
have its usual significance because of the special demands of the event. A why question indicates that a particular event is sur-
situation. We propose that both in the theater and in these ex- prising and requests the explanation of an effect, denned as a
periments two traces are laid down: (a) a literal memory of the contrast between an observation and a more normal alternative.
person expressing pro-Castro sentiments and (b) a memory of A successful explanation will eliminate the state of surprise. This
the behavior in the context of the situational constraints. Both is commonly done in one of three ways. First, A may deny the
memories are elements of the set that will be evoked by further implied assertion that X is abnormal in the light of what Q
observations of the actor's political opinions or by a question already knows. This is the why-not answer, which invites Q to
 NORM THEORY 149

change his opinion that X is abnormal. Second, A may inform partners usually trip over her feet, there was special reason to
Q of some fact of which Q was previously ignorant, which makes expect Ralph to be more fortunate.
the occurrence of X normal. For example, Q may not have known An experimental demonstration of the ambiguity of why
that Joan used the Brown text, which he knows to be excellent. questions was described by Miller (1981). Several groups of stu-
Third, A may indicate that there is a causal link, of which Q was dent and graduate nurses were asked to explain their decision to
presumably ignorant, between the effect X and some known as- enter the nursing profession. Different versions of the same basic
pect of the situation. For example, Q may know that Joan had question were used in the different groups. The basic question
used the Brown textbook, but he may need to be told that it is was "Why did you go into nursing?" An analysis of the answers
excellent. to this question indicated that student nurses cited significant
The choice of causal feature is constrained in many ways, aspects of nursing (e.g., "it is a respected profession") more often
which have been extensively discussed by philosophers (for par- than did graduate nurses. On the other hand, the graduate nurses
ticularly relevant treatments, see Hart & Honore, 1959; Mackie, were more likely to cite personal qualities (e.g., "I like people").
1974) and by psychologists (Einhorn & Hogarth, 1986; Kelley, The critical finding of Miller's study was that the differences
1967; Schustack & Sternberg, 1981). We shall not be concerned between students and graduate nurses vanished when they were
with the factors that determine impressions of causal efficacy. asked questions that explicitly specified the relevant norm: "Why
We focus here on a constraint that relates directly to the notion did you decide to go into nursing rather than some other profes-
of norm: A cause must be an event that could easily have been sion?" or "Why did you decide to go into nursing when most of
otherwise. In particular, a cause cannot be a default value among your friends did not?" As expected, the former elaboration
the elements that the event X has evoked. The rule that a default yielded a majority of answers for both groups that referred to
value cannot be presented as a cause was noted by Hart and nursing, whereas the answers to the second question referred
Honore (1959), who observed that the statement "It was the predominantly to personal dispositions. In view of this result,
presence of oxygen that caused the fire" makes sense only if the contrasting answers of the two groups to the unelaborated
there were reasons to view the presence of oxygen as abnormal. why question appear to reflect different interpretations of an
It is important to note, however, that a property need not be ambiguous question rather than different causal beliefs.
statistically unusual to serve as an explanation; it is only pre- Questioners convey cues that broadly specify the content of
cluded from being a default value. Peculiar behaviors of cars the causal answers that they wish to receive (Lehnert, 1978). For
observed on the road are frequently "explained" by reference to example, the questions "Why did Carter lose the 1980 election?"
the drivers being young, elderly, or female, although these are and "Why did Reagan win the 1980 election?" refer to the same
hardly unusual cases. The default value for an automobile driver event but differ in the explanation that they request: some note-
appears to be middle-aged male, and driving behavior is rarely worthy fact about Carter in the first question, about Reagan in
explained by it. the second. In the absence of indications to the contrary, the
subject of the sentence is supposed to be its focus (Pryor & Kriss,
Ambiguities in Causal Questions 1977), and the syntactical form of the question suggests an
equivalent form for the answer.
Conversations in general, and answers to questions in partic-
ular, are governed by subtle rules that determine what is said
Perspective Differences
and what is presupposed or implicated (Clark, 1979;Grice, 1975).
