Affect, Generalization, and the Perception of Risk

1 Journal of Personality and Social Psychology1983, Vol. 45, No. 1, 20-31 Copyright 1983 by theAmerican Psychological Association, , generalization , and the Perception of RiskEric J. JohnsonCarnegie-Mellon UniversityAmos TverskyStanford UniversityExperimental manipulations of affect induced by a brief newspaper report of atragic event produced a pervasive increase in subjects' estimates of the frequencyof many risks and other undesirable events. Contrary to expectation, the effectwas independent of the similarity between the report arid the estimated risk. Anaccount of a fatal stabbing did not increase the frequency estimate of a closelyrelated risk, homicide, more than the estimates of unrelated risks such as naturalhazards. An account of a happy event that created positive affect produced acomparable global decrease in judged frequency of a society, we have never been more con-cerned with the assessment, the management,and the regulation of risk.

2 Because publicreaction to hazards from pesticides, nuclearpower, or food additives appears to influencethe regulation and management of these tech-nologies, it is important to understand howthe lay person perceives and evaluates is particularly true for hazards such asterrorism, nuclear power, or genetic engi-neering for which the available statistical dataare very limited and where the assessmentsof the risks are based on subjective and in-tuitive , psychologists and other researchershave shown increasing interest in the mannerin which people perceive and estimate theseverity of various risks. Lichtenstein, Slovic,Fischhoff, Layman, and Combs (1978) askedlay people to estimate the number of deathsper year that are due to various investigators found that the subjectiveestimates were consistent across severalmethods of elicitation and correlated reason-ably well (median r =.)

3 7) with actuarial es-timates of frequency. A comparison of ob-jective and subjective estimates revealed twoThis work has been supported in part by a NationalScience Foundation postdoctoral fellowship to the firstauthor and Office of Naval Research Grant NR 197-058to Stanford University. We thank Baruch Fischhoff, DonMacGregor, and Paul Slovic for their useful commentsand their generous help in conducting Experiments 1and for reprints should be sent to Eric J. John-son, Graduate School of Industrial Administration, Car-negie-Mellon University, Pittsburgh, Pennsylvania The primary bias refers to the com-mon tendency to overestimate infrequentcauses of death while underestimating morefrequent causes. Thus, people estimated thata relatively rare cause of fatality, polio, oc-curred about 10 times more frequently thanit actually does and underestimated the oc-currence of a common cause, lung cancer, byabout a factor of 10.

4 The secondary bias re-fers to the observation that "overestimatedcauses of death were generally dramatic andsensational whereas underestimated causestended to be unspectacular events, whichclaim one victim at a time" (Slovic, Fisch-hoff, & Lichtenstein, 1982, p. 467). This biascan be attributed to the availability heuristicin which one estimates the frequency of aclass by the ease with which its instances arebrought to mind (Tversky & Kahneman,1973). For example, homicide is perceivedto occur more frequently than suicide, andtwice as many people are thought to die fromfire than from drowning. Actually, this is nottrue. About 6,000 more people die in suicidesthan in homicides, and drowning causesdeath about as frequently as fire. Lichtensteinet al. suggested that these errors could reflectan availability bias induced by the media,which report homicide and fire more fre-quently than suicide and drowning.

5 A con-tent analysis of newspaper reports (Combs& Slovic, 1979) supported this and affect in Risk PerceptionOne characteristic that distinguishes judg-ments about risks from other estimates, suchas letter frequency, is that they seldom occur20 affect , generalization , AND THE Perception OF RISK21in an emotionally neutral context. When wewitness an accident or read a newspaper re-port about a natural disaster, we do notmerely revise our subjective probabilities; weare also shaken and disturbed. Our encoun-ters with news about risk and death com-monly generate fear, anxiety, and that you have just read a disturbingnewspaper account _of, say, the senselessshooting of an innocent passerby. Such anaccount, and the emotional response it elicits,may well increase your estimate of the num-ber of deaths due to homicide more than isjustified by the information contained in investigate the role of affect in judg-ments of risk we constructed brief accountsof tragic deaths of a single person from a spe-cific cause.

