Affect, Generalization, and the Perception of Riskheatherlench.com/wp-content/uploads/2008/07/johnson-tversky.pdfExperimental manipulations of affect induced by a brief newspaper report

Journal of Personality and Social Psychology1983, Vol. 45, No. 1, 20-31

Copyright 1983 by theAmerican Psychological Association, Inc.

Affect, Generalization, and the Perception of Risk

Eric J. JohnsonCarnegie-Mellon University

Amos TverskyStanford University

Experimental 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 risks.

As a society, we have never been more con-cerned with the assessment, the management,and the regulation of risk. 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 risks.This 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 judgments.

Indeed, 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 hazards.These investigators found that the subjectiveestimates were consistent across severalmethods of elicitation and correlated reason-ably well (median r = .7) with actuarial es-timates of frequency. A comparison of ob-jective and subjective estimates revealed two

This 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 2.

Requests for reprints should be sent to Eric J. John-son, Graduate School of Industrial Administration, Car-negie-Mellon University, Pittsburgh, Pennsylvania 15213.

biases. 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. 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. A con-tent analysis of newspaper reports (Combs& Slovic, 1979) supported this interpretation.

Cognition and Affect in Risk Perception

One characteristic that distinguishes judg-ments about risks from other estimates, suchas letter frequency, is that they seldom occur

20

AFFECT, GENERALIZATION, AND THE PERCEPTION OF RISK 21

in 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 worry.Imagine 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 thereport.

To 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. 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 not.

We 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 estimates.

2. The stories may produce a local effect.That is, a story about homicide may increasethe estimated frequency of homicide, and soforth. 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 story.

3. The effect of the story could also gen-eralize to other risks according to their sim-ilarity to the risk described in the story.Hence, 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 (e.g., tornado orplane crashes). 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 target.

4. 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 increase.The four possible effects are not necessarilyincompatible; both local and global increasescould be found in the same set of data.

The Set of Risks

To 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. 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 subjects.

To 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 1.

In 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. (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 .92. It also yieldsseveral distinct and interpretable clusters, in-cluding diseases, violent acts, and hazards. Tostudy the impact of mood on risk estimates,we selected one risk from each of these clus-ters (leukemia, homicide, fire) as target risksand constructed a newspaperlike account foreach. We have also identified the two nearestneighbors of these risks as near-target risks.In Figure 1, the three target risks are under-lined, and the respective near-target risks arestarred.

Our first study includes three experimentalgroups, each of which read a single storyabout a death caused either by leukemia,homicide, or fire. The fourth group served asa control and did not read stories related todeath. Each experimental condition con-tained three types of risks: the target (de-scribed in the story), the near-target (its twoclosest neighbors), and the nontarget risks(the remainder).

Experiment 1

Method

Subjects. Seventy-two paid subjects, about equallydivided between men and women, participated in theexperiment. They were recruited through an advertise-ment placed in the University of Oregon newspaper, and

HAZARD

TRANSPORACCIDENTS

VIOLEN

TECHNOLOG

DISEASES

NATURAL TORNADOS

WMiltW. '1 FLOOD"

LIGHTNING*

1 , ELECTROCUTION*

rATION, ACCIDENTAL FALLS

I . inni n MI- a i-i- rii-MTi-

„.„„ HOMICIOFT ACTS \ - F

\ , , , , „„ _ __ JOXtC CHEMICAL SPILLS

WHONAHY J" ' STROKE

DISEASE

M

^ CANCER*TAMPPD* .

Figure 1. Additive tree representation of risks used in Experiment 1.


tested in moderate-sized groups in a classroom setting.Subjects completed the two relevant questionnaires alongwith several other, unrelated experimental tasks.

