Decisions from Experience and the Effect of Rare Events in Risky Choice

Author(s): Ralph Hertwig, Greg Barron, Elke U. Weber, Ido Erev

Reviewed work(s):
Source: Psychological Science, Vol. 15, No. 8 (Aug., 2004), pp. 534-539
Published by: Sage Publications, Inc. on behalf of the Association for Psychological Science
Stable URL: http://www.jstor.org/stable/40064012 .
Accessed: 17/01/2012 16:27

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.

Sage Publications, Inc. and Association for Psychological Science are collaborating with JSTOR to digitize,
preserve and extend access to Psychological Science.

http://www.jstor.org
 PSYCHOLOGICALSCIENCE
Research Article
Decisions From
Experience
the Effect of Rare
Risky Choice
Ralph Hertwig,1 Greg Barron,2 Elke U. Weber,3 and Ido Erev4
University of Basel, Basel, Switzerland; Harvard University; Columbia University;and Technion,Haifa, Israel
ABSTRACT - When people have access to information sources
such as newspaper weather forecasts, drug-package inserts,
and mutual-fund brochures, all of which provide convenient
descriptions of risky prospects, they can make decisions from
description. When people must decide whether to back up their
computer's hard drive, cross a busy street, or go out on a date,
however, they typically do not have any summary description of
the possible outcomes or their likelihoods. For such decisions,
people can call only on their own encounters with such pros-
pects, making decisions from experience. Decisions from expe-
rience and decisions from description can lead to dramatically
different choice behavior. In the case of decisions from de-
scription, people make choices as if they overweight the prob-
ability of rare events, as described by prospect theory. Wefound
that in the case of decisions from experience, in contrast, people
make choices as if they underweight the probability of rare
events, and we explored the impact of two possible causes of this
underweighting- reliance on relatively small samples of infor-
mation and overweighting of recently sampled information. We
conclude with a call for two different theories of risky choice.
Whyare doctorsand patientsoften at odds with one another?Rushed
office visits, poor interpersonalskills on the part of doctors,and in-
transigenceon the partof patientsmayall contributeto disagreement
and misunderstandings.Here we focus on another factor that can
strainthe relationship:Patients'and doctors'decisionsare oftenbased
on informationthat, though equivalent in content, comes from dif-
ferent sources. Consider,for example, the decision whetherto vac-
cinate a child against diphtheria, tetanus, and pertussis (DTaP).
Parents who research the side effects of the DTaP vaccine on the
NationalImmunizationProgramWebsite will find that up to 1 child
out of 1,000 will develop high fever and about 1 child out of 14,000
will experienceseizuresas a resultof immunization.Althoughdoctors
have these same statistics at their disposal, they also have access to
Address correspondence to Ralph Hertwig, University of Basel,
Department of Psychology, Missionsstrasse 60/62, 4055 Basel,
Switzerland.
534 Copyright 2004 AmericanPsychologicalSociety
and
Events in
informationnot easily available to parents- namely,personalexpe-
rience, gatheredacross manypatients,that vaccinationrarelyresults
in side effects;few doctorshave encounteredone of the unusualcases
in which high fever or seizuresfollow vaccination.If the importance
assigned to rare events differsas a functionof how one learns about
their likelihood,then doctorsand patientsmightwell disagreeabout
whethervaccinationis advised.
Does the impactof rare events on risky decisions depend on how
knowledgeabout their likelihood was obtained?Before addressing
this question- the focus of this article- we introducethe dominant
decision-theoreticframework,according to which people evaluate
possible outcomesof their decisions in termsof beliefs and values.
EXPECTATIONS: FROM PROBABILITIES TO
DECISION WEIGHTS
The fundamentalprinciplesof probabilitywerefirstformulatedin the
mid-1600s in an exchangeof letters betweenthe Frenchmathemati-
cians Blaise Pascal and PierreFermat.Dealingwithvariousgambling
problems, their epistolary discussion gave rise to the concept of
mathematicalexpectation,which at the time was believed to capture
the basis for rationalchoice. Choosingrationallymeantchoosingthe
optionwith the higherexpectedvalue (EV),which in modernnotation
is definedas
EV = IpM,
where /?, and %,are the probabilityand the amount of money,
respectively,associatedwith each possible outcome(i = 1, . . . , n) of
that option.
Despite its elegance, the definitionof rationaldecision makingas
maximizationof expected value ran into difficulty when Nicholas
Bernoulli,a SwissmathematicianfromBasel, posedthe St. Petersburg
paradox. The paradox is that, contraryto expected-value theory,
people are willing to pay only relatively small amountsto take a
gamble that is assumed to have an infinite expected value (but see
Jorland's,1987, analysis). To reconcile the theorywith people'sbe-
havior,Nicholas'scousin Daniel Bernoulliproposedretainingthe core
of expected-valuetheorywhile replacingobjectivemonetaryoutcomes
withthe notionof subjectiveutility.Specifically,he suggestedthatthe
Volume15- Number8
 R. Hertwiget al.