The situation is especially complicated when the conversation is The coordination of the norms that apply to an effect and to
actually a test, as is frequently the case in psychological experi- a proposed cause is illustrated in an example discussed by the
ments. The unique feature of a test is that the questioner is not legal philosophers Hart and Honore in their classic Causation
ignorant or puzzled, as questioners usually are. When the question in the Law (1959):
is ambiguous, the respondent faces the bewildering task of
choosing a state of ignorance for the questioner. A woman married to a man who suffers from an ulcerated condition
The why question appears to be especially susceptible to am- of the stomach might identify eating parsnips as the cause of his
indigestion. The doctor might identify the ulcerated condition as the
biguities. Consider a perennial favorite of attribution research: cause and the meal as a mere occasion, (p. 33)
"Why did Ralph trip on Joan's feet?" (McArthur, 1972). The
event in question is clearly specified, but the effect—defined as The causes chosen by his wife and by the physician refer to the
a contrast between the event and its norm—is not. To answer same event but explain different effects. It is evident from her
this question, the respondent must first identify what it is that answer that the wife is concerned with an exception to an intra-
the experimenter considers surprising. In everyday conversations personal norm: "Why does he have indigestion today but not on
intonation provides a potent cue to the intended meaning of a other days?" The physician, on the other hand, is concerned with
question and to the violated presupposition that underlies it. It an interpersonal norm: "Why does this patient suffer from in-
is instructive to read the question about Ralph and Joan aloud digestion when others do not?" The difference could reflect the
several times, each time stressing a different word. The location role of availability in the recruitment of norm elements: The
of the major stress substantially reduces the number of possible wife is likely to retrieve many memories of recent days on which
interpretations, although it does not suffice to disambiguate the her husband, although known to have an ulcer, did not suffer
question completely. For example, the reading "Why did Ralph indigestion. These memories, which resemble the present oc-
trip over Joan's feet?" suggests either that (a) it is unusual for casion in most respects, will define a norm for it. The physician,
Joan's partners to trip over her feet or that (b) although Joan's of course, is unlikely to have had the same amount of exposure
150 DANIEL KAHNEMAN AND DALE T. MILLER

to the patient. According to the rule that coordinates causes to spontaneously when a significant event violates a norm that it
effects (and explanations to questions), the wife chooses as a evokes (see also Hastie, 1984; Weiner, 1985). Causal search can
cause a property that distinguishes this particular day from other also be prompted by a question that presupposes a violated norm
days, and the physician selects a feature that distinguishes this (see Lalljee & Abelson, 1983). Second, the present approach
patient from other patients. draws attention to the difficulty of assessing the accuracy of at-
The situational attribution made by the wife and the dispo- tributers who differ in their perspectives (Funder, 1982; Monson
sitional attribution made by the physician in Hart and Honore's & Snyder, 1977). It is important to distinguish real disagreements
example recall the actor-observer differences described by Jones in causal attributions from specious disagreements that arise
and Nisbett (1971). Actors often explain their actions and atti- when people answer different questions. Finally, the present
tudes by reference to eliciting properties of situations, whereas analysis identifies a necessary feature of any factor that is con-
observers of the same actions and attitudes attribute them to the sidered a possible cause of a surprising event: A cause cannot
actor's distinctive characteristics. As in Hart and Honore's ex- be a default value of the norm that the consequence has evoked.
ample, the situational attribution corresponds to an intrapersonal
norm, whereas a dispositional attribution of the same behavior Concluding Remarks
relates it to an interpersonal norm. The intuitions about differ-
ential availability that make the indigestion example so com- This essay has proposed a theory of norms. The two main
pelling apply as well to the case of actors and observers. The functions of norms are the representation of knowledge of cat-
question "Why do you like this particular girl?" appears to favor egories and the interpretation of experience. We have challenged
the recruitment of thoughts about the respondent's attitude to- the conception of norms as precomputed structures and have
ward other girls. The question "Why does he like this particular suggested that norms—and sometimes even their elements—are
girl?" is more likely to evoke in an observer thoughts of the constructed on the fly in a backward process that is guided by
attitudes of other people toward that girl (Nisbett, Caputo, Legant, the characteristics of the evoking stimulus and by the momentary
& Maracek, 1973). The different elements that are evoked pro- context. In this regard our treatment resembles other approaches
duce quite different questions: "Why do you like this girl more that emphasize the role of specific episodes and exemplars in
than most other girls?" and "Why does he like this girl more the representation of categories (Barsalou, in press; Brooks, 1978;
than most others do?" Each question, in turn, constrains appro- Hintzman, in press; Jacoby & Brooks, 1984; McClelland & Ru-
priate answers to factors that vary among the elements of the melhart, 1985; Medin & Schaffer, 1978; Schank, 1982). A dis-
evoked set—other girls for the actor, other individuals for the tinctive aspect of the present analysis is the separation of nor-
observer. mality and post hoc interpretation, on the one hand, from prob-
The intuitions about the wife-physician example cannot be ability and anticipation, on the other. Another distinctive aspect
reduced to the accounts commonly offered for actor-observer is our attempt to identify the rules that determine which attri-
differences. The contrast could not be explained by difference butes of experience are immutable and which are allowed to
of knowledge (Jones & Nisbett, 1971) or of perceptual salience vary in the construction of counterfactual alternatives to reality.