Procedure. In the first questionnaire, entitled "News-paper Reporting Study," the subjects were given the fol-lowing instructions:

A recent trend in journalism has been the increase ofpersonal interest and feature stories . . . describingboth good news such as personal successes, as well asaccounts of death and disaster,

The subjects read a few stories of this type and ratedeach story on a 9-point scale with respect to interest andquality of writing. They indicated how they would feelif they had read the story in their local newspaper on amood scale ranging from "positive, uplifted" to "nega-tive, depressed." This last scale served as a manipulationcheck.

All four groups were presented with two brief andmundane items, which were two paragraphs long, mod-eled after "People in the News" columns in local papers.

The three experimental groups each received an ad-ditional story about the death of a single person. Theexperimental stories that describe the fatal events con-sisted of three paragraphs. The portrayal of the deathwas detailed, designed to induce anxiety and worry. Thevictim in each case was a young male who appeared tobe an unsuspecting normal undergraduate prior to theonset of the fatal event.

Following the completion of this questionnaire, theparticipants were presented with a second booklet, en-titled "Perception of Risk Questionnaire," containing thedependent measures. In the first part of this question-naire, the subjects were asked to express, on a 9-pointscale, their level of worry and concern for each of the18 causes of death. In the second part, the subjects wereasked to estimate the frequency of various fatalities. Asin the study of Lichtenstein et al. (1978), subjects weretold that each year about 50,000 people in the UnitedStates die in motor vehicle accidents. They were thenasked to estimate the number of annual fatalities due toeach of the remaining 17 causes of death. They wereurged to be as accurate as possible and to check theiranswers for consistency.

Data analysis. The frequency estimates spanned sev-eral orders of magnitude, producing very skewed distri-butions. Consequently, all observed estimates were sub-jected to a logarithmic transformation, yielding approx-imately normal distributions. Note that the mean of thetransformed values is the logarithm of the geometricmean.

All the experiments reported in this article use a sim-ilar design. The stories represent a between-subjects fac-tor, whereas the risks are a within-subjects factor. Sig-nificance tests were performed using the subjects factoras the error term for the story effect and the Subjects XRisk X Story interaction for the effects of risk and theRisk X Story interaction.

Results

Manipulation checks. As expected, thestories presented to the experimental groups

were judged more depressing than the fillerstories. The means of the two filler storieswere 4.27 and 5.67 on the 9-point scale,whereas the means of the critical stories were6.90 for the leukemia story, 6.89 for the firestory, and 7.12 for the crime story. This com-parison of filler versus critical stories is sig-nificant by an a priori comparison, F(l,120)= 114.6, p<, 001.

Worry scales. The stories were also effec-tive in raising the level of concern for variouscauses of death: In each condition the meanof reported worry increased relative to thecontrol (mean increase = .75 scale unit). Ofa total of 54 comparisons (3 conditions X 18risks) between the experimental and controlgroups, 46 were higher in the experimentalgroup. The increase in worry, however, wasnot related to the target risk, or to the matchbetween the assessed risk and the risk men-tioned in the story.

Frequency estimates. Our major depen-dent measure, the frequency estimates, re-veals a strong global increase. Of 51 com-parisons (3 experimental groups with 17 riskseach), 41 show increases relative to control.Overall, the frequency estimates of the threeexperimental groups were 74% higher than

' those of the control group. The crime ac-count was the most effective of the three sto-ries, increasing estimates by an average of133%. The leukemia and the fire account in-creased the average estimate by 56% and 50%,respectively.

lb test for a local effect, in the presenceof a global effect, we compared the estimateof each target risk across each of the threeexperimental conditions. A local increase im-plies a larger increase for targets when theymatch the topic of the story than when thestory and target do not match. Gradient gen-eralization implies, in addition, a larger in-crease for the near-target risks when theymatch the topic of the story. Table 1 presentsthe ratios of the geometric means of the ex-perimental estimates to the geometric meansof the control estimates for target, near-target,and nontarget risks when matched and notmatched by the story. Note that this ratio ishigher in the unmatched condition than inthe matched condition for leukemia and fire,contrary to the prediction of the local andgradient hypotheses. For crime, the reverse


