Académique Documents
Professionnel Documents
Culture Documents
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp
.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.
American Economic Association is collaborating with JSTOR to digitize, preserve and extend access to Journal
of Economic Literature.
http://www.jstor.org
MATTHEW RABIN
University of California at Berkeley
For- commnents oni. this e:sosy or discussion o01 r-ela.ted topics, I thanik Henry Aari-on.,Geor-ge
Akerlof, James And1reon.i, Steveni. Blatt, Gary Charn-^1ess,Eddie Dekel, Peter Diamond Jon.
Elster, Erik Eyster, Ernst Fehr, David I. Levine, George Loewenistein, Rob MacCotwi-, Jamnes
Montgomniery, Vali-La-ini.Mid, Ted O'Donoghute, Dra^zen.Pr^elec, llya Segal, Eldai Sha fir Genie
Smiiolensky,Joel Sobel, AmitosTversky, foutr anonymous referees, anid especially ColilhnCamiterer-,
Danny Kahnenman, and Richiari-dThaller. Co-authors oni, resear^chii-elated to the topics of this
essay include David Bowman, Deborah Minehart, Ted O'Donoghtue, anid Joel Sch t-ag. Helpfill
research assistance wcaspr-ovided by Gadi Barlevy, Nikki Blasberg, Gail Brennan, Paul Ellick-
son, April Franco, Marcus Henzg, Br-uce Hstu, Jini, Woo Jtunig, and especially Ste-veui Blatt,
V
Jinumty Chani, Erik Eyster-, and Clara Wang. I amitextremitely grateftul to the Nationial Scienice,
Ruissell Sage, Alfred P. Sloain, and MacArthur Fouindatiols for finian.cial stuppor-t oni t-esearch
verbose an.d citationi-he7avy version of this essay, covering a few
related to thi.s essay. A mitori-e
additional topics, is avvailable as Uniiversity of Californtia. at Berkeley Departmiienitof Econiomit-
ics Working Paper No. 97-251.
an underlying premise of this essay is what research shows some of those de-
that we should strive to understand psy- partures are.
chological findings in light of these
methods.3 These methods raise prob- 2. Preferences
lems for doing full justice to behavioral
reality: Because of the h-igh premium This section discusses how psycho-
economics places on the logic and pre- logical findings suggest we modify the
cision of arguments and the quantifica- utility functions economists employ.
tion of evidence, attending to all facets
of human nature is neither feasible nor A. Reference Levels, Adaptation,
desirable. The realization that many de- and Losses
tails of human behavior must be ig- Overwhelming evidence shows that
nored, however, should not license in- humans are often more sensitive to how
stitutionalized complacency about the their current situation differs from
behavioral validity of our assumptions; some reference level than to the abso-
"tractability" and "parsimony" should be lute characteristics of the situation
guiding principles in our efforts to (Harry Helson 1964). For instance, the
make our research more realistic, not same temperature that feels cold when
pretexts for avoiding this task. As it now we are adapted to hot temperatures
stands, some important psychological may appear hot when we are adapted to
findings seem tractable and parsimoni- cold temperatures. Understanding that
ous enough that we should begin the people are often more sensitive to
process of integrating them into eco- changes than to absolute levels suggests
nomics. Incorporating other findings that we ought incorporate into utility
will take longer. But even in those analysis such factors as habitual levels
cases, economists ought to become of consumption. Instead of utility at
aware of the shortcomings of our mod- time t depending solely on present con-
els, regret these shortcomings, and keep sumption, ct, it may also depend on a
our eyes open for ways to remedy them. "reference level," rt, determined by fac-
While I conclude in Section 5 by briefly tors like past consumption or expecta-
discussing the general endeavor of tions of future consumption. Hence, in-
incorporating psychological findings stead of a utility function of the form
into economics, I have omitted from ut(rt;ct), utility should be written in a
this essay many of the classical meta- more general form, ut(rt;ct). Wlile somie
arguments about whether it is "possi- economnists have over the years incorpo-
ble" that behavioral research identifies rated reference dependence into their
important departures from economists' economic analysis, it is fair to say that
habitual assumptions about human be- the ways and degrees to which refer-
havior. Of course it is possible, and in ence points influence behavior have not
fact it is true. In this essay, I review fully been appreciated by economists.4
Researchers have identified a pervasive
3 Indeed, I have organized the essay to reflect
feature of reference dependence: In a
this premise. At times, difficulties in organization
belie deeper difficulties of how to fit this research wide variety of domains, people are sig-
into the economics framework. For instance, while nificantly more averse to losses than
Section 2 discusses approaches to modeling loss
aversion fully within tihe rational-choice frame- 4 There have been some earlier examples of
work, material in Section 4 suggests loss aversion economists considering reference dependence
isn't always best conceived of as compatible with (James Duesenberiy 1949; Richard Easterlin 1974;
utility maximization. and Harl Ryder and Geoffrey Heal 1973).
they are attracted to same-sized gains it more than before she possessed it.
(Kahneman, Jack Knetsch, and Thaler Kahneman, Knetsch, and Thaler (1990)
1990). One realm where such loss aver- nicely illustrate this phenomenon. They
sion plays out is in preferences over randomly gave mugs worth about $5
wealth levels. Tversky and Kahneman each to one group of students. Minimal
(1991) suggest that in the domain of selling prices were elicited from those
money (and in others where the sizes of given the mugs (with an incentive-com-
losses and gains can be measured), patible procedure that ensured honest
people value modest losses roughly reports). Minimal "prices"-sums of
twice as much as equal-sized gains. money such that they would choose that
That the displeasure from a monetary sum rather than the mug-were elicited
loss is greater than the pleasure from a from another group of subjects not
same-sized gain is also implied by a given mugs. These two groups, "sellers"
concave utility function, which econo- and "choosers," faced precisely the
mists typically use as the explanation same choice between money and mugs,
for risk aversion. But loss aversion says but their reference points differed:
that the value function abruptly changes Those who were randomly given mugs
slope at the reference level, so that peo- treated the mugs as part of their refer-
ple dislike even small-scale risk. For in- ence levels or endowments, and consid-
stance, most people prefer their status ered leaving without a mug to be a loss,
quo to a 50/50 bet of losing $10 or gain- whereas individuals not given mugs con-
ing $11. The standard concave-utility- sidered leaving without a mug as re-
function explanation for risk aversion is maining at their reference point. In one
simply not a plausible explanation of experiment, the median value placed on
such risk attitudes. Rajnish Mehra and the mug was $3.50 by choosers but
Edward Prescott (1985) and Larry Ep- $7.00 by sellers. Such results have been
stein and Stanley Zin (1990) have, for replicated repeatedly by many re-
instance, observed that the expected- searchers in many contexts.
utility framework cannot simultaneously As established by Knetsch and John
explain both the small-scale and large- Sinden (1984), William Samuelson and
scale risk attitudes implied by macro Richard Zeckhauser (1988), and Knetsch
data, and Rabin (1997) provides a "cali- (1989), a comparable phenomenon-the
bration theorem" that indeed shows status quo bias-holds in multiple-good
that no concave utility function can si- choice problems. Here, loss aversion
multaneously explain plausible small- implies that individuals tend to prefer
scale and large-scale risk attitudes. A the status quo to changes that involve
reference-based kink in the utility func- losses of some goods, even when these
tion is required to explain such risk atti- losses are offset by gains of other goods.
tudes within the expected-utility frame- Knetsch and Sinden (1984) and Knetsch
work.5 (1989), for instance, randomly gave stu-
Loss aversion is related to the strik- dents either candy bars or decorated
ing endowmnent effect identified by mugs. Later, each student was offered
Thaler (1980): Once a person comes to the opportunity to exchange her gift for
possess a good, she immediately values the other one-a mug for a candy bar or
vice versa. Ninety percent of both mug-
5Uzi Segal and Avia Spivak (1990), and others,
owners and candy-owners chose not to
however, develop a model of such first-order risk trade. Because the goods were allocated
aversion outside the expected-utility framework. randomly and transaction costs were
minimal, the different behavior for the with the standard concavity assump-
two groups of subjects presumably re- tion.6
flected preferences that were induced In order to study the effects of refer-
by the initial allocation. Knetsch (1989) ence points in a dynamic utility-maximi-
and Tversky and Kahneman (1991) zation framework, we need to take into
show that such preferences can be use- account how people feel about the ef-
fully captured by utility functions de- fects their current choices have on their
fined over reference levels as well as future reference points. To maximize
consumption levels. their long-run utilities when reference
In addition to loss aversion, another points matter, people must determine
important reference-level effect is di- two things beyond how they feel about
minishing sensitivity: The marginal ef- departures from reference points: How
fects in perceived well-being are current behavior affects future refer-
greater for changes close to one's refer- ence points and how they feel about
ence level than for changes further changes in their reference points.