utility of money increases nonlinearlywith its amount,rising at a We propose that the answer lies in how decision makers learn about
decreasing rate as absolute monetaryvalue increases. In modern the likelihood with which rare (and other) events can be expected to
terms,Bernoulli'sconcept of expectedutility (EU) is definedas occur. Research on risky choice typically provides respondents with
a summary description of each option, for example:
EU = Zpmixi),

where u(xi) is a positive but deceleratingfunction of the monetary
amount#;.
After von Neumannand Morgenstern(1947) put expected-utility
theoryon axiomaticgrounds,it quickly became the most influential
theoryof individualchoice behavior.Before long, however,experi-
mentsrevealedsystematicviolationsof the theory'saxioms.Perhaps
the most prominentviolationis the Allais paradox(Allais, 1953), in
which decision makers choosing between risky prospects do not
conformto the independenceaxiom, according to which outcomes
commonto all prospects(and with knownprobabilities)should have
no influenceon the decision. In responseto this and otheranomalies,
various modificationsof the expected-utilityframeworkhave been
proposedin recent decades. These modificationsneverthelessretain
the theory'skey idea, namely, the Bernoullianmultiplication.The
most influentialmodification,prospecttheory,assumes that the value
(a formof utility)of each outcomeis multipliedby a decision weight
(Kahneman& Tversky,1979).
In the presentcontext,the notionof decisionweightis key.Prospect
theory'sdecision weights are not assumed to reflect any explicit
judgment of the subjective probabilityof outcomes that decision
makerscould be asked to provide.Instead,weightsare inferredfrom
choices and providea measureof the impactthatan outcomehas on a
decision. Prospecttheoryposits that people choose as if small-prob-
ability events receive more weight than they deserve accordingto
their objectiveprobabilitiesof occurrenceand as if large-probability
events receive less weightthan they deserve. Prospecttheory'sdeci-
sion-weightfunctionplots weights (inferredfromchoices) that range
from0 to 1 againstobjectiveprobabilities:Points above the 45 di-
agonalsignal overweighting;that is, they representweights that ex-
ceed the outcomes'objectiveprobabilityof occurrence.Points below
the 45 diagonalsignal the oppositepattern.Prospecttheory'sover-
weightingof rareevents has been crucial in accountingfor manyvi-
olations of expected-utility theory, including the classic Allais
paradox(Camerer,2000; Kahneman& Tversky, 1979; Tversky &
Kahneman,1992).
DECISIONSFROMDESCRIPTIONAND DECISIONS
FROMEXPERIENCE
In light of extensive empirical evidence consistent with prospect
theory's weighting function, the results of several recent studies
(Barkan, Zohar, & Erev, 1998; Barron & Erev, 2003; Erev, 1998;
Weber, Shafir, & Blais, 2004) came as a surprise. Observed choices
indicated not overweighting of small-probability outcomes (hence-
forth, rare events), but rather the opposite: People made choices as if
they underweighted rare events; that is, rare events received less
weight than their objective probability of occurrence warranted. What
makes people choose as if they underweight rare events in some
studies and as if they overweight them in other studies?