Table 1Ratios of Experiment to Control Means for Matched and Nonmatched Stories: Experiment

Type of risk

TargetNear target

. Nontarget

TargetNear targetNontarget

TargetNear targetNon target

Risk

DiseasesLeukemiaLung cancer, stomach cancerAll other risks

HazardsFireElectrocution, lightningAll other risks

ViolenceHomicideWar, terrorismAll other risks

Story

Matched Nonmatched

3.52 3.653.45 5.021.37 1.55

.941.341.52

1.22.75

2.86

.50

.75

.97

.16

.91

.53

M

4.20

1.49

1.51

1.67

1.41

1.97

Note. Means are geometric.

is true, but the effect does not approach sta-tistical significance by an a priori contrast,F(\, 1,474) = .10, p > .50. Thus, there is noevidence of either a local or gradient effect.

These conclusions are confirmed by theresults of an analysis of variance (ANOVA).The a priori contrast of the three experimen-tal groups to the control is significant, F(3,69) = 2.76, p < .05, but the overall effect ofthe Story X Risk interaction is not, /r(48,1,078)= 1.04,p> .40.

To obtain an alternative test of the gradienthypothesis that does not depend on thechoice of near-target risks, we correlated, foreach experimental story, the change in riskestimate (relative to control) induced by thatstory with the similarity of each risk to thetarget of that story. The gradient hypothesisimplies a positive correlation between themagnitude of the effect and the similarity tothe target. The product-moment correlationswere -.06, -.32, and .21, for fire, crime, andleukemia, respectively, none of which are sig-nificantly different from zero.

Although we found no significant differ-ences among the experimental stories, wefound sizable differences among the risks.Table 1 shows that, on average, negativemood induces a larger increase for estimatesof diseases than for hazards and for violence,a pattern confirmed by a post hoc compari-son, F(2, 1,078) = 4.23, p < .05.

Accuracy of estimates. As a sidelight, itis interesting to examine the accuracy of thesubjects' estimates. Figure 2 plots the avail-able statistical estimates for 13 of the 18causes of death against the values estimatedby the control group, along with the best-fit-ting linear regression line. The correlationbetween the subjective and the objective val-ues is .48. Figure 2 reveals a strong tendencyto underestimate large risks and a weak ten-dency to overestimate small risks. Becausethe current set of risks included fewer ex-tremely rare causes of death (less than 100fatalities a year) than did the study by Lich-tenstein et al. (1978), the overestimation ofinfrequent risks is less pronounced in ourdata. For similar reasons, the regression ofsubjective against objective estimates is fairlylinear in the present study and quadratic inthe study of Lichtenstein et al.

Discussion

Much to our surprise, we found a largeglobal increase in estimated frequency butfailed to detect either local or gradient effects.The story influenced the estimates for theentire group of risks regardless of the simi-larity between the estimated risk and thetopic of the story.

Experiment 2 was designed to obtain localand gradient effects by constructing a more


sensitive test. First, we added a hew task thatrequires the subjects in each condition torank the risks with respect to the number offatalities. This procedure induces a directcomparison between target and nontargetrisks that does not require numerical esti-mates of frequency. Second, we reduced theset of risks to seven: the three target risks,each target's most similar (near-target) riskfrom Experiment 1, and one nontarget risk.A local or a gradient effect may be more pro-nounced when the task is made simpler byreducing the number of estimates. Third, westrengthened the experimental manipulationby placing the stories and the frequency es-timates in the same questionnaire. Finally,we increased the statistical power of the testby doubling the total number of subjects.

Experiment 2

Method

Subjects. A new group of 186 subjects were recruitedthrough an advertisement placed in the University ofOregon newspaper. As in Experiment 1, they completedseveral unrelated tasks along with the current question-naire.

Procedure. The three experimental groups receivedthe two filler stories plus one of the critical accountsfrom Experiment 1. As before, they rated the stories for

interest, quality of writing, and the mood evoked. Thecontrol group did not read these accounts and proceededto the next section of the questionnaire.