(Arkin & Duval, 1975; Storms, 1973). It is not explained by the Our closest neighbor in this enterprise is Hofstadter (1979, 1985),
distinction between a state of self-consciousness and other states with his highly evocative treatment of what he calls slippability.
of consciousness (Duval & Wicklund, 1972). Nor is it compatible Like him, we believe that it is "very hard to make a counterfactual
with the hypothesis that the focus of attention is assigned a dom- world in which counterfactuals were not a key ingredient of
inant causal role (Fiske, Kenny, & Taylor, 1982; Ross, 1977; thought" (Hofstadter, 1985, p. 239).
Taylor & Fiske, 1978), inasmuch as the husband surely plays a Our current understanding of the rules for the retrieval of
more focal emotional role for the wife than for the physician. norm elements far exceeds our understanding of the rules for
The hypothesis of the present treatment is that the same event the construction of unrealized alternatives. We believe that the
evokes different norms in the wife and the physician of the ex- roles of presuppositions and mutability in counterfactual thought
ample, and in actors and observers in other situations. Different define a promising area for future research. Norms and cognate
descriptions of the same event can appear to provide conflicting concepts have often been applied to the study of comparative
answers to the same question, when in fact they are concerned judgment and personality description, but we have argued that
with different questions. This proposal can be subjected to a the concept is also central to numerous other processes, including
simple test: Do the observers actually disagree? A negative answer affective reactions and causal reasoning.
is suggested by several studies. Nisbett et al. (1973) found that
subjects easily adopt a typical observer perspective in reporting References
how their choice of a girlfriend would be described by a close Alba, J. W., & Hasher, L. (1983). Is memory schematic? Psychological
friend. Other data confirm the ability of actors to mimic observers Bulletin, 93, 203-231.
(Miller, Baer, & Schenberg, 1979). On the other hand, observers Anderson, C. A. (1983). Abstract and concrete data in the perseverance
of social theories: When weak data lead to unshakeable beliefs. Journal
instructed to empathize with one of the participants in a dialogue
of Experimental Social Psychology, 19, 93-108.
tend to adopt an actor perspective in explaining that person's Anderson, J. R. (1983). A spreading activation theory of memory. Journal
behavior (Regan & Totten, 1975). of Verbal Learning and Verbal Behavior, 22, 261-295.
In summary, we contend that there are a number of advantages Anderson, N. H. (1981). Foundations of information integration theory.
to viewing the process of causal reasoning from the perspective New York: Academic Press.
of norms. First, our analysis provides an account of the ante- Appley, M. H. (Ed.). (1971). Adaptation-level theory: A symposium. New
cedents of causal reasoning. A search for explanation may occur \brk: Academic Press.
 NORM THEORY 151

Arkin, R. A., & Duval, S. (1975). Effects of focus of attention on the A framework for classification learning. Journal of Experimental Psy-
causal attributions of actors and observers. Journal of Experimental chology: Learning, Memory, and Cognition, 10, 234-257.
Social Psychology, 11, 427-438. Funder, D. C. (1982). On the accuracy of dispositional vs. situational
Avant, L. L., & Kelson, H. (1973). Theories of perception. In B. B. Watson attributions. Social Cognition, 1, 205-222.
(Ed.), Handbook of general psychology (pp. 419-448). Englewood Cliffs, Garner, W. R. (1962). Uncertainty and structure as psychological concepts.
NJ: Prentice Hall. New York: Wiley.
Barsalou, L. W. (1983). Ad hoc categories. Memory and Cognition, 11, Garner, W. R. (1970). Good patterns have few alternatives. American
211-227. Scientist, 58, 34-42.