^ote that we referto the weightingof events in the same as if sense as in
prospecttheory.Moreover,we assumethatthe weightingof outcomeswithlarge
probabilitiesis the mirrorimage of the weightingpatternfor rareevents.
A: Get $4 with probability .8, $0 otherwise.
or
B: Get $3 for sure.
The outcomes of each option and their probabilities are provided, and
the information is conveyed visually (e.g., using a pie chart or fre-
quency distribution) or numerically. Moreover, respondents are often
required to make only one choice per problem and rarely receive
feedback. We refer to such decisions as decisions from description.
Studies of human risky choice almost exclusively examine decisions
from description. In a recent meta-analysis of all studies involving
decisions between a two-outcome risky prospect and a sure thing (with
equal expected value), Weber et al. (2004) found that all 226 choice
situations called for decisions from description.
Outside the laboratory, however, people often must make choices
without a description of possible choice outcomes, let alone their
probabilities. Because people can rely only on personal experience
under such conditions, we refer to these as decisions from experience.
Only a few studies have investigated decisions from experience in
humans. In one (Barron & Erev, 2003), decision makers acquired in-
formation about outcomes and probabilities by making repeated choices
(initially under ignorance) and receiving feedback about the outcomes
of their choices. Specifically, decision makers were repeatedly asked to
choose between two buttons projected on the computer screen. Each
button selection initiated a random draw from the initially unknown
payoff distribution associated with that button. Given the payoff dis-
tributions in the aforementioned example, selecting option A could lead
to outcomes of $4 or $0. Selecting option B, in contrast, would always
result in an outcome of $3. By choosing between the buttons (and
associated payoff distributions) and experiencing the contingency
between choices and outcomes over repeated trials, respondents in this
feedback paradigm gradually acquired knowledge about the two payoff
distributions. Across trials, 63% of respondents chose option A. In
contrast, only 20% of respondents in Kahneman and Tversky's (1979)
study, which provided a description of the two options, chose option A.
The differences between the two studies are consistent with under-
weighting the rare event (the SO outcome of option A) in decisions from
experience and overweighting it in decisions from description: Un-
derweighting the rare event increases the attractiveness of option A,
whereas overweighting decreases A's attractiveness. What causes un-
derweighting of rare events in decisions from experience?
UNDERWEIGHTING: THE RESULT OF DIRECT
EXPERIENCE OR REPEATED DECISIONS?
In Kahneman and Tversky's (1979) study, respondents made a single
choice in a problem that presented outcomes and probabilities. In
^here are differentways to definea rareevent. Given that our study is one
of the first attemptsto systematicallystudy decisions from experience, we
somewhatarbitrarilydefinedrareevents as those with a probabilityof .20 or
less. Futurestudies mayfine-tuneor replacethis definitionwithone according
to which "the rarerthe event, the greaterthe underweighting."