6 Like many newspaper stories, ouraccounts described details of the tragic inci-dent but gave no information about its prev-alence. Thus the stories were designed to gen-erate affect while conveying minimal dataabout the frequency of the relevant hazardin the general population. The effect of thesestories on the perceived frequency of deathscaused by various risks was studied by com-paring estimates made by two groups: onewho had read the stories and another whohad consider four possible effects of theexperimental treatment, ordered by the rangeof their impact:1. The stories may have no effect on es-timates of fatalities. This appears to be a nor-matively appropriate response because thestories, which describe the death of a singleperson, do not justify significant changes infrequency The stories may produce a local is, a story about homicide may increasethe estimated frequency of homicide, and soforth.

7 This could be caused by several mech-anisms. The incident described in the storycould serve as a retrieval cue, making similarinstances more available, thereby increasingthe judged frequency of the relevant risk. Thesubject in the study, like the reader of a news-paper, could also make the inference that theselection of a story about a particular causeof death indicates that this risk is especiallyserious or prevalent. Either mechanism re-sults in a local increase,'that is, a higher fre-quency estimate for the cause of death men-tioned in the The effect of the story could also gen-eralize to other risks according to their sim-ilarity to the risk described in the , a story about a leukemia victim, forexample, will be expected to have (a) a con-siderable effect on the estimated frequencyof other cancers, (b) a weaker effect on theestimated frequency of other diseases, and(c) little or no effect on the estimated fre-quency of unrelated risks ( , tornado orplane crashes).

8 This hypothesis (which ap-peared most probable to us) is suggested bythe classical notion that the gradient of gen-eralization is determined by the similaritybetween the critical stimulus and the The story could produce a global effect;that is, an increase in the judged frequencyof both related and unrelated risks. There isevidence in the social psychological literaturethat changes in affect could influence a widerange of unrelated behaviors and judgmentssuch as the frequency of helping (Isen &Levin, 1972) and the evaluations of the per-formance of household appliances (Isen,Shalker, Clark, & Karp, 1978). (For a reviewof some of the relevant literature, see Clarkand Isen, 1982).Thus, we consider four possible effects ofthe exposure to an affect -arousing story: (a)no increase in fatality estimates; (b) a localincrease, limited to the target risk; (c) a gen-eralization gradient; and (d) a global four possible effects are not necessarilyincompatible; both local and global increasescould be found in the same set of Set of RisksTo generate a set of risks for the presentstudies, we asked a group of 68 undergrad-uates from Stanford University to list themajor risks, including diseases, accidents,and disasters, that are important causes ofdeath in the United States.

9 Of the 36 riskslisted by the subjects, we selected a set of 18risks that were either the most frequentcauses of death or were listed most frequentlyby the ascertain the perceived structure ofthese risks, we presented all pairs of 18 risksto 110 subjects recruited through an adver-22 ERIC J. JOHNSON AND AMOS TVERSKY tisement in the University of Oregon studentnewspaper. They were asked to rate the sim-ilarity between risks on a scale from 1 (lowsimilarity) to 10 (high similarity). Threegroups of subjects each rated one third of thetotal of 153 pairs. The average similaritiesbetween risks were analyzed using the ADD-TREE algorithm (Sattath & Tversky, 1977).The resulting tree representation of the prox-imity among the 18 risks is shown in Fig-ure an ADDTREE (or additive tree) repre-sentation, the objects (risks) appear as theterminal nodes of the tree, and the distancebetween objects is the length of the horizontalpart of the path that connects them.

10 (Thevertical part is included for graphical con-venience.) Thus, leukemia is close to stomachcancer, less close to stroke, and quite distantfrom war and from terrorism. The length ofeach horizontal link in the tree can be inter-preted as the measure of the set of featuresthat are shared by all objects that originatefrom this link. Hence, the terminal links rep-resent the unique features of the respectiveobjects, whereas the nonterminal links cor-respond to the clusters induced by the judgedsimilarities. A suggested interpretation ofthese clusters, or links, is provided by itali-cized labels. (For a discussion of tree andother representations of risks, see Johnson& Tversky, in press.)The tree displayed in Figure 1 fit the judgedsimilarities quite well: The product-momentcorrelation between them is.