chapter 12 a model of heuristic judgmenta model of heuristic judgment 269 in an optimal regression...

P1 : IKB-IRK/KAB P2 : IKB-IYP/KAB QC: IYP/JZO T1 : KOD0521824176c1 2 .xml CB798B/Holyoak 0 521 82417 6 February 2 , 2005 0:41

C H A P T E R 1 2

A Model of Heuristic Judgment

Daniel KahnemanShane Frederick

The program of research now known as theheuristics and biases approach began with astudy of the statistical intuitions of experts,who were found to be excessively confi-dent in the replicability of results from smallsamples (Tversky & Kahneman, 1971 ). Thepersistence of such systematic errors in theintuitions of experts implied that their intu-itive judgments may be governed by funda-mentally different processes than the slower,more deliberate computations they had beentrained to execute.

From its earliest days, the heuristics andbiases program was guided by the idea thatintuitive judgments occupy a position – per-haps corresponding to evolutionary history –between the automatic parallel operationsof perception and the controlled serial op-erations of reasoning. Intuitive judgmentswere viewed as an extension of percep-tion to judgment objects that are not cur-rently present, including mental represen-tations that are evoked by language. Themental representations on which intuitivejudgments operate are similar to percepts.Indeed, the distinction between perceptionand judgment is often blurry: The perception

of a stranger as menacing entails a predictionof future harm.

The ancient idea that cognitive processescan be partitioned into two main families –traditionally called intuition and reason –is now widely embraced under the generallabel of dual-process theories (Chaiken &Trope, 1999; Evans and Over, 1996; Ham-mond, 1996; Sloman, 1996, 2002 ; see Evans,Chap. 8). Dual-process models come inmany flavors, but all distinguish cognitiveoperations that are quick and associativefrom others that are slow and governed byrules (Gilbert, 1999).

To represent intuitive and deliberate rea-soning, we borrow the terms “system 1 ” and“system 2” from Stanovich and West (2002).Although suggesting two autonomous ho-munculi, such a meaning is not intended.We use the term “system” only as a label forcollections of cognitive processes that canbe distinguished by their speed, their con-trollability, and the contents on which theyoperate. In the particular dual-process modelwe assume, system 1 quickly proposes intu-itive answers to judgment problems as theyarise, and system 2 monitors the quality of

2 67


2 68 the cambridge handbook of thinking and reasoning

these proposals, which it may endorse, cor-rect, or override. The judgments that areeventually expressed are called intuitive ifthey retain the hypothesized initial proposalwith little modification.

The effect of concurrent cognitive tasksprovides the most useful indication ofwhether a given mental process belongs tosystem 1 or system 2 . Because the over-all capacity for mental effort is limited, ef-fortful processes tend to disrupt each other,whereas effortless processes neither causenor suffer much interference when com-bined with other tasks (Kahneman, 1973 ;Pashler, 1998). It is by this criterion that weassign the monitoring function to system 2 :People who are occupied by a demandingmental activity (e.g., attempting to hold inmind several digits) are much more likelyto respond to another task by blurting outwhatever comes to mind (Gilbert, 1989). Bythe same criterion, the acquisition of highlyskilled performances – whether perceptualor motor – involves the transformation of anactivity from effortful (system 2) to effort-less (system 1 ). The proverbial chess masterwho strolls past a game and quips, “Whitemates in three” is performing intuitively(Simon & Chase, 1973).

Our views about the two systems aresimilar to the “correction model” proposedby Gilbert (1989, 1991 ) and to other dual-process models (Epstein, 1994 ; Hammond,1996; Sloman, 1996; see also Shweder,1977). We assume system 1 and system 2

can be active concurrently, that automaticand controlled cognitive operations competefor the control of overt responses, and thatdeliberate judgments are likely to remainanchored on initial impressions. We alsoassume that the contribution of the twosystems in determining stated judgmentsdepends on both task features and individ-ual characteristics, including the time avail-able for deliberation (Finucane et al., 2000),mood (Bless et al., 1996; Isen, Nygren, &Ashby, 1988), intelligence (Stanovich &West, 2002), cognitive impulsiveness (Fred-erick, 2004), and exposure to statisticalthinking (Agnoli, 1991 ; Agnoli & Krantz,1989; Nisbett et al., 1983).

In the context of a dual-system view,errors of intuitive judgment raise twoquestions: “What features of system 1 cre-ated the error?” and “Why was the error notdetected and corrected by system 2?” (cf.Kahneman & Tversky, 1982). The first ques-tion is more basic, of course, but the secondis also relevant and ought not be overlooked.Consider, for example, the paragraph thatTversky and Kahneman (1974 ; p. 3 inKahneman, Slovic, & Tversky, 1982) used tointroduced the notions of heuristic and bias:

The subjective assessment of probability re-sembles the subjective assessment of physi-cal quantities such as distance or size. Thesejudgments are all based on data of lim-ited validity, which are processed accord-ing to heuristic rules. For example, the ap-parent distance of an object is determinedin part by its clarity. The more sharplythe object is seen, the closer it appears tobe. This rule has some validity, because inany given scene the more distant objectsare seen less sharply than nearer objects.However, the reliance on this rule leads tosystematic errors in the estimation of dis-tance. Specifically, distances are often over-estimated when visibility is poor becausethe contours of objects are blurred. On theother hand, distances are often underesti-mated when visibility is good because theobjects are seen sharply. Thus the relianceon clarity as an indication leads to com-mon biases. Such biases are also found inintuitive judgments of probability.

This statement was intended to extendBrunswik’s (1943) analysis of the percep-tion of distance to the domain of intuitivethinking and to provide a rationale for us-ing biases to diagnose heuristics. However,the analysis of the effect of haze is flawed:It neglects the fact that an observer lookingat a distant mountain possesses two relevantcues, not one. The first cue is the blur of thecontours of the target mountain, which ispositively correlated with its distance, whenall else is equal. This cue should be givenpositive weight in a judgment of distance,and it is. The second relevant cue, whichthe observer can readily assess by lookingaround, is the ambient or general haziness.


a model of heuristic judgment 2 69

In an optimal regression model for estimat-ing distance, general haziness is a suppressorvariable, which must be weighted negativelybecause it contributes to blur but is uncor-related with distance. Contrary to the argu-ment made in 1974 , using blur as a cue doesnot inevitably lead to bias in the judgmentof distance – the illusion could just as wellbe described as a failure to assign adequatenegative weight to ambient haze. The effectof haziness on impressions of distance is afailing of system 1 : The perceptual system isnot designed to correct for this variable. Theeffect of haziness on judgments of distanceis a separate failure of system 2 . Althoughpeople are capable of consciously correctingtheir impressions of distance for the effectsof ambient haze, they commonly fail to doso. A similar analysis applies to some of thejudgmental biases we discuss later, in whicherrors and biases only occur when both sys-tems fail.

In the following section, we presentan attribute-substitution model of heuris-tic judgment, which assumes that difficultquestions are often answered by substi-tuting an answer to an easier one. Thiselaborates and extends earlier treatmentsof the topic (Kahneman & Tversky, 1982 ;Tversky & Kahneman, 1974 , 1983). Fol-lowing sections introduce a research designfor studying attribute substitution, as wellas discuss the controversy over the repre-sentativeness heuristic in the context of adual-system view that we endorse. The finalsection situates representativeness withina broad family of prototype heuristics, inwhich properties of a prototypical exemplardominate global judgments concerning anentire set.

Attribute Substitution

The early research on judgment heuris-tics was guided by a simple and generalhypothesis: When confronted with a diffi-cult question, people may answer an eas-ier one instead and are often unaware ofthe substitution. A person who is asked“What proportion of long-distance relation-

ships break up within a year?” may answeras if she had been asked “Do instances offailed long-distance relationships come read-ily to mind?” This would be an applica-tion of the availability heuristic. A profes-sor who has heard a candidate’s job talk andnow considers the question “How likely is itthat this candidate could be tenured in ourdepartment?” may answer the much easierquestion: “How impressive was the talk?”.This would be an example of one form ofthe representativeness heuristic.

The heuristics and biases research pro-gram has focused primarily on representa-tiveness and availability – two versatile at-tributes that are automatically computedand can serve as candidate answers to manydifferent questions. It has also focused prin-cipally on thinking under uncertainty. How-ever, the restriction to particular heuristicsand to a specific context is largely arbitrary.Kahneman and Frederick (2002) argued thatthis process of attribute substitution is ageneral feature of heuristic judgment; thatwhenever the aspect of the judgmental ob-ject that one intends to judge (the target at-tribute) is less readily assessed than a relatedproperty that yields a plausible answer (theheuristic attribute), individuals may unwit-tingly substitute the simpler assessment. Foran example, consider the well-known studyby Strack, Martin, and Schwarz (1988) inwhich college students answered a surveythat included these two questions: “Howhappy are you with your life in general?” and“How many dates did you have last month?”The correlation between the two questionswas negligible when they occurred in theorder shown, but rose to .66 if the datingquestion was asked first. We suggest that thequestion about dating frequency automati-cally evokes an evaluation of one’s romanticsatisfaction and that this evaluation lingersto become the heuristic attribute when theglobal happiness question is subsequentlyencountered.

To further illustrate the process of at-tribute substitution, consider a question ina study by Frederick and Nelson (2004):“If a sphere were dropped into a opencube, such that it just fit (the diameter



of the sphere is the same as the interiorwidth of the cube), what proportion ofthe volume of the cube would the sphereoccupy?” The target attribute in this judg-ment (the volumetric relation between acube and sphere) is simple enough to be un-derstood but complicated enough to accom-modate a wide range of estimates as plau-sible answers. Thus, if a relevant simplercomputation or perceptual impression ex-ists, respondents will have no strong basis forrejecting it as their “final answer.” Frederickand Nelson (2004) proposed that the arealratio of the respective cross-sections servesthat function; that is, that respondents an-swer as if they were asked the simpler two-dimensional analog of this problem (“If acircle were drawn inside a square, what pro-portion of the area of the square does thecircle occupy?”). As evidence, they notedthat the mean estimate of the “sphere insidecube” problem (74%) is scarcely differentfrom the mean estimate of the “circle insidesquare” problem (77%) and greatly exceedsthe correct answer (52%) – a correct an-swer that most people, not surprisingly, aresurprised by.

Biases

Whenever the heuristic attribute differsfrom the target attribute, the substitutionof one for the other inevitably introducessystematic biases. In this treatment, weare mostly concerned with weighting bi-ases, which arise when cues available tothe judge are given either too much ortoo little weight. Criteria for determiningoptimal weights can be drawn from sev-eral sources. In the classic lens model, theoptimal weights associated with differentcues are the regression weights that opti-mize the prediction of an external criterion,such as physical distance or the grade pointaverage that a college applicant will attain(Brunswik, 1943 ; Hammond, 1955). Ouranalysis of weighting biases applies to suchcases, but it also extends to attributes forwhich no objective criterion is available,such as an individual’s overall happinessor the probability that a particular patientwill survive surgery. Normative standards for

these attributes must be drawn from the con-straints of ordinary language and are oftenimprecise. For example, the conventional in-terpretation of overall happiness does notspecify how much weight ought to be givento various life domains. However, it certainlydoes require that substantial weight be givento every important domain of life and thatno weight at all be given to the currentweather or to the recent consumption of acookie. Similar rules of common sense ap-ply to judgments of probability. For example,the statement “John is more likely to survivea week than a month” is clearly true, and,thus, implies a rule that people would wanttheir probability judgments to follow. Ac-cordingly, neglect of duration in assessmentsof survival probabilities would be properlydescribed as a weighting bias, even if therewere no way to establish a normative prob-ability for individual cases (Kahneman &Tversky, 1996).