15- Number
Volume 8 535
 Decisions From Experience

contrast,Barronand Erev's(2003) respondentsmade many choices
and discoveredoutcomesand probabilitiesonly fromfeedback.Both
the informationmode(symbolicdescriptionvs. directexperience)and
the numberof decisions (single vs. repeated)have been demonstrated
to affect people's judgments and decisions. For example, repeated
decisions (withoutcomefeedback)can eliminatepreferencereversals
(Chu & Chu, 1990; see Hertwig& Ortmann,2001), and direct ex-
perience of base rates can strongly improve Bayesian reasoning
(Koehler,1996; Hertwig& Ortmann,2001). Forinstance,doctorsuse
base rates acquired through personal experience in a normative
fashion,but do not use numericallydescribedbase rates normatively
(Weber,Bockenholt,Hilton, & Wallace,1993).
Whichpropertiesof decisions fromexperienceare thus responsible
for the underweightingof rare events? To disentanglethe two candi-
date properties - direct experience and repeateddecisions- we ex-
ploited the fact that personalknowledgeaboutrisky prospectscan be
acquired in different ways. In the samplingparadigmof Weberet al.
(2004), respondents experienced outcomes and their frequencies
without making repeated consequential decisions (unlike in the
feedback paradigm).If underweightingof rare events originates in
repeateddecisions, then the samplingparadigmshould not produce
underweightingof rare events. If underweightingof rare events orig-
inates in direct experience of the occurrence(or nonoccurrence)of
rare events, however,then the sampling paradigmshould result in
underweighting.
METHOD
One hundredstudentsat the Technion(Haifa,Israel)were presented
with the six decision problems displayed in Table 1 (taken from
Barron& Erev's,2003, feedback design). All six problemspresent
options that differ with respect to expected value; four of them offer
positive and two offer negative prospects.Half the participants,the
descriptiongroup, saw the problemsdescribed (as in Table 1) on a
computerscreen.The otherhalf, the experiencegroup,saw twobuttons
on the computerscreen and were told thateach buttonwas associated
with a payoffdistribution.In each group,25 respondentswere pre-
sented with three of the six problems, and the remaining25 re-
spondentswerepresentedwiththe otherthreeproblems.Clickingon a
given buttonelicited the samplingof an outcome(withreplacement)
from its distribution.Respondentscould sample in whateverorder
they desired,and howeveroftenthey wished;they wereencouragedto
sample until they felt confidentenoughto decide fromwhich box to
drawfor a real payoff.Once they had stoppedsamplingand indicated
their preferredoption, they turnedto the next problem.Finally,re-
spondentsplayed out the selected optionsand received real payoffs.
In both groups,participantsreceived a $4.50 fee for showingup
and 20 for each point won (e.g., outcomeof 32 in Problem6 was
equivalentto 640).
RESULTS
Decisions from experience clearly differedfrom decisions fromde-
scription.As shown in Table 1, the percentageof respondentswho
chose option H (i.e., the option with the higher expected value, as
calculated by probabilitytimes monetaryvalue) in each problem
differedmarkedlybetweenthe twogroups.Forinstance,in Problem1,
88% of participantsin the experience group selected option H,
whereasonly 36% of participantsin the descriptiongroupselected
this option(these resultsare similarto Kahneman& Tversky's,1979,
findingof 20% H choices).Acrossall problems,the average(absolute)
differencebetweenthe percentageof respondentschoosingoptionH
in the experience and descriptiongroupswas 36 percentagepoints,
and between-groupsdifferenceswere statisticallysignificantfor all
problemsexcept Problem2. Equallyimportant,the directionof each
difference was consistent with that predicted by the assumption
that rareevents are underweightedin decisions fromexperience(see
Table1). Forexample,fewerrespondentsin the experiencegroupthan
in the descriptiongroupwerepredictedto select optionH in Problem
2, and the oppositepredictionwas madefor Problem4.
To summarize,respondents in the experience and description
groupsfaced structurallyidentical problems.Yet their choices were

TABLE 1
Summary of the Decision Problems and Results

Expected
Options8 value Percentagechoosing H
Decision Description Experience Rare Prediction Difference
problem H L H L group group event for H choices betweengroups0
1 4^8 3,1.0 3.2 3 36 88 0, .2 Higher +52 (z = 3.79, p = .000)
2 4^2 3, .25 0.8 0.75 64 44 4, .2 Lower -20 (z = 1.42, p = .176)
3 -3,1.0 -32, .1 -3 -3.2 64 28 -32, .1 Lower -36 (z = 2.55, p = .005)
4 -3,1.0 -4, .8 -3 -3.2 28 56 0, .2 Higher +28 (z = 2.01, p = .022)
5 32, .1 3,1.0 3.2 3 48 20 32, .1 Lower -28 (z = 2.09, p = .018)
6 32, .025 3, .25 0.8 0.75 64 12 32, .025 Lower -52 (z = 3.79, p = .000)

Note. Underlining indicates the options including rare events. H = option with the higher expected value; L = option with the lower expected value.
aFor each option, only one outcome is given, followed by its probability; the second outcome, which is not stated, was 0 and occurred with a
probability complementary to the stated one. For instance, the outcomes of the H option in Problem 2 were 4 with a probability of .2 and 0 with a
probability of .8; the outcomes of the L option in Problem 2 were 3 with a probability of .25 and 0 with a probability of .75. bThe entries in this
column indicate whether the percentage of respondents choosing the H option was expected to be higher or lower in the experience group than in
the description group, assuming underweighting of the rare event in the experience group. cThis column shows the percentage of H choices in the
experience group minus the percentage of H choices in the description group, along with the z statistic testing whether the difference between the
two sample proportions (relative frequencies) is significantly different from zero.