All groups then ranked the seven risks in terms of thelikelihood that the risk would be the cause of their owndeath. Afterward, they were given the same instructionas in Experiment 1 and asked to estimate the frequency'of deaths for the seven causes,

Results

Manipulation checks. Again, the storieseffectively induced a negative mood. Themeans on the mood scale (6.82 for crime,6.67 for leukemia, and 6.63 for fire) werehigher than the values for the filler stories andclose to the corresponding values obtainedin Experiment 1.

Ranking of risks. An examination of themean rank in each condition shows no ten-dency to assign a higher rank to a risk whenit matches the topic of the story than whenit does not. For leukemia, the matched storyproduced a .30 decrease in mean rank. Whenmatched by the story, fire and street crimeshow increases of .50 and .22, respectively,but neither is statistically significant. TheRisk X Story interaction, which should re-flect a local or generalization effect, does notapproach significance, F(IS, 1,092) = .71,

100,000

10,000 •

Scc

8

01

1000 •

100 1000 10,000 100.000

ACTUAL FREQUENCY

Figure 2. Subjective versus objective estimates of frequency in Experiment 1.


p > .50. Furthermore, the a priori contrastcomparing the targets' ranking in the controlto the targets' ranking in the appropriate ex-perimental condition is also not significant,F(\, 1,092) = .01, p> .50. Ranking data, ofcourse, do not provide a test of a global effect.

Frequency estimates. Again the resultsshow a global increase and ho evidence fora local increase or for a generalization gra-dient. The crime story was the most effective,raising the mean estimate 144%, followed bythe leukemia story, which raised the esti-mates by 73%. The fire story was the leasteffective, showing an overall increase of only14%. The experimental groups' estimateswere larger than the control groups' for 19of the 21 comparisons.

Table 2 presents the ratios of the geometricmeans of experimental to control estimatesfor target, near-target, and nontarget risk es-timates. Recall that a local effect implies thatthe ratio is higher when the target matchesthe topic of the story than when it does not.This pattern appears to hold for leukemiaand crime, whereas fire reveals the oppositepattern. Closer examination shows that theeffect for leukemia is far from significant.Although crime produces a substantial effect,the increase is due to the greater effectivenessof the crime story. As seen in Table 2, thenontarget risks show larger increases (2.83)than either the target (2.38) or near-target

(1.16) risks. An a priori contrast comparingthe increase in the target does not differ fromthe overall increase, F(l, 1,083) = 1.15,p > .30.

An ANOVA confirms the effect of the threestories; the main effect of story is significant,F(3, 182) = 4.11, p < .01, as is the contrastcomparing the control to the three experi-mental groups, F(\, 182) = 5.86, p < .025.The lack of local and gradient effects is againdemonstrated by the insignificant Risk XStory interaction, F( 18,1,083)= ,80,;? > .65.

Table 2 also snows a differential mood ef-fect on the various risks. In particular, theimpact of the stories is largest for diseases(as in Experiment 1), smallest for violence,and intermediate for hazards. This orderingis significant, as shown by a post hoc contrast,F(2, 1,083) = 9.85, p < .001. Evidently, es-timates of diseases tend to be more volatilethan estimates of other risks.

Discussion

The major conclusion of these analyses isthe robustness of the global increases in per-ceived frequency of death. Several changesin the experimental procedure designed toinduce a local effect did not produce the ex-pected result. These data demonstrate thatbad mood created by brief stories have per-vasive global effects on estimates of fatalities.