For some judgmental tasks, informationthat could serve to supplement or correct theheuristic is not neglected or underweightedbut simply lacking. If asked to judge the rela-tive frequency of words beginning with K orR (Tversky & Kahneman, 1973) or to com-pare the population of a familiar foreign citywith one that is unfamiliar (Gigerenzer &Goldstein, 1996), respondents have little re-course but to base their judgments on easeof retrieval or recognition. The necessary re-liance on these heuristic attributes renderssuch judgments susceptible to biasing factors(e.g., the amount of media coverage). How-ever, unlike weighting biases, such biases ofinsufficient information cannot be describedas errors of judgment because there is no wayto avoid them.

Accessibility and Substitution

The intent to judge a target attribute initi-ates a search for a reasonable value. Some-times this search ends quickly because therequired value can be read from a storedmemory (e.g., the answer to the question“How tall are you?”) or a current experience(e.g., the answer to the question “How muchdo you like this cake?”). For other judg-ments, however, the target attribute does



not readily come to mind, but the searchfor it evokes other attributes that are con-ceptually and associatively related. For ex-ample, a question about overall happinessmay retrieve the answer to a related ques-tion about satisfaction with a particular as-pect of life upon which one is currentlyreflecting.

We adopt the term accessibility to referto the ease (or effort) with which particu-lar mental contents come to mind (see, e.g.,Higgins, 1996; Tulving & Pearlstone, 1966).The question of why thoughts become ac-cessible – why particular ideas come to mindat particular times – has a long history in psy-chology and encompasses notions of stimu-lus salience, associative activation, selectiveattention, specific training, and priming. Inthe present usage, accessibility is determinedjointly by the characteristics of the cogni-tive mechanisms that produce it and by thecharacteristics of the stimuli and events thatevoke it, and it may refer to different aspectsand elements of a situation, different ob-jects in a scene, or different attributes of anobject.

Attribute substitution occurs when a rela-tively inaccessible target attribute is assessedby mapping a relatively accessible and re-lated heuristic attribute onto the target scale.Some attributes are permanent candidatesfor the heuristic role because they are rou-tinely evaluated as part of perception andcomprehension and therefore always acces-sible (Tversky & Kahneman, 1983). Thesenatural assessments include physical prop-erties such as size and distance and moreabstract properties such as similarity (e.g.,Tversky & Kahneman, 1983 ; see Goldstone& Son, Chap. 2), cognitive fluency in per-ception and memory (e.g., Jacoby & Dallas,1991 ; Schwarz & Vaughn, 2002 ; Tversky &Kahneman, 1973), causal propensity (Hei-der, 1944 ; Kahneman & Varey, 1990; Mi-chotte, 1963), surprisingness (Kahneman &Miller, 1986), mood (Schwarz & Clore,1983), and affective valence (e.g., Bargh,1997; Cacioppo, Priester, & Berntson, 1993 ;Kahneman, Ritov, & Schkade, 1999; Slovicet al., 2002 ; Zajonc, 1980, 1997).

Because affective valence is a natural as-sessment, it is a candidate for attribute sub-

stitution in a wide variety of affect-ladenjudgments. Indeed, the evidence suggeststhat a list of major general-purpose heuris-tics should include an affect heuristic (Slovicet al., 2002). Slovic and colleagues (2002)show that a basic affective reaction gov-erns a wide variety of more complex evalua-tions such as the cost–benefit ratio of varioustechnologies, the safe level of chemicals, oreven the predicted economic performanceof various industries. In the same vein, Kah-neman and Ritov (1994) and Kahneman,Ritov, and Schkade (1999) proposed that anautomatic affective valuation is the principaldeterminant of willingness to pay for publicgoods, and Kahneman, Schkade, and Sun-stein (1998) interpreted jurors’ assessmentsof punitive awards as a mapping of outrageonto a dollar scale of punishments.

Attributes that are not naturally assessedcan become accessible if they have been re-cently evoked or primed (see, e.g., Bargh etal., 1986; Higgins & Brendl, 1995). The ef-fect of temporary accessibility is illustratedby the “romantic satisfaction heuristic” forjudging happiness. The mechanism of at-tribute substitution is the same, however,whether the heuristic attribute is chronicallyor temporarily accessible.

There is sometimes more than one can-didate for the role of heuristic attribute. Foran example that we borrow from Anderson(1991 ), consider the question “Are moredeaths caused by rattlesnakes or bees?” A re-spondent who has recently read about some-one who died from a snakebite or bee stingmay use the relative availability of instancesof the two categories as a heuristic. If noinstances come to mind, that person mightconsult his or her impressions of the “dan-gerousness” of the typical snake or bee, anapplication of representativeness. Indeed, itis possible that the question initiates botha search for instances and an assessment ofdangerousness, and that a contest of accessi-bility determines the role of the two heuris-tics in the final response. As Anderson ob-served, it is not always possible to determinea priori which heuristic will govern the re-sponse to a particular problem.

The original list of heuristics (Tver-sky & Kahneman, 1974) also included an



“anchoring heuristic.” An anchoring effect,however, does not involve the substitution ofa heuristic attribute for a target attribute: Itis due to the temporary salience of a particu-lar value of the target attribute. However, an-choring and attribute substitution are bothinstances of a broader family of accessibilityeffects (Kahneman, 2003). In attribute sub-stitution, a highly accessible attribute con-trols the evaluation of a less accessible one.In anchoring, a highly accessible value ofthe target attribute dominates its judgment.This conception is compatible with morerecent theoretical treatments of anchor-ing (see, e.g., Chapman & Johnson, 1994 ,2002 ; Mussweiler & Strack 1999; Strack &Mussweiler, 1997).

Cross-Dimensional Mapping

The process of attribute substitution in-volves the mapping of the heuristic at-tribute of the judgment object onto thescale of the target attribute. Our notion ofcross-dimensional mapping extends Stevens’(1975) concept of cross-modality matching.Stevens postulated that intensive attributes(e.g., brightness, loudness, the severity ofcrimes) can be mapped onto a common scaleof sensory strength, allowing direct matchingof intensity across modalities – permitting,for example, respondents to match the loud-ness of sounds to the severity of crimes. Ourconception allows other ways of compar-ing values across dimensions, such as match-ing relative positions (e.g., percentiles)in the frequency distributions or ranges ofdifferent attributes (Parducci, 1965). An im-pression of a student’s position in the dis-tribution of aptitude may be mapped di-rectly onto a corresponding position in thedistribution of academic achievement andthen translated into a letter grade. Notethat cross-dimensional matching is inher-ently nonregressive: A judgment or predic-tion is just as extreme as the impressionmapped onto it. Ganzach and Krantz (1990)applied the term “univariate matching” to aclosely related notion.

Cross-dimensional mapping presents spe-cial problems when the scale of the tar-

get attribute has no upper bound. Kahne-man, Ritov, and Schkade (1999) discussedtwo situations in which an attitude (or af-fective valuation) is mapped onto an un-bounded scale of dollars: when respondentsin surveys are required to indicate how muchmoney they would contribute for a cause,and when jurors are required to specify anamount of punitive damages against a neg-ligent firm. The mapping of attitudes ontodollars is a variant of direct scaling in psy-chophysics, where respondents assign num-bers to indicate the intensity of sensations(Stevens, 1975). The normal practice of di-rect scaling is for the experimenter to pro-vide a modulus – a specified number thatis to be associated with a standard stimu-lus. For example, respondents may be askedto assign the number 10 to the loudness ofa standard sound and judge the loudnessof other sounds relative to that standard.Stevens (1975) observed that when the ex-perimenter fails to provide a modulus, re-spondents spontaneously adopt one. How-ever, different respondents may pick modulithat differ greatly (sometimes varying by afactor of 100 or more); thus, the variabilityin judgments of particular stimuli is domi-nated by arbitrary individual differences inthe choice of modulus. A similar analysisapplies to situations in which respondentsare required to use the dollar scale to ex-press affection for a species or outrage to-ward a defendant. Just as Stevens’ observershad no principled way to assign a number toa moderately loud sound, survey participantsand jurors have no principled way to scaleaffection or outrage into dollars. The anal-ogy of scaling without a modulus has beenused to explain the notorious variability ofdollar responses in surveys of willingness topay and in jury awards (Kahneman, Ritov,& Schkade, 1999; Kahneman, Schkade, &Sunstein, 1998).

System 2 : The Supervision ofIntuitive Judgments

Our model assumes that an intuitive judg-ment is expressed overtly only if it isendorsed by system 2 . The Stroop task



illustrates this two-system structure. Ob-servers who are instructed to report the colorin which words are printed tend to stum-ble when the word is the name of anothercolor (e.g., the word BLUE printed in greenink). The difficulty arises because the word isautomatically read, and activates a response(“blue” in this case) that competes with therequired response (“green”). Errors are rarein the Stroop test, indicating generally suc-cessful monitoring and control of the overtresponse, but the conflict produces delaysand hesitations. The successful suppressionof erroneous responses is effortful, and itsefficacy is reduced by stress and distraction.

Gilbert (1989) described a correctionmodel in which initial impulses are oftenwrong and normally overridden. He arguedthat people initially believe whatever theyare told (e.g., “Whitefish love grapes”) andthat it takes some time and mental effort to“unbelieve” such dubious statements. Hereagain, cognitive load disrupts the control-ling operations of system 2 , increasing therate of errors and revealing aspects of intu-itive thinking that are normally suppressed.In an ingenious extension of this approach,Bodenhausen (1990) exploited natural tem-poral variability in alertness. He found that“morning people” were substantially moresusceptible to a judgment bias (the conjunc-tion fallacy) in the evening and that “eveningpeople” were more likely to commit the fal-lacy in the morning.

Because system 2 is relatively slow, its op-erations can be disrupted by time pressure.Finucane et al. (2000) reported a study inwhich respondents judged the risks and ben-efits of various products and technologies(e.g., nuclear power, chemical plants, cellu-lar phones). When participants were forcedto respond within 5 seconds, the correlationsbetween their judgments of risks and theirjudgments of benefits were strongly nega-tive. The negative correlations were muchweaker (although still pronounced) when re-spondents were given more time to pondera response. When time is short, the sameaffective evaluation apparently serves as aheuristic attribute for assessments of bothbenefits and risks. Respondents can move

beyond this simple strategy, but they needmore than 5 seconds to do so. As this exam-ple illustrates, judgment by heuristic oftenyields simplistic assessments, which system 2

sometimes corrects by bringing additionalconsiderations to bear.

Attribute substitution can be preventedby alerting respondents to the possibilitythat their judgment could be contaminatedby an irrelevant variable. For example, al-though sunny or rainy weather typically af-fects reports of well-being, Schwarz andClore (1983) found that weather has noeffect if respondents are asked about theweather just before answering the well-being question. Apparently, this question re-minds respondents that their current mood(a candidate heuristic attribute) is influ-enced by a factor (current weather) that isirrelevant to the requested target attribute(overall well-being). Schwarz (1996) alsofound that asking people to describe theirsatisfaction with some particular domain oflife reduces the weight this domain receivesin a subsequent judgment of overall well be-ing. As these examples illustrate, althoughpriming typically increases the weight of thatvariable on judgment (a system 1 effect), thisdoes not occur if the prime is a sufficientlyexplicit reminder that brings the self-criticaloperations of system 2 into play.

We suspect that system 2 endorsements ofintuitive judgments are granted quite casu-ally under normal circumstances. Considerthe puzzle “A bat and a ball cost $1 .10 in to-tal. The bat costs $1 more than the ball. Howmuch does the ball cost?” Almost everyonewe ask reports an initial tendency to answer“10 cents” because the sum $1 .10 separatesnaturally into $1 and 10 cents, and 10 centsis about the right magnitude. Many peo-ple yield to this immediate impulse. Evenamong undergraduates at elite institutions,about half get this problem wrong when itis included in a short IQ test (Frederick,2004). The critical feature of this problemis that anyone who reports 10 cents has ob-viously not taken the trouble to check hisor her answer. The surprisingly high rateof errors in this easy problem illustrateshow lightly system 2 monitors the output of



system 1 : People are often content to trusta plausible judgment that quickly comes tomind. (The correct answer, by the way, is5 cents.)

