Maps of Bounded Rationality: Psychology for Behavioral Economics
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Maps of Bounded Rationality:
Psychology for Behavioral Economics†
By D ANIEL KAHNEMAN*
The work cited by the Nobel committee was hopes have been realized to some extent, giving
done jointly with Amos Tversky (1937–1996) rise to an active program of research by behav-
during a long and unusually close collaboration. ioral economists (Thaler, 2000; Colin Camerer
Together, we explored the psychology of intu- et al., forthcoming; for other examples, see
itive beliefs and choices and examined their Kahneman and Tversky, 2000).
bounded rationality. Herbert A. Simon (1955, My work with Tversky comprised three sep-
1979) had proposed much earlier that decision arate programs of research, some aspects of
makers should be viewed as boundedly rational, which were carried out with other collaborators.
and had offered a model in which utility maxi- The rst explored the heuristics that people use
mization was replaced by satis cing. Our re- and the biases to which they are prone in vari-
search attempted to obtain a map of bounded ous tasks of judgment under uncertainty, includ-
rationality, by exploring the systematic biases ing predictions and evaluations of evidence
that separate the beliefs that people have and the (Kahneman and Tversky, 1973; Tversky and
choices they make from the optimal beliefs and Kahneman, 1974; Kahneman et al., 1982). The
choices assumed in rational-agent models. The second was concerned with prospect theory, a
rational-agent model was our starting point and model of choice under risk (Kahneman and
the main source of our null hypotheses, but Tversky, 1979; Tversky and Kahneman, 1992)
Tversky and I viewed our research primarily as and with loss aversion in riskless choice (Kah-
a contribution to psychology, with a possible neman et al., 1990, 1991; Tversky and Kahne-
contribution to economics as a secondary ben- man, 1991). The third line of research dealt with
e t. We were drawn into the interdisciplinary framing effects and with their implications for
conversation by economists who hoped that rational-agent models (Tversky and Kahneman,
psychology could be a useful source of assump- 1981, 1986). The present essay revisits these
tions for economic theorizing, and indirectly a three lines of research in light of recent ad-
source of hypotheses for economic research vances in the psychology of intuitive judgment
(Richard H. Thaler, 1980, 1991, 1992). These and choice. Many of the ideas presented here
were anticipated informally decades ago, but
the attempt to integrate them into a coherent
approach to judgment and choice is recent.
†
This article is a revised version of the lecture Daniel Economists often criticize psychological re-
Kahneman delivered in Stockholm, Sweden, on December search for its propensity to generate lists of
8, 2002, when he received the Bank of Sweden Prize in
Economic Sciences in Memory of Alfred Nobel. The article errors and biases, and for its failure to offer a
is copyright © The Nobel Foundation 2002 and is published coherent alternative to the rational-agent model.
here with the permission of the Nobel Foundation. This complaint is only partly justi ed: psycho-
* Woodrow Wilson School, Princeton University, logical theories of intuitive thinking cannot
Princeton, NJ 08544 (e-mail: Kahneman@princeton.edu). match the elegance and precision of formal nor-
This essay revisits problems that Amos Tversky and I mative models of belief and choice, but this is
studied together many years ago, and continued to discuss in just another way of saying that rational models
a conversation that spanned several decades. It builds on an
analysis of judgment heuristics that was developed in col- are psychologically unrealistic. Furthermore,
laboration with Shane Frederick (Kahneman and Frederick, the alternative to simple and precise models is
2002). A different version was published in American Psy- not chaos. Psychology offers integrative con-
chologist in September 2003. For detailed comments on this cepts and mid-level generalizations, which gain
version I am grateful to Angus Deaton, David Laibson,
Michael Rothschild, and Richard Thaler. The usual caveats
credibility from their ability to explain ostensi-
apply. Geoffrey Goodwin, Amir Goren, and Kurt Schoppe bly different phenomena in diverse domains. In
provided helpful research assistance. this spirit, the present essay offers a uni ed
14491450 THE AMERICAN ECONOMIC REVIEW DECEMBER 2003
treatment of intuitive judgment and choice, ations and overt behavior also goes on. We do
which builds on an earlier study of the relation- not express every passing thought or act on
ship between preferences and attitudes (Kahne- every impulse. But the monitoring is normally
man et al., 1999) and extends a model of lax, and allows many intuitive judgments to be
judgment heuristics recently proposed by Kah- expressed, including some that are erroneous
neman and Shane Frederick (2002). The guid- (Kahneman and Frederick, 2002). Ellen J.
ing ideas are (i) that most judgments and most Langer et al. (1978) provided a well-known
choices are made intuitively; (ii) that the rules example of what she called “mindless behav-
that govern intuition are generally similar to the ior.” In her experiment, a confederate tried to
rules of perception. Accordingly, the discussion cut in line at a copying machine, using various
of the rules of intuitive judgments and choices preset “excuses.” The conclusion was that state-
will rely extensively on visual analogies. ments that had the form of an unquali ed re-
Section I introduces a distinction between quest were rejected (e.g., “Excuse me, may I use
two generic modes of cognitive function, corre- the Xerox machine?”), but almost any statement
sponding roughly to intuition and reasoning. that had the general form of an explanation was
Section II describes the factors that determine accepted, including “Excuse me, may I use the
the relative accessibility of different judgments Xerox machine because I want to make cop-
and responses. Section III relates prospect the- ies?” The super ciality is striking.