Barsalou, L. W. (1985). Ideals, central tendency, and frequency of in- Goldstein, W. M. (1982). Inconsistent assessments of preference: Effects
stantiation as determinants of graded structure in categories. Journal of stimulus presentation method on the preference reversal phenomenon.
of Experimental Psychology: Learning, Memory, and Cognition, 11, Unpublished manuscript, University of Michigan.
629-654. Grice, H. P. (1975). Logic and conversation. In P. Cole & J. L. Morgan
Barsalou, L. W. (in press). The instability of graded structure: Implications (Eds.), Syntax and semantics: Speech acts (pp. 41-58). New York: Ac-
for the nature of concepts. In U. Neisser (Ed.), Concepts reconsidered: ademic Press.
The ecological and intellectual bases of categories. New "Vbrk: Cam- Hamill, R., Wilson, T. D., & Nisbett, R. E. (1980). Insensitivity to sample
bridge University Press. bias: Generalizing from atypical cases. Journal of Personality and Social
Birnbaum, M. H. (1982). Controversies in psychological measurement. Psychology, 39, 578-589.
In B. Wegener (Ed.), Social attitudes andpsychophysical measurement Hansen, R. D., & Donoghue, J. M. (1977). The power of consensus:
(pp. 401-485). Hillsdale, NJ: Erlbaum. Information derived from one's own and other's behavior. Journal of
Brooks, L. R. (1978). Nonanalytic concept formation and memory for Personality and Social Psychology, 35, 294-302.
instances. In E. Rosch & B. B. Lloyd (Eds.), Cognition and categori- Hart, H. L. A., & Honore, A. M. (1959). Causation in the law. London:
zation (pp. 169-215). Hillsdale, NJ: Erlbaum. Oxford University Press.
Brooks, L. R. (in press). Decentralized control of categorization: The Hastie, R. (1981). Schematic principles in human memory. In E. T. Hig-
role of prior processing episodes. In U. Neisser (Ed.), Concepts recon- gins, C. P. Herman, & M. P. Zanna (Eds.), Social cognition: The Ontario
sidered: The ecological and intellectual bases of categories. New "Vbrk: symposium (Vol. 1, pp. 39-88). Hillsdale, NJ: Erlbaum.
Cambridge University Press. Hastie, R. (1984). Causes and effects of causal attribution. Journal of
Brown, D. R., & Reich, C. M. (1971). Individual differences in adaptation- Personality and Social Psychology, 46, 44-56.
level theory. In M. H. Appley (Ed.), Adaptation-level theory: A sym- Hastie, R., & Kumar, P. A. (1979). Person memory: Personality traits as
posium (pp. 215-231). New York: Academic Press. organizing principles in memory for behaviors. Journal of Personality
Clark, H. H. (1979). Responding to indirect speech acts. Cognitive Psy- and Social Psychology, 37, 25-38.
chology, 11, 430-474. Heider, F. (1958). The psychology of interpersonal relations. New York:
Clark, H. H., & Haviland, S. E. (1977). Comprehension and the given- Wiley.
new contract. In R. O. Freedle (Ed.), Discourse production and com- Helson, H. (1964). Adaptation level theory: An experimental and system-
prehension (pp. 1-40). Norwood, NJ: Ablex. atic approach to behavior. New \brk: Harper.
Coren, S., & Girgus, J. S. (1978). Seeing is deceiving: The psychology of Helson, H., & Kozaki, A. (1968). Effects of duration of series and anchor
visual illusions. Hillsdale, NJ: Erlbaum. stimuli on judgments of perceived size. American Journal of Psychology,
Coren, S., & Miller, J. (1974). Size contrast as a function of figural sim- 81, 291-302.
ilarity. Perception and Psychophysics, 16, 355-357. Higgins, E. T, & King, G. (1981). Accessibility of social constructs: In-
Davis, J. A. (1966). The campus as a frog pond: An application of the formation-processing consequences of individual and contextual vari-
theory of relative deprivation to career decisions of college men. Amer- ability. In N. Cantor & J. F. Kihlstrom (Eds.), Personality, cognition
ican Journal of Sociology, 72, 17-31. and social interaction (pp. 69-121). Hillsdale, NJ: Erlbaum.