The bat and ball problem elicits many er-rors, although it is not really difficult andcertainly not ambiguous. A moral of thisexample is that people often make quickintuitive judgments to which they are notdeeply committed. A related moral is thatwe should be suspicious of analyses that ex-plain apparent errors by attributing to re-spondents a bizarre interpretation of thequestion. Consider someone who answers aquestion about happiness by reporting hersatisfaction with her romantic life. The re-spondent is surely not committed to the ab-surdly narrow interpretation of happinessthat her response seemingly implies. Morelikely, at the time of answering, she thinksthat she is reporting happiness: A judgmentcomes quickly to mind and is not obviouslymistaken – end of story. Similarly, we pro-pose that respondents who judge probabil-ity by representativeness do not seriously be-lieve that the questions “How likely is X tobe a Y?” and “How much does X resemblethe stereotype of Y?” are synonymous. Peo-ple who make a casual intuitive judgmentnormally know little about how their judg-ment came about and know even less aboutits logical entailments. Attempts to recon-struct the meaning of intuitive judgments byinterviewing respondents (see, e.g., Hertwig& Gigerenzer, 1999) are therefore unlikelyto succeed because such probes require bet-ter introspective access and more coherentbeliefs than people normally muster.

Identifying a Heuristic

Hypotheses about judgment heuristics havemost often been studied by examiningweighting biases and deviations from nor-mative rules. However, the hypothesis thatone attribute is substituted for another in ajudgment task – for example, representative-ness for probability – can also be tested moredirectly. In the heuristic elicitation design,

one group of respondents provides judg-ments of a target attribute for a set of ob-jects and another group evaluates the hy-pothesized heuristic attribute for the sameobjects. The substitution hypothesis im-plies that the judgments of the two groups,when expressed in comparable units (e.g.,percentiles), will be identical. This sectionexamines several applications of heuristicelicitation.

Eliciting Representativeness

Figure 1 2 .1 displays the results of two ex-periments in which a measure of represen-tativeness was elicited. These results werepublished long ago, but we repeat them herebecause they still provide the most directevidence for both attribute substitution andthe representativeness heuristic. For a morerecent application of a similar design, seeBar-Hillel and Neter (1993).

The object of judgment in the study fromwhich Figure 1 2 .1 (a) is drawn (Kahneman &Tversky, 1973 ; p. 1 27 in Kahneman, Slovic,& Tversky, 1982) was the following descrip-tion of a fictitious graduate student, whichwas shown along with a list of nine fields ofgraduate specialization:

Tom W. is of high intelligence, althoughlacking in true creativity. He has a needfor order and clarity and for neat and tidysystems in which every detail finds its ap-propriate place. His writing is rather dulland mechanical, occasionally enlivened bysomewhat corny puns and by flashes ofimagination of the sci-fi type. He has astrong drive for competence. He seems tohave little feel and little sympathy for otherpeople and does not enjoy interacting withothers. Self-centered, he nonetheless has adeep moral sense.

Participants in a representativeness groupranked the nine fields of specialization bythe degree to which Tom W. “resembles atypical graduate student.” Participants in theprobability group ranked the nine fields ac-cording to the likelihood of Tom W.’s spe-cializing in each. Figure 1 2 .1 (a) plots themean judgments of the two groups. Thecorrelation between representativeness and



Linda

1

2

3

4

5

6

7

1 2 3 4 5 6 7

mean rank (similarity)

mea

n r

ank

(lik

elih

oo

d)

Tom W.

1

2

3

4

5

6

7

8

9

1 2 3 4 5 6 7 8

mean rank (similarity)

mea

n r

ank

(lik

elih

oo

d)

9

a b

Figure 1 2 .1 . (a) Plot of average ranks for nine outcomes for Tom W. ranked by probability and bysimilarity to stereotypes of graduate students in various fields. (b) Plot of average ranks for eightoutcomes for Linda ranked by probability and by representativeness.

probability is nearly perfect (.97). Nostronger support for attribute-substitutioncould be imagined. However, interpretingrepresentativeness as the heuristic attributein these judgments does require two addi-tional plausible assumptions – that represen-tativeness is more accessible than probabil-ity, and that there is no third attribute thatcould explain both judgments.

The Tom W. study was also intended toexamine the effect of the base rates of out-comes on categorical prediction. For thatpurpose, respondents in a third group esti-mated the proportion of graduate studentsenrolled in each of the nine fields. By design,some outcomes were defined quite broadly,whereas others were defined more narrowly.As intended, estimates of base rates var-ied markedly across fields, ranging from 3%for Library Science to 20% for Humanitiesand Education. Also by design, the descrip-tion of Tom W. included characteristics (e.g.,introversion) that were intended to makehim fit the stereotypes of the smaller fields(library science, computer science) betterthan the larger fields (humanities and socialsciences).1 As intended, the correlation be-tween the average judgments of representa-tiveness and of base rates was strongly nega-tive (−.65 ).

The logic of probabilistic prediction inthis task suggests that the ranking of out-comes by their probabilities should be in-termediate between their rankings by rep-resentativeness and by base rate frequencies.Indeed, if the personality description is takento be a poor source of information, proba-bility judgments should stay quite close tothe base rates. The description of Tom W.was designed to allow considerable scopefor judgments of probability to diverge fromjudgments of representativeness, as this logicrequires. Figure 1 2 .1 (a) shows no such di-vergence. Thus, the results of the Tom W.study simultaneously demonstrate the sub-stitution of representativeness for probabil-ity and the neglect of known (but not explic-itly mentioned) base rates.

Figure 1 2 .1 (b) is drawn from an earlystudy of the Linda problem, the best-knownand most controversial example in the rep-resentativeness literature (Tversky & Kahne-man, 1982) in which a woman named Lindawas described as follows:

Linda is 31 years old, single, outspokenand very bright. She majored in philoso-phy. As a student she was deeply concernedwith issues of discrimination and social jus-tice and also participated in antinucleardemonstrations.



As in the Tom W. study, separate groupsof respondents were asked to rank a set ofeight outcomes by representativeness andprobability. The results are shown in Fig-ure 1 2 .1 (b). Again the correlation betweenthese rankings was almost perfect (.99).1

Six of the eight outcomes that subjectswere asked to rank were fillers (e.g., ele-mentary school teacher, psychiatric socialworker). The two critical outcomes were #6

(bank teller) and the so-called conjunctionitem #8 (bank teller and active in the fem-inist movement). Most subjects ranked theconjunction higher than its constituent, bothin representativeness (85%) and probabil-ity (89%). The observed ranking of the twoitems is quite reasonable for judgments ofsimilarity, but not for probability: Linda mayresemble a feminist bank teller more thanshe resembles a bank teller, but she cannotbe more likely to be a feminist bank tellerthan to be a bank teller. In this problem, re-liance on representativeness yields probabil-ity judgments that violate a basic logical rule.As in the Tom W. study, the results make twopoints: They support the hypothesis of at-tribute substitution and also illustrate a pre-dictable judgment error.

The Representativeness Controversy

The experiments summarized in Figure 1 2 .1provided direct evidence for the represen-tativeness heuristic and two concomitantbiases: neglect of base rates and conjunc-tion errors. In the terminology introducedby Tversky and Kahneman (1983), the de-sign of these experiments was “subtle”: Ad-equate information was available for partic-ipants to avoid the error, but no effort wasmade to call their attention to that informa-tion. For example, participants in the TomW. experiment had general knowledge of therelative base rates of the various fields of spe-cialization, but these base rates were not ex-plicitly mentioned in the problem. Similarly,both critical items in the Linda experimentwere included in the list of outcomes, but

they were separated by a filler so respondentswould not feel compelled to compare them.In the anthropomorphic language used here,system 2 was given a chance to correct thejudgment but was not prompted to do so.

In view of the confusing controversy thatfollowed, it is perhaps unfortunate that thearticles documenting base rate neglect andconjunction errors did not stop with subtletests. Each article also contained an experi-mental flourish – a demonstration in whichthe error occurred in spite of a manipula-tion that called participants’ attention to thecritical variable. The engineer–lawyer prob-lem (Kahneman & Tversky, 1973) includedspecial instructions to ensure that respon-dents would notice the base rates of theoutcomes. The brief personality descriptionsshown to respondents were reported to havebeen drawn from a set containing descrip-tions of 30 lawyers and 70 engineers (or viceversa), and respondents were asked “Whatis the probability that this description be-longs to one of the 30 lawyers in the sampleof 100?” To the authors’ surprise, base rateswere largely neglected in the responses, de-spite their salience in the instructions. Sim-ilarly, the authors were later shocked to dis-cover that more than 80% of undergraduatescommitted a conjunction error even whenasked point blank whether Linda was morelikely to be “a bank teller” or “a bank tellerwho is active in the feminist movement”(Tversky & Kahneman, 1983). The noveltyof these additional direct or “transparent”tests was the finding that respondents con-tinued to show the biases associated withrepresentativeness even in the presence ofstrong cues pointing to the normative re-sponse. The errors that people make in trans-parent judgment problems are analogous toobservers’ failure to allow for ambient hazein estimating distances: A correct responseis within reach, but not chosen, and the fail-ure involves an unexpected weakness of thecorrective operations of system 2 .

Discussions of the heuristics and biasesapproach have focused almost exclusivelyon the direct conjunction fallacy and onthe engineer–lawyer problems. These are



also the only studies that have been exten-sively replicated with varying parameters.The amount of critical attention is remark-able because the studies were not, in fact,essential to the authors’ central claim. Interms of the present treatment, the claimwas that intuitive prediction is an operationof system 1 , which is susceptible to both baserate neglect and conjunction fallacies. Therewas no intent to deny the possibility of sys-tem 2 interventions that would modify oroverride intuitive predictions. Thus, the ar-ticles in which these studies appeared wouldhave been substantially the same, althoughfar less provocative, if respondents had over-come base rate neglect and conjunction er-rors in transparent tests.

To appreciate why the strong forms ofbase rate neglect and of the conjunction fal-lacy sparked so much controversy, it is use-ful to distinguish two conceptions of humanrationality (Kahneman, 2000b). Coherencerationality is the strict conception that re-quires the agent’s entire system of beliefsand preferences to be internally consistentand immune to effects of framing and con-text. For example, an individual’s probabil-ity p (“Linda is a bank teller”) should be thesum of the probabilities p (“Linda is a bankteller and a feminist”), and p (“Linda is a bankteller and not a feminist”). A subtle test ofcoherence rationality could be conducted byasking individuals to assess these three prob-abilities on separate occasions under circum-stances that minimize recall. Coherence canalso be tested in a between-groups design. Ifrandom assignment is assumed, the sum ofthe average probabilities assigned to the twocomponent events should equal the averagejudged probability of “Linda is a bank teller.”If this prediction fails, then at least someindividuals are incoherent. Demonstrationsof incoherence present a significant chal-lenge to important models of decision the-ory and economics, which attribute to agentsa very strict form of rationality (Tversky &Kahneman, 1986). Failures of perfect coher-ence are less provocative to psychologists,who have a more realistic view of humancapabilities.

A more lenient concept, reasoning ra-tionality, only requires an ability to reasoncorrectly about the information currentlyat hand without demanding perfect consis-tency among beliefs that are not simulta-neously evoked. The best known violationof reasoning rationality is the famous “fourcard” problem (Wason, 1960). The failure ofintelligent adults to reason their way throughthis problem is surprising because the prob-lem is “easy” in the sense of being easilyunderstood once explained. What everyonelearns, when first told that intelligent peo-ple fail to solve the four-card problem, isthat one’s expectations about human rea-soning abilities had not been adequately cal-ibrated. There is, of course, no well-definedmetric of reasoning rationality, but whatevermetric one uses, the Wason problem callsfor a downward adjustment. The surprisingresults of the Linda and engineer–lawyerproblems led Tversky and Kahneman to asimilar realization: The reasoning of theirsubjects was less proficient than they had an-ticipated. Many readers of the work sharedthis conclusion, but many others stronglyresisted it.