ory to the general proposition that changes and Frederick (2003, personal communication)
differences are more accessible than absolute has used simple puzzles to study cognitive self-
values. Section IV explains framing effects in monitoring, as in the following example: “A bat
terms of differential salience and accessibility. and a ball cost $1.10 in total. The bat costs $1
Section V reviews an attribute substitution more than the ball. How much does the ball
model of heuristic judgment. Section VI de- cost?” Almost everyone reports an initial ten-
scribes a particular family of heuristics, called dency to answer “10 cents” because the sum
prototype heuristics. Section VII discusses the $1.10 separates naturally into $1 and 10 cents,
interactions between intuitive and deliberate and 10 cents is about the right magnitude. Fred-
thought. Section VIII concludes. erick found that many intelligent people yield to
this immediate impulse: 50 percent (47/93) of a
I. The Architecture of Cognition: Two Systems group of Princeton students and 56 percent
(164/293) of students at the University of Mich-
The present treatment distinguishes two igan gave the wrong answer. Clearly, these re-
modes of thinking and deciding, which corre- spondents offered their response without rst
spond roughly to the everyday concepts of rea- checking it. The surprisingly high rate of errors
soning and intuition. Reasoning is what we do in this easy problem illustrates how lightly the
when we compute the product of 17 by 258, ll output of effortless associative thinking is mon-
an income tax form, or consult a map. Intuition itored: people are not accustomed to thinking
is at work when we read the sentence “Bill hard, and are often content to trust a plausible
Clinton is a shy man” as mildly amusing, or judgment that quickly comes to mind. Re-
when we nd ourselves reluctant to eat a piece markably, Frederick has found that errors in
of what we know to be chocolate that has been this puzzle and in others of the same type
formed in the shape of a cockroach (Paul Rozin were signi cant predictors of high discount
and Carol Nemeroff, 2002). Reasoning is done rates.
deliberately and effortfully, but intuitive thoughts In the examples discussed so far, intuition
seem to come spontaneously to mind, without was associated with poor performance, but in-
conscious search or computation, and without tuitive thinking can also be powerful and accu-
effort. Casual observation and systematic re- rate. High skill is acquired by prolonged
search indicate that most thoughts and actions practice, and the performance of skills is rapid
are normally intuitive in this sense (Daniel T. and effortless. The proverbial master chess
Gilbert, 1989, 2002; Timothy D. Wilson, 2002; player who walks past a game and declares
Seymour Epstein, 2003). “white mates in three” without slowing is per-
Although effortless thought is the norm, forming intuitively (Simon and William G.
some monitoring of the quality of mental oper- Chase, 1973), as is the experienced nurse whoVOL. 93 NO. 5 KAHNEMAN: MAPS OF BOUNDED RATIONALITY 1451
FIGURE 1. THREE COGNITIVE SYSTEMS
detects subtle signs of impending heart failure neither cause nor suffer much interference when
(Gary Klein, 1998; Atul Gawande, 2002). combined with other tasks. For example, a driv-
The distinction between intuition and reason- er’s ability to conduct a conversation is a sen-
ing has recently been a topic of considerable sitive indicator of the amount of attention
interest to psychologists (see, e.g., Shelly currently demanded by the driving task. Dual
Chaiken and Yaacov Trope, 1999; Gilbert, tasks have been used in hundreds of psycholog-
2002; Steven A. Sloman, 2002; Keith E. ical experiments to measure the attentional de-
Stanovich and Richard F. West, 2002). There is mands of different mental activities (for a
substantial agreement on the characteristics that review, see Harold E. Pashler, 1998). Studies
distinguish the two types of cognitive processes, using the dual-task method suggest that the self-
for which Stanovich and West (2000) proposed monitoring function belongs with the effortful
the neutral labels of System 1 and System 2. operations of System 2. People who are occu-
The scheme shown in Figure 1 summarizes pied by a demanding mental activity (e.g., at-
these characteristics. The operations of System tempting to hold in mind several digits) are
1 are fast, automatic, effortless, associative, and much more likely to respond to another task by
often emotionally charged; they are also gov- blurting out whatever comes to mind (Gilbert,
erned by habit, and are therefore dif cult to 1989). The phrase that “System 2 monitors the
control or modify. The operations of System 2 activities of System 1” will be used here as
are slower, serial, effortful, and deliberately shorthand for a hypothesis about what would
controlled; they are also relatively exible and happen if the operations of System 2 were dis-
potentially rule-governed. rupted. For example, it is safe to predict that the
The difference in effort provides the most percentage of errors in the bat-and-ball question
useful indications of whether a given mental will increase, if the respondents are asked this
process should be assigned to System 1 or Sys- question while attempting to keep a list of
tem 2. Because the overall capacity for mental words in their active memory.
effort is limited, effortful processes tend to dis- In the language that will be used here, the
rupt each other, whereas effortless processes perceptual system and the intuitive operations1452 THE AMERICAN ECONOMIC REVIEW DECEMBER 2003
of System 1 generate impressions of the at-
tributes of objects of perception and thought.
These impressions are not voluntary and need
not be verbally explicit. In contrast, judgments
are always explicit and intentional, whether or
not they are overtly expressed. Thus, System 2
is involved in all judgments, whether they orig-
inate in impressions or in deliberate reasoning.
The label “intuitive” is applied to judgments
that directly re ect impressions.
Figure 1 illustrates an idea that guided the
research that Tversky and I conducted from its
early days: that intuitive judgments occupy a FIGURE 2. EXAMPLES OF DIFFERENTIAL ACCESSIBILITY
position—perhaps corresponding to evolution-
ary history—between the automatic operations
of perception and the deliberate operations of dismantled is not perceptually accessible,
reasoning. All the characteristics that students though it can be estimated by a deliberate pro-
of intuition have attributed to System 1 are also cedure, such as multiplying the area of a block
properties of perceptual operations. Unlike per- by the number of blocks. Of course, the situa-
ception, however, the operations of System 1 tion is reversed with Figure 2b. Now the blocks
are not restricted to the processing of current are laid out and an impression of total area is
stimulation. Like System 2, the operations of immediately accessible, but the height of the
System 1 deal with stored concepts as well as tower that could be constructed with these
with percepts, and can be evoked by language. blocks is not.