Duval, S., & Wicklund, R. A. (1972). A theory of objective self-awareness. Higgins, E. T, & Lurie, L. (1983). Context, categorization, and memory:
New \brk: Academic Press. The "change-of-standard" effect. Cognitive Psychology, 15, 525-547.
Eich, J. M. (1982). A composite holographic associative recall model. Hinton, G. E., & Anderson, J. A. (Eds.). (1981). Parallel models of as-
Psychological Review, 89, 627-661. sociative memory. Hillsdale, NJ: Erlbaum.
Einhorn, H. J., & Hogarth, R. M. (1982). A theory of diagnostic inference: Hintzman, D. L. (in press). "Schema abstraction" in a multiple-trace
I. Imagination and the psychophysics of evidence. Center for Decision memory model. Psychological Review.
Research, Graduate School of Business, University of Chicago. Hintzman, D. L., & Ludlam, G. (1980). Differential forgetting of pro-
Einhorn, H. J., & Hogarth, R. M. (1986). Probable cause: A decision totypes and old instances: Simulation by an exemplar-based classifi-
making framework. Psychological Bulletin, 99, 000-000. cation model. Memory and Cognition, 8, 378-382.
Fischhoff, B. (1975). Hindsight is not equal to foresight: The effects of Hofstadter, D. R. (1979). Godel, Escher, Bach: An eternal golden braid.
outcome knowledge on judgment under uncertainty. Journal of Ex- New \ork: Basic Books.
perimental Psychology: Human Perception and Performance, 1, 288- Hofstadter, D. R. (1985). Metamagical themas: Questing for the essence
299. of mind and pattern. New York: Basic Books.
FischhofF, B. (1982). For those condemned to study the past: Heuristics Jacoby, L. L., & Brooks, L. R. (1984). Nonanalytic cognition: Memory,
and biases in hindsight. In D. Kahneman, P. Slovic, & A. Tversky perception, and concept learning. In G. H. Bower (Ed.), The psychology
(Eds.), Judgment under uncertainty: Heuristics and biases (pp. 335- of learning and motivation: Advances in research and theory (Vol. 18,
351). New York: Cambridge University Press. pp. 1-47). New York: Academic Press.
Fiske, S. T, Kenny, D. A., & Taylor, S. E. (1982). Structural models for Johnson, M. K. (1983). A modular model of memory. In G. H. Bower
the mediation of salience effects on attribution. Journal of Experimental (Ed.), The psychology of learning and motivation: Advances in research
Social Psychology, 18, 105-127. and theory (Vol. 17, pp. 81-123). New York: Academic Press.
Fleming, J., & Arrowood, A. J. (1979). Information processing and the Jones, E. E. (1979). The rocky road from acts to dispositions. American
perseverance of discredited self-perceptions. Personality and Social Psychologist, 34, 107-117.
Psychology Bulletin, 5, 201-205. Jones, E. E., & Harris, V. A. (1967). The attribution of attitudes. Journal
Fried, L. S., & Holyoak, K. J. (1984). Induction of category distributions: of Experimental Social Psychology, 3, 1-24.
152 DANIEL KAHNEMAN AND DALE T. MILLER

Jones, E. E., & McGillis, D. (1976). Correspondent inference and the Medin, D. L., & Schaffer, M. M. (1978). Context theory of classification
attribution cube: A comparative appraisal. In J. Harvey, W. Ickes, & learning. Psychological Review, 85, 207-238.
R. Kidd (Eds.), New directions in attribution research (Vol. 1, pp. 389- Mellers, B. A. (1982). Equity judgments: A revision of Aristotelian views.
420). Hillsdale, NJ: Erlbaum. Journals of Experimental Psychology: General, 111, 242-270.
Jones, E. E., & Nisbett, R. E. (1971). The actor and the observer: Divergent Mellers, B. A., & Birnbaum, M. H. (1983). Contextual effects in social
perceptions of the causes of behavior. Morristown, NJ: General Learning judgment. Journal of Experimental Social Psychology, 19, 157-171.
Press. Metcalfe Eich, J. (1985). Levels of processing, encoding specificity, elab-
Kahneman, D., & Tversky, A. (1973). On the psychology of prediction. oration, and CHARM. Psychological Review, 92, 1-38.
Psychological Review, 80, 237-251. Miller, A. G., Baer, R., & Schenberg, P. (1979). The bias phenomenon
Kahneman, D., & Tversky, A. (1982). The simulation heuristic. In D. in attitude attribution: Actor and observer perspectives. Journal of Per-
Kahneman, P. Slovic, & A. Tversky (Eds.), Judgment under uncertainty: sonality and Social Psychology, 37, 1421-1431.