536 Volume15- Number8

 R. Hertwig et al.

Fig. 1 . Median number of draws in the experience group for each of the six decision prob-
lems (see Table 1). Results are shown separately for the total number of draws and for draws
of the options with the higher (H) and lower (L) expected values (i.e., probability times
monetary value). For each option, only one outcome is given, followed by its probability; the
second outcome, which is not stated, was 0 and occurred with a probability complementaryto
the stated one. The standard deviations for the total number of draws per problem were 4.1,
21.4, 37.8, 20.1, 10.4, and 14.9 for Problems 1 through 6, respectively.

dramaticallydifferent.Differencesin choices wereconsistentwiththe Weberet al. (2004). Across problems,the numberof drawsforoption
assumptionthat in decisions from experience, rare events had less H was roughlyequal to the numberof drawsfor optionL.
impactthanthey deservedon the basis of objectiveprobability(andin Althoughone may speculateaboutthe reasonsfor people'slimited
decisions from description,rare events had more impact than they search effort,3it exacts an obviousprice:The smallerthe numberof
deserved).Moreover,the underweightingof rare events in decisions drawsfrom a payoff distribution,the largerthe probabilitythat the
fromexperienceappearsto be robustacross experimentalparadigms: respondentwill not come acrossthe rareevent and, consequently,will
The choice percentagesin the experience groupin the presentstudy remainignorantof its existence. For illustrationconsiderProblem5,
were close to the percentagesBarronand Erev (2003) observedpre- in which the medianrespondentsampledseven cardsfromthe payoff
viously(r = .93, p < .01). The similarpatternof results suggests that
distributionthat offered "32" with a probabilityof .1 (and "0" oth-
it is indeed directexperienceof outcomesand their likelihoods- and erwise).Therefore,mostrespondents(18 out of 25) neverencountered
not repeatedchoices- that accounts for the underweightingof rare "32."In additionto increasingthe chanceof not encounteringthe rare
events in decisions fromexperience. But how does direct experience event at all, small samplescause the rareevent to be encounteredless
lead to underweighting? frequentlythan expected (given its objective likelihood).This is be-
cause the binomialdistributionfor the numberof times a particular
outcomewill be observedin n independenttrialswhen/?is small(i.e.,
DECISIONS FROM EXPERIENCE: LIMITED the event is rare)and n is small (i.e., few draws)is skewed. In such
INFORMATION SEARCH skeweddistributions,one is morelikely to encounterthe rareevent in
small samples less frequentlythan expected (i.e., np) than to en-
Because decisions fromexperience depend on the sampled informa- counterit morefrequentlythanexpected. Forexample,let us assume
tion, any accountof how such decisions are made ought to consider that each of 1,000 people draws20 times froma distributionin which
how people search for informationand how the results of the search
affect subsequentdecisions (see also Fiedler, 2000; Kareev,2000).
Figure 1 displays the median numberof draws per problemin the *0ne explanation involves short-term memory limits that provide a natural
For our the most observationis stopping rule for information search (Kareev, 2000). More than half of re-
experiencegroup. purposes, important
spondents sampled one choice option exclusively before they switched to the
thatthe totalnumberof drawsper problemwas relativelysmall, witha other one, and the median number of draws from each option was around
medianof 15- a result close to the medianof 17 drawsobservedby seven - a number often associated with the capacity of short-term memory.