away. As Kahneman and Tversky (1979) Economists Ryder and Heal (1973)
note, diminishing sensitivity is a perva- model the process by which reference
sive pattern of human perception, points change with the formula
where our perceptions are a concave +
rt- occt-i (1-oc)rt-1, where cc e (0,1) is a
function of the magnitudes of change. parameter measuring how quickly peo-
For instance, we are more likely to dis- ple adjust their reference points. In a
criminate between 30 and 6? changes rational-expectations model, people will
in room temperature than between take this formula into account when
23? and 26? changes. This applies to maximizing their long run well being.
both increases and decreases in tem- Such an account of how reference levels
perature. In the context of preferences are determined seems intuitive, though
over uncertain monetary outcomes, there seems to be little evidence on this
diminishing sensitivity implies that topic.7 Evidence is similarly sparse
the slope of a person's utility function about how people's preferences depend
over wealth becomes flatter as her on chang;es in reference points.8 With-
wealth gets further away from her 6 Because this "risk-loving" tendency is in con-
reference level. Because for losses flict with the diminishing marginal utility of in-
relative to the reference level "further come, however, it has often been conjectured that
away" implies lower wealth levels, di- risk aversion may reappear for large losses that
might push a person to extremely Tow coinsump-
minishing sensitivity has a provocative tion levels.
implication: While people are likely to 7 Bowmnan,Minehart, and Rabin (1997) combine the
be risk averse over gains, they are ofter, Ryder and Heal approach of rational-expectations, ref-
erence-dependent utilities with a utility function that
risk-loving over losses. Kahneman and incorporates loss aversion and di- minishing sensitivity.
Tversky (1979) found that 70 percent Duesenberry (1949) implicitly posited a reference
of subjects report that they would pre- function closer to rt Max {cT t < t-that is, a per-
son's reference level was her highest past consumption
fer a 3/4 chance of losing nothing and level.
1/4 chance of losing $6,000 to a 2/4 8 For exceptions, see Loewenstein and Nachum
chance of losing nothing and 1/4 chance Sicherman (1991), Robert Frank (1985 ch. 15,
1989), and Frank and Robert Hutchens (1993),
each of losing $4,000 or $2,000. Be- who have identified a tendency for people to pre-
cause the preferred lottery here is a fer income profiles that are steady or increasing
mean-preserving spread of the less-pre- over time to same-sized decreasing profiles,
strongly indicating that people prefer not to be-
ferred lottery, the responses of 70 per- come accustomed to levels of consumption they
cent of the subjects are inconsistent know they cannot maintain. For consideration of
of industrial relations (Stacy Adams preferences.9 To get a sense for how so-
1963; Akerlof 1982; and Akerlof and cial preferences differ from simple al-
Janet Yellen 1990). truism, consider what can be called "be-
Experimental research makes clear havioral distributive justice": How do
that preferences depart from pure self- people choose to divide resources?10
interest in non-trivial ways: Subjects There are two aspects to this question:
contribute to public goods more than First, what do people, when disinter-
can be explained by pure self-interest; ested, feel are proper rules for alloca-
they often share money when they tion? Second, to what degree do people
could readily grab it for themselves; and sacrifice self-interest for the sake of
they often sacrifice money to retaliate these principles? Very roughly, imagine
against unfair treatment. Debates re- that person l's utility function takes the
main, as some researchers have empha- form Ul - (1 - r).Inl + r.W1(n1J 12), where
sized the possibility of producing W, is person l's view of the proper allo-
laboratory environments (e.g., well-or- cation, and Ill is (as above) her self-in-
ganized, private-information double- terested payoff, and 0O?<r?1 measures
auction spot markets) that induce be- the weight this person puts on self-in-
havior that is closer to purely terest versus proper allocation. To un-
self-interested than is behavior in other derstand the implications of fairness
settings. But, disentangling debates and justice for behavior, we need to
over the nature of preferences from know both the nature of the W, func-
strong ancillary assumptions about tion, and the level of r.
which institutions and environments To address the question of disinter-
matter in the real world, there does not ested assessments, suppose two people
appear to be debate among behavioral together find $10 worth of money or
researchers about whether underlying other goods on the ground. How would
preferences depart non-trivially from the average person, acting as a third
pure self-interest.
One form of social motivations on 9 See Dennis Krebs (1970, 1982) for some psy-
which economists have focused is al- chological evidence on altruism and need-based
truism, in the sense that people put posi- helping motives. Even where simple altruism may
adequately describe behavior (e.g., in donating to
tive value on the well-being of others. charity), psychological research may be of value
Roughly, this approach says that person for welfare economics. In particular, research has
1 acts as if she is maximizing prefer- explored whether people who help others do so
for "truly" altruistic reasons, in the sense that the
ences of the form U1(x)- (1- r)Il(x) actions they take will lower their experienced
+ rH2(x),where Ill(x) is person I's "ma- well-being, or whether they do so "only" to allevi-
terial well-being" from outcome x, and ate painful guilt, etc. Do you help a stranger when
inconvenient because it is the right thing to do
H2(x) is person 2's material well-being. even though it makes you worse off? Or do you
By letting r be small, we can capture like bringing joy to others in the same way you like
the idea that people are mostly self-in- eating apples? Or do you know you will have un-
pleasant, guilt-induced nightmares if you don't
terested; by assuming r>0, we can in- help? These distinctions will be important in wel-
vestigate when and how concern for fare analysis: Even if we think behavior is de-
others affects behavior and welfare. scribed by Max Ul(x) - (1 - r)-.lH(x) + rH2(x), we
must still resolve whether Person l's experienced
Simple altruism may parsimoniously well-being is described by Ui(x) or fli(x).
capture important phenomena in many 10 I use the term "behavioral distributive jus-
contexts. But there is a mass of experi- tice" to emphasize that I am reviewing evidence
on how people actually feel about distributive jus-
mental evidence that indicates it is tice, not normative or philosophical questions of
often a very wrong description of social what is the proper notion of distributive justice.
party, decide to split the surplus be- - Doctors have determined that
tween the two?11 The simple-altruism Jones's metabolism is such that his
perspective would suggest solutions body derives 100 milligrammes of vita-
such as giving the surplus to the person min F from each grapefruit con-
who is poorer-or for non-money sumed, while it derives no vitamin F
goods-the person who values the whatsoever from avocado.
goods more. But research shows that - Doctors have also determined that
many people in many contexts do not Smith's metabolism is such that his
find the "maximal-benefits" criterion at- body derives 50 milligrammes of vita-
tractive. One simple alternative norm is min F from each grapefruit consumed
prevalent: Resources should be split and also from each avocado con-
50-50. Except in extreme cases, often sumed.
we ignore issues of relative usefulness - Both persons, Jones and Smith, are
and feel that goods should be divided interested in the consumption of
equally. grapefruit or avocados only insofar as
But the maximal-benefits criterion such consumption provides vitamin
fares even worse than this. Many people F-and the more the better. All the
feel goods should be allocated accord- other traits of the two fruits (such as
ing to a "maximin" criterion which equal- taste, calorie content, etc.) are of no
izes welfare improvements between the consequence to them.
two people. That is, disinterested peo-
- No trades can be made after the di-
ple often seem to maximize preferences
vision takes place.
of the form Wo- Min[rli,n2], where Ill
and 112 are the gains in utility from di- How should the fruits be divided be-
viding resources. Because it takes a tween Jones and Smith, if the division is
greater allocation to increase the utility to be just?
of a person who values a resource less, Grapefruits are worth more to Jones
the maximnin criterion typically implies than to Smith, and avocados are worth-
that more than half the resources are al- less to Jones. The "socially efficient,"
located to the person who values those metabolism-maximizing allocation,
resources less. Consider the following therefore, is for Jones to get all the
hypothetical situation that Maya Yaari grapefruits, and Smith to get all the
and Menahem Bar-Hillel (1984, p. 8) avocados. (Such an allocation would
posed to 163 subjects: seem to accord somewhat to 50-50
norms.) The maximin allocation, how-
Q1:A shipment containing 12 grape-
ever, would be to give eight grapefruits
fruits and 12 avocados is to be distrib-
to Jones, and give the remaining grape-
uted between Jones and Smith. The
fruits and all the avocados to Smith.
following information is given, and is
Subjects were given a menu of five
known also to the two recipients:
different allocations, and asked to state
which allocation they found the most
III assume the surplus is found so as to con- "just." Yaari and Bar-Hillel (1984, p.
sider the thought experiment that neither party 10) report the percentage of respon-
deserves the money more than the other. Desert
will obviously be relevant in many situations-and dents who chose each of the five alloca-
the massive psychological literature on "equity tions. The five allocations are denoted
theory" shows that people feel that those who have by the number of grapefruits and avoca-
put more effort into creating resources have more
claim on those resources (Ellen Berscheid, David dos to be distributed to Jones and
Boye, and Elaine Walster 1968). Smith (denoted by their initials):
suggest that the maximin rule resonates that the prevailing market price is fair.