Heuristics and biases (pp. 201-208). New \brk: Cambridge University Miller, D. T. (1981, August). Changes over time in the selection of causes
Press. and the definitions of effects. Paper presented at the meeting of the
Kelley, H. H. (1967). Attribution theory in social psychology. In D. Levine American Psychological Association, Los Angeles.
(Ed.), Nebraska symposium on motivation (Vol. 15, pp. 192-240). Lin- Miller, D. T, & McFarland, C. (1986). Counterfactual thinking and victim
coln: University of Nebraska Press. compensation. Manuscript in preparation.
Kraut, R. E., & Higgins, E. T. (1984). Communication and social cog- Monson, T. C., & Snyder, M. (1977). Actors, observers and the attribution
nition. In R. S. Wyer& T. K. Srull (Eds.), Handbook of social cognition process: Toward a reconceptualization. Journal of Experimental Social
(Vol. 3, pp. 81-127). Hillsdale, NJ: Erlbaum. Psychology, 13, 89-111.
Lakoff, G. (in press). Women, fire, and dangerous things. Chicago: Uni- Murdock, B. B. (1982). A theory for the storage and retrieval of item
versity of Chicago Press. and associative information. Psychological Review, 89, 609-626.
Lalljee, M, & Abelson, R. P. (1983). The organization of explanations. Murphy, G. L., & Medin, D. L. (1985). The role of theories in conceptual
In M. Hewstone (Ed.), Attribution theory: Social and functional exten- coherence. Psychological Review, 92, 289-316.
sions (pp. 65-80). Oxford, England: Blackwell. Nisbett, R. E., & Borgida, E. (1975). Attribution and the psychology of
Landman, J. (1984). Regret and undoing: Retrospective assessment of prediction. Journal of Personality and Social Psychology, 32, 932-943.
hypothetical and real life events. Unpublished doctoral dissertation, Nisbett, R. E., Borgida, E., Crandall, R., & Reed, H. (1976). Popular
University of Michigan. induction: Information is not always informative. In J. Carrol & J.
Lehnert, W. (1978). The process of question answering. Hillsdale, NJ: Payne (Eds.), Cognition and social behavior (pp. 227-236). Hillsdale,
Erlbaum. NJ: Erlbaum.
Lerner, M. J. (1980). The belief in a just world. New York: Plenum. Nisbett, R. E., Caputo, C., Legant, P., & Maracek, J. (1973). Behavior
Lerner, M. J., & Miller, D. T. (1978). Just world research and the attri- as seen by the actor and as seen by the observer. Journal of Personality
bution process: Looking back and ahead. Psychological Bulletin, 85, and Social Psychology, 27, 154-164.
1030-1051. Nisbett, R. E., & Ross, L. (1980). Human inference: Strategies and short-
Lewis, D. (1973). Counterfactuals. Cambridge, MA: Harvard University comings of social judgment. Englewood Cliffs, NJ: Prentice-Hall.
Press. Osherson, D. N., & Smith, E. E. (1982). Gradedness and conceptual
Lockhead, G. R., & King, M. C. (1983). A memory model of sequential combination. Cognition,,12, 299-318.
effects in scaling. Journal of Experimental Psychology: Human Per- Parducci, A. (1956). Direction of shift in the judgment of single stimuli.
ception and Performance, 9, 461-473. Journal of Experimental Psychology, 51, 169-178.
Mackie, J. L. (1974). The cement of the universe: A study of causation. Parducci, A. (1965). Category judgment: A range-frequency model. Psy-
Oxford, England: Clarendon Press. chological Review, 72, 407-418.
Mandler, G. (1967). Organization and memory. In K. W. Spence & J. A. Parducci, A. (1974). Contextual effects: A range-frequency analysis. In
Spence (Eds.), The psychology of learning and motivation: Advances E. C. Carterette & M. P. Friedman (Eds.), Handbook of perception:
in research and theory (Vol. 1, pp. 328-372). New \brk: Academic Psychophysical judgment and measurement (Vol. 7, pp. 127-141). New
Press. \brk: Academic Press.