Volume15- Number8 537

 Decisions From Experience

TABLE 2
Percentage of Respondents in the Experience Group Who Selected the Option Involving the Rare Event as
a Function of How Often the Rare Event Was Encountered During Sampling

Percentagechoosingoptionwith rareevent
~ . a Encountered Encounteredas
Decision 1 less frequently frequentlyas or more
problem H L Rare event than expected frequentlythan expected
1 4^8 3, 1.0 0, .2 Negative 88 (21/24) - (1/1)
2 4^2 3, .25 4, .2 Positive 33 (6/18) 71 (5/7)
3 -3, 1.0 -32, .1 -32, .1 Negative 100 (12/12) 46 (6/13)
4 -3,1.0 -4, .8 0, .2 Positive 44(11/25) -(0/0)
5 32, .1 3, 1.0 32, .1 Positive 5 (1/19) 67 (4/6)
6 32, .025 3, .25 32, .025 Positive 5 (1/19) 33 (2/6)

Note. Underliningindicatesthe optionsincludingrare events. H = optionwith the higherexpectedvalue;L = optionwiththe lower

expectedvalue.aForeach option, only one outcomeis given, followedby its probability;the secondoutcome,whichis not stated,
was 0 and occurredwith a probabilitycomplementaryto the statedone. For instance, the outcomesof the H optionin Problem2
were 4 with a probabilityof .2 and 0 with a probabilityof .8; the outcomesof the L optionin Problem2 were 3 with a probability
of .25 and 0 with a probabilityof .75. bRelativefrequenciesare given in parentheses.

the criticalevent occurswith a probabilityof .1. Of the 1,000 people,
285 will samplethe critical event 2 times and thus could estimateits
probabilityaccurately,if asked explicitly. Another323 of them will
encounterthe event 3, 4, 5, ... , or 20 times and thus would most
likely overestimateits probability.But 392 people- almosttwo fifths
of the total- will not sample the event at all or will sample it only 1
time, and thus would most likely underestimatep.
The outcomesof samplingin the presentstudywereconsistentwith
the skewedbinomialdistributionfor small samples.Averagingacross
all problemsin the experiencegroupshowedthat 78% of respondents
sampled the critical rare event less frequentlythan expected (np),
whereas22% sampledit as frequentlyas expectedor morefrequently
than expected. Moreover,the experiencedfrequencyof the critical
event had a clear impact on choices. Across all problems,when the
rareevent representeda positive event (e.g., "32" in Problem5) and
was encounteredless frequentlythan its expected numberof occur-
rences, people selected the option involvingthat event 23% of the
time. When it was encounteredas frequentlyas expected or more
frequentlythan expected, people selected the option involving the
rare, positive event 58% of the time. Similarly,when the rare event
representeda negativeevent (e.g., "0" in Problem1), it was selected
92% of the time when it was encounteredless frequentlythan its
expected numberoccurrences,and only 50% of the time when it was
encountered as frequently as or more frequently than expected.
Table2 shows that the patternis similarfor the individualproblems.
(Note that the pattern cannot be tested in Problems 1 and 4, as
everybodyor almost everybodyencounteredthe rare event in these
problemsless frequentlythan expected.)
Reliance on small samplesof experiencenot only plays a key role
in decisions fromexperiencebut also contributesto perceptionof the
worldas less variablethan it actuallyis. In fact, underestimatingthe
varianceof a populationis equivalentto underweightinga rareevent.
For instance, in Problem6, about two thirds of the respondentsen-
countereda uniformsequence of one "0"afteranotherwhen drawing
fromthe payoffdistributioninvolvingthe rareevent. If they had used
sample variability,or the lack thereof,to estimate populationvaria-
bility (withoutcorrectingfor sample size), then they would have un-
derestimatedthe true variabilityin this payoffdistribution.Kareev,
Arnon,and Horwitz-Zeliger (2002) discoveredthatpeopleindeedtend
to misperceivevariabilityand that using experiencesamplesof lim-
ited size (withthe size often being relatedto the capacityof working
memory)appearsto cause variabilityto be underestimated.
DECISIONS FROM EXPERIENCE: RECENCY EFFECTS
Although small samples have the result that decision makers
explicitly and hence presumablyalso implicitly underestimatethe
probabilityof rare events, underweightingof rare events is likely to
emerge in decisions fromexperienceeven if people can provideac-
curateexplicit estimatesof the probabilities.Hereis why.In decisions
fromexperience,respondentsneed to updatetheir impressionof the
options' attractivenessby combiningnewly sampled outcomeswith
their knowledgefrompreviousdraws.Such updatingcan give rise to
recencyeffects (e.g., Hogarth& Einhorn,1992), that is, to judgments
in which recently sampled outcomes receive greater weight than
earlier sampledones. Note that a recency effect wouldresult in the
underweightingof rare events even in large samples:Owingto their
rarity,they are less likely than morecommonevents to have occurred
recentlyand thus less likely to affectthe decision. By the same logic,
commonevents will tend to be overweightedbecause they are more
likely to have occurredrecently.
To examine whetherrecency affecteddecisions in the experience
group,we split the sequenceof drawsfromeach optionintotwohalves
foreach problemandeach respondent.Thenwe computedthe options'
average payoffs obtained for the first and second halves of the
samples, predictedeach person'schoice on the basis of the payoffs,
and analyzed how many of the actual choices coincided with the
predicted choices (for the first and second halves of the samples,
separately).The second half of samplesclearlyhad greaterpredictive
powerthan the firsthalf. Whereasthe firsthalf predicted,on average,
59% of the final choices, the second half predicted 75% of the
choices, (49)= - 3.1, p = .003, two-tailed. In other words, rare
events have less impactthan they deserve not only because decision
makershave not encounteredthem, or have encounteredthem less
frequentlythanexpected,but also becausethey have not encountered
them recently.