with many people even when allocating Other experiments find virtually no
money, and even when self-interest is at change in either behavior or percep-
stake. Allowing proportions of a "pie" to tions of fairness over time (Fehr and
have different monetary values to dif- Armin Falk 1996). In any event, be-
ferent parties has also been a theme in cause adjustments of fairness judgments
the experimental bargaining research are not immediate, fairness considera-
(Roth and J. Keith Murnighan 1982; tions may help explain the sort of
and Kagel, Chung Kim, and Donald medium-run wage and price stickiness
Moser 1996). In these settings, dis- studied by macroeconomists, and evi-
entangling self-interested bargaining dence that market outcomes are likely
strategies from preferences for just allo- to be self-interested exists only for com-
cations is quite difficult, but results petitive spot markets. See, for example,
seem comparable to Andreoni and Fehr, Erich Kirchler, and Andreas
Miller's. Weichbold (1994) for an experimental
In moving from abstract, context-free study of labor markets where behavior
allocation problems to everyday eco- never converges to the self-interested
nomic fairness judgments, things be- outcome.
come significantly more complicated. Finally, in attempting to capture
First, as elsewhere, reference levels are behavioral findings with models of
crucial. Thaler (1985) and Kahneman, social preferences, it is important to
Knetsch, and Thaler (1986a, 1986b) note that people seem to implicitly
demonstrate that loss aversion plays a (but pervasively) consider equitable
very strong role in people's notion of sharing over changes in total endow-
fairness; firms have more of an obliga- ments, not total endowments them-
tion not to hurt workers or customers selves. Preferences defined over final
relative to reference transactions than wealth states cannot plausibly explain
they have to improve the terms of rules such as 50/50 sharing or the maxi-
trade. Relatedly, people's general per- min criterion. With plausible assump-
ceptions of fair behavior may adjust tions about initial endowments entering
over time. Kahneman, Knetsch, and any moderate size division-of-the-pie
Thaler (1986a, p. 730) argue that, situation, any social welfare function
"Terms of exchange that are initially defined with respect to overall con-
seen as unfair may in time acquire sumption levels will almost always yield
the status of a reference transaction. all-or-nothing allocations. Apparently,
Thus, the gap between the behavior people generally have a one-pie-at-a-
that people consider fair and the behav- time conception of fair-division prob-
ior that they expect in the market- lems. This is not an insurmountable
place tends to be rather small." Robert tendency: If people are presented
Franciosi et al. (1995) experimentally with several allocation problems to-
support this hypothesis by testing gether, they will likely attend to the
reactions to unfair price increases in a overall implications of their several
laboratory posted-offers market; they choices. Nevertheless, any attempt to
show that the role of fairness con- capture behavioral norms of fairness
siderations in price-determination di- and dis- tributional justice with formal
minishes with repetition, suggesting models of social preference must con-
that in competitive spot markets front the "piecemeal" nature of these
people may eventually come to believe norms.
therefore, learning that others were that their null hypothesis of completely self-inter-
ested behavior corresponds to zero contributions,
not conserving would cause you to in- where zero contributions is also the most extreme
tensify your conservation efforts. This behavior subjects could exhibit. Therefore, all de-
prediction is inconsistent with intuition partures from full rationality are necessarily in the
direction of "generous" behavior. Andreoni (1995)
and empirical evidence: People are shows that, by a very conservative estimate, at
more inclined to conserve water if they least half the contributions to public goods are in-
think other people are conserving, not tentional rather than "errors." See also Claudia
Keser (1996) for evidence that generous contribu-
if they think others are splurging. Peo- tions are not merely an artifact of experimental
ple reciprocate the lack of public spirit- design.
the inability of the two captured prison- erates posing as subjects were in a posi-
ers to communicate with each other in- tion to help real subjects fill out some
dicates that we implicitly believe that worksheets. One third of the subjects
the prisoner's dilemma also really were told that the confederate had vol-
amounts to a coordination game. If de- untarily offered to help; one third were
fecting were truly a dominant strategy, told that the experimenter had in-
pre-game communication would not structed the confederate to help; and
matter. More direct evidence of recip- one third were told that the confederate
rocal altruism, in the context of the m-ight be willing to help, but the con-
prisoner's dilemma, comes from Shafir federate was instructed to refuse to
and Tversky (1992). When subjects help. When the subjects were later
were told that their anonymous partner given an opportunity to assist the con-
in a prisoners' dilemma had cooperated, federates, they reciprocated earlier
16 percent also cooperated; when sub- help, but did sa significantly more when
jects were told that their partner did it was voluntary than when it was invol-
not cooperate, only 3 percent cooper- untary.
ated. Volition is also central to the propen-
Reciprocity motives manifest them- sity to retaliate against negative actions.
selves not only in people's refusal to Sally Blount (1995) asked subjects
cooperate with others who are being about their willingness to accept take-
uncooperative, but also in their willing- it-or-leave-it offers made by anonymous
ness to sacrifice to hurt others who are other parties on how to split $10.15 One
being unfair. A consumer may refuse to group of subjects was told that the "ulti-
buy a product sold by a monopolist at matum" was coming from anonymous
an "unfair" price, even if she hurts her- other students, and that their responses
self by foregoing the product. An em- would determine the division between
ployee who feels he has been mis- them and these anonymous other stu-
treated by a firm may engage in costly dents. Another group was told that a
acts of sabotage, perhaps to the point of third party (also an anonymous student)
violently retaliating against his employ- was to determine the offer made. In
ers. Members of a striking labor union this variant, the person who would be
may strike longer than is in their mate- hurt by a subject's decision to reject an
rial interests because they want to pun- offer did not participate in the offer,
ish a firm for being unfair. and the third party who made the offers
A crucial feature of the psychology of would not be affected by the subject's
reciprocity is that people determine decision. A final group of subjects were
their dispositions toward others accord- told that the offer would be generated
ing to motives attributed to these oth- randomly by a computer-simulated rou-
ers, not solely according to actions lette wheel. In one study, the average
taken. When motivated by reciprocal al- minimal acceptable offers for those
truism, for instance, people differenti- groups were $2.91, $2.08, and $1.20.
ate between those who take a generous That is, people did reject very low of-
action by choice and those who are
forced to do so. Demonstrating both the 15 The "ultimatum game" of the sort studied b
basic principle of reciprocity and the Blount was first developed by Werner Gtith, Roif
role of volition, Richard Goranson and Schmittberger, and Bernd Schwarze (1982). For
reviews of the (massive) literature developed
Leonard Berkowitz (1966, p. 229) con- since, see Thaler (1988), Guth and Reinhard Tietz
ducted an experiment in which confed- (1990), and Camerer and Thaler (1995).
cause we expect close to the same prob- that people exaggerate the likelihood
ability distribution of types in small that a short sequence of flips of a fair
groups as in large groups, for example, coin will yield roughly the same number
we tend to view it as comparably likely of heads as tails. What is commonly
that at least 80 percent of 20 coin flips known as "the gambler's fallacy" is a
will come up heads than that at least 80 manifestation of this bias: If a fair coin
percent of 5 coin flips will come up has not (say) come up tails for a while,
heads; in fact, the probabilities are then on the next flip it is "due" for a
about 1 percent and 19 percent, respec- tails, because a sequence of flips of a
tively. Kahneman and Tversky (1982a, fair coin ought to include about as many
p. 44) asked undergraduates the follow- tails as heads.
ing question: When the underlying probability dis-
tribution generating observed se-
A certain town is served by two hospitals. In
the larger hospital about 45 babies are born
quences is uncertain, the gambler's fal-
each day, and in the smaller hospital about 15 lacy leads people to over-infer the
babies are born each day. As you know, about probability distribution from short se-
50 percent of all babies are boys. However, quences. Because we underestimate the
the exact percentage varies from day to day. frequency of a mediocre financial ana-
Sometimes it may be higher than 50 percent,
sometimes lower.
lyst making lucky guesses three times in
For a period of 1 year, each hospital re- a row, we exaggerate the likelihood that
corded the days on which more than 60 per- an analyst is good if she is right three
cent of the babies born were boys. Which times in a row. This tendency to over-
hospital do you think recorded more such infer from short sequences, in turn,
days?