Table 2Ratios of Experimental to Control Means for Matched and Nonmatched Stories: Experiment 2

Story

Type of risk




Risk

DiseasesLeukemiaLung cancer, stomach cancerAll other risks

HazardsFireElectrocution, lightningAH other risks

ViolenceHomicideWar, terrorismAll other risks

Matched

2.702.291.45

1.621.131.17

2.381.162.83

Nonmatched

2.491.951.49

3.021.661.34

1.31.76

1.64

M '

") ">i

1.48

I 84

1.28

1 1 *

1.87



However, we have found no connection be-tween the information contained in a storyand its impact on the estimated frequency ofdeath. The overriding factor in these in-creases is not the story ,told but rather themood it induces in the reader. Hence, it isnatural to inquire whether a sad story, un-related to risk, would also increase estimatesof fatality, and whether the effect of the pres-ent stories extends to the judged frequencyof undesirable events that do not introducethe risk of death. In the next experiment weinclude a nonrisk story and extend the list ofestimates to include nonfatal events like dam-age from floods and adverse life events suchas divorce and bankruptcy.

Experiment 3

MethodSubjects. One hundred ninety-one students in an in-

troductory psychology class at the University of Califor-nia, Berkeley, completed this questionnaire, along withseveral others at the beginning of a regularly scheduledclass meeting. Eight subjects were discarded from theanalysis because of incomplete data or nonsensical re-sponses.

Procedure. Experiment 3 used the Newspaper Re-porting Study questionnaire, which required subjects toread several stories and rate them on three scales. Thecontrol group received the two filler stories used in Ex-periment 1. There were two experimental groups: Oneread the crime story, which was the most effective in thetwo previous studies; the second read a depressing storyunrelated to risk. The story described several unfortu-nate events that occurred to a young male who had justbroken up with his girlfriend, was under stress at his job,and was further pressured by his family.

Upon completion of the Newspaper Reporting Study,subjects received a Perception of Risk questionnaire inwhich they were asked to estimate the frequency of some21 undesirable events. The instructions and the formatwere the same as in Experiment 1. The list included, inaddition to the seven risks from Experiment 2, two newtypes of events. The first involves nonfatal risks, that is,the number of people who suffer, but do not necessarilydie, from various causes. The second involves events notnormally viewed as risks that have a negative impact onthe quality of life. The entire set of events used in thisstudy is presented in the left-hand column of Figure 3.Thus, Experiment 3 extends our previous studies by us-ing a nonrisk story and by investigating nonfatal events.

Results

Manipulation checks. The depression storywas as effective as the crime story in inducingnegative affect, The means on the mood scalewere 6.92 and 6.93, respectively, which are

Table 3Ratios of Experimental to Control Means:Experiment 3

Story

Type of risk

Fatal risksNonfatal risksLife problems

Crime

1.301.732.63

Depression

1.231.632.82

M

1.271.652.75


comparable to the results of the first two stud-ies. The mean for the filler stories in the con-trol group was 5.42. The differences betweenthe means are highly significant by an a prioricontrast of the experimental groups to thecontrol, F(l, 183) = 31.61, p < .001.

Frequency estimates. The street crimeand the depression stories increased thejudged frequencies by 76% and 75%, respec-tively. An ANOVA confirms the main effect ofstory, F(2, 183) = 3.06, p < .05, and the pat-tern of the increases is quite similar in thetwo experimental groups. There was no hintof a local increase or a generalization gra-dient. For example, the increase in the esti-mated frequency of crime and chronicdepression,-which could be considered thematched risks for the two stories, are not sig-nificantly different, F(l, 3,575) = .01, andF(l, 3,575) = .69 (both ps > .40). Again, itappears that negative mood increases thejudged frequency of negative events, regard-less of the relation between the content of thestory and the estimated quantity.

Table 3 displays, for each story, the ratioof the experimental group geometric meansof experimental to control estimates for eachtype of event. The table indicates that themagnitude of the mood effect varies for thethree types of event. Life problems yieldedlarger ratios than did nonfatal risks, whichin turn were larger than the ratios for thefatalities. This is confirmed by a post hoccontrast, F(2, 3,575) = 11.81, p < .01.