This view of intuition suggests that the vast Some relational properties are accessible.
store of scienti c knowledge available about Thus, it is obvious at a glance that Figures 2a
perceptual phenomena can be a source of useful and 2c are different, but also that they are more
hypotheses about the workings of intuition. The similar to each other than either is to Figure
strategy of drawing on analogies from percep- 2b. And some statistical properties of ensembles
tion is applied in the following section. are accessible, while others are not. For an
example, consider the question “What is the
II. The Accessibility Dimension average length of the lines in Figure 3?” This
question is easy. When a set of objects of the
A de ning property of intuitive thoughts is same general kind is presented to an observer—
that they come to mind spontaneously, like per- whether simultaneously or successively—a rep-
cepts. The technical term for the ease with resentation of the set is computed automatically,
which mental contents come to mind is acces- which includes quite precise information about
sibility (E. Tory Higgins, 1996). To understand the average (Dan Ariely, 2001; Sang-Chul
intuition, we must understand why some Chong and Anne Treisman, 2003). The repre-
thoughts are accessible and others are not. The sentation of the prototype is highly accessible,
remainder of this section introduces the concept and it has the character of a percept: we form an
of accessibility by examples drawn from visual impression of the typical line without choosing
perception. to do so. The only role for System 2 in this task
Consider Figures 2a and 2b. As we look at is to map the impression of typical length onto
the object in Figure 2a, we have immediate the appropriate scale. In contrast, the answer to
impressions of the height of the tower, the area the question “What is the total length of the
of the top block, and perhaps the volume of the lines in the display?” does not come to mind
tower. Translating these impressions into units without considerable effort.
of height or volume requires a deliberate oper- As the example of averages and sums illus-
ation, but the impressions themselves are highly trates, some attributes are more accessible than
accessible. For other attributes, no perceptual others, both in perception and in judgment. At-
impression exists. For example, the total area tributes that are routinely and automatically
that the blocks would cover if the tower were produced by the perceptual system or by SystemVOL. 93 NO. 5 KAHNEMAN: MAPS OF BOUNDED RATIONALITY 1453
was initiated within a fraction of a second, well
before the meaning of the stimulus was con-
sciously registered, the emotional valence of the
word had a substantial effect. Participants were
relatively faster in pulling a lever toward them-
selves (approach) for positive words, and rela-
tively faster pushing the lever away when the
word was aversive. The tendencies to approach
or avoid were evoked by an automatic process
that was not under conscious voluntary control.
Several psychologists have commented on the
in uence of this primordial evaluative system
(here included in System 1) on the attitudes and
preferences that people adopt consciously and
deliberately (Zajonc, 1998; Kahneman et al.,
1999; Paul Slovic et al., 2002; Epstein, 2003).
The preceding discussion establishes a di-
mension of accessibility. At one end of this
FIGURE 3. DIFFERENTIAL ACCESSIBILITY dimension we nd operations that have the
OF STATISTICAL PROPERTIES
characteristics of perception and of the intuitive
System 1: they are rapid, automatic, and effort-
less. At the other end are slow, serial, and
1, without intention or effort, have been called effortful operations that people need a special
natural assessments (Tversky and Kahneman, reason to undertake. Accessibility is a contin-
1983). Kahneman and Frederick (2002) com- uum, not a dichotomy, and some effortful op-
piled a partial list of these natural assessments. erations demand more effort than others. Some
In addition to physical properties such as size, of the determinants of accessibility are probably
distance, and loudness, the list includes more genetic; others develop through experience. The
abstract properties such as similarity, causal acquisition of skill gradually increases the ac-
propensity, surprisingness, affective valence, cessibility of useful responses and of productive
and mood. ways to organize information, until skilled per-
The evaluation of stimuli as good or bad is a formance becomes almost effortless. This effect
particularly important natural assessment. The of practice is not limited to motor skills. A
evidence, both behavioral (John A. Bargh, master chess player does not see the same board
1997; Robert B. Zajonc, 1998) and neurophys- as the novice, and visualizing the tower in an
iological (e.g., Joseph E. LeDoux, 2000), is array of blocks would also become virtually
consistent with the idea that the assessment of effortless with prolonged practice.
whether objects are good (and should be ap- The impressions that become accessible in
proached) or bad (should be avoided) is carried any particular situation are mainly determined,
out quickly and ef ciently by specialized neural of course, by the actual properties of the object
circuitry. A remarkable experiment reported by of judgment: it is easier to see a tower in Figure
Bargh (1997) illustrates the speed of the evalu- 2a than in Figure 2b, because the tower in the
ation process, and its direct link to approach and latter is only virtual. Physical salience also de-
avoidance. Participants were shown a series of termines accessibility: if a large green letter and
stimuli on a screen, and instructed to respond to a small blue letter are shown at the same time,
each stimulus as soon as it appeared, by moving “green” will come to mind rst. However, sa-
a lever that blanked the screen. The stimuli were lience can be overcome by deliberate attention:
affectively charged words, some positive (e.g., an instruction to look for the small object will
LOVE) and some aversive (e.g., VOMIT), but enhance the accessibility of all its features.
this feature was irrelevant to the participant’s Analogous effects of salience and of sponta-
task. Half the participants responded by pulling neous and voluntary attention occur with more
the lever toward themselves, half responded by abstract stimuli. For example, the statements
pushing the lever away. Although the response “Team A beat team B” and “Team B lost to1454 THE AMERICAN ECONOMIC REVIEW DECEMBER 2003
instead seen as a number when placed within a
context of numbers. More generally, expecta-
tions (conscious or not) are a powerful determi-
nant of accessibility.
Another important point that Figure 4 illus-
trates is the complete suppression of ambiguity
in conscious perception. This aspect of the dem-
onstration is spoiled for the reader who sees the
two versions in close proximity, but when the
two lines are shown separately, observers will
not spontaneously become aware of the alterna-
tive interpretation. They “see” the interpretation
of the object that is the most likely in its con-
FIGURE 4. A N EFFECT OF CONTEXT ON ACCESSIBILITY
text, but have no subjective indication that it
could be seen differently. Ambiguity and uncer-
tainty are suppressed in intuitive judgment as
well as in perception. Doubt is a phenomenon of
team A” convey the same information, but be- System 2, an awareness of one’s ability to think
cause each sentence draws attention to its gram- incompatible thoughts about the same thing.
matical subject, they make different thoughts The central nding in studies of intuitive deci-
accessible. Accessibility also re ects temporary sions, as described by Klein (1998), is that
states of associative activation. For example, the experienced decision makers working under
mention of a familiar social category temporarily pressure (e.g., re ghting company captains)
increases the accessibility of the traits associated rarely need to choose between options because,
with the category stereotype, as indicated by a in most cases, only a single option comes to mind.