leads to misperception of regression to
Twenty-two percent of the subjects the mean. Because we read too much
said that they thought that it was more into patterns that depart from the
likely that the larger hospital recorded norm, we don't expect that further ob-
more such days, and 56 percent said servations will look more normal. As
that they thought the number of days teachers, we exaggerate the extent to
would be about the same. Only 22 per- which one good or bad performance on
cent of subjects correctly answered that a test is a sign of good or bad aptitude,
the smaller hospital would report more so we don't expect exceptional perfor-
such days. This is the same fraction as mances to be followed by unexceptional
guessed exactly wrong. Apparently, the performances as often as they are. Mis-
subjects simply did not see the rele- understanding regression to the mean
vance of the number of child births per gives rise to spurious explanations for
day.'9 observed regression. When a student
The law of small numbers implies performs poorly on the midterm but
well on the final, teachers infer that the
19While people believe in the law of small num-
student has worked harder; if the stu-
bers, they apparently don't believe in the law of
large numbers: We underestimate the resemblance dent performs well on a midterm but
that large samples will have to the overall popula- poorly on the final, teachers infer that
tion. Kahneman and Tversky (1982a), for instance, the student has slacked off. Tversky and
found that subjects on average thought that there
was a more than 1/10 chance that more than 750 Kahneman (1974) give another exam-
of 1000 babies born on a given day would be male. ple. Flight-training instructors observed
The actual likelihood is way less than 1 percent. that when they praised pilots for
To overstate it a bit, people seem to have a univer-
sal probability distribution over sample means that smooth landings, performance usually
is insensitive to the sample size. deteriorated on the next landing, but
when they criticized pilots for poor ent subjects began viewing the pictures
landings, performance improved the at different points in the focusing pro-
next time. But random performance will cess, but the pace and final degree of
lead to "deterioration" following a good focus were identical for all subjects. Of
landing and "improvement" following a those subjects who began their viewing
poor landing. These flight instructors at a severe-blur stage, less than a quar-
developed a wrong theory of incentives ter eventually identified the pictures
based on erroneous statistical reason- correctly, whereas over half of those
ing. who began viewing at a light-blur stage
Another implication of the law of were able to correctly identify the pic-
small numbers is that people expect too tures. Bruner and Potter (1964, p. 424)
few lengthy streaks in sequences of ran- conclude that "Interference may be ac-
dom events. As with regression to the counted for partly by the difficulty of
mean, therefore, people tend to gener- rejecting incorrect hypotheses based on
ate spurious explanations for long substandard cues." That is, people who
streaks that are determined by chance. use weak evidence to form initial hy-
For instance, there is widespread belief potheses have difficulty correctly inter-
in the "hot hand" in basketball-that preting subsequent, better information
particular basketball players are streak that contradicts those initial hypothe-
shooters who have "on" nights and "off' ses. David Perkins (1981) argues that
nights which cannot be explained by such experiments provide support for
randomness. Thomas Gilovich, Robert the perspective that "fresh" thinkers
Vallone, and Tversky (1985) and Tver- may be better at seeing solutions to
sky and Gilovich (1989a, 1989b) have problems than people who have medi-
argued that the almost universally ac- tated at length on the problems, be-
cepted phenomenon of the hot hand is cause the fresh thinkers are not over-
non-existent in basketball. The exagger- whelmed by the "interference" of old
ated belief in hot hands seems partly ex- hypotheses.
plained by the misperception that This form of anchoring does not nec-
purely random streaks are too long to essarily imply that people misinterpret
be purely random. additional evidence, only that they ig-
nore additional evidence. Psychological
B. Belief Perseverance and
evidence reveals a stronger and more
Confirmatory Bias
provocative phenomenon: People tend
A range of research suggests that to misread evidence as additional sup-
once forming strong hypotheses, people port for initial hypotheses.20 If a
are often too inattentive to new infor- teacher initially believes that one stu-
mation contradicting their hypotheses. dent is smarter than another, she has
Once you become convinced that one the propensity to confirm that hypothe-
investment strategy is more lucrative sis when interpreting later perfor-
than another, you may not sufficiently mance.
attend to evidence suggesting the strat- Some evidence for confirmatory bias
egy is flawed. A particularly elegant is a series of experiments demonstrating
demonstration of such "anchoring" is how providing the same ambiguous in-
found in Jerome Bruner and Mary Pot- formation to people who differ in their
ter (1964). About 90 subjects were
shown blurred pictures that were gradu- 20 For a formal model of confirmatory bias, see
ally brought into sharper focus. Differ- Rabin and Schrag (1997).
initial beliefs on some topic can move John Darley and Paget Gross (1983)
their beliefs further apart. To illustrate demonstrate a related and similarly
such polarization, Charles Lord, Lee striking form of polarization. Seventy
Ross, and Mark Lepper (1979, pp. undergraduates were asked to assess a
2102) asked 151 undergraduates to nine-year-old girl's academic skills in
complete a questionnaire that included several different academic areas. Before
three questions on capital punishment. completing this task, the students re-
Later, 48 of these students were re- ceived information about the girl and
cruited to participate in another experi- her family and viewed a video tape of
ment. Twenty-four of them were se- the girl playing in a playground. One
lected because their answers to the group of subjects was given a fact sheet
earlier questionnaire indicated that they that described the girl's parents as col-
were "'proponents' who favored capital lege graduates who held white-collar
punishment, believed it to have a deter- jobs; these students viewed a video of
rent effect, and thought most of the the girl playing in what appeared to be
relevant research supported their own a well-to-do suburban neighborhood.
beliefs. Twenty-four were opponents of The other group of subjects was given a
capital punishment, doubted its deter- fact sheet that described the girl's par-
rent effect and thought that the rele- ents as high school graduates who held
vant research supported their views." blue-collar jobs; these students viewed
These subjects were then asked to a video of the same girl playing in what
judge the merits of randomly selected appeared to be an impoverished inner
studies on the deterrent efficacy of the city neighborhood. Without being sup-
death penalty, and to state whether a plied any more information, half of each
given study (along with criticisms of group of subjects was then asked to
that study) provided evidence for or evaluate the girl's reading level, meas-
against the deterrence hypothesis. Sub- ured in terms of equivalent grade level.
jects were then asked to rate, on 16 There was a small difference in the two
point scales ranging from -8 to +8, how groups' estimates-those subjects who
the studies they had read moved their had viewed the "inner-city" video rated
attitudes toward the death penalty, and the girl's skill level at an average of 3.90
how they had changed their beliefs re- (i.e., 9/10 through 3rd grade) while
garding its deterrent efficacy. At confi- those who had viewed the "suburban"
dence levels of p < .01 or stronger, video rated the girl's skill level at an av-
Lord, Ross, and Lepper found that pro-
ponents of the death penalty became on
average more in favor of the death pen- and other academics have probably observed how
differing schools of thought interpret ambiguous
alty and believed more in its deterrent evidence. A colleague saw the same model-cali-
efficacy, while opponents became even brating the elasticity of demand facing a Cournot
less in favor of the death penalty and oligopolist as a function of the number of firms in
an industry-described at the University of Chi-
believed even less in its deterrent effi- cago and at M.I.T. A Chicago economist derived
cacy. Scott Plous (1991) replicates the the formula and said, "Look how few firms you
Lord, Ross, Lepper results in the con- need to get close to infinite elasticities and perfect
competition." An M.I.T. economist derived the
text of judgment about the safety of nu- same formula and said, "Look at how large n has
clear technology.21 to be before you get anywhere close to an infinite
elasticity and perfect competition." These differ-
21 Lord, Ross, and Lepper posit that even pro- ent schools each interpreted the same mathemati-
fessional scientists are susceptible to such "same- cal formula as evidence reinforcing their respec-
evidence polarization." Indeed, many economists tive views.
erage of 4.29. The remaining subjects in sound with those observing economists'
each group were shown a second video reactions to behavioral evidence that
of the girl answering (with mixed suc- might be damaging to habitual econom-
cess) a series of questions. Afterwards, ics assumptions:
they were asked to evaluate the girl's With confirming evidence, we suspect that
reading level. The inner-city video both lay and professional scientists rapidly re-
group rated the girl's skill level at an duce the complexity of the information and
average of 3.71, significantly below the remember only a few well-chosen supportive
impressions. With disconfirming evidence,
3.90 estimate of the inner-city subjects
they continue to reflect upon any information
who did not view the question-answer that suggests less damaging "alternative inter-
video. Meanwhile, the suburban video pretations." Indeed, they may even come to
group rated the girl's skill level at an regard the ambiguities and conceptual flaws
average of 4.67, significantly above the in the data opposing their hypotheses as
somehow suggestive of the fundamental cor-
4.29 estimate of the suburban subjects
rectness of those hypotheses.
who did not view the second video.