The implicit challenge to reasoning ra-tionality was met by numerous attempts todismiss the findings of the engineer–lawyerand the Linda studies as artifacts of ambigu-ous language, confusing instructions, conver-sational norms, or inappropriate normativestandards. Doubts have been raised aboutthe proper interpretation of almost everyword in the conjunction problem, including“bank teller,” “probability,” and even “and”(see, e.g., Dulany & Hilton, 1991 ; Hilton &Slugoski, 2001 ). These claims are not dis-cussed in detail here. We suspect that mostof them have some validity and that theyidentified mechanisms that may have madethe results in the engineer–lawyer and Lindastudies exceptionally strong. However, wenote a significant weakness shared by allthese critical discussions: They provide noexplanation of the essentially perfect con-sistency of the judgments observed in di-rect tests of the conjunction rule and inthree other types of experiments: subtle



comparisons, between-Ss comparisons, andmost important, judgments of representa-tiveness (see also Bar-Hillel & Neter, 1993).Interpretations of the conjunction fallacyas an artifact implicitly dismiss the resultsof Figure 1 2 .1 (b) as a coincidence (for anexception, see Ayton, 1998). The story ofthe engineer-lawyer problem is similar. Hereagain, multiple demonstrations in whichbase rate information was used (see Koehler,1996, for a review) invited the inference thatthere is no general problem of base rate ne-glect. Again, the data of prediction by repre-sentativeness in Figure 1 2 .1 (a) (and relatedresults reported by Kahneman & Tversky,1973) were ignored.

The demonstrations that under some con-ditions people avoid the conjunction fallacyin direct tests, or use explicit base rate in-formation, led some scholars to the blanketconclusion that judgment biases are artifi-cial and fragile and that there is no need forjudgment heuristics to explain them. Thisposition was promoted most vigorously byGigerenzer (1991 ). Kahneman and Tversky(1996) argued in response that the heuris-tics and biases position does not preclude thepossibility of people’s performing flawlesslyin particular variants of the Linda and theengineer–lawyer problems. Because laypeo-ple readily acknowledge the validity ofthe conjunction rule and the relevance ofbase rate information, the fact that theysometimes obey these principles is neither asurprise nor an argument against the role ofrepresentativeness in routine intuitive pre-diction. However, the study of conditionsunder which errors are avoided can help usunderstand the capabilities and limitationsof system 2 . We develop this argument fur-ther in the next section.

Making Biases Disappear: A Taskfor System 2

Much has been learned over the years aboutvariables and experimental procedures thatreduce or eliminate the biases associatedwith representativeness. We next discussconditions under which errors of intuition

are successfully overcome and some circum-stances under which intuitions may not beevoked at all.

statistical sophistication

The performance of statistically sophisti-cated groups of respondents in different ver-sions of the Linda problem illustrates the ef-fects of both expertise and research design(Tversky & Kahneman, 1983). Statistical ex-pertise provided no advantage in the eight-item version in which the critical items wereseparated by a filler and were presumablyconsidered separately. In the two-item ver-sion, in contrast, respondents were effec-tively compelled to compare “bank teller”with “bank teller and is active in the femi-nist movement.” The incidence of conjunc-tion errors remained essentially unchangedamong the statistically naive in this condi-tion but dropped dramatically for the statis-tically sophisticated. Most of the experts fol-lowed logic rather than intuition when theyrecognized that one of the categories con-tained the other. In the absence of a promptto compare the items, however, the statis-tically sophisticated made their predictionsin the same way as everyone else does – byrepresentativeness. As Stephen Jay Gould(1991 , p. 469) noted, knowledge of the truthdoes not dislodge the feeling that Linda is afeminist bank teller: “I know [the right an-swer], yet a little homunculus in my headcontinues to jump up and down, shouting atme – ‘but she can’t just be a bank teller; readthe description.’”

intelligence

Stanovich (1999) and Stanovich and West(2002) observed a generally negative corre-lation between conventional measures of in-telligence and susceptibility to judgment bi-ases. They used transparent versions of theproblems, which include adequate cues tothe correct answer and therefore providea test of reasoning rationality. Not surpris-ingly, intelligent people are more likely topossess the relevant logical rules and also torecognize the applicability of these rules in



particular situations. In the terms ofthe present analysis, high-IQ respondentsbenefit from relatively efficient system 2 op-erations that enable them to overcome er-roneous intuitions when adequate informa-tion is available. (However, when a problemis too difficult for everyone, the correlationmay reverse because the more intelligent re-spondents are more likely to agree on a plau-sible error than to respond randomly, as dis-cussed in Kahneman, 2000b.)

frequency format

Relative frequencies (e.g., 1 in 10) are morevividly represented and more easily under-stood than equivalent probabilities (.10) orpercentages (10%). For example, the emo-tional impact of statements of risk is en-hanced by the frequency format: “1 personin 1000 will die” is more frightening than aprobability of .001 (Slovic et al., 2002). Thefrequency representation also makes it eas-ier to visualize partitions of sets and detectthat one set is contained in another. As aconsequence, the conjunction fallacy is gen-erally avoided in direct tests in which thefrequency format makes it easy to recog-nize that feminist bank tellers are a subset ofbank tellers (Gigerenzer & Hoffrage, 1995 ;Tversky & Kahneman, 1983). For similar rea-sons, some base rate problems are more eas-ily solved when couched in frequencies thanin probabilities or percentages (Cosmides &Tooby, 1996). However, there is little sup-port for the more general claims about theevolutionary adaptation of the mind to dealwith frequencies (Evans et al., 2000). Fur-thermore, the ranking of outcomes by pre-dicted relative frequency is very similar tothe ranking of the same outcomes by rep-resentativeness (Mellers, Hertwig, & Kahne-man, 2001 ). We conclude that the frequencyformat affects the corrective operations ofsystem 2 , not the intuitive operations of sys-tem 1 . The language of frequencies improvesrespondents’ ability to impose the logic ofset inclusion on their considered judgmentsbut does not reduce the role of representa-tiveness in their intuitions.

manipulations of attention

The weight of neglected variables can be in-creased by drawing attention to them, andexperimenters have devised many ingeniousways to do so. Schwarz et al. (1991 ) foundthat respondents pay more attention to baserate information when they are instructedto think as statisticians rather than clini-cal psychologists. Krosnick, Li, and Lehman(1990) exploited conversational conventionsabout the sequencing of information andconfirmed that the impact of base rate in-formation was enhanced by presenting thatinformation after the personality descrip-tion rather than before it. Attention to thebase rate is also enhanced when partici-pants observe the drawing of descriptionsfrom an urn (Gigerenzer, Hell, & Blank,1988) perhaps because watching the draw-ing induces conscious expectations that re-flect the known proportions of possible out-comes. The conjunction fallacy can alsobe reduced or eliminated by manipulationsthat increase the accessibility of the rel-evant rule, including some linguistic vari-ations (Macchi, 1995), and practice withlogical problems (Agnoli, 1991 ; Agnoli &Krantz, 1989).

The interpretation of these attentional ef-fects is straightforward. We assume mostparticipants in judgment studies know, atleast vaguely, that the base rate is rele-vant and that the conjunction rule is valid(Kahneman & Tversky, 1982). Whether theyapply this knowledge to override an intu-itive judgment depends on their cognitiveskills (education, intelligence) and on for-mulations that make the applicability of arule apparent (frequency format) or a rel-evant factor more salient (manipulations ofattention). We assume intuitions are less sen-sitive to these factors and that the appear-ance or disappearance of biases mainly re-flects variations in the efficacy of correctiveoperations. This conclusion would be circu-lar, of course, if the corrective operationswere both inferred from the observation ofcorrect performance and used to explain thatperformance. Fortunately, the circularity canbe avoided because the role of system 2



can be verified – for example, by using ma-nipulations of time pressure, cognitive load,or mood to interfere with its operations.

within-subjects factorial designs

The relative virtues of between-subjects andwithin-subject designs in studies of judg-ment are a highly contentious issue. Facto-rial designs have their dismissive critics (e.g.,Poulton, 1989) and their vigorous defenders(e.g., Birnbaum, 1999). We do not attemptto adjudicate this controversy here. Our nar-rower point is that between-subjects designsare more appropriate for the study of heuris-tics of judgment. The following argumentsfavor this conclusion:

� Factorial designs are transparent. Partici-pants are likely to identify the variablesthat are manipulated, especially if thereare many trials and especially in a fullyfactorial design in which the same stimu-lus attributes are repeated in varying com-binations. The message that the designconveys to the participants is that the ex-perimenter expects to find effects of ev-ery factor that is manipulated (Bar-Hillel& Fischhoff, 1981 ; Schwarz, 1996).

� Studies that apply a factorial designto judgment tasks commonly involveschematic and impoverished stimuli. Thetasks are also highly repetitive. Thesefeatures encourage participants to adoptsimple mechanical rules that will allowthem to respond quickly without formingan individuated impression of each stim-ulus. For example, Ordonez and Benson(1997) required respondents to judge theattractiveness of gambles on a 100-pointscale. They found that under time pres-sure many respondents computed or esti-mated the expected values of the gamblesand used the results as attractiveness rat-ings (e.g., a rating of 1 5 for a 52% chanceto win $31 .50).

� Factorial designs often yield judgmentsthat are linear combinations of the ma-nipulated variables. This is a centralconclusion of a massive research effortconducted by Anderson (1996), who

observed that people often average or addwhere they should multiply.

In summary, the factorial design is notappropriate for testing hypotheses about bi-ases of neglect because it effectively guaran-tees that no manipulated factor is neglected.Figure 1 2 .2 illustrates this claim by sev-eral examples of an additive extension effectthat we discuss further in the next section.The experiments summarized in the differ-ent panels share three important features:(1 ) In each case, the quantitative variableplotted on the abscissa was completely ne-glected in similar experiments conducted ina between-subjects or subtle design; (2) ineach case, the quantitative variable com-bines additively with other information; (3)in each case, a compelling normative ar-gument can be made for a quasimulti-plicative rule in which the lines shown inFigure 1 2 .2 should fan out. For example, Fig-ure 1 2 .2(c) presents a study of categoricalprediction (Novemsky & Kronzon, 1999) inwhich respondent 5 judged the relative like-lihood that a person was a member of oneoccupation rather than another (e.g., com-puter programmer vs. flight attendant) onthe basis of short personality sketches (e.g.,“shy, serious, organized, and sarcastic”) andone of three specified base rates (10%, 50%,or 90%). Representativeness and base ratewere varied factorially within subjects. Theeffect of base rate is clearly significant in thisdesign (see also Birnbaum & Mellers, 1983).Furthermore, the effects of representative-ness and base rate are strictly additive. AsAnderson (1996) argued, averaging (a spe-cial case of additive combination) is the mostobvious way to combine the effects of twovariables that are recognized as relevant (e.g.,“She looks like a bank teller, but the base-rateis low.”). Additivity is not normatively ap-propriate in this case – any Bayes-like com-bination would produce curves that initiallyfan out from the origin and converge againat high values. Similar considerations applyto the other three panels of Figure 1 2 .2 dis-cussed later. Between-subjects and factorialdesigns often yield different results in stud-ies of intuitive judgment. Why should we



Ariely Data

3035

4045

5055

6065

7075

80

0 10 20 30 40 50

Duration in Seconds

Ave

rsiv

enes

s Down

Down&Up

Up

Up&Down

Schreiber & Kahneman Data

012345

6789

10

0 5 10 15 20 25 30

Duration in Seconds

Ave

rsiv

enes

s 71 dB

75 dB

78 dB

80 dB

Novemsky & Kronzon Data

0

10

20

30

40

50

60

70

80

90

0 20 40 60 80 100

Base-rate (%)

Po

ster

ior

Pro

bab

ility

(%

)

Programmer

Surgeon

Accountant

Engineer`

Kahneman, Ritov, & Schkade Data

0

5

10

15

20

25

30

35

0 20 40 60 80

Percentage Population Decline

Mea

n co

ntri

buti

on in

$

Low

Medium

High

a b

c d

Figure 1 2 .2 . (a) Willingness to pay to restore damage to species that differ in popularity as a functionof the damage they have suffered (from Kahneman, Ritov, & Schkade 2000); (b) global evaluations ofaversive sounds of different loudness as a function of duration for subjects selected for their highsensitivity to duration (from Schreiber & Kahneman, 2000); (c) ratings of probability for predictionsthat differ in representativeness as a function of base rate frequency (from Novemsky & Kronzon,1999); (d) global evaluations of episodes of painful pressure that differ in temporal profile as afunction of duration (Ariely, 1998).

believe one design rather than the other?The main argument against the factorial de-sign is its poor ecological validity. Encounter-ing multiple judgment objects in rapid suc-cession in a rigidly controlled structure isunique to the laboratory, and the solutionsthat they evoke are not likely to be typical.Direct comparisons among concepts thatdiffer in only one variable – such as bankteller and feminist bank tellers – also providea powerful hint and a highly unusual oppor-tunity to overcome intuitions. The between-subjects design, in contrast, mimics the hap-hazard encounters in which most judgments

are made and is more likely to evoke the ca-sually intuitive mode of judgment that gov-erns much of mental life in routine situations(e.g., Langer, 1978).