538 Volume15- Number8

 R. Hertwiget al.
Elsewhere(Hertwig,Barron,Weber,& Erev, in press), we have
proposedand tested a modelthat successfullycapturesthe effects of
both recency and sample size on risky choice by specifyinghow old
and new informationis integrated.
DECISIONS FROM EXPERIENCE IN HUMANS AND OTHER
ANIMALS
Not only probabilitiesand outcomesbut many kinds of information
can be learnedthroughexperienceor description.Base rates,distribu-
tionalinformation,and degreesof causal strengthare a few examples.
Thus, one might expect the way in which informationis learned to
influencecognitiveprocessesin manydomains.Indeed,in researchon
Bayesianreasoning,for instance,there is evidence that performance
dependson whetherbase rates are directlyexperiencedor symboli-
cally described(see, e.g., Koehler,1996, and Weberet al., 1993).
Because animalsdo not share humans'ability to process symbolic
representationsof risky prospects, all their decisions (e.g., about
where to forage)are decisions fromexperience. Weberet al. (2004)
reportedsome striking similarities in the behaviorof humans and
lower animalswhen humansare placed in situationsin which they,
too, must make decisions from experience. Human decisions from
description,however,differ markedlyfrom human decisions from
experience,provokingthese authors- and us- to call for two differ-
ent theoriesof riskychoice. In a studyof foragingdecisions made by
bees, Real (1991) observed that "bumblebeesunderperceiverare
events and overperceivecommonevents" (p. 985), when the events
are instancesof food.Toexplainwhybees' "probabilitybias"diverges
fromthat observedin humans,Real cited, amongother factors,the
factthatbees'samplesfrompayoffdistributionsare truncatedbecause
of memoryconstraints.Although humans and bumblebees do not
sharea recentevolutionaryhistory,ourresultssuggestthatthere is no
contradictionbetweenthe decisions of bumblebeesand those of hu-
manswhen humansmust also rely on experience.
Acknowledgments- We are gratefulto EduardBrandstatter,Valerie
Chase,EricJohnson,YaakovKareev,BarbaraSpellman,and Masanori
Takezawafor helpfulcomments.Wealso thankAnitaToddfor editing
the manuscript,the GermanResearch Foundation(Deutsche For-
schungsgemeinschaft) fora grantto the firstauthor(He 2768/6-1), and
the Max WertheimerMinerva Center for technical and financial
support.Preparationof the manuscriptwas facilitatedby a fellowship
fromthe Institutefor AdvancedStudyin Berlin to the thirdauthor.
REFERENCES
Allais, M. (1953). Le comportementde l'hommerationneldevant le risque,
critiquedes postulatset axiomesde l'ecole Americaine.Econometrica,
21, 503-546.
15- Number
Volume 8 539
Barkan,R., Zohar,D., & Erev,I. (1998). Accidentsand decision makingunder
uncertainty:A comparisonof four models. OrganizationalBehaviorand
HumanDecisionProcesses,74, 118-144.