Even though the two groups viewed the The above passages hint at the role
identical question-and-answer video, that selective scrutiny of evidence plays
the additional information further po- in confirmatory bias. Another form of
larized their assessments of the girl's "scrutiny-based" confirmatory bias is
skill level. Darley and Gross (1983) in- what I shall call hypothesis-based filter-
terpret this result as evidence of confir- ing. While it is sensible to interpret am-
matory bias-subjects were influenced biguous data according to current hy-
by the girl's background in their initial potheses, people tend to use the
judgments, but their beliefs were evi- consequent "filtered" evidence inappro-
dently influenced even more strongly by priately as further evidence for these
the effect their initial hypotheses had hypotheses. If a student gives an un-
on their interpretation of further evi- clear answer to an exam question, it is
dence. perfectly reasonable for a teacher to be
Certain types of evidence flows seem influenced in his evaluation of the an-
to be most conducive to confirmatory swer by his prior perceptions of that
bias. Ambiguity of evidence is widely student's mastery of the material. How-
recognized to be an important mediat- ever, after assigning differential grades
ing factor in both confirmatory bias to students according differential inter-
and overconfidence (see, e.g., Gideon pretation of comparable answers, it is a
Keren 1987; and Dale Griffin and Tver- mistake to then use differential grades
sky 1992). Keren (1988) notes the lack on the exam as further evidence of the
of confirmatory bias in visual percep- differences in the students' abilities.
tions, and concludes that confirmatory Lord, Ross, and Lepper note a similar
tendency depends on some degree of distinction in reflecting on the bias in
abstraction and the need for interpreta- their experiment: It is legitimate for
tion not present in simple visual tasks. people to differentially assess proba-
Lord, Ross, and Lepper (1979, p. 2099) tiveness of different studies according
posit that when faced with complex and to their current beliefs about the merits
ambiguous evidence, we emphasize the of the death penalty. The "sin" is in us-
strength and reliability of confirming ing their hypothesis-based interpreta-
evidence but the weaknesses and unreli- tions of the strength of different studies
ability of disconfirming evidence. They as further support for their beliefs.
also report an impression that may re- Even when each individual datum is
p. 288) first proposed this bias by they would have thought without that
observing that "(a) Reporting an information.25
outcome's occurrence increases its A pervasive fact about human judg-
perceived probability of occurrence; ment is that people disproportionately
and (b) people who have received out- wveight salient, memorable, or vivid evi-
come knowledge are largely unaware dence even when they have better
of its having changed their percep- sources of information.26 For instance,
tions [along the lines of (a)]." Combin- our assessment of a given city's crime
ing these, the literature on the hind- rate is likely to be too influenced by
sight bias shows that people exaggerate whether we personally know somebody
the degree to which their beliefs before who has been assaulted, even if we are
an informative event would be similar familiar with much more relevant gen-
to their current beliefs. We tend to eral statistics. Likewise, dramatic sto-
think we "knew it would happen all ries by people we know about difficul-
along." After a politician wins election, ties with a brand of car are likely to be
people label it as inevitable-and be- overly influential even if we are famil-
lieve that they always thought it was in- iar, via Consumer Reports, with general
evitable. statistics of the reliability of different
One example of Fischhoff s (1975) brands. In both these cases, and in
original demonstration of this effect was many others, the more salient informa-
to give subjects a historical passage re- tion should have virtually no influence
garding British intrusion into India and on our beliefs in the face of much more
military interaction with the Gurkas of pertinent statistical information. Tver-
Nepal. Without being told the outcome sky and Kahneman (1973) discuss, for
of this interaction, some subjects were example, how salience may distort clini-
asked to predict the likelihood of each cians' assessments of the relationship
of four possible outcomes: 1) British between severe depression and suicide.
victory, 2) Gurka victory, 3) military Incidents in which patients commit sui-
stalemate with a peace settlement, 4)
military stalemate without a peace set- 25 The definition of hindsight bias regards peo-
tlement. Four other sets of subjects ple's perceptions of how they themselves would
have answered a particular question absent infor-
were each told a different one of the mation they now have. As economists, we are
four outcomes was the true one (the likely to care mostly about a person's beliefs about
real true outcome is that the two sides other people, not about herself. In general, it is
hard to control experimentally for the fact that
fought to a stalemate without reaching a people have different information, and hard to iso-
peace settlement). For each reported late the hindsight bias when asking subjects what
outcome, when compared to a control others would have believed absent certain infor-
mation. Subjects may believe that others have dif-
group not told any outcome, subjects' ferent beliefs for a variety of reasons (e.g., you
average ex post guesses of their hypo- could believe that other people are not as smart as
thetical ex ante estimates were 15 per- you). I feel, however, that the evidence suggests
that we have a tendency to think that other people
cent higher than those of the control "should have known" as well (Hawkins and Hastie
group. People don't sufficiently "sub- 1990, p. 319).
26 In Tversky and Kahneman's (1973) formula-
tract" information they currently have
tion: "[A] person is said to employ the availability
about an outcome in imagining what heuristic whenever he estimates frequency or
probability by the ease with which instances or as-
sociations could be brought to mind." For more
for the existence of the bias. They conclude that general reviews of the ro1e of salience and vivid-
the bias is very real. (They also argue that the ef- ness, see Susan Fiske and Shelley Taylor (1991,
fects are "small.") chs. 5,7).
cide are much more likely to be remem- reduces or eliminates observed biases.
bered than are instances where patients The results are surprisingly negative.
do not commit suicide. This is likely to More generally, the research leads to
lead to an exaggerated assessment of mixed conclusions about when and how
the probability that depressed patients learning takes place, but very much
will commit suicide. does not support the strong versions of
Finally, there is a mass of psychologi- the experts-get-things-right and in-the-
cal research that finds people are prone real-world-people-learn hypotheses.
toward overconfidence in their judg- Research also suggests we should use
ments. The vast majority of researchers extreme caution in defining the relevant
argue that such overconfidence is per- notion of learning, because many peo-
vasive, and most of the research con- ple who do learn general principles do
cerns possible explanations (of which not apply those principles in particular
confirmatory bias discussed above is situations. In the context of overconfi-
one). 27 dence, for instance, Griffin and Tversky
(1992) and Andrea Baumann, Raisa De-
D. Do Learning and Expertise ber, and Gail Thompson (1991) con-
Eliminate Biases? clude that people who are aware of
The conjecture that experience helps their own accuracy overall are overcon-
overcome biases often leads economists fident on a case-by-case basis. When
to doubt the relevance of laboratory evi- people understand the limits in their
dence from inexperienced subjects. It is abilities to predict events accurately,
commonly argued that if important eco- they tend not to apply this general
nomic activity is performed by special- knowledge in calibrating the appropri-
ists and experts, or consists of tasks ate confidence in individual cases:
done repeatedly by the same individu- Kahneman and Tversky (1982b, p. 495)
als, the assumption of full rationality call such errors errors of applications,
fares much better than some of the psy- and note that "An error of application is
chological evidence indicates. most convincingly demonstrated when a
Do experience, expertise, and learn- person, spontaneously or with minimal
ing virtually eliminate biases? These are prompting, clutches his head and ex-
reasonable conjectures, and such fac- claims: 'How could I have missed
tors probably do on average moderate that?'" Even if people learn the rele-
biases. But the conjectures do not ap- vant statistical truths of their environ-
pear to be nearly as valid as economists ment, they may continue to make errors
imagine. Kahneman and Tversky in their judgnmentsand decision making
(1982a) and Tversky and Kahneman in every single case. One fears that
(1982), for instance, present experi- economists may sloppily interpret such
ments with subjects who vary in their head-clutching as evidence for the ra-
level of statistical sophistication, to test tionality hypothesis, rather than against
whether general knowledge of statistics it. But evidence that people see their
errors when confronted with them does
27 See, e.g., Stuart Oskamp (1965), jayashree not boost the rationality assumption as
Mahajan (1992), and Paul Paese and Maryellen
Kinnaly (1993). An early paper arguing this is economists use it. Our models use the
Fischhoff, Slovic, and Lichtenstein (1977), who rationality assumption as a realized fea-
also tested the robustness of overconfidence with ture of human behavior, not merely a
monetary stakes rather than reported judgments.
No decrease in overconfidence was found relative human potential.
to the no-money-stakes condition. As was demonstrated in the context
of confirmation bias discussed above, ning of the semester to predict their own
"learning" can even sometimes tend to future behaviors toward fellow sorority
exacerbate errors. Relatedly, Griffin members. Each subject was asked to pre-
and Tversky (1992), address the rela- dict "yes" or "no" whether she would en-
tionship between expertise and over- gage in each of six different behaviors-
confidence. When certain forms of pre- and to assess her confidence in her
dictability are high and when feedback prediction. Randomly, some of the mem-
takes the form of unambiguous statisti- bers were asked to list "why they might
cal evidence, experts tend to have a or might not perform each of the . . be-
pretty good sense of how accurate their haviors ... People were given a separate
predictions are. In such cases, experts page on which to list their reasons for
not only know more, but are more real- each behavior. They were told that the
istic than laypersons about how much purpose of this task was to 'organize
they know. But when predictability is their thoughts,' that 'no one will actually
low, experts are often more susceptible read what you have written,' and that
to overconfidence than are laypersons. Iyour reasons will be discarded.' They
Griffin and Tversky (1992, p. 430) pro- were urged to 'list as many reasons as
vide illustrations: you can think of, filling up this page if
you can'." Other subjects were asked to
If the future state of a mental patient, the make the same six predictions, but not
Russian economy, or the stock market cannot
be predicted from present data, then experts asked to think of reasons for their behav-
who have rich models of the system in ques- ior.
tion are more likely to exhibit overconfidence At the end of the semester, subjects
than lay people who have a very limited un- were asked if they had actually per-
derstanding of these systems. Studies of clini- formed each of the six activities. The
cal psychologists (e.g., Oskamp 1965) and
stock market analysts (e.g., Yates 1990) are results from this experiment showed
consistent with this hypothesis. that the act of reasoning increased sub-
jects' overconfidence regarding their
While a reasonable conjecture is that predictions of their own behavior.