The finding that negative mood increasesrisk estimates, independent of the relationbetween the risk and the story, suggests thatpositive mood could produce a similar globalreduction in risk estimates. This hypothesisis tested in our last experiment.


Experiment 4

Method

Subjects. A total of 108 subjects participated in thisexperiment. Eighty-eight were recruited through signsplaced on the Carnegie-Mellon University campus. Thesubjects were paid for their participation in the study,which included several additional experimental tasks. Anadditional 20 subjects were undergraduates who com-pleted the questionnaire at the beginning of regularlyscheduled class meetings.

Procedure. The study was similar to Experiment 3but used a single experimental group, who read a storydesigned to produce positive affect. It described a seriesof fortunate events occurring to a young male, includingadmission to medical school and success on a difficultexam.

Results

Manipulation checks. The positive storyshows a significant decrease in reported worrycompared to the control, 5.56 versus 4.70.

This effect is significant, F(l, 106) = 12.25,p< .001.

Frequency estimates. Positive mood re-sulted in lower estimated frequencies for 20of the 21 estimates. The average decrease was89%, which is comparable in magnitude tothe increase in estimates caused by the neg-ative story. An ANOVA confirms the signifi-cance of the effect of the story, F(\, 106) =5.78, p < .01, and the lack of an interactionbetween story and risk, F(2Q, 2,120) = 1.06;p > .60.

Figure 3 displays the difference (in logunits) between experimental and control es-timates for the 21 events estimated in Ex-periments 3 and 4. The data from Experi-ment 3, which induced negative mood, isshown by black horizontal bars, and the datafrom Experiment 4, which induced positivemood, is shown in white. If the mood ma-nipulation had no effect, the bars would van-ish, representing no change in the estimates.

a:"5

Airplane AccidentsFire

Leukemia

War and Terrorism

Stomach Cancer

Electrocution

Street Crime

TornadoFloods

Toxic ChemicalsStroke

Heart Attacks

Lightning

Accidental Falls

Malnutrition

Bankruptcy

AlcoholismDepression

UnemploymentDivorce

Sexual Impotence

LegendNegative A f f e c tPositive Af fect

Decreases Increases

Figure 3. Increases and decreases (log scale) in estimated frequency, relative to control, induced by positiveor negative affect for each of 21 risks.


• Figure 3 highlights the effectiveness of themanipulation of mood. With few exceptionsthe black bars are to the right of the zero line,representing increases.in estimated fre-quency, whereas the white bars are to the left,representing decreases. Furthermore, Figure3 reveals that the effect of mood, both posi-tive and negative, tends to be larger for lifeproblems than for fatal and nonfatal risks.Note that the life problems used in this studyare generally more frequent than the risks.The data in Figure 3 suggest the hypothesisthat frequent events are more labile than lessfrequent events. This hypothesis is consistentwith the results of Experiments 1 and 2 inwhich the most frequent causes of death—diseases—were more susceptible to the ex-perimental manipulation than were otherrisks.

To explore the relationship between per-ceived frequency and lability of estimates, letXt and d, respectively, be the geometricmeans of the estimated frequency of event/ in the experimental and control groups. Theratio, EJ = Xi/C,, measures the effect of theexperimental manipulation. The correlationbetween Et and C, across all events is .74 forExperiment 3 and —.46 for Experiment 4.

Note that proportional increases or de-creases in frequency estimates as a conse-quence of a mood effect should produce acorrelation of zero between C, and /?,-. Thepresence of substantial positive and negativecorrelations indicates that the impact ofmood is larger than predicted by a propor-tional effect. Other variables besides per-ceived frequency, such as knowledge, ambi-guity, and the nature of the risk, can un-doubtedly influence the volatility of theestimates. The investigation of these factorsrequires a design in which (perceived) fre-quency of risk is not confounded with thesefactors.