Barron,G., & Erev,I. (2003). Smallfeedback-baseddecisions and theirlimited
correspondenceto description-baseddecisions. Journal of Behavioral
DecisionMaking,16, 215-233.
Camerer,C.F.(2000). Prospecttheoryin the wild: Evidence fromthe field. In
D. Kahneman& A. Tversky(Eds.),Choices,values,andframes (pp. 288-
300). Cambridge,England:CambridgeUniversityPress.
Chu, Y.P., & Chu, R.L. (1990). The subsidence of preference reversals in
simplifiedand marketlikeexperimentalsettings:A note. AmericanEco-
nomicReview,80, 902-911.
Erev, I. (1998). Signal detection by humanobservers:A cutoff reinforcement
learning model of categorizationdecisions under uncertainty.Psycho-
logical Review,105, 280-298.
Fiedler, K. (2000). Bewareof samples! A cognitive-ecologicalsamplingap-
proachto judgmentbiases. PsychologicalReview,107, 659-676.
Hertwig,R., Barron,G., Weber,E.U., & Erev, I. (in press). The role of in-
formationsampling in risky choice. In K. Fiedler & P. Juslin (Eds.),
Informationsamplingas a key to understandingadaptivecognition.New
York:CambridgeUniversityPress.
Hertwig,R., & Ortmann,A. (2001). Experimentalpractices in economics:A
methodologicalchallenge for psychologists?Behavioraland Brain Sci-
ences,24, 383-451.
Hogarth,R.M., & Einhorn,H.J. (1992). Ordereffects in belief updating:The
belief-adjustmentmodel. CognitivePsychology,24, 1-55.
Jorland,G. (1987). The Saint Petersburgparadox1713-1937. In L. Krueger,
L.J.Daston,& M. Heidelberger(Eds.),Theprobabilisticrevolution:Vol.I.
Ideas in history(pp. 157-190). Cambridge,MA:MITPress.
Kahneman,D., & Tversky,A. (1979). Prospecttheory:An analysisof decision
underrisk. Econometrica,47, 263-291.
Kareev,Y. (2000). Seven (indeed, plus or minus two) and the detection of
correlations.PsychologicalReview,107, 397^02.
Kareev,Y., Arnon,S., & Horwitz-Zeliger, R. (2002). On the misperceptionof
variability.Journalof ExperimentalPsychology:General,131, 287-297.
Koehler,J.J. (1996). The base rate fallacy reconsidered:Descriptive,norma-
tive, and methodologicalchallenges.Behavioraland Brain Sciences,19,
1-53.
Real, L.A. (1991). Animal choice behavior and the evolution of cognitive
architecture.Science,253, 980-986.
Tversky,A., & Kahneman,D. (1992). Advancesin prospecttheory:Cumula-
tive representationof uncertainty.Journal of Risk and Uncertainty,5,
297-323.
von Neuman,J., & Morgenstern,O. (1947). Theoryof games and economic
behavior(2nd ed.). Princeton,NJ: PrincetonUniversityPress.
Weber,E.U., Bockenholt,U., Hilton,D.J., & Wallace,B. (1993). Determinants
of diagnostichypothesisgeneration:Effects of information,base rates,
and experience.Journalof ExperimentalPsychology:Learning,Memory,
and Cognition,19, 1151-1164.
Weber,E.U., Shafir,S., & Blais, A.-R. (2004). Predictingrisk sensitivity in
humansand lower animals:Risk as varianceor coefficientof variation.
PsychologicalReview,111, 430-^45.
(Received 6/24/02; Revisionaccepted 11/20/03)