lowered threshold for recognizing behaviors as The compound cognitive system that has
indications of these traits (Susan T. Fiske, 1998). been sketched here is an impressive computa-
As designers of billboards know well, moti- tional device. It is well-adapted to its environ-
vationally relevant and emotionally arousing ment and has two ways of adjusting to changes:
stimuli spontaneously attract attention. Bill- a short-term process that is exible and effort-
boards are useful to advertisers because paying ful, and a long-term process of skill acquisition
attention to an object makes all its features that eventually produces highly effective re-
accessible—including those that are not linked sponses at low cost. The system tends to see
to its primary motivational or emotional signif- what it expects to see—a form of Bayesian
icance. The “hot” states of high emotional and adaptation—and it is also capable of responding
motivational arousal greatly increase the acces- effectively to surprises. However, this marvel-
sibility of thoughts that relate to the immediate ous creation differs in important respects from
emotion and to the current needs, and reduce the another paragon, the rational agent assumed in
accessibility of other thoughts (George Loe- economic theory. Some of these differences are
wenstein, 1996, 2000; Jon Elster, 1998). An explored in the following sections, which review
effect of emotional signi cance on accessibility several familiar results as effects of accessibility.
was demonstrated in an important study by Yu- Possible implications for theorizing in behavioral
val Rottenstreich and Christopher K. Hsee economics are explored along the way.
(2001), which showed that people are less sen-
sitive to variations of probability when valuing III. Changes or States: Prospect Theory
chances to receive emotionally loaded out-
comes (kisses and electric shocks) than when A general property of perceptual systems is
the outcomes are monetary. that they are designed to enhance the accessi-
Figure 4 (adapted from Jerome S. Bruner and bility of changes and differences. Perception is
A. Leigh Minturn, 1955) includes a standard reference-dependent: the perceived attributes
demonstration of the effect of context on acces- of a focal stimulus re ect the contrast between
sibility. An ambiguous stimulus that is per- that stimulus and a context of prior and con-
ceived as a letter within a context of letters is current stimuli. This section will show thatVOL. 93 NO. 5 KAHNEMAN: MAPS OF BOUNDED RATIONALITY 1455
FIGURE 5. REFERENCE-DEPENDENCE IN THE PERCEPTION OF BRIGHTNESS
intuitive evaluations of outcomes are also expected utility of wealth (the moral expecta-
reference-dependent. tion). The language of Bernoulli’s essay is pre-
The role of prior stimulation is familiar in the scriptive—it speaks of what is sensible or
domain of temperature. Immersing the hand in reasonable to do—but the theory was also in-
water at 20°C will feel pleasantly warm after tended as a description of the choices of reason-
prolonged immersion in much colder water, and able men (Gerd Gigerenzer et al., 1989). As in
pleasantly cool after immersion in much most modern treatments of decision-making,
warmer water. Figure 5 illustrates reference- Bernoulli’s essay does not acknowledge any
dependence in vision. The two enclosed squares tension between prescription and description.
have the same luminance, but they do not ap- The proposition that decision makers evaluate
pear equally bright. The point of the demonstra- outcomes by the utility of nal asset positions
tion is that the brightness of an area is not a has been retained in economic analyses for al-
single-parameter function of the light energy most 300 years. This is rather remarkable, be-
that reaches the eye from that area, just as the cause the idea is easily shown to be wrong; I
experience of temperature is not a single-param- call it Bernoulli’s error.
eter function of the temperature to which one is Tversky and I constructed numerous thought
currently exposed. An account of perceived experiments when we began the study of risky
brightness or temperature also requires a param- choice that led to the formulation of prospect
eter for a reference value (often called adapta- theory (Kahneman and Tversky, 1979). Exam-
tion level), which is in uenced by the context of ples such as Problems 1 and 2 below convinced
current and prior stimulation. us of the inadequacy of the utility function for
From the vantage point of a student of per- wealth as an explanation of choice.
ception, it is quite surprising that in standard
economic analyses the utility of decision out-
comes is assumed to be determined entirely by
Problem 1
the nal state of endowment, and is therefore Would you accept this gamble?
reference-independent. In the context of risky
choice, this assumption can be traced to the 50% chance to win $150
brilliant essay that rst de ned a theory of ex- 50% chance to lose $100
pected utility (Daniel Bernoulli, 1738). Ber-
noulli assumed that states of wealth have a Would your choice change if your
speci ed utility, and proposed that the decision overall wealth were lower by $100?
rule for choice under risk is to maximize the1456 THE AMERICAN ECONOMIC REVIEW DECEMBER 2003
There will be few takers of the gamble in Prob-
lem 1. The experimental evidence shows that
most people will reject a gamble with even
chances to win and lose, unless the possible win
is at least twice the size of the possible loss
(e.g., Tversky and Kahneman, 1992). The an-
swer to the second question is, of course, neg-
ative. Next consider Problem 2:
Problem 2
Which would you choose?
lose $100 with certainty
or
50% chance to win $50
50% chance to lose $200 FIGURE 6. A SCHEMATIC V ALUE FUNCTION FOR CHANGES
Would your choice change if your
overall wealth were higher by $100? characterized by three features: (1) it is con-
cave in the domain of gains, favoring risk
aversion; (2) it is convex in the domain of
In Problem 2, the gamble appears much more losses, favoring risk seeking; (3) most impor-
attractive than the sure loss. Experimental re- tant, the function is sharply kinked at the
sults indicate that risk-seeking preferences are reference point, and loss-averse—steeper for
held by a large majority of respondents in prob- losses than for gains by a factor of about
lems of this kind (Kahneman and Tversky, 2–2.5 (Kahneman et al., 1991; Tversky and
1979). Here again, the idea that a change of Kahneman, 1992).