Prototype Heuristics and the Neglectof Extension

In this section, we offer a common accountof three superficially dissimilar judgmentaltasks: (1 ) categorical prediction (e.g., “In aset of 30 lawyers and 70 engineers, what is the



probability that someone described as ‘charm-ing, talkative, clever, and cynical’ is one of thelawyers?”); (2) summary evaluations of pastevents (e.g., “Overall, how aversive was it tobe exposed for 30 minutes to your neighbor’scar alarm?”); and (3) economic valuationsof public goods (e.g., “What is the most youwould be willing to pay to prevent 2 00,000 mi-grating birds from drowning in uncovered oilponds?”). We propose that a generalizationof the representativeness heuristic accountsfor the remarkably similar biases that are ob-served in these diverse tasks.

The original analysis of categorical pre-diction by representativeness (Kahneman &Tversky 1973 ; Tversky & Kahneman, 1983)invoked two assumptions in which the word“representative” was used in different ways:(1 ) A prototype (a representative exemplar)is used to represent categories (e.g., banktellers) in the prediction task, and (2) theprobability that the individual belongs to acategory is judged by the degree to which theindividual resembles (is representative of) thecategory stereotype. Thus, categorical pre-diction by representativeness involves twoseparate acts of substitution – the substitu-tion of a representative exemplar for a cat-egory and the substitution of the heuris-tic attribute of representativeness for thetarget attribute of probability. Perhaps be-cause they share a label, the two pro-cesses have not been distinguished in dis-cussions of the representativeness heuristic.We separate them here by describing proto-type heuristics in which a prototype is sub-stituted for its category, but in which repre-sentativeness is not necessarily the heuristicattribute.

The target attributes to which prototypeheuristics are applied are extensional. An ex-tensional attribute pertains to an aggregatedproperty of a set or category for which anextension is specified – the probability thata set of 30 lawyers includes Jack, the over-all unpleasantness of a set of moments ofhearing a neighbor’s car alarm, and the per-sonal dollar value of saving a certain numberof birds from drowning in oil ponds. Nor-mative judgments of extensional attributesare governed by a general principle of con-ditional adding, which dictates that each el-

ement of the set adds to the overall judg-ment an amount that depends on the el-ements already included. In simple cases,conditional adding is just regular adding –the total weight of a collection of chairs isthe sum of their individual weights. In othercases, each element of the set contributesto the overall judgment, but the combina-tion rule is not simple addition and is mosttypically subadditive. For example, the eco-nomic value of protecting X birds should beincreasing in X, but the value of saving 2000

birds is for most people less than twice aslarge as the value of saving 1000 birds.

The logic of categorical prediction entailsthat the probability of membership in a cat-egory should vary with its relative size, orbase rate. In prediction by representative-ness, however, the representation of out-comes by prototypical exemplars effectivelydiscards base rates because the prototype of acategory (e.g., lawyers) contains no informa-tion about the size of its membership. Next,we show that phenomena analogous to theneglect of base rate are observed in otherprototype heuristics: The monetary value at-tached to a public good is often insensitiveto its scope, and the global evaluation of atemporally extended experience is often in-sensitive to its duration. These various in-stantiations of extension neglect (neglect ofbase rates, scope, and duration) have beendiscussed in separate literatures, but all canbe explained by the two-part process thatdefines prototype heuristics: (1 ) A categoryis represented by a prototypical exemplar,and (2) a (nonextensional) property of theprototype is then used as a heuristic attributeto evaluate an extensional target attribute ofthe category. As might be expected from theearlier discussion of base rate neglect, exten-sion neglect in all its forms is most likely to beobserved in between-subjects experiments.Within-subject factorial designs consistentlyyield the additive extension effect illustratedin Figure 1 2 .2 .

Scope Neglect in Willingness to Pay

The contingent valuation method (CVM)was developed by resource economists (seeMitchell & Carson, 1989) as a tool for



assessing the value of public goods for pur-poses of litigation or cost–benefit analysis.Participants in contingent valuation (CV)surveys are asked to indicate their willing-ness to pay (WTP) for specified public goods,and their responses are used to estimate thetotal amount that the community would payto obtain these goods. The economists whodesign contingent valuation surveys inter-pret WTP as a valid measure of economicvalue and assume that statements of WTPconform to the extensional logic of con-sumer theory. The relevant logic has beendescribed by a critic of CVM (Diamond,1996), who illustrates the conditional addingrule by the following example: In the ab-sence of income effects, WTP for saving Xbirds should equal WTP for saving (X − k)birds, plus WTP to save k birds, where thelast value is contingent on the costless priorprovision of safety for (X − k) birds.

Strict adherence to Bayes’ rule may bean excessively demanding standard for intu-itive predictions; similarly, it would be toomuch to ask for WTP responses that strictlyconform to the “add-up rule.” In both cases,however, it seems reasonable to expect somesensitivity to extension – to the base rateof outcomes in categorical prediction and tothe scope of the good in WTP. In fact, severalstudies have documented nearly completeneglect of scope in CV surveys. The best-known demonstration of scope neglect is anexperiment by Desvouges et al. (1993), whoused the scenario of migratory birds thatdrown in oil ponds. The number of birds saidto die each year was varied across groups.The WTP responses were completely insen-sitive to this variable; the mean WTPs forsaving 2000, 20,000, or 200,000 birds were$80, $78, and $88, respectively.

A straightforward interpretation of thisresult involves the two acts of substitutionthat characterize prototype heuristics. Thedeaths of numerous birds are first repre-sented by a prototypical instance – perhapsan image of a bird soaked in oil and drown-ing. The prototype automatically evokesan affective response, and the intensity ofthat emotion is then mapped onto the dol-lar scale – substituting the readily accessi-ble heuristic attribute of affective intensity

for the more complex target attribute ofeconomic value. Other examples of radicalinsensitivity to scope lend themselves to asimilar interpretation. Among others, Kah-neman (1986) found that Toronto residentswere willing to pay almost as much to cleanup polluted lakes in a small region of On-tario as to clean up all the polluted lakes inOntario, and McFadden and Leonard (1993)reported that residents in four western stateswere willing to pay only 28% more to protect57 wilderness areas than to protect a singlearea (for more discussion of scope insensitiv-ity, see Frederick & Fischhoff, 1998).

The similarity between WTP statementsand categorical predictions is not limitedto such demonstrations of almost completeextension neglect. The two responses alsoyield similar results when extension andprototype information are varied factori-ally within subjects. Figure 1 2 .2(a) showsthe results of a study of WTP for pro-grams that prevented different levels ofdamage to species of varying popularity(Ritov & Kahneman, unpublished observa-tions, cited in Kahneman, Ritov, & Schkade,1999). As in the case of base rate [Figure1 2 .2(c)], extensional information (levels ofdamage) combines additively with nonex-tensional information. This rule of combina-tion is unreasonable; in any plausible theoryof value, the lines would fan out.

Finally, the role of the emotion evokedby a prototypical instance was also exam-ined directly in the same experiment, us-ing the heuristic elicitation paradigm intro-duced earlier: Some respondents were askedto imagine that they saw a television pro-gram documenting the effect of adverse eco-logical circumstances on individual mem-bers of different species. The respondentsindicated, for each species, how much con-cern they expected to feel while watchingsuch a documentary. The correlation be-tween this measure of affect and willingnessto pay, computed across species, was .97.

Duration Neglect in the Evaluationof Experiences

We next discuss experimental studies of theglobal evaluation of experiences that extend



over some time, such as a pleasant or ahorrific film clip (Fredrickson & Kahne-man, 1993), a prolonged unpleasant noise(Schreiber & Kahneman, 2000), pressurefrom a vise (Ariely, 1998), or a painful med-ical procedure (Redelmeier & Kahneman,1996). Participants in these studies provideda continuous or intermittent report of hedo-nic or affective state, using a designated scaleof momentary affect (Figure 1 2 .3). Whenthe episode had ended, they indicated aglobal evaluation of “the total pain or dis-comfort” associated with the entire episode.

We first examine the normative rules thatapply to this task. The global evaluation ofa temporally extended outcome is an exten-sional attribute, which is governed by a dis-tinctive logic. The most obvious rule is tem-poral monotonicity: There is a compellingintuition that adding an extra period of painto an episode of discomfort can only makeit worse overall. Thus, there are two waysof making a bad episode worse – makingthe discomfort more intense or prolongingit. It must therefore be possible to trade offintensity against duration. Formal analyseshave identified conditions under which thetotal utility of an episode is equal to thetemporal integral of a suitably transformedmeasure of the instantaneous utility associ-ated with each moment (Kahneman, 2000a;Kahneman, Wakker, & Sarin, 1997).

Next, we turn to the psychology.Fredrickson and Kahneman (1993) proposeda “snapshot model” for the retrospectiveevaluation of episodes, which again involvestwo acts of substitution: First, the episode isrepresented by a prototypical moment; next,the affective value attached to the represen-tative moment is substituted for the exten-sional target attribute of global evaluation.The snapshot model was tested in an exper-iment in which participants provided con-tinuous ratings of their affect while watch-ing plotless films that varied in duration andaffective value (e.g., fish swimming in coralreefs, pigs being beaten to death with clubs),and later reported global evaluations of theirexperiences. The central finding was that theretrospective evaluations of these observerswere predicted with substantial accuracy by

a simple average of the peak affect recordedduring a film and the end affect reported asthe film was about to end. This has beencalled the peak/end rule. However, the cor-relation between retrospective evaluationsand the duration of the films was negligible –a finding that Fredrickson and Kahneman la-beled duration neglect. The resemblance ofduration neglect to the neglect of scope andbase rate is striking and unlikely to be ac-cidental. In this analysis, all three are mani-festations of extension neglect caused by theuse of a prototype heuristic.

The peak/end rule and duration neglecthave both been confirmed on multiple oc-casions. Figure 1 2 .3 presents raw data froma study reported by Redelmeier and Kahne-man (1996), in which patients undergoingcolonoscopy reported their current level ofpain every 60 seconds throughout the proce-dure. Here again, an average of peak and endpain quite accurately predicted subsequentglobal evaluations and choices. The durationof the procedure varied considerably amongpatients (from 4 to 69 minutes), but thesedifferences were not reflected in subsequentglobal evaluations in accord with durationneglect. The implications of these psycho-logical rules of evaluation are paradoxical. InFigure 1 2 .3 , for example, it appears evidentthat patient B had a worse colonoscopy thanpatient A (on the assumption they used thescale similarly). However, it is also appar-ent that the peak/end average was worsefor patient A, whose procedure ended ata moment of relatively intense pain. Thepeak/end rule prediction for these two pro-files is that patient A would evaluate theprocedure more negatively than patient Band would be more likely to prefer to un-dergo a barium enema rather than a repeatcolonoscopy. The prediction was correct forthese two individuals and confirmed by thedata of a large group of patients.