General Discussion

The results of the preceding studies dem-onstrate that mood induced by brief reportshas a large and pervasive impact on estimatesof the frequency of risks and other undesir-able events. Furthermore, the effect is inde-pendent of the similarity between the story

Table 4Percentage of Subjects Increasing Estimates forRisks that Match and Do Not Match anUnderestimated Target

Under-estimated

risk

Leukemia

Fire

Crime

Estimate type

Near-Target(lung cancer,stomach cancer)

Nontarget(all others)

Near-Target(lightning,electrocution)


Near-Target(war, terrorism)


Matched

72

19

51

31

60

16

Nonmatched

13

23

10

17

22

25

and the risk. The latter result contrastssharply with the typical finding that the de-gree of generalization is largely determinedby similarity. For example, we have obtaineda highly ordered generalization gradient overthe set of 18 risks from Experiment 1, usinga conditional revision procedure (Johnson& Tversky, in press). In this study, subjectsfirst estimated the frequency of death due tothe 18 risks. They were then told to assumethat they had underestimated a particular tar-get risk and asked to indicate the other esti-mates that they would increase given this in-formation. The gradient hypothesis impliesthat the more similar, near-target risks willbe revised more often than the unrelated,nontarget risks. Table 4 summarizes the testof this hypothesis for the three target risksused in Experiments 1 and 2. The matchedcolumn presents the percentage of subjectswho increased their estimates of near-targetrisks. The nonmatched column presents thepercentage of subjects who increased theirestimates for the corresponding risks whenthey do not match the target. Table 4 showsa small overall increase for nontarget risks,and a large increase in the matched near-tar-get risks. These results reveal a small globaleffect and a strong generalization gradient.Evidently the effect of the information pro-vided by the stories in Experiment 1 and 2


was different than the effect of the informa-tion that a particular risk was underesti-mated. The former effect produced a largeglobal effect, whereas the latter yielded theexpected gradient of generalization.

The pervasive global effect of mood andthe absence of a local effect pose a seriousproblem to memory-based models of thiseffect, such as spreading activation within asemantic network. In such models, the im-pact of an experience is largely determinedby the strength of association between theinput (e.g., the story) and the target (e.g., therisks). Risks that are closely linked to thestory should be influenced more than unre-lated risks, .contrary to the present findings.The novel aspect of these results is not theglobal mood effect, which has been observedby several investigators (see, e.g., Bower,1981; Feather, 1966; Clark & Isen, 1982;Bower & Wright, Note 1), but rather the pres-ence of a pervasive global effect in the absenceof any local effect of similarity or association.This combination is particularly surprisingbecause the effect of mood was produced notby an arousing experience but rather by abrief mundane account of a specific event.Evidently, people dissociate the affective im-pact of the account from its content. Theseobservations are consistent with the view thatthe influence of affect is at least partially in-dependent of semantic association (Zajonc,1980). The results give rise to the hypothesisthat we tend to make judgments that arecompatible with our current mood, evenwhen the subject matter is unrelated to thecause of that mood. If we attend the theaterin a bad mood, for example, we are likely tobe critical of the play, and although we donot normally attribute our mood to the play,we often perceive characteristics in the playthat could have produced our current mood.

This hypothesis suggests that people areoften not aware of the linkage between thestimulus that induces the mood and the re-sponse it elicits even when they know of thecause of the mood and are aware of its effects.Although we know that a bounced check mayput us in a bad mood, which in turn canmake us short tempered, we rarely attributea refusal to help a friend to a bounced check.When such attributions are made, they are

usually limited to events that are similar tothe cause of mood. Financial difficulties mayexplain why a new bill is particularly annoy-ing, but they are rarely used to explain cyn-ical responses to a colleague's suggestion.