greater thoughtfulness and intelligently While those who reasoned about their
searching for patterns in the world would predictions were roughly as confident
help douse biases, the quote specifically as those who did not reason (reasoners
targets (economists take note . . .) as sus- predicted their own Yes/No predictions
ceptible to overconfidence those "ex- would be accurate 80 percent of the
perts who have rich models of the system time, while the control group predicted
in question." Indeed, many authors have their accuracy at 82 percent), the rea-
hypothesized the role of the reasoning soners were in fact significantly less ac-
process itself in exacerbating the confir- curate than the control group in their
matory bias discussed above, which in predictions; reasoners' predictions were
turn leads to overconfidence (see, e.g., accurate 71 percent of the time, com-
Craig Anderson, Lepper, and Ross 1980; pared to 79 percent for the control
Ross et al. 1977; and Timothy Wilson group.
and Jonathan Schooler 1991). Wilson and
Suzanne LaFleur (1995, pp. 23-24) con- 4. Is "Maximnizing
Utility" the Right
duct an experiment on the role of "rea- Model?
soning" in strengthening confidence:
Members of six sororities at the Univer- The varied material of this section
sity of Virginia were asked at the begin- suggests that it may be wrong to con-
ceptualize some types of economic be- the measure of the hedonic experience
havior in terms of an agent who maxi- of that outcome. . . . The decision util-
mizes a stable, coherent utility function. ity of an outcome . . . is the weight as-
Indeed, some of the material here also signed to that outcome in a decision."
calls into question some of the interpre- The realization that decision and expe-
tations presented earlier. For instance, rienced utility may be systematically
loss-averse behavior may often reflect a different cuts to the core of our models
flawed rule of thumb employed because of choice. It also cuts to the core of our
people misperceive their own long-run methods of research, requiring us to
well-being, rather than a modified util- formulate ways of inferring and eliciting
ity function as suggested in Section 2. preferences that go beyond a "revealed
By calling into question the utility- preference" method to attempt to infer
maximization interpretation of behavior people's hedonic experiences, through
given in Section 2, such material sug- such methods as self reports of satisfac-
gests greater difficulty in improving the tion and even psychological measure-
behavioral realism of formal economics. ments.
On the other hand, the material at the How do people misperceive their
end of the section on self-control prob- utilities? One pattern is that we tend to
lems provides a great opportunity to im- underestimate how quickly and how
prove the realism of economics: Re- fully we will adjust to changes, not fore-
searchers have shown that a (relatively) seeing that our reference points will
simple multiple-self model of time-in- change. In a classic study, Philip Brick-
consistent discounting tractably modi- man, Dan Coates, and Ronnie Janoff-
fies our familiar exponential model to Bulman (1978) interviewed both lottery
yield a model that is manifestly more winners (with average winnings of about
realistic behaviorally and surely has im- $479,545) and a control group; they
portant economic consequences. found virtually no difference in rated
happiness of lottery non-winners and
A. Do We Know What Makes Us Happy?
winners. While such interview evidence
The research on biases reviewed in is inconclusive, the researchers con-
Section 3 indicates that people mis- trolled for alternative explanations
judge the probabilistic consequences of (such as selection bias or biased presen-
their decisions. But other research sug- tation by interviewers). Two effects
gests that, even when they correctly seemed to explain why lottery winners
perceive the physical consequences of would be less happy than the winners
their decisions, people systematically had presumably anticipated. First, mun-
misperceive the well-being they derive dane experiences become less satisfying
from such outcomes. We often system- by contrast to the "peak" experience of
atically mispredict our future experi- winning the lottery. Second, we become
enced utility, even when those predic- habituated to our circumstances: Along
tions rely only on accurate assessments the lines of the material presented in
of our past experienced utility (Kahne- Section 2, eventually the main carriers
man 1994; and Kahneman, Peter Wak- of utility become not the absolute levels
ker, and Rakesh Sarin 1997). As Kahne- of consumption, but departures from
man (1994, p. 21) puts it, "These our (new) reference level.28
considerations suggest an explicit dis- 28 Brickman, Coates, and janoff-Bulman (1978)
tinction between two notions of utility. also found more equivocal evidence for these ef-
The experienced utility of an outcome is fects by interviewing 29 paraplegics and quadri-
People do not anticipate the degree even though they will not be spending
of such adaptation, and hence exagger- their investment in the short term.
ate expected changes in utility caused Camerer et al. (1997) argue that New
by changes in their lives.29 This sug- York taxi drivers decide when to quit
gests that the "decision-utility" aversion driving for the day by a rule of thumb
people have to losses is not consonant that says they should make sure to
with "experienced utility." This realiza- match their usual take for the day. In
tion, in turn, calls for a reexamination some more extreme examples of loss
of the first topic of Section 2: Are loss aversion it is hard to believe that the
aversion, the endowment effect, and "transition utility" can rationally rank
other reference effects rational or irra- high relative to long-term utility. For
tional? If people experience losses rela- instance, Thaler (1980) compared sub-
tive to a status quo as quite unpleasant, jects' willingness to pay for a cure for a
then loss-averse behavior is rational, be- disease that leads to a quick and pain-
cause people are correctly anticipating less death with probability .001 versus
and avoiding unpleasant sensations. the minimum price you would accept to
And, the remembered "loss" of an voluntarily subject yourself to the same
owned mug may carry over time, or in disease. Subjects often required an or-
any event be substantial relative to the der of magnitude more money to expose
long-termn utility consequences of own- themselves to the disease than they
ing the mug. would pay for a cure. People charge
Yet loss aversion often seems to be a heavy premiums for losses relative to
judgmental bias: In decisions with sig- their status quo, even when it is hard
nificant long-run consequences, people to imagine that any experienced "tran-
should put less weight than they do on sition utility" is significant relative
their initial experience of losses. In- to long-term utility consequences-
deed, some researchers invoke loss here, whether or not you live beyond a
aversion more as an irrational rule of week.
thumb than as a rational utility func- Another example of how people
tion. Shlomo Benartzi and Thaler misperceive utility consequences of
(1995) argue that the equity-premium their choices is Richard Herrnstein and
puzzle can be explained by investors' Prelec's (1992b) theory of meliora-
111741frcir%n c-- ,hrI rrn f i r'n il 1 Nccpc tion. Based on a mass of evidence gath-
ered from people and other animals,
they argue that people tend to make
plegics about their happiness before their acci- current choices according to which
dents, their current happiness, and their expected choice directly yields the highest utility,
future happiness. Paraplegics felt less well off on
average than they felt before, and rated their hap- without taking into account the choice's
piness as lower than did lottery winners or the effect on the utilities from future
control group. Moreover, they did not currently choices. That is, people often ignore
enjoy mundane activities more than lottery win-
ners or the control group. However, accident vic- "internalities"-the effects a current
tims put more emphasis on, and took a more posi- choice has on the utilities of later
tive view of, mundane pleasures-they rated both choices. Say you eat at one of two
past pleasure and anticipated future pleasure from
such mundane activities slightly higher than either restaurants every night, either Blon-
the lottery winners or the control group. die's or Fat Slice. You enjoy Fat Slice
29 In the simple model of reference-point ad- more, but because you also enjoy vari-
justment discussed in Section 2, this can be trans-
lated as saying that people systematically underes- ety, your utility each evening is as fol-
timate the parameter cc. lows:
dures should yield the same prefer- selves whether they have a "reason" to
ences. choose one option over another (Shafir,
Itamar Simonson and Tversky (1992) Simonson, and Tversky 1993).
provide examples of context effects,
D. Time-Variant Preferences
where the addition of a new option to a
menu of choices may actually increase People have a taste for immediate
the proportion of consumers who gratification. We procrastinate on tasks
choose one of the existing options. For such as mowing the lawn that involve
example, the proportion of consumers immediate costs and delayed rewards
who chose a particular model of micro- and do soon things such as seeing a
wave oven increased when a second, movie that involve immediate rewards
more expensive model was added to and delayed costs. Economists tradi-
their choice set. (Subjects were first tionally model such tastes by assuming
asked to look at a catalogue containing that people discount streams of utility
the prices and descriptions of all the over time exponentially. An important
relevant choices from which their even- qualitative feature of exponential dis-
tual choice sets were drawn, so the re- counting is that it implies that a per-
sults seem unlikely to be due to any in- son's intertemporal preferences are
formation revealed by the choice sets.) time-consistent: A person feels the
As another example, Simonson and same about a given intertemporal trade-
Tversky (1992) ran an experiment that off no matter when she is asked.