$100 in total wealth would affect preferences If Bernoulli’s formulation is transparently
cannot be taken seriously. incorrect as a descriptive model of risky
We examined many choice pairs of this choices, as has been argued here, why
type in our early explorations, and concluded has this model been retained for so long?
that the very abrupt switch from risk aversion The answer appears to be that the assign-
to risk seeking could not plausibly be ex- ment of utility to wealth is an aspect of ra-
plained by a utility function for wealth. Pref- tionality, and therefore compatible with the
erences appeared to be determined by general assumption of rationality in economic
attitudes to gains and losses, de ned relative theorizing (Kahneman, 2003a). Consider
to a reference point, but Bernoulli’s theory Problem 3:
and its successors did not incorporate a ref-
erence point. We therefore proposed an alter-
native theory of risk, in which the carriers of
utility are gains and losses— changes of Problem 3
wealth rather than states of wealth. One nov- Two persons get their monthly report
elty of prospect theory was that it was explic- from a broker:
itly presented as a formal descriptive theory A is told that her wealth went from
of the choices that people actually make, not 4M to 3M
as a normative model. This was a departure B is told that her wealth went from
from a long history of choice models that 1M to 1.1M
served double duty as normative logics and as Who of the two individuals has more
idealized descriptive models. reason to be satis ed with her nancial
The distinctive predictions of prospect the- situation?
ory follow from the shape of the value func-
tion, which is shown in Figure 6. The value Who is happier today?
function is de ned on gains and losses and isVOL. 93 NO. 5 KAHNEMAN: MAPS OF BOUNDED RATIONALITY 1457 Problem 3 highlights the contrasting interpre- when Thaler (1980) used it to explain riskless tations of utility in theories that de ne outcomes choices. In particular, loss aversion explained a as states or as changes. In Bernoulli’s analysis violation of consumer theory that Thaler identi ed only the rst of the two questions of Problem 3 and labeled the “endowment effect”: the selling is relevant, and only long-term consequences price for consumption goods is much higher than matter. Prospect theory, in contrast, is con- the buying price, often by a factor of 2 or more. cerned with short-term outcomes, and the value The value of a good to an individual appears to be function presumably re ects an anticipation of higher when the good is viewed as something that the valence and intensity of the emotions that could be lost or given up than when the same good will be experienced at moments of transition is evaluated as a potential gain (Kahneman et al., from one state to another (Kahneman, 2000a, b; 1990, 1991; Tversky and Kahneman, 1991). Barbara Mellers, 2000). Which of these con- When half the participants in an experimental cepts of utility is more useful? The cultural market were randomly chosen to be endowed norm of reasonable decision-making favors the with a good (a mug) and trade was allowed, the long-term view over a concern with transient emo- volume of trade was about half the amount that tions. Indeed, the adoption of a broad perspective would be predicted by assuming that value was and a long-term view is an aspect of the meaning independent of initial endowment (Kahneman of rationality in everyday language. The nal- et al., 1990). Transaction costs did not explain states interpretation of the utility of outcomes is this counterexample to the Coase theorem, be- therefore a good t for a rational-agent model. cause the same institution produced no indica- These considerations support the normative tion of reluctance to trade when the objects of and prescriptive status of the Bernoullian de - trade were money tokens. The results suggest nition of outcomes. On the other hand, an ex- that the participants in these experiments did not clusive concern with the long term may be value the mug as an object they could have and prescriptively sterile, because the long term is consume, but as something they could get, or not where life is lived. Utility cannot be di- give up. Interestingly, John A. List (2003a, b) vorced from emotion, and emotions are trig- found that the magnitude of the endowment gered by changes. A theory of choice that effect was substantially reduced for participants completely ignores feelings such as the pain of with intense experience in the trading of sports- losses and the regret of mistakes is not only cards. Experienced traders (who are also con- descriptively unrealistic, it also leads to pre- sumers) showed less reluctance to trade one scriptions that do not maximize the utility of good for another—not only sportscards, but also outcomes as they are actually experienced— mugs and other goods—as if they had learned to that is, utility as Bentham conceived it (Kahne- base their choice on long-term value, rather than man, 1994, 2000a; Kahneman et al., 1997). on the immediate emotions associated with get- Bernoulli’s error—the idea that the carriers ting or giving up objects. of utility are nal states—is not restricted to Reference-dependence and loss aversion help decision-making under risk. Indeed, the incor- account for several phenomena of choice. The rect assumption that initial endowments do not familiar observation that out-of-pocket losses matter is the basis of Coase’s theorem and of its are valued much more than opportunity costs is multiple applications (Kahneman et al., 1990). readily explained, if these outcomes are evalu- The error of reference-independence is built ated on different limbs of the value function. into the standard representation of indifference The distinction between “actual” losses and maps. It is puzzling to a psychologist that these losses of opportunities is recognized in many maps do not include a representation of the ways in the law (David Cohen and Jack L. decision maker’s current holdings of various Knetsch, 1992) and in lay intuitions about rules goods—the counterpart of the reference point in of fairness in the market (Kahneman et al., prospect theory. The parameter is not included, 1986). Loss aversion also contributes to the of course, because consumer theory assumes well-documented status-quo bias (William that it does not matter. Samuelson and Richard Zeckhauser, 1988). Be- The core idea of prospect theory—that the cause the reference point is usually the status value function is kinked at the reference point quo, the properties of alternative options are and loss averse— became useful to economics evaluated as advantages or disadvantages
1458 THE AMERICAN ECONOMIC REVIEW DECEMBER 2003
relative to the current situation, and the disad- In this version of the problem, a substantial
vantages of the alternatives loom larger than majority of respondents favor Program A, indi-
their advantages. Other applications of the con- cating risk aversion. Other respondents, se-
cept of loss aversion are documented in several lected at random, receive a question in which
chapters in Kahneman and Tversky (2000). the same cover story is followed by a different
description of the options:
IV. Framing Effects
In the display of blocks in Figure 2, the same If Program A9 is adopted, 400 people will
property (the total height of a set of blocks) was die
highly accessible in one display and not in an-
other, although both displays contained the If Program B9 is adopted, there is a one-
same information. This observation is entirely third probability that nobody will die and
unremarkable—it does not seem shocking that a two-thirds probability that 600 people
some attributes of a stimulus are automatically will die
perceived while others must be computed, or
that the same attribute is perceived in one dis-
play of an object but must be computed in A substantial majority of respondents now
another. In the context of decision-making, favor Program B9, the risk-seeking option. Al-
however, similar observations raise a signi cant though there is no substantive difference be-
challenge to the rational-agent model. tween the versions, they evoke different
The assumption that preferences are not af- associations and evaluations. This is easiest to
fected by inconsequential variations in the see in the certain option, because outcomes that
description of outcomes has been called exten- are certain are overweighted relative to out-
sionality (Kenneth J. Arrow, 1982) and invari- comes of high or intermediate probability (Kah-
ance (Tversky and Kahneman, 1986), and is neman and Tversky, 1979). Thus, the certainty
considered an essential aspect of rationality. of saving people is disproportionately attractive,
Invariance is violated in framing effects, where while accepting the certain death of people is
extensionally equivalent descriptions lead to disproportionately aversive. These immediate
different choices by altering the relative salience affective responses respectively favor A over B
of different aspects of the problem. Tversky and and B9 over A9. As in Figures 2a and 2b, the
Kahneman (1981) introduced their discussion of different representations of the outcomes high-
framing effects with the following problem: light some features of the situation and mask
others.