A postexperimental questionnaire, whichfollowed Experiment 3, supported these hy-potheses. Only 3% of the experimental sub-jects (n = 67) acknowledged the possibilitythat the depressing story influenced their es-timates of fire or leukemia, although 50%agreed that the story influenced their mood,and 40% of the control subjects («. = 57)agreed that a negative mood could affect es-timates of unrelated risks. Hence subjectsacknowledged both the effect of the story ontheir mood and the effect of mood upon theirestimates, but they did not link the cause oftheir mood to its subsequent effect. The re-sponses of the (experimental) subjects alsoshowed that awareness of the effects of moodis enhanced when the cause of the mood (i.e.,the story) matches the relevant judgment.Although 40% of the subjects thought thatthe depressing story had changed their esti-mate for depression, only 14% thought it hadchanged their estimate for street crime. Inreality, the critical story had an equally pro-nounced effect for both related events, suchas depression, and unrelated events, such asstreet crime or toxic chemical spills.

The interpretation of public perception ofthe threat and the prevalence of risks shouldtake into account the susceptibility of layjudgments to manipulations of mood, andthe apparent lack of awareness of this effect.

Reference Note

1. Bower, G. B., & Wright, W. F. Mood effects on sub-jective probability assessment (Working Paper 1981-3). Unpublished manuscript, University of Minne-sota, 1981.

References

Bower, G. H. Mood and memory. American Psychologist,1981, 36, 129-148.

Clark, M. S., & Isen, A. M. Toward understanding therelationship between feeling states and social behavior.In A. Hastorf & A. M. Isen (Eds.), Cognitive socialpsychology. Amsterdam: Elsevier, 1982.


Combs, B., & Slovic, P. Newspaper coverage of causesof death. Journalism Quarterly, 1979, 56, 837-849.

Feather, N. T. Effects of prior success and failure on ex-pectations of success and subsequent performance.Journal of Personality and Social Psychology, 1966,3, 287-298.

Isen, A. M., & Levin, P. F. The effect of feeling good onhelping. Cookies and kindness. Journal of Personalityand Social Psychology, 1972, 21, 384-388,

Isen, A. M., & Shalker, T. E., Clark, M., & Karp, L.Affect, accessibility of material in memory, and be-havior: A cognitive loop? Journal of Personality andSocial Psychology, 1978, 3 7, 1-14.

Johnson, E. J., & Tversky, A. Representations of per-ceptions of risk. Journal of Experimental Psychology:General, in press.

Lichtenstein, S., Slovic, P,, Fischhoff, B., Layman, M.,& Combs, B, Judged frequency of lethal events. Jour-

nal of Experimental Psychology: Human Learning andMemory, 1978, 4, 551-578.

Sattath, S., & Tversky, A. Additive similarity trees. Psy-chometrika, 1977, 42, 319-345.

Slovic, P., Fischhoff, B., & Lichtenstein, S. Facts versusfears: Understanding perceived risk. In D. Kahneman,P. Slovic, & A. Tversky (Eds.), Judgment under un-certainty: Heuristics and biases. Cambridge, England:Cambridge University Press, 1982.

Tversky, A., & Kahneman, D. Availability: A heuristicfor judging frequency and probability. Cognitive Psy-chology, 1973, 5, 207-232.

Zajonc, R. B. Feeling and thinking: Preferences need noinferences, American Psychologist, 1980,35, 151-185.

Received December 31, 1981Revision received July 22, 1982

Search Opens for Section Editor of JPSP

The Publications and Communications Board has opened nominations for the ed-itorship of the Attitudes and Social Cognition section of the Journal of Personalityand Social Psychology. The nominee will complete outgoing section editor MelvinManis's term: Candidates should be available to start receiving manuscripts in early1984 to prepare for issues published in 1985, the final year of the term. Candidatesmust be members of APA. To nominate candidates, prepare a statement of onepage or less in support of each nomination. Submit nominations no later thanAugust 1, 1983 to:

Robert PerloffGraduate School of BusinessUniversity of PittsburghPittsburgh, Pennsylvania 15260(412)624-6438

Documents

Affect, Generalization, and the Perception of Riskheatherlench.com/wp-content/uploads/2008/07/johnson-tversky.pdfExperimental manipulations of affect induced by a brief newspaper report