The effects of substantial variations of du-ration remained small (although statisticallyrobust) even in studies conducted in a fac-torial design. Figure 1 2 .2(d) is drawn from astudy of responses to ischemic pain (Ariely,1998), in which duration varied by a factor of4 . The peak/end average accounted for 98%



Patient A

0

1

2

3

4

5

6

7

8

0 10 20

Time (minutes)

Pai

n In

ten

sity

Patient B

0

1

2

3

4

5

6

7

8

0 10 20Time (minutes)

Pai

n In

ten

sity

Figure 1 2 .3 . Pain intensity reported by twocolonoscopy patients.

of the systematic variance of global evalua-tions in that study and for 88% of the vari-ance in a similar factorial study of responsesto loud unpleasant sounds [Schreiber & Kah-neman, 2000, Figure 1 2 .2(b)]. Contrary tothe normative standard for an extensional at-tribute, the effects of duration and of otherdeterminants of evaluation were additive[Figures 1 2 .2(b) and 1 2 .2(d)].

The participants in these studies werewell aware of the relative duration of theirexperiences and did not consciously de-cide to ignore duration in their evalua-tions. As Fredrickson and Kahneman (1993)noted, duration neglect is an attentionalphenomenon:

. . . duration neglect does not implythat duration information is lost, northat people believe that duration is

unimportant . . . people may be aware ofduration and consider it important in theabstract [but] what comes most readily tomind in evaluating episodes are the salientmoments of those episodes and the affectassociated with those moments. Durationneglect might be overcome, we suppose, bydrawing attention more explicitly to theattribute of time. (p. 54)

This comment applies equally well toother instances of extension neglect: The ne-glect of base rate in categorical prediction,the neglect of scope in willingness to pay, theneglect of sample size in evaluations of ev-idence (Griffin & Tversky, 1992 ; Tversky &Kahneman, 1971 ), and the neglect of prob-ability of success in evaluating a program ofspecies preservation (DeKay & McClelland,1995). More generally, inattention plays asimilar role in any situation in which the in-tuitive judgments generated by system 1 vio-late rules that would be accepted as valid bythe more deliberate reasoning that we asso-ciate with system 2 . As we noted earlier, theresponsibility for these judgmental mishapsis properly shared by the two systems: Sys-tem 1 produces the initial error, and system2 fails to correct it, although it could.

Violations of Dominance

The conjunction fallacy observed in theLinda problem is an example of a domi-nance violation in judgment: Linda must beat least as likely to be a bank teller as tobe a feminist bank teller, but people be-lieve the opposite. Insensitivity to extension(in this case, base rate) effectively guaran-tees the existence of such dominance viola-tions. For another illustration, consider thequestion: “How many murders were therelast year in [Detroit/Michigan]?” Althoughthere cannot be more murders in Detroitthan in Michigan, because Michigan con-tains Detroit, the word “Detroit” evokes amore violent image than the word “Michi-gan” (except of course for people who im-mediately think of Detroit when Michiganis mentioned). If people use an impres-sion of violence as a heuristic and neglectgeographic extension, their estimates of



murders in the city may exceed their esti-mates for the state. In a large sample of Uni-versity of Arizona students, this hypothesiswas confirmed – the median estimate of thenumber of murders was 200 for Detroit and100 for Michigan.

Violations of dominance akin to the con-junction fallacy have been observed in sev-eral other experiments involving both indi-rect (between-subjects) and direct tests. In aclinical experiment reported by Redelmeier,Katz, and Kahneman (2001 ), half of a largegroup of patients (N = 682 ) undergoing acolonoscopy were randomly assigned to acondition that made the actual experiencestrictly worse. Unbeknownst to the patient,the physician deliberately delayed the re-moval of the colonoscope for approximately1 minute beyond the normal time. The in-strument was not moved during the extra pe-riod. For many patients, the mild discomfortof the added period was an improvementrelative to the pain than they had just ex-perienced. For these patients, of course, pro-longing the procedure reduced the peak/endaverage of discomfort. As expected, retro-spective evaluations were less negative inthe experimental group, and a 5 -year follow-up showed that participants in that groupwere also somewhat more likely to complywith recommendations to undergo a repeatcolonoscopy (Redelmeier, Katz, & Kahne-man, 2001 ).

In an experiment that is directly analo-gous to the demonstrations of the conjunc-tion fallacy, Kahneman et al. (1993) exposedparticipants to two cold-pressor experiences,one with each hand: a “short” episode (im-mersion of one hand in 1 4

◦C water for60 seconds), and a “long” episode (the shortepisode, plus an additional 30 seconds duringwhich the water was gradually warmed to1 5

◦C). The participants indicated the inten-sity of their pain throughout the experience.When they were later asked which of thetwo experiences they preferred to repeat,a substantial majority chose the long trial.These choices violate dominance, becauseafter 60 seconds in cold water anyone willprefer the immediate experience of a warmtowel to 30 extra seconds of slowly dimin-

ishing pain. In a replication, Schreiber andKahneman (2000, experiment 2) exposedparticipants to pairs of unpleasant noises inimmediate succession. The participants lis-tened to both sounds and chose one to be re-peated at the end of the session. The “short”noise lasted 8 seconds at 77 db. The “long”noise consisted of the short noise plus anextra period (of up to 24 seconds) at 66 db(less aversive, but still unpleasant and cer-tainly worse than silence). Here again, thelonger noise was preferred most of the time,and this unlikely preference persisted over aseries of five choices.

The violations of dominance in these di-rect tests are particularly surprising becausethe situation is completely transparent. Theparticipants in the experiments could eas-ily retrieve the durations of the two experi-ences between which they had to choose,but the results suggest that they simplyignored duration. A simple explanation isthat the results reflect “choosing by liking”(see Frederick, 2002). The participants inthe experiments simply followed the nor-mal strategy of choice: “When choosing be-tween two familiar options, consult your ret-rospective evaluations and choose the onethat you like most (or dislike least).” Lik-ing and disliking are products of system 1 ,which do not conform to the rules of ex-tensional logic. System 2 could have inter-vened, but in these experiments it generallydid not. Kahneman et al. (1993) described aparticipant in their study, who chose to re-peat the long cold-pressor experience. Soonafter the choice was recorded, the partic-ipant was asked which of the two expe-riences was longer. As he correctly identi-fied the long trial, the participant was heardto mutter “the choice I made doesn’t seemto make much sense.” Choosing by likingis a form of mindlessness (Langer, 1978),which illustrates the casual governance ofsystem 2 .

Like the conjunction fallacy in directtests, which we discussed earlier, violationsof temporal monotonicity in choices shouldbe viewed as an expendable flourish. Be-cause the two aversive experiences occurredwithin a few minutes of each other and



respondents could accurately recall the dura-tion of the two events, system 2 had enoughinformation to override choosing by liking.Its failure to do so is analogous to the fail-ures observed in direct tests of the Lindaproblem. In both cases, the violations ofdominance tell us nothing new about sys-tem 1 ; they only illustrate an unexpectedweakness of system 2 . Just as the theory ofintuitive categorical prediction would haveremained intact if the conjunction fallacyhad not “worked” in a direct test, the modelof evaluation by moments would have sur-vived even if violations of dominance hadbeen eliminated in highly transparent situa-tions. The same methodological issues arisein both contexts. Between-subjects experi-ments or subtle tests are most appropriatefor studying the basic intuitive evaluations ofsystem 1 , and also most likely to reveal com-plete extension neglect. Factorial designs inwhich extension is manipulated practicallyguarantee an effect of this variable, and al-most guarantee that it will be additive, asin Figures 1 2 .2(b) and 1 2 .2(d) (Ariely, 1998;Ariely, Kahneman, & Loewenstein, 2000;Schreiber & Kahneman, 2000). Finally, al-though direct choices sometimes yield sys-tematic violations of dominance, these vio-lations can be avoided by manipulations thatprompt system 2 to take control.

In our view, the similarity of the re-sults obtained in diverse contexts is a com-pelling argument for a unified interpreta-tion, and a significant challenge to critiquesthat pertain only to selected subsets of thisbody of evidence. A number of commenta-tors have offered competing interpretationsof base rate neglect (Cosmides & Tooby,1996; Koehler, 1996), insensitivity to scopein WTP (Kopp, 1992), and duration ne-glect (Ariely & Loewenstein, 2000). How-ever, these interpretations are generally spe-cific to a particular task and would not carryover to analogous findings in other domains.Similarly, the various attempts to explain theconjunction fallacy as an artifact do not ex-plain analogous violations of dominance inthe cold-pressor experiment. The accountwe have offered is, in contrast, equally ap-plicable to all three contexts and possibly

others (see also Kahneman, Ritov, &Schkade, 1999). We attribute extension ne-glect and violations of dominance to a lazysystem 2 , and to a prototype heuristic thatcombines two processes of system 1 : the rep-resentation of categories by prototypes andthe substitution of a nonextensional heuris-tic attribute for an extensional target at-tribute. We also propose that people havesome appreciation of the role of extensionin the various judgment tasks. Consequently,they will incorporate extension in their judg-ments when their attention is drawn to thisfactor – most reliably in factorial experi-ments, and sometimes (although not always)in direct tests. The challenge for compet-ing interpretations is to provide a unified ac-count of the diverse phenomena that havebeen considered in this section.

Conclusions and Future Directions

The original goal of the heuristics and biasesprogram was to understand intuitive judg-ment under uncertainty. Heuristics were de-scribed as a collection of disparate cognitiveprocedures, related only by their commonfunction in a particular judgmental domain –choice under uncertainty. It now appears,however, that judgment heuristics are ap-plied in a wide variety of domains and sharea common process of attribute substitution,in which difficult judgments are made bysubstituting conceptually or semantically re-lated assessments that are simpler and morereadily accessible.

The current treatment explicitly ad-dresses the conditions under which intu-itive judgments are modified or overridden.Although attribute substitution provides aninitial input into many judgments, it neednot be the sole basis for them. Initial impres-sions are often supplemented, moderated, oroverridden by other considerations, includ-ing the recognition of relevant logical rulesand the deliberate execution of learned al-gorithms. The role of these supplemental oralternative inputs depends on characteristicsof the judge and the judgment task.



Our use of the dual-process terminologydoes not entail a belief that every mentaloperation (including each postulated heuris-tic) can be definitively assigned to one sys-tem or the other. The placement of di-viding lines between “systems” is arbitrarybecause the bases by which we characterizemental operations (difficulty of acquisition,accessibility to introspection, and disrupt-ability) are all continua. However, this doesnot make distinctions less meaningful; thereis broad agreement that mental operationsrange from rapid, automatic, perception-likeimpressions to deliberate computations thatapply explicit rules or external aids.

Many have questioned the usefulnessof the notion of heuristics and biases bypointing to inconsistencies in the degree towhich illusions are manifested across differ-ent studies. However, there is no mysteryhere to explain. Experimental studies of “thesame” cognitive illusions can yield differentresults for two reasons: (1 ) because of vari-ation in factors that determine the accessi-bility of the intuitive illusion, and (2) be-cause they vary in factors that determine theaccessibility of the corrective thoughts thatare associated with system 2 . Both types ofvariation can often be anticipated becauseof the vast amount of psychological knowl-edge that has accumulated about the differ-ent sets of factors that determine the easewith which thoughts come to mind – fromprinciples of grouping in perception to prin-ciples that govern transfer of training in rulelearning (Kahneman, 2003). Experimentalsurprises will occur, of course, and shouldlead to refinements in the understanding ofthe rules of accessibility.

The argument that system 1 will be ex-pressed unless it is overridden by system 2

sounds circular, but it is not, because empir-ical criteria can be used to test whether a par-ticular characterization of the two systems isaccurate. For example, a feature of the situ-ation will be associated with system 2 if it isshown to influence judgments only when at-tention is explicitly directed to it (through,say, a within-subjects design). In contrast, avariable will be associated with system 1 if itcan be shown to influence even those judg-

ments that are made in a split second. Thus,one need not be committed, a priori, to as-signing a process to a particular system; thedata will dictate the best characterization.