Mandler, G. (1975). Memory storage and retrieval: Some limits on the Posner, M. I., & Keele, S. W. (1968). On the genesis of abstract ideas.
reach of attention and consciousness. In P. M. A. Rabbitt & S. Dornic Journal of Experimental Psychology, 77, 353-363.
(Eds.), Attention and performance K(pp. 499-516). New \brk: Aca- Pryor, J. B., & Kriss, N. (1977). The cognitive dynamics of salience in
demic Press. the attribution process. Journal of Personality and Social Psychology,
Mandler, J. M. (1984). Stories, scripts, and scenes: Aspects of schema 35, 49-55.
theory. Hillsdale, NJ: Erlbaum. Ratcliff, R., & McKoon, G. (1981). Does activation really spread? Psy-
McArthur, L. A. (1972). The how and what of why: Some determinants chological Review, 88, 454-457.
and consequences of causal attribution. Journal of Personality and Read, D. (1985). Determinants of relative mutability. Unpublished re-
Social Psychology, 22, 171-193. search, University of British Columbia, Vancouver, Canada.
McClelland, J. L. (1985). Putting knowledge in its place: A scheme for Read, S. J. (1984). Analogical reasoning in social judgment: The impor-
programming processing structures on the fly. Cognitive Science, 9, tance of causal theories. Journal of Personality and Social Psychology,
113-146. 46, 14-25. ,
McClelland, J. L., & Rumelhart, D. E. (1985). Distributed memory and Regan, D. X, & Totten, J. (1975). Empathy and attribution: Turning
the representation of general and specific information. Journal of Ex- observers into actors. Journal of Personality and Social Psychology,
perimental Psychology: General, 114, 159-188. 32, 850-856.
Medin, D. L. (1983). Structural principles in categorization. In T. J. Restle, F. (1978a). Assimilation predicted by adaptation-level theory with
Tighe & B. E. Shepp (Eds.), Perception, cognition, and development: variable weights. In N. J. Castellan, Jr., & F. Restle (Eds.), Cognitive
Interactional analyses (pp. 203-230). Hillsdale, NJ: Erlbaum. theory (Vol. 3, pp. 75-91). Hillsdale, NJ: Erlbaum.
Medin, D. L., Altom, M. W., Edelson, S. M., & Freko, D. (1982). Cor- Restle, F. (1978b). Relativity and organization in visual size judgments.
related symptoms and simulated medical classification. Journal of Ex- In E. Leeuwenberg & H. Buffart (Eds.), Formal theories of visual per-
perimental Psychology: Learning, Memory, and Cognition, 8, 37-50. ception (pp. 247-263). New York: Wiley.
 NORM THEORY 153

Rips, L. R., & Turnbull, W. (1980). How big is big? Relative and absolute Smith, E. E., & Medin, D. L. (1981). Categories and concepts. Cambridge,
properties in memory. Cognition, 8, 145-174. MA: Harvard University Press.
Rosch, E. (1973). On the internal structure of perceptual and semantic Snyder, C. R., Shenkel, R. J., & Lowery, C. R. (1977). Acceptance of
categories. In T. E. Moore (Ed.), Cognitive development and the ac- personality interpretations: The "Barnum effect" and beyond. Journal
quisition of language (pp. 111-144). New York: Academic Press. of Consulting and Clinical Psychology, 45, 104-114.
Rosch, E. (1978). Principles of categorization. In E. Rosch & B. B. Lloyd Storms, M. (1973). Videotape and the attribution process: Reversing ac-
(Eds.), Cognition and categorization (pp. 27-48). Hillsdale, NJ: Erl- tors' and observers' points of view. Journal of Personality and Social
baum. Psychology, 27, 165-175.
Rosch, E., & Mervis, C. B. (1975). Family resemblances: Studies in the Suls, J. M., & Miller, R. L. (Eds.). (1977). Social comparison processes:
internal structure of categories. Cognitive Psychology, 7, 573-605. Theoretical and empirical perspectives. Washington, DC: Halsted-Wiley.
Rosch, E., Mervis, C. B., Gray, W. D., Johnson, D. M., & Boyes-Braem, Taylor, S. E., & Fiske, S. T. (1978). Salience, attention and attribution:
P. (1976). Basic objects in natural categories. Cognitive Psychology, 8, Top of the head phenomena. In L. Berkowitz (Ed.), Advances in ex-
382-439. perimental social psychology (Vol. 11, pp. 249-288). New York: Aca-
Ross, L. (1977). The intuitive psychologist and his shortcomings. In L. demic Press.