In an essay about the ethics of policy,
Thomas C. Schelling (1984) presented a com-
The Asian disease pellingly realistic example of the dilemmas
Imagine that the United States is pre- raised by framing. Schelling reports asking his
paring for the outbreak of an unusual students to evaluate a tax policy that would
Asian disease, which is expected to kill allow a larger child exemption to the rich than
600 people. Two alternative programs to to the poor. Not surprisingly, his students found
combat the disease have been proposed. this proposal outrageous. Schelling then pointed
Assume that the exact scienti c estimates
of the consequences of the programs are out that the default case in the standard tax table
as follows: is a childless family, with special adjustments
for families with children, and led his class to
If Program A is adopted, 200 people agree that the existing tax schedule could be
will be saved rewritten with a family with two children as the
default case. In this formulation, childless fam-
If Program B is adopted, there is a ilies would pay a surcharge. Should this sur-
one-third probability that 600 people will charge be as large for the poor as for the rich?
be saved and a two-thirds probability that Of course not. The two versions of the question
no people will be saved about how to treat the rich and the poor both
trigger an intuitive preference for protecting theVOL. 93 NO. 5 KAHNEMAN: MAPS OF BOUNDED RATIONALITY 1459 poor, but these preferences are incoherent. blocks, and they do not spontaneously trans- Schelling’s problem highlights an important form the representation of puzzles or decision point. Framing effects are not a laboratory cu- problems. Obviously, no one is able to recog- riosity, but a ubiquitous reality. The tax table nize “137 3 24” and “3,288” as “the same” must be framed one way or another, and each number without going through some elaborate frame will increase the accessibility of some computations. Invariance cannot be achieved by responses and make other responses less likely. a nite mind. There has been considerable interest among The impossibility of invariance raises signif- behavioral economists in a particular type of icant doubts about the descriptive realism of framing effect, where a choice between two rational-choice models (Tversky and Kahne- options A and B is affected by designating man, 1986). Absent a system that reliably gen- either A or B as a default option. The option erates appropriate canonical representations, designated as the default has a large advantage intuitive decisions will be shaped by the factors in such choices, even for decisions that have that determine the accessibility of different fea- considerable signi cance. Eric J. Johnson et al. tures of the situation. Highly accessible features (1993) described a compelling example. The will in uence decisions, while features of low states of Pennsylvania and New Jersey both accessibility will be largely ignored—and the offer drivers a choice between an insurance correlation between accessibility and re ective policy that allows an unconstrained right to sue, judgments of relevance in a state of complete and a less expensive policy that restricts the information is not necessarily high. right to sue. The unconstrained right to sue is A particularly unrealistic assumption of the the default in Pennsylvania, the opposite is the rational-agent model is that agents make their default in New Jersey, and the takeup of full choices in a comprehensively inclusive context, coverage is 79 percent and 30 percent in the two which incorporates all the relevant details of the states, respectively. Johnson and Daniel G. present situation, as well as expectations about Goldstein (2003) estimate that Pennsylvania all future opportunities and risks. Much evi- drivers spend 450 million dollars annually on dence supports the contrasting claim that peo- full coverage that they would not purchase if ple’s views of decisions and outcomes are their choice were framed as it is for New Jersey normally characterized by “narrow framing” drivers. (Kahneman and Daniel Lovallo, 1993), and by Johnson and Goldstein (2003) also compared the related notions of “mental accounting” the proportions of the population enrolled in (Thaler, 1985, 1999) and “decision bracketing” organ donation programs in seven European (Daniel Read et al., 1999). countries in which enrollment was the default The following are some examples of the and four in which nonenrollment was the de- prevalence of narrow framing. The decision of fault. Averaging over countries, enrollment in whether or not to accept a gamble is normally donor programs was 97.4 percent when this considered as a response to a single opportunity, was the default option, 18 percent otherwise. not as an occasion to apply a general policy The passive acceptance of the formulation (Gideon Keren and Willem A. Wagenaar, 1987; given has signi cant consequences in this Tversky and Donald A. Redelmeier, 1992; Kah- case, as it does in other recent studies where neman and Lovallo, 1993; Shlomo Benartzi and the selection of the default on the form that Thaler, 1999). Investors’ decisions about partic- workers completed to set their 401(k) contri- ular investments appear to be considered in butions dominated their ultimate choice isolation from the remainder of the investor’s (Brigitte Madrian and Dennis Shea, 2001; portfolio (Nicholas Barberis et al., 2003). The James J. Choi et al., 2002). time horizon that investors adopt for evaluating The basic principle of framing is the passive their investments appears to be unreasonably acceptance of the formulation given. Because of short—an observation that helps explain the this passivity, people fail to construct a canon- equity-premium puzzle (Benartzi and Thaler, ical representation for all extensionally equiva- 1995). Finally, the prevalence of the gain/loss lent descriptions of a state of affairs. They do framing of outcomes over the wealth frame, not spontaneously compute the height of a which was discussed in the previous sec- tower that could be built from an array of tion, can now be seen as an instance of narrow
1460 THE AMERICAN ECONOMIC REVIEW DECEMBER 2003
framing. A shared feature of all these examples
is that decisions made in narrow frames depart
far more from risk neutrality than decisions that
are made in a more inclusive context.