illustrates that elicited subjects' prefer- Casual observation, introspection,
ence for an elegant Cross pen versus re- and psychological research all suggest
ceiving $6. While only 36 percent of that the assumption of time-consistency
subjects choosing only between the is importantly wrong. Our short term
Cross pen and the $6 chose the Cross tendency to pursue immediate gratifica-
pen, 46 percent of subjects who were tion is inconsistent with our long term
also given the choice of a less attractive preferences. While today we feel that it
pen chose the Cross pen. In both these is best that we not overeat tomorrow,
examples, the addition of an option that tomorrow we tend to overeat; while to-
compared unfavorably (as more expen- day we feel we should write a referee
sive or lower quality) to an existing op- report tomorrow, tomorrow we tend to
tion enhanced the perceived attractive- put it off. More generally, when consid-
ness of the existing option. ering tradeoffs between two future mo-
While people are often unaware that ments, we give stronger relative weight
the menu of choices influences their de- to the earlier moment as it gets closer.
cisions, Simonson and Tversky note that Kris Kirby and Herrnstein (1995), for
at other times decision makers explicitly instance, asked subjects to state their
rationalize their choices with refer- preferences among a series of pairs, in
ences to their choice sets. For instance, each case choosing between a smaller,
people may state explicitly that a given earlier reward and a larger, later re-
choice is a compromise between two ward. Subjects were (truthfully) told
other choices. Indeed, such. findings that one of their choices would be im-
suggest an alternative to the utility- plemented. In two experiments with
maximization framework that may help monetary rewards, 23 of 24 subjects
explain framing effects, preference re- "consistently reversed their choices
versals, and context effects: People may from the smaller, earlier reward to the
make choices in part by asking them- later, larger reward as the delay to both
rewards increased." Both the monetary April 14, versus spending eleven hours
stakes and the delays were substantial- to complete the same task on April 15.
subjects received an average of about Assume that your instantaneous disutil-
$21.50, with an average delay of about ity from doing work is simply the num-
21/2 weeks.32 ber of hours of work- ut(I0) = -10 and
Hence, a person's preferences today ut( 1) = -11 for all t. Suppose that 6 =
over her future delays in rewards are 1, but that f = .8 for a one-day delay:
different than her future preferences You are willing to suffer a given loss in
over those same delays, so that prefer- utility tomorrow for a gain in utility to-
ences are not time consistent. Formal day that is 80 percent as large.
models of such time-variant prefer- Suppose that April 14 has arrived and
ences have been developed.33 Edmund you are considering whether or not to
Phelps and Robert Pollak (1968) cap- work. You can experience a disutility of
ture the taste for immediate gratifica- -10 by working today, or experience a
tion with a simple two-parameter model discounted utility of .8.(-11) = -8.8 by
that slightly modifies exponential dis- delaying the work until tomorrow. You
counting. Let ut be the instantaneous will, therefore, delay work. Contrast
utility a person gets in period t. Then this with what your decision would be
her intertemporal preferences at time t, if, instead of choosing when to work on
Ut, can be represented by the following April 14, you are told by your boss that
utility function, where both f and 6 lie you must decide on February 1. Be-
between 0 and 1: cause from February 1 you discount
both dates by ,B, you will choose to work
For all t, T
10 hours on April 14 rather than 11
Ut(utput + 1,...,Ut) (8)t.ut + 3.X(8)t U. hours on April 15. From the February 1
t=t+ 1
point of view, you find procrastinating
The parameter 6 determines how
in April an undesirable thing. For the
"time-consistently patient" a person is,
exact same problem, your choice on
just as in exponential discounting. If
February 1 is different than your choice
3 = 1, then these preferences are simply
on April 14. Irrespective of its specific
exponential discounting. But for ,6 < 1,
prediction, exponential discounting
these preferences capture in a parsimo-
would predict that your choice would
nious way the type of time-inconsistent
be the same whether you made that
preferences so widely observed. To see
choice on February 1 or April 14. This
how these preferences capture the pref-
example seems well-calibrated: On
erence for immediate gratification, sup-
April 14, most of us are apt to put off
pose that you had a choice between do-
the work until April 15, even if it means
ing ten hours of an unpleasant task on
a little more work. Absent a substantive
32 These numbers are calculated from the data difference between the two dates, virtu-
presented by Kirby and Herrnstein (1995, p. 85- ally no one would choose the delay if
86). Other psychological research showing prefer-
ences are not time-consistent includes Shin-Ho asked on February 1.
Chung and Herrnstein (1967), George Ainslie To examine dynamic choice given
(1991), Ainslie and Herrnstein (1981), Thaler time-variant preferences given these
(1981), and Loewenstein and Prelec (1992).
33 For economics papers on time-inconsistent preferences, for each point in time, a
discounting, see, e.g., Robert Strotz (1955), person is modeled as a separate "agent"
Steven Goldman (1979, 1980), Schelling (1978), who chooses her current behavior to
Thaler and Hersh Shefrin (1981), David Laibson
(1994, 1997), and O'Donoghue and Rabin (1997a, maximize her current long-run prefer-
1997b). ences, whereas each of her future
selves, with her own preferences, will from which you cannot check out, not
choose her future behavior to maximize owning a television, contributing to a
her preferences. On one level, this idea "Christmas Club" from which you are
of multiple selves-that a single human not allowed to withdraw money until
does not have unified preferences that Christmas, or buying only small pack-
are stable over time-is a radical de- ages of enticing foods so that you won't
parture from the utility-maximization overeat when you get home. More sub-
framework. But because this conceptu- tly, you may try to control yourself
alization of intertemporal choice uses a through a variety of internal "rules"
familiar tool-dynamic game theory-it (e.g., never drink alcohol), even if you
is ready-made for adoption by econo- have no external mechanisms of self-
mists interested in improving the be- control.
havioral realism of our models. Attempts to control our own future
The behavior predicted by models of behavior indicate an awareness that we
time-variant preferences often differs may not behave as we would wish to be-
dramatically from the behavior pre- have. This raises the question of how
dicted by the exponential model. The aware people are of their time-inconsis-
most notorious examples are efforts at tency. You may have expectations about
self control: Because you may not like your propensity to misbehave, or you
the way you will behave in the future, may naively believe that your prefer-
you may scheme to manipulate your fu- ences in the future will match your cur-
ture options. Consider again the work rent preferences. If today you prefer
example. Instead of your boss telling not to overeat tomorrow, you may
you that you must choose on February 1 naively believe that you will feel the
when to work, suppose now she gives same way when facing an enticing bowl
you three options: You commit to do of ice cream tomorrow. If on February
the task on April 14; you commit to do 1 you prefer less work on April 14 to
the task on April 15; or you wait until more work on April 15, you may believe
April 14 and then choose on which day you'll feel the same way in April.
to do the task. Which would you Strotz (1955) labels people who are
choose? The advantage of waiting is fully aware of their future self-control
manifest: By not precluding either of problems as sophisticated, and people
your options, if there are any uncertain- who are fully unaware that they will
ties that may be resolved between now have a self-control problem as naive.
and April, the flexibility you have re- While some degree of sophistication is
tained may be valuable. Yet we some- implied by the existence of some of the
times engage in behavior precisely to self-commitment devices illustrated
restrict our own future flexibility. If above, it does appear that people un-
there were few uncertainties, you might derestimate the degree to which their
want to commit on February 1 to the future behavior will not match their
April 14 date. Given your current pref- current preferences over future behav-
erence to do the task earlier, you wish ior. This accords with the evidence dis-
to restrict your future self from procras- cussed earlier, that people often incor-
tinating. More generally, researchers rectly predict their own future
have explored many self-commitment preferences: As with predicting the ef-
devices we employ to limit our future fects of changes in reference points,
choices. Such self-commitment devices here too knowing your future prefer-
include alcohol clinics and fat farms ences means that you know your prefer-
ences won't accord with your current assumptions, or because it fails to bear
preferences. For example, people may sufficiently great burdens of proof, or
repeatedly not have the "will power" to because the implied behavior is unlikely
forego tempting foods or quit smoking to matter in the types of (market) set-
while predicting that tomorrow they tings that economists care about. Be-
will have this will power. While behav- cause behavioral research is so often as-
ioral evidence that calibrates the degree sessed in light of such arguments, it is
of sophistication seems sparse, Loewen- common when presenting psychological
stein (1996, pp. 281-82) reaches the findings to discuss broad methodologi-
conclusion that people may be naive in- cal objections and attempt to rebut
directly from psychological findings them.
such as the evidence of people mispre- I refrain from doing so. It is my
dicting changes in utility. strong impression that many of the ar-
Whether they are sophisticated or guments invoked against the reality or
naive, people's time-inconsistent pro- relevance of behavioral research derive
pensity for immediate gratification is from unfamiliarity with the details of
important in a variety of economic this research. Hence, my hope and
realms. As investigated by several re- guess is that as economists become
searchers (see, e.g., Thaler and Shefrin more familiar with this research, such
1981; and Laibson 1997), such prefer- arguments will dissipate. And as the ag-
ences may be important to savings be- gressive uncuriosity shown in the past
havior because the benefits of current toward behavioral research continues to
consumption are immediate, whereas diminish, we can look forward to focus-
the increased future consumption that ing entirely on its substance.