Berkowitz (Ed.), Advances in experimental social psychology (Vol. 10, Tversky, A., & Kahneman, D. (1971). Belief in the law of small numbers.
pp. 173-220). New York: Academic Press. Psychological Bulletin, 76, 105-110.
Ross, L., & Anderson, C. A. (1982). Shortcomings in the attribution
Tversky, A., & Kahneman, D. (1973). Availability: A heuristic for judging
process: On the origins and maintenance of erroneous social assess-
frequency and probability. Cognitive Psychology, 5, 207-232.
ments. In D. Kahneman, P. Slovic, & A. Tversky (Eds.), Judgment
under uncertainly: Heuristics and biases (pp. 129-152). New \brk: Tversky, A., & Kahneman, D. (1982). Causal schemas in judgments under
Cambridge University Press. uncertainty. In D. Kahneman, P. Slovic, & A. Tversky (Eds.), Judgment
Ross, L., & Lepper, M. R. (1980). The perseverance of beliefs: Empirical under uncertainty: Heuristics and biases (pp. 117-128). New York:
and normative considerations. In R. A. Shweder & D. Fiske (Eds.), Cambridge University Press.
New directions for methodology of social and behavioral science: Fallible Tversky, A., & Kahneman, D. (1983). Extensional versus intuitive rea-
judgment in behavioral research (pp. 17-36). San Francisco: Jossey- soning: The conjunction fallacy in probability judgment. Psychological
Bass. Review, 90, 293-315.
Roth, E. M., & Shoben, E. J. (1983). The effect of context on the structure Ward, L. M. (1979). Stimulus information and sequential dependencies
of categories. Cognitive Psychology, 15, 346-378. in magnitude estimation and cross-modality matching. Journal of Ex-
Rumelhart, D. E. (1980). Schemata: The building blocks of cognition. perimental Psychology: Human Perception and Performance, 5, 444-
In R. J. Spiro, B. C. Bruce, & W. F. Brewer (Eds.), Theoretical issues 459.
in reading comprehension: Perspectives from cognitive psychology, lin- Weiner, B. (1985). "Spontaneous" causal thinking. Psychological Bulletin,
guistics, artificial intelligence, and education (pp. 33-58). Hillsdale, 97, 74_84.
NJ: Erlbaum. Wyer, R. S., & Gordon, S. E. (1984). The cognitive representation of
Sarris, V. (1976). Effects of stimulus-range and anchor value on psycho- social information. In R. S. Wyer & T. K. Srull (Eds.), Handbook of
physical judgment. In H. G. Geissler & Y. M. Zabrodin (Eds.), Advances social cognition (Vol. 2, pp. 73-150). Hillsdale, NJ: Erlbaum.
in psychophysics (pp. 253-268). Berlin, East Germany: UEB Deutscher Wyer, R. S., & Srull, T. K. (1981). Category accessibility: Some theoretical
Verlag der Wissenschaften. and empirical issues concerning the processing of social stimulus in-
Schank, R. C. (1982). Dynamic memory: Learning in computers and formation. In E. T. Higgins, C. P. Herman, & M. P. Zanna (Eds.),
people. New York: Cambridge University Press. Social cognition: The Ontario symposium (Vol. l,pp. 161-197). Hills-
Schul, Y, & Bernstein, E. (1985). When discounting fails: Conditions dale, NJ: Erlbaum.
under which individuals use discredited information in making a judg- Zadeh, L. A. (1965). Fuzzy sets. Information and control, 8, 338-353.
ment. Journal of Personality and Social Psychology, 49, 894-903. Zadeh, L. A. (1976). A fuzzy-algorithmic approach to the definition of
Schustack, M. W., & Sternberg, R. J. (1981). Evaluation of evidence in complex or imprecise concepts. International Journal of Man-Machine
causal inference. Journal of Experimental Psychology: General, 110, Studies, 8, 249-291.
101-120.
Shafer, G. (1976). A mathematical theory of evidence. Princeton, NJ:
Princeton University Press.
Slovic, P. (1985). Violations of dominance in rated attractiveness of playing Received October 12, 1984
bets. Decision Research Report 85-6. Eugene, OR: Decision Research. Revision received November 6, 1985