The prevalence of narrow frames is an effect
of accessibility, which can be understood by
referring to the displays of blocks in Figure
2. The same set of blocks is framed as a tower
in Figure 2a, and as a at array in Figure 2b. Al-
though it is possible to “see” a tower in Figure
2b, it is much easier to do so in Figure 2a. Nar-
row frames generally re ect the structure of the
environment in which decisions are made. The
choices that people face arise one at a time, and
the principle of passive acceptance suggests that
they will be considered as they arise. The prob-
lem at hand and the immediate consequences of
the choice will be far more accessible than all
other considerations, and as a result decision
problems will be framed far more narrowly than
the rational model assumes. FIGURE 7. AN ILLUSION OF ATTRIBUTE SUBSTITUTION
Source: Photo by Lenore Shoham, 2003.
V. Attribute Substitution: A Model of Judgment
Heuristics Kahneman and Frederick (2002) recently re-
visited the early studies of judgment heuristics,
The rst research program that Tversky and I and proposed a formulation in which the reduc-
undertook together consisted of a series of stud- tion of complex tasks to simpler operations is
ies of various types of judgment about uncertain achieved by an operation of attribute substitu-
events, including numerical predictions and as- tion. “Judgment is said to be mediated by a
sessments of the probabilities of hypotheses. heuristic when the individual assesses a speci-
Our conclusion in a review of this work was that ed target attribute of a judgment object by
“people rely on a limited number of heuristic substituting another property of that object—the
principles which reduce the complex tasks of heuristic attribute—which comes more readily
assessing probabilities and predicting values to to mind” (p. 53). Unlike the early work, Kah-
simpler judgmental operations. In general, these neman and Frederick’s conception of heuristics
heuristics are quite useful, but sometimes they is not restricted to the domain of judgment
lead to severe and systematic errors” (Tversky under uncertainty.
and Kahneman, 1974, p. 1124). The article in- For a perceptual example of attribute substi-
troduced three heuristics—representativeness, tution, consider the question: “What are the
availability, and anchoring—that were used to sizes of the two horses in Figure 7, as they are
explain a dozen systematic biases in judgment drawn on the page?” The images are in fact
under uncertainty, including nonregressive pre- identical in size, but the gure produces a com-
diction, neglect of base-rate information, over- pelling illusion. The target attribute that observ-
con dence, and overestimates of the frequency ers intend to evaluate is objective two-
of events that are easy to recall. Some of the dimensional size, but they are unable to do this
biases were identi ed by systematic errors in veridically. Their judgments map an impression
estimates of known quantities and statistical of three-dimensional size (the heuristic at-
facts. Other biases were de ned by discrep- tribute) onto units of length that are appropriate
ancies between the regularities of intuitive to the target attribute, and scaled to the size
judgments and the principles of probability of the page. This illusion is caused by the
theory, Bayesian inference, and regression differential accessibility of competing interpreta-
analysis. tions of the image. An impression of three-VOL. 93 NO. 5 KAHNEMAN: MAPS OF BOUNDED RATIONALITY 1461
dimensional size is the only impression of size
that comes to mind for na‡¨ve observers—paint- Tom W. is of high intelligence, although
ers and experienced photographers are able to lacking in true creativity. He has a need
do better—and it produces an illusion in the for order and clarity, and for neat and
perception of picture size. tidy systems in which every detail nds
A study by Fritz Strack et al. (1988) illus- its appropriate place. His writing is
trates the role of attribute substitution in a dif- rather dull and mechanical, occasion-
ferent context. College students responded to a ally enlivened by somewhat corny puns
survey which included the two following ques- and by ashes of imagination of the
tions in immediate succession: “How happy are sci- type. He has a strong drive for
you with your life in general?” and “How many competence. He seems to have little feel
and little sympathy for other people and
dates did you have last month?” The correlation does not enjoy interacting with others.
between the two questions was 0.12 when they Self-centered, he nonetheless has a deep
appeared in the order shown. Among respon- moral sense.
dents who received the same questions in re-
verse order, the correlation was 0.66. The
psychological interpretation of the high correla-
tion1 is inferential, but straightforward. The dat- Participants in a similarity group ranked the
ing question undoubtedly evoked in many nine elds by the degree to which Tom W.
respondents an emotionally charged evaluation “resembles a typical graduate student” (in that
of their romantic life. This evaluation was eld). The description of Tom W. was deliber-
highly accessible when the question about ately constructed to make him more representa-
happiness was encountered next, and it was tive of the less populated elds, and this
mapped onto the scale of general happiness. manipulation was successful: the correlation be-
In the interpretation offered here, the respon- tween the averages of representativeness rank-
dents answered the happiness question by re- ings and of estimated base rates was 20.62.
porting what came to their mind, and failed to Participants in the probability group ranked the
notice that they were answering a question nine elds according to the likelihood that Tom
that had not been asked—a cognitive illusion W. would have specialized in each. The respon-
that is analogous to the visual illusion of dents in the latter group were graduate students
Figure 7. in psychology at major universities. They were
The most direct evidence for attribute substi- told that the personality sketch had been written
tution was reported by Kahneman and Tversky by a psychologist when Tom W. was in high
(1973), in a task of categorical prediction. There school, on the basis of personality tests of du-
were three experimental groups in their experi- bious validity. This information was intended to
ment. Participants in a base-rate group evalu- discredit the description as a source of valid
ated the relative frequencies of graduate information.
students in nine categories of specialization.2 The statistical logic is straightforward. A de-
Mean estimates ranged from 20 percent for Hu- scription based on unreliable information must
manities and Education to 3 percent for Library be given little weight, and predictions made in
Science. the absence of valid evidence must revert to
Two other groups of participants were shown base rates. This reasoning implies that judg-
the same list of areas of graduate specialization, ments of probability should be highly correlated
and the following description of a ctitious with the corresponding base rates in the Tom
graduate student. W. problem.