saving allows is delayed. Self-control While none of the broad-stroke argu-
problems are also clearly important in ments for inattention to psychological
the demand for addictive goods and research are compelling, obviously not
fatty foods. Similarly, the role of self- all psychological research will be both
control in purchasing decisions is well confirmed by field data and proven to
known among marketing experts be of great economic importance. In-
(Stephen Hoch and Loewenstein 1991). deed, abandoning the view that hy-
Naughty goods are sold in small pack- potheses departing from rationality,
ages because people tend to avoid large self-interest, or other habitual assump-
packages of such goods to prevent over- tions are methodologically illicit can
consumption. free us to evaluate these hypotheses
with the same rigorous standards that
5. Conclusion our discipline, at its best, applies else-
where. We can confront plausible hy-
Over the years, economists have prof- potheses about human behavior with
fered many reasons for downplaying the both healthy skepticism and genuine cu-
relevance of behavioral research chal- riosity, empirically test their validity,
lenging our habitual assumptions. and carefully draw out their economic
Claims abound that evidence inconsis- implications. And, as we apply these rig-
tent with our traditional model of hu- orous standards, we can then begin to
man behavior can be neglected because treat claims that (say) investors irration-
the evidence derives from observations ally infer too much from short-term
of people insufficiently motivated to be- performance, or that employees feel re-
have themselves according to economic sentful when mistreated, as presump-
tively plausible and presumptively rele- BRICKMAN, PHILIP; COATES, DAN AND JANOFF-
BULMAN, RONNIE. "Lottery Winners and Acci-
vant hypotheses worth keeping in mind dent Victims: Is Happiness Relative?" J. Person-
in our economic analysis. ality & Social Psychology, Aug. 1978, 36(8), pp.
917-27.
BRUNER, JEROME S. AND POTTER, MARY C. "In-
REFERENCES
terference in Visual Recognition," Science, Apr.
ADAMS, J. STACY. "Wage Inequities, Productivity 24, 1964, 144(3617), pp. 424-25.
and Work Quality," Ind. Relat., Oct. 1963, 3, CAMERER, COLIN F. "Individual Decision Mak-
pp. 9-16. ing, in Handbook of experimental economics.
AINSLIE, GEORGE W. "Derivation of 'Rational' Eds.: JOHN KAGEL AND ALVIN E. ROTH.
Economic Behavior from Hyperbolic Discount Princeton: Princeton U. Press, 1995, pp. 587-
Curves," Amer. Econ. Rev., May 1991, 81(2), 703.
pp. 334-40. CAMERER, COLIN ET AL. "Labor Supply of New
AINSLIE, GEORGE W. AND HERRNSTEIN, RICH- York City Cabdrivers: One Day at a Time,"
ARD J. "Preference Reversal and Delayed Rein- Quart. J.Econ.,
May 1997,112(2), pp. 407-41.
forcemnent,"Animal Learning & Behavior, Nov. CAMERER, COLIN F. AND THALER, RICHARD H.
1981, 9(4), pp. 476-82. "Ultilmatums, Dictators and Manners," J. Econ.
AKERLOF, GEORGE A. "Labor Contracts as Partial Perspectives, Spring 1995, 9(2), pp. 209-19.
Gift Exchange," Quart. J. Econ., Nov. 1982, CHAPMAN, LOREN J. AND CHAPMAN, JEAN P.
97(4), pp. 543-69. "Genesis of Popular But Erroneous Psychodiag-
. "Gift Exchange and Efficiency-Wage The- nostic Observations," J. Abnormal Psychology,
ory: Four Views," Amer. Econ. Rev., May 1984, 1967, 72(3), pp. 193-204.
74(2), pp. 79-83. . "Illusory Correlation as an Obstacle to the
AKERLOF, GEORGE A. AND YELLEN, JANET L. Use of Valid Psychodiagnostic Signs," J. Abnor-
"The Fair Wage-Effort Hypothesis and Unem- mal Psychology, 1969, 74(3), pp. 271-80.
ployment," Quart. J. Econ., May 1990, 105(2), - . "Test Results Are What You Think They
pp. 255-83. Are," Psychology Today, Nov. 1971, 5(6), pp.
ANDERSON, CRAIG A.; LEPPER, MARK R. AND 18-22.
Ross, LEE. "Perseverance of Social Theories: CHARNESS, GARY. "Attribution and Reciprocity
The Role of Explanation in the Persistence of in a Silnulated Labor Market: An Experimen-
Discredited Information,"J. Personality & So- tal Investigation." Mimeo, U. California at
cial Psychology, Dec. 1980, 39(1-6), pp. 1037- Berkeley, 1996.
49. CHRISTENSEN-SZALANSKI, JAY J. AND WILLHAM,
ANDREONI, JAMES. "Cooperation in Public-Goods CYNTHIA. "The Hindsight Bias: A Meta-Analy-
Experiments: Kindness or Confusion?" Amer. sis," Organizational Behavior & Human Deci-
Econ. Rev., Sept. 1995, 85(4), pp. 891-904. sion Processes, Feb. 1991, 48(1), pp. 147-68.
ANDREONI, JAMES AND MILLER, JOHN H. "Giv- CHUNG, SHIN-HO AND HERRNSTEIN, RICHARD J.
ing According to GARP: An Experimental Study "Choice and Delay of Reinforcement," J. Ex-
of Rationality and Altruismn." Working Paper perimental Analysis of Behavior, 1967, 10(l),
No. 9601. Social Systems Research Institute, U. pp. 67-74.
Wisconsin at Madison, 1996. CONLISK, JOHN. "Why Bounded Rationality?" J.
BAUMANN, ANDREA O.; DEBER, RAISA B. AND Econ. Lit., June 1996, 34(2), pp. 669-700.
THOMPSON, GAIL G. "Overconfidence Among CROSON, RACHEL T. A. "Expectations in Volun-
Physicians and Nurses: The 'Micro-Certainty, tary Contribution Mechanisms." Working Paper
Macro-Uncertainty' Phenomenon," Social Sci- No. 95-03-02, Department of Operations and
ence & Medicine, 1991, 32(2), pp. 167-74. Information Management, The Wharton
BENARTZI, SHLOMO AND THALER, RICHARD H. School, U. Penn., 1995.
"Myopic Loss Aversion and the Equity Pre- DARLEY, JOHN M. AND GROSS, PAGET H. "A Hy-
mium Puzzle," Quart. J. Econ., Feb. 1995, pothesis-Confirming Bias in Labeling Effects,"
110(1), pp. 73-92. J. Personality & Social Psychology, Jan. 1983,
BERSCHEID, ELLEN; BOYE, DAVID AND WAL- 44(1), pp. 20-33.
STER, ELAINE. "Retaliation as a Means of Re- DAWES, ROBYN M. AND THALER, RICHARD H.
storing Equity,"J. Personality & Social Psychol- "Anomalies: Cooperation," J. Econ. Perspec-
ogy, 1968, 10(4), pp. 370-76. tives, Summer 1988, 2(3), pp. 187-97.
BLOUNT, SALLY. "When Social Outcomes Aren't DIAMOND, PETER A. AND HAUSMAN, JERRY A.
Fair: The Effect of Causal Attributions on Pref- "Contingent Valuation: Is Some Number Better
erences," Organizational Behavior & Human than No Number?" J. Econ. Perspectives, Fall
Decision Processes, Aug. 1995, 63(2), pp. 131- 1994, 8(4), pp. 45-64.
44. DUESENBERRY, JAMES S. Income, saving and the
BOWMAN, DAVID; MINEHART, DEBORAH AND theory of consumer behavior. Cambridge: Har-
RABIN, MATTHEW. "Loss Aversion in a Con- vard U. Press, 1949.
sumnption-Savings Model." Milneo, U. Califor- EASTERLIN, RICHARD A. "Does Economic
nia at Berkeley, 1997. Growth Improve the Human Lot? Some Em-
. "Rational Choice and the Framing of De- & Social Psychology, Jan. 1995, 68(1), pp. 21-
cisions," J. Bus., Oct. 1986, 59(4, Part 2), pp. 35.
S251-78. WILSON, TIMOTHY D. AND SCHOOLER,
-. "Loss Aversion in Riskless Choice: A Ref- JONATHAN W. "Thinking Too Much: Introspec-
erence-Dependent Model," Quart. J. Econ., tion Can Reduce the Quality of Preferences and
Nov. 1991, 106(4), pp. 1039-61. Decisions," J. Personality & Social Psychology,
TVERSKY, AMOS AND THALER, RICHARD H. Feb. 1991, 60(2), pp. 181-92.
"Anomalies: Preference Reversals," J. Econ. YAARI, MAYA E. AND BAR-HILLEL, MENAHEM.
Perspectives, Spring 1990, 4(2), pp. 201-11. "On Dividing Justly," Social Choice Welfare,
WILSON, TIMOTHY D. AND LAFLEUR, SUZANNE 1984, 1(1), pp. 1-24.
J. "Knowing What You'll Do: Effects of Analyz- YATES, J. FRANK. Judgment and decision making.
ing Reasons on Self-Prediction," J. Personality Englewood Cliffs, NJ: Prentice-Hall, 1990.