The two-system model is a frameworkthat combines a set of empirical generaliza-tions about cognitive operations with a setof tests for diagnosing the types of cognitiveoperations that underlie judgments in spe-cific situations. The generalizations and thespecific predictions are testable and can berecognized as true or false. The frameworkitself will be judged by its usefulness as aheuristic for research.

Acknowledgments

This chapter is a modified version of achapter by Kahneman and Frederick (2002).Preparation of this chapter was supported bygrant SES-0213481 from the National Sci-ence Foundation.

Note

1 . The entries plotted in Figure 1 2 .1 are averagesof multiple judgments, and the correlations arecomputed over a set of judgment objects. Itshould be noted that correlations between av-erages are generally much higher than corre-sponding correlations within the data of indi-vidual respondents (Nickerson, 1995). Indeed,group results may even be unrepresentative ifthey are dominated by a few individuals whoproduce more variance than others and havean atypical pattern of responses. Fortunately,this particular hypothesis is not applicable tothe experiments of Figure 1 2 .1 , in which all re-sponses were ranks.

References

Agnoli, F. (1991 ). Development of judgmentalheuristics and logical reasoning: Training coun-teracts the representativeness heuristic. Cogni-tive Development, 6, 195–217.

Agnoli, F., & Krantz, D. H. (1989). Suppressingnatural heuristics by formal instruction: The



case of the conjunction fallacy. Cognitive Psy-chology, 2 1 , 515–550.

Anderson, N. H. (1991 ). Contributions to informa-tion integration theory. Vol. I: cognition. Hillsdale,NJ: Erlbaum.

Anderson, N. H. (1996). A functional theory ofcognition. Mahwah, NJ: Erlbaum.

Ariely, D. (1998). Combining experiences overtime: The effects of duration, intensity changesand on-line measurements on retrospectivepain evaluations. Journal of Behavioral DecisionMaking, 1 1 , 19–45 .

Ariely, D. (2001 ). Seeing sets: Representationby statistical properties. Psychological Science,1 2 (2), 1 57–162 .

Ariely, D., Kahneman, D., & Loewenstein, G.(2000). Joint comment on “When does dura-tion matter in judgment and decision making?”(Ariely & Loewenstein, 2000). Journal ofExperimental Psychology: General, 1 2 9, 524–529.

Ariely, D., & Loewenstein, G. (2000). When doesduration matter in judgment and decision mak-ing. Journal of Experimental Psychology: Gen-eral, 1 2 9, 508–523 .

Ayton, P. (1998). How bad is human judgment?In G. Wright & P. Goodwin (Eds.), Forecastingwith judgment. West Sussex, UK: Wiley.

Bargh, J. A. (1997). The automaticity of everydaylife. In R. S. Wyer, Jr. (Ed.), Advances in socialcognition (Vol. 10). Mahwah, NJ: Erlbaum.

Bargh, J. A., Bond, R. N., Lombardi, W. J., & Tota,M. E. (1986). The additive nature of chronicand temporary sources of construct accessibil-ity. Journal of Personality and Social Psychology,50(5), 869–878.

Bar-Hillel, M., & Fischhoff, B. (1981 ). When dobase-rates affect predictions? Journal of Per-sonality and Social Psychology, 41 (4), 671–680.

Bar-Hillel, M., & Neter, E. (1993). How alike isit versus how likely is it: A disjunction fallacyin probability judgments. Journal of Personalityand Social Psychology, 41 (4), 671–680.

Begg, I., Anas, A., & Farinacci, S. (1992). Disso-ciation of processes in belief: Source recollec-tion, statement familiarity, and the illusion oftruth. Journal of Experimental Psychology: Gen-eral, 1 2 1 (4), 446–458.

Begg, I., & Armour, V. (1991 ). Repetition andthe ring of truth: Biasing comments. Cana-dian Journal of Behavioural Science, 2 3 , 195–21 3 .

Birnbaum, M. H. (1999). How to show that 9 >

221 : Collect judgments in a between-subjectsdesign. Psychological Methods, 4(3), 243–249.

Birnbaum, M. H., & Mellers, B. A. (1983).Bayesian inference: Combining base rates withopinions of sources who vary in credibility.Journal of Personality and Social Psychology, 45 ,792–804 .

Bless, H., Clore, G. L., Schwarz, N., Golisano,V., Rabe, C., & Wolk, M. (1996). Mood andthe use of scripts: Does a happy mood reallylead to mindlessness? Journal of Personality andSocial Psychology, 71 (4), 665–679.

Bodenhausen, G. V. (1990). Stereotypes as judg-mental heuristics: Evidence of circadian vari-ations in discrimination. Psychological Science,1 (5), 319–322 .

Brunswik, E. (1943). Organismic achievementand environmental probability. PsychologicalReview, 50, 255–272 .

Cacioppo, J. T., Priester, J. R., & Berntson,G. G. (1993). Rudimentary determinants ofattitudes: II. Arm flexion and extension havedifferential effects on attitudes. Journal of Per-sonality and Social Psychology, 65 , 5–1 7.

Chaiken, S., & Trope, Y. (1999). Dual-process the-ories in social psychology. New York: GuilfordPress.

Chapman, G., & Johnson, E. (2002). Incorpo-rating the irrelevant: Anchors in judgments ofbelief and value. In T. Gilovich, D. Griffin, &D. Kahneman (eds.), Heuristics and biases: Thepsychology of intuitive judgment (pp. 1 20–1 38).New York: Cambridge University Press.

Chase, W. G., & Simon, H. A. (1973). Perceptionin chess. Cognitive Psychology, 4 , 55–81 .

Cohen, J. (1969). Statistical power analysis for thebehavioral sciences. San Diego: Academic Press.

Cohen, J. (1992). A power primer. PsychologicalBulletin, 1 1 2 (1 ), 1 55–1 59.

Cosmides L., & Tooby J. (1996). Are humansgood intuitive statisticians after all? Rethinkingsome conclusions from the literature on judg-ment under uncertainty. Cognition, 58(1 ), 1–73 .

DeKay, M. L., & McClelland, G. H. (1995). Prob-ability and utility components of endangeredspecies preservation programs. Journal of Ex-perimental Psychology: Applied, 2 , 60–83 .

Desvouges, W. H., Johnson, F., Dunford, R.,Hudson, S., Wilson, K., & Boyle, K. (1993).Measuring resource damages with contingentvaluation: Tests of validity and reliability.



In Contingent valuation: A critical assessment.Amsterdam: North Holland.

Diamond, P. (1996). Testing the internal consis-tency of contingent valuation surveys. Journalof Environmental Economics and Management,2 8, 1 55–1 73 .

Dulany, D. E. & Hilton, D. J. (1991 ). Conver-sational implicature, conscious representation,and the conjunction fallacy. Social Cognition, 9,85–1 10.

Epstein, S. (1994). Integration of the cognitiveand the psychodynamic unconscious. AmericanPsychologist, 49(8), 709–724 .

Evans, J., Handley, S. J., Perham, N., Over, D. E.,& Thompson, V. A. (2000). Frequency versusprobability formats in statistical word prob-lems. Cognition, 77, 197–21 3 .

Evans, J. St. B. T., & Over, D. E. (1996). Rationalityand reasoning. Hove, UK: Psychology Press.

Finucane, M. L., Alhakami, A., Slovic, P., &Johnson, S. M. (2000). The affect heuristic injudgments of risks and benefits. Journal of Be-havioral Decision Making, 1 3 , 1–1 7.

Frederick, S. W. (2002). Automated choiceheuristics. In T. Gilovich, D. Griffin, & D.Kahneman (eds.), Heuristics and biases: Thepsychology of intuitive judgment (pp. 548–558).New York: Cambridge University Press.

Frederick, S. W. (2004). (Im)patient, (un)-adventurous, (in)sensitive, and (im)moral: Cog-nitive (in)ability as a predictor of time preference,risk preference, magnitude sensitivity, and cheat-ing. Working paper, Massachusetts Institute ofTechnology, Cambridge, MA.

Frederick, S. W., & Fischhoff, B. (1998). Scope(in)sensitivity in elicited valuations. Risk, Deci-sion, and Policy, 3 , 109–1 23 .

Frederick, S. W., & Nelson, L. (2004). Attributesubstitution in physical judgments, Working pa-per, Massachusetts Institute of Technology,Cambridge, MA.

Fredrickson, B. L., & Kahneman, D. (1993). Du-ration neglect in retrospective evaluations ofaffective episodes. Journal of Personality and So-cial Psychology, 65(1 ), 45–55 .

Ganzach, Y., & Krantz, D. H. (1990). The psy-chology of moderate prediction: I. Experiencewith multiple determination. OrganizationalBehavior and Human Decision Processes, 47,1 77–204 .

Gigerenzer, G. (1991 ). How to make cognitiveillusions disappear: Beyond “heuristics and bi-ases.” In W. Stroebe & M. Hewstone (Eds.),

European review of social psychology (Vol. 2 ,pp. 83–1 1 5). Chichester, UK: Wiley.

Gigerenzer, G., Czerlinski, J., & Martignon,L. (2002). How good are fast and frugalheuristics? In T. Gilovich, D. Griffin, & D.Kahneman (Eds.), Heuristics and biases: Thepsychology of intuitive judgment (pp. 559–581 ).New York: Cambridge University Press.

Gigerenzer, G., & Goldstein, D. G. (1996).Reasoning the fast and frugal way: Modelsof bounded rationality. Psychological Review,1 03(4), 650–669.

Gigerenzer, G., Hell, W., & Blank, H. (1988). Pre-sentation and content – The use of base ratesas a continuous variable. Journal of Experimen-tal Psychology – Human Perception and Perfor-mance, 1 4(3), 513–525 .

Gigerenzer, G., & Hoffrage, U. (1995). Howto improve Bayesian reasoning withoutinstruction: Frequency formats. PsychologicalReview, 1 02 , 684–704 .

Gigerenzer, G., Todd, P. M., & the ABC Group.(1999). Simple heuristics that make us smart.New York: Oxford University Press.

Gilbert, D. (1989). Thinking lightly about others:Automatic components of the social inferenceprocess. In J. Uleman & J. A. Bargh (Eds.),Unintended thought (pp. 1 89–21 1 ). New York:Guilford Press.

Gilbert, D. (1991 ). How mental systems believe.American Psychologist, 46(2), 107–1 19.

Gilbert, D. (1999). What the mind’s not. In S.Chaiken & Y. Trope (Eds.), Dual-process theo-ries in social psychology (pp. 3–1 1 ). New York:Guilford Press.

Gould, S. J. (1991 ). Bully for brontosaurus. Reflec-tions in natural history. New York: Norton.

Griffin, D., & Tversky, A. (1992). The weighing ofevidence and the determinants of confidence.Cognitive Psychology, 2 4(3), 41 1–435 .

Hammond, K. R. (1955). Probabilistic function-ing and the clinical method. Psychological Re-view, 62 , 255–262 .

Hammond, K. R. (1996). Human judgment andsocial policy. New York: Oxford UniversityPress.

Heider, F. (1944). Social perception and phenom-enal causality. Psychological Review, 51 , 358–374 .

Hertwig, R., & Gigerenzer, G. (1999). The ‘con-junction fallacy’ revisited: How intelligent in-ferences look like reasoning errors. Journal ofBehavioral Decision Making, 1 2 (4), 275–305 .



Higgins, E. T., & Brendl, C. M. (1995). Accessi-bility and applicability: Some “activation rules”influencing judgment. Journal of ExperimentalSocial Psychology, 31 , 218–243 .

Hilton, D. J., & Slugoski, B. R. (2001 ). Conversa-tional processes in reasoning and explanation.In A. Tesser & N. Schwarz (Eds.), Blackwellhandbook of social psychology. Vol. 1 : Intrain-dividual processes (pp. 1 81–206). Oxford, UK:Blackwell.