The psychology of the task is also straight-
forward. The similarity of Tom W. to various
1
The observed value of 0.66 underestimates the true stereotypes is a highly accessible natural assess-
correlation between the variables of interest, because of ment, whereas judgments of probability are dif-
measurement error in all variables. cult. The respondents are therefore expected to
2
The categories were Business Administration; Com-
puter Science; Engineering; Humanities and Education;
substitute a judgment of similarity (representa-
Law; Library Science; Medicine; Physical and Life Sci- tiveness) for the required judgment of probabil-
ences; Social Sciences and Social Work. ity. The two instructions—to rate similarity or1462 THE AMERICAN ECONOMIC REVIEW DECEMBER 2003
FIGURE 8. TWO TESTS OF ATTRIBUTE SUBSTITUTION IN A PREDICTION TASK
probability—should therefore elicit similar by the similarity of Linda to the category
judgments. prototypes; others ranked the same outcomes
The scatterplot of the mean judgments of the by probability.
two groups is presented in Figure 8a. As the
gure shows, the correlation between judg-
ments of probability and similarity is nearly Linda is 31 years old, single, outspoken
perfect (0.98). The correlation between judg- and very bright. She majored in philoso-
ments of probability and base rates is 20.63. phy. As a student she was deeply con-
The results are in perfect accord with the hy- cerned with issues of discrimination and
pothesis of attribute substitution. They also con- social justice and also participated in an-
rm a bias of base-rate neglect in this tinuclear demonstrations.
prediction task. The results are especially com-
pelling because the responses were rankings.
The large variability of the average rankings of As might be expected, 85 percent of respon-
both attributes indicates highly consensual re- dents in the similarity group ranked the con-
sponses, and nearly total overlap in the system- junction item (#8) higher than its constituent,
atic variance. indicating that Linda resembles the image of a
Figure 8b shows the results of another study feminist bank teller more than she resembles a
in the same design, in which respondents were stereotypical bank teller. This ordering of the
shown the description of a woman named two items is quite reasonable for judgments of
Linda, and a list of eight possible outcomes similarity. However, it is much more problem-
describing her present employment and activi- atic that 89 percent of respondents in the prob-
ties. The two critical items in the list were #6 ability group also ranked the conjunction higher
(“Linda is a bank teller”) and the conjunction than its constituent. This pattern of probability
item #8 (“Linda is a bank teller and active in judgments violates monotonicity, and has been
the feminist movement”). The other six pos- called the “conjunction fallacy” (Tversky and
sibilities were unrelated and miscellaneous Kahneman, 1983).
(e.g., elementary school teacher, psychiatric The observation that biases of judgment are
social worker). As in the Tom W. problem, systematic was quickly recognized as relevant
some respondents ranked the eight outcomes to the debate about the assumption of rationalityVOL. 93 NO. 5 KAHNEMAN: MAPS OF BOUNDED RATIONALITY 1463
in economics (see, e.g., Peter A. Diamond, ple’s decisions often express affective evalua-
1977; David M. Grether, 1978; Howard Kun- tions (attitudes), which do not conform to the
reuther, 1979; Arrow, 1982). There has also logic of economic preferences. To understand
been some discussion of the role of speci c preferences, then, we may need to understand
judgment biases in economic phenomena, espe- the psychology of emotions. And we cannot
cially in nance (e.g., Werner F. M. De Bondt take it for granted that preferences that are con-
and Thaler, 1985; Robert J. Shiller, 2000; An- trolled by the emotion of the moment will be
drei Shleifer, 2000; Matthew Rabin, 2002). Re- internally coherent, or even reasonable by the
cent extensions of the notion of heuristics to the cooler criteria of re ective reasoning. In other
domain of affect may be of particular relevance words, the preferences of System 1 are not
to the conversation between psychology and necessarily consistent with the preferences of
economics, because they bear on the core con- System 2. The next section will show that some
cept of a preference. As was noted earlier, af- choices are not appropriately sensitive to vari-
fective valence is a natural assessment, which is ations of quantity and cost—and are better de-
automatically computed and always accessible. scribed as expressions of an affective response
This basic evaluative attribute (good/bad, like/ than as economic preferences.
dislike, approach/avoid) is therefore a candidate
for substitution in any task that calls for a fa- VI. Prototype Heuristics
vorable or unfavorable response. Slovic and his
colleagues (see, e.g., Slovic et al., 2002) intro- The results summarized in Figure 8 showed
duced the concept of an affect heuristic. They that the judgments that subjects made about the
showed that affect (liking or disliking) is the Tom W. and Linda problems substituted the
heuristic attribute for numerous target at- more accessible attribute of similarity (repre-
tributes, including the evaluation of the costs sentativeness) for the required target attribute of
and bene ts of various technologies, the safe probability. The goal of the present section is to
concentration of chemicals, and even the pre- embed the representativeness heuristic in a
dicted economic performance of various indus- broader class of prototype heuristics, which
tries. In an article aptly titled “Risk as share a common psychological mechanism—
Feelings,” Loewenstein et al. (2001) docu- the representation of categories by their proto-
mented the related proposition that beliefs about types—and a remarkably consistent pattern of
risk are often expressions of emotion. biases.
If different target attributes are strongly in- In the display of lines in Figure 3, the average
uenced by the same affective reaction, the (typical) length of the lines was highly accessi-
dimensionality of decisions and judgments ble, but the sum of their lengths was not. Both
about valued objects may be expected to be observations are quite general. Classic psycho-
unreasonably low. Indeed, Melissa L. Finucane logical experiments have established the fol-
et al. (2000) found that people’s judgments of lowing proposition: whenever we look at or
the costs and bene ts of various technologies think about a set (ensemble, category) which is
are negatively correlated, especially when the suf ciently homogeneous to have a prototype,
judgments are made under time pressure. A information about the prototype is automati-
technology that is liked is judged to have low cally accessible (Michael I. Posner and Stephen
costs and large bene ts. These judgments are W. Keele, 1968; Eleanor Rosch and Carolyn B.
surely biased, because the correlation between Mervis, 1975). The prototype of a set is char-
costs and bene ts is generally positive in the acterized by the average values of the salient
world of real choices. In the same vein, Kahne- properties of its members. The high accessibil-
man et al. (1997) presented evidence that dif- ity of prototype information serves an important
ferent responses to public goods (e.g., adaptive function. It allows new stimuli to be
willingness to pay, ratings of moral satisfaction categorized ef ciently, by comparing their fea-
for contributing) yielded essentially inter- tures to those of category prototypes.3 For
changeable rankings of a set of policy issues.
Here again, a basic affective response appeared
to be the common factor. 3
Stored information about individual exemplars also
Kahneman et al. (1997) suggested that peo- contributes to categorization.You can also read