Isen, A. M., Nygren, T. E., & Ashby, F. G. (1988).Influence of positive affect on the subjectiveutility of gains and losses – It is just not worththe risk. Journal of Personality and Social Psy-chology, 55(5), 710–717.

Jacoby, L. L., & Dallas, M. (1981 ). On the re-lationship between autobiographical memoryand perceptual learning. Journal of Experimen-tal Psychology: General, 3 , 306–340.

Kahneman, D. (1986). Valuing environmen-tal goods: An assessment of the contin-gent valuation method. In R. Cummings, D.Brookshire, & W. Schulze (Eds.), Valuing envi-ronmental goods: An assessment of the contingentvaluation method. Totowa, NJ: Rowman andAllanheld.

Kahneman, D. (2000a). Experienced utility andobjective happiness: A moment-based ap-proach. In D. Kahneman & A. Tversky (Eds.),Choices, values, and frames. New York: Cam-bridge University Press and the Russell SageFoundation.

Kahneman, D. (2000b). A psychological point ofview: Violations of rational rules as a diagnosticof mental processes (commentary on Stanovichand West). Behavioral and Brain Sciences, 2 3 ,681–683 .

Kahneman, D. (2003). A perspective on judg-ment and choice: Mapping bounded rational-ity. American Psychologist, 58, 697–720.

Kahneman, D., & Frederick, (2002). In T.Gilovich, D. Griffin, & D. Kahneman (Eds.),Heuristics of intuitive judgment. New York:Cambridge University Press.

Kahneman, D., Fredrickson, D. L., Schreiber,C. A., & Redelmeier, D. A. (1993). When morepain is preferred to less: Adding a better end.Psychological Science, 4 , 401–405 .

Kahneman, D., & Miller, D. T. (1986). Normtheory: Comparing reality with its alternatives.Psychological Review, 93 , 1 36–1 53 .

Kahneman, D., & Ritov, I. (1994). Determinantsof stated willingness to pay for public goods: A

study in the headline method. Journal of Riskand Uncertainty, 9, 5–38.

Kahneman, D., Ritov, I., & Schkade, D. (1999).Economic preferences or attitude expressions?An analysis of dollar responses to public issues.Journal of Risk and Uncertainty, 1 9, 203–235 .

Kahneman, D., Schkade, D. A., & Sunstein, C. R.(1998). Shared outrage and erratic awards: Thepsychology of punitive damages. Journal of Riskand Uncertainty, 1 6, 49–86.

Kahneman, D., Slovic, P., & Tversky, A. E. (1982).Judgment under uncertainty: Heuristics and bi-ases. New York: Cambridge University Press.

Kahneman, D., & Tversky, A. (1972). Subjectiveprobability: A judgment of representativeness.Cognitive Psychology, 3 , 430–454 .

Kahneman, D., & Tversky, A. (1973). On the psy-chology of prediction. Psychological Review, 80,237–251 .

Kahneman, D., & Tversky, A. (1982). On thestudy of statistical intuitions. Cognition, 1 1 ,1 23–1 41 .

Kahneman, D., & Tversky, A. (1996). On the re-ality of cognitive illusions: A reply to Gigeren-zer’s critique. Psychological Review, 1 03 , 582–591 .

Kahneman, D., & Varey, C. A. (1990). Propensi-ties and counterfactuals: The loser that almostwon. Journal of Personality and Social Psychol-ogy, 59(6), 1 101–1 1 10.

Kahneman, D., Wakker, P. P., & Sarin, R. (1997).Back to Bentham? Explorations of experiencedutility. Quarterly Journal of Economics, 1 1 2 ,375–405 .

Keysar, B. (1989). On the functional equivalenceof literal and metaphorical interpretations indiscourse. Journal of Memory and Language, 2 8,375–385 .

Koehler, J. (1996). The base-rate fallacy reconsi-dered: Descriptive, normative, and method-ological challenges. Behavioral and Brain Sci-ences, 1 9, 1–53 .

Kopp, R. (1992). Why existence value should beused in cost–benefit analysis. Journal of PolicyAnalysis and Management, 1 1 , 1 23–1 30.

Krosnick, J. A., Li, F., & Lehman, D. R. (1990).Conversational conventions, order of informa-tion acquisition, and the effect of base rates andindividuating information on social judgment.Journal of Personality and Social Psychology, 59,1 1 40–1 1 52 .

Langer, E. J. (1978). Rethinking the role ofthought in social interaction. In Harvey, Ickes,



& Kidd (Eds.), New directions in attribution re-search (Vol. 2). Potomac, MD: Erlbaum.

Macchi, L. (1995). Pragmatic aspects of the base-rate fallacy. Quarterly Journal of ExperimentalPsychology, 48A, 1 88–207.

Mandler, G., Hamson, C., & Dorfman, J. (1990).Tests of dual process theory – Word primingand recognition. Quarterly Journal of Experi-mental Psychology, 42 (4), 713–739.

Margolis, H. (1987). Patterns, thinking & cognition.Chicago: University of Chicago Press.

McFadden, D., & Leonard, G. K. (1993). Issuesin the contingent valuation of environmentalgoods: Methodologies for data collection andanalysis. In Hausman (Ed.), Contingent valua-tion. A critical assessment. Amsterdam: North-Holland.

Mellers, B., Hertwig, R., & Kahneman, D. (2001 ).Do frequency representations eliminate con-junction effects? An exercise in adversarial col-laboration. Psychological Science, 1 2 , 269–275 .

Michotte, A. (1963). The perception of causality.New York: Basic Books.

Mitchell, R., & Carson, R. (1989). Using surveysto value public goods: The contingent valuationmethod. Washington, DC: Resources for theFuture.

Nickerson, C. (1995). Does willingness-to-payreflect the purchase of moral satisfaction? A re-consideration of Kahneman and Knetsch. Jour-nal of Environmental Economics and Manage-ment, 2 8, 1 26–1 33 .

Nisbett, R. E., Krantz, D. H., Jepson, C., & Kunda,Z. (1983). The use of statistical heuristics ineveryday inductive reasoning. Psychological Re-view, 90(4), 339–363 .

Novemsky, N., & Kronzon, S. (1999). How arebase-rates used, when they are used: A compar-ison of Bayesian and additive models of base-rate use. Journal of Behavioral Decision Making,1 2 , 55–69.

Ordonez, L., & Benson, L., III. (1997). Decisionsunder time pressure: How time constraint af-fects risky decision making. Organizational Be-havior and Human Decision Processes, 71 (2),1 21–1 40.

Parducci, A. (1965). Category judgment: A range-frequency model. Psychological Review, 72 ,407–418.

Pashler, H. E. (1998). The psychology of attention.Cambridge, MA: MIT Press.

Poulton, (1989). Bias in quantifying judgments.London: Erlbaum.

Redelmeier, D., Katz, J., & Kahneman, D. (2001 ).Memories of colonoscopy: A randomized trial.

Redelmeier, D., & Kahneman, D. (1996). Pa-tients’ memories of painful medical treatments:Real-time and retrospective evaluations oftwo minimally invasive procedures. Pain, 66,3–8.

Schreiber, C., & Kahneman, D. (2000). Deter-minants of the remembered utility of aver-sive sounds. Journal of Experimental Psychology:General, 1 2 9, 27–42 .

Schwarz, N. (1996). Cognition and communi-cation: Judgmental biases, research methods, andthe logic of conversation. Mahwah, NJ: Erlbaum.

Schwarz, N., Bless, H., Strack, F., Klumpp, G.,Rittenauer-Schatka, H., & Simons, A. (1991 ).Ease of retrieval as information: Another lookat the availability heuristic. Journal of Personal-ity and Social Psychology, 61 , 195–202 .

Schwarz, N., & Clore, G. L. (1983). Mood, mis-attribution, and judgments of well-being: In-formative and directive functions of affectivestates. Journal of Personality and Social Psychol-ogy, 45(3), 513–523 .

Schwarz, N., Strack, F., Hilton, D., & Naderer, G.(1991 ). Base rates, representativeness, and thelogic of conversation: The contextual relevanceof “irrelevant” information. Social Cognition, 9,67–84 .

Schwarz, N., & Vaughn, L. A. (2002). The avail-ability heuristic revisited: Ease of recall andcontent of recall as distinct sources of infor-mation. In T. Gilovich, D. Griffin, & D. Kahne-man (Eds.), Heuristics & biases: The psychologyof intuitive judgment (pp. 103–1 19). New York:Cambridge University Press.

Shweder, R. (1977). Likeness and likelihood ineveryday thought – Magical thinking in judg-ments about personality. Current Anthropology,1 8(4), 637–658.

Simon, H. A., & Chase, W. G. (1973). Skill inchess. American Scientist, 61, 394–403 .

Sloman, S. A. (1996). The empirical case fortwo systems of reasoning. Psychological Bulletin,1 1 9, 3–22 .

Sloman, S. A. (2002). Two systems of reason-ing. In T. Gilovich, D. Griffin, & D. Kahne-man (Eds.), Heuristics & biases: The psychologyof intuitive judgment (pp. 379–396). New York:Cambridge University Press.

Slovic, P., Finucane, M., Peters, E., & MacGre-gor, D. G. (2002). The affect heuristic. In T.Gilovich, D. Griffin, & D. Kahneman (Eds.),



Heuristics & biases: The psychology of intu-itive judgment (pp. 397–420). New York: Cam-bridge University Press.

Stanovich, K. E. (1999). Who is rational?: Studiesof individual differences in reasoning. Mahwah,NJ: Erlbaum.

Stanovich, K. E., & West, R. (2002). Individualdifferences in reasoning: Implications for therationality debate? In T. Gilovich, D. Griffin,& D. Kahneman (Eds.), Heuristics & biases:The psychology of intuitive judgment (pp. 421–440). New York: Cambridge UniversityPress.

Stevens, S. S. (1957). On the psychophysical law.Psychological Review, 64 , 1 53–1 81 .

Stevens, S. S. (1975). Psychophysics: Introductionto its perceptual, neural, and social prospects.New York: Wiley.

Strack, F., Martin, L. L., & Schwarz, N. (1988).Priming and communication: The social deter-minants of information use in judgments of life-satisfaction. European Journal of Social Psychol-ogy, 1 8, 429–442 .

Tversky, A., & Kahneman, D. (1971 ). Belief in thelaw of small numbers. Psychological Bulletin, 76,105–1 10.

Tversky, A., & Kahneman, D. (1973). Availa-bility: A heuristic for judging frequency and

probability. Cognitive Psychology, 5(2), 207–232 .

Tversky, A., & Kahneman, D. (1974). Judgmentunder uncertainty: Heuristics and biases. Sci-ence, 1 85(4157), 1 1 24–1 1 3 1 .

Tversky, A., & Kahneman, D. (1982). Judgmentsof and by representativeness. In D. Kahneman,P. Slovic, & A. Tversky (Eds.), Judgment underuncertainty: Heuristics and biases (pp. 84–98).New York: Cambridge University Press.

Tversky, A., & Kahneman, D. (1983). Extensionalvs. intuitive reasoning: The conjunction fallacyin probability judgment. Psychological Review,90, 293–315 .

Tversky, A., & Kahneman, D. (1986). Rationalchoice and the framing of decisions. Journal ofBusiness, 59, S251–S278.

Wason, P. C. (1960). On the failure to elimi-nate hypotheses in a conceptual task. QuarterlyJournal of Experimental Psychology, 1 2 , 1 29–1 40.

Zajonc, R. B. (1980). Feeling and thinking: Pref-erences need no inferences. American Psychol-ogist, 35(2), 1 5 1–1 75 .

Zajonc, R. B. (1997). Emotions. In D. T. Gilbert,S. T. Fiske, & G. Lindzey (Eds.), Handbook ofsocial psychology (4th ed., pp. 591–632). NewYork: Oxford University Press.


294

chapter 12 a model of heuristic judgmenta model of heuristic judgment 269 in an optimal regression...

Documents