Perception of Social Distributions Richard E

Perception of Social Distributions Richard E. Nisbett and Ziva Kunda University of Michigan

Accurate representation of the distribution of social attitudes and behaviors can guide effective social behavior and is often essential for correct inferences. We examined the accuracy of people's beliefs about the distributions of a large number of attitudinal and behavioral dimensions. In two studies we measured actual attitudes and behaviors in a student population, and we assessed beliefs by asking subjects to estimate the distribution of 100 students on these dimensions. We examined the accuracy of subjects' perceptions of the means, standard deviations, and distribution shapes. Subjects showed a number of systematic biases, including overestimation of dispersion and overestimation of the means of behavioral distributions and a false consensus bias, but their overall accuracy was impressive. Approximate knowledge of the distribu- ences about the meaning of social behavior. tions of important social variables would In fact, Kelley (1967) considered "consensus seem to be essential for effective social func- information," that is, information about the tioning. Reasonably accurate guesses about distributions of people's attitudes and behav- the dispersion and central tendency of im- ior, to be a major factor in his analysis of portant behavioral and attitudinal dimensions variance (ANOVA) scheme for inferences about could serve to guide one's behavior and self- causality in the social domain. More recently, presentation in unfamiliar social situations. Thagard and Nisbett (1982) have argued that Such knowledge could allow one to adhere knowledge about the variability of kinds of to accepted norms of behavior and conver- objects, including social objects, is re- sation and to steer away from controversial quired for appropriate generalizations. Nis- topics so as to avoid offending others who bett, Krantz, Jepson, and Kunda (1983) have hold different opinions. Indeed, one's effec- demonstrated that such knowledge mediates tiveness in group situations often depends on other types of statistical inference as well. the ability to gauge the extent of support for Despite the importance of people's percep- one's own position. Politicians have long ap- tions of social distributions, there are very preciated the value of knowledge about others' few studies that have examined the degree of opinions, and contemporary ones are willing people's accuracy in perceiving the central to spend large sums of money on polls de- tendency, dispersion, or shape of behavioral signed to obtain very precise knowledge of and attitudinal distributions. opinions. Some indirect evidence comes from work Approximate knowledge of the parameters on perception of physical objects, which in- of social distributions also would seem to be dicates that people are highly sensitive to all essential in order to make appropriate infer- three of these parameters. It has long been known that people are quite able to estimate the central tendencies of perceptual distri- This research was supported by National Science butions (e.g., Peterson & Beach, 1967; Posner Foundation Grant SES-8218846 and National Institute & Keele, 1968). More recent work by Fried of Mental Health Grant 1RO1 MH38466-01. We thank J. E. Keith Smith for advice on the analysis of the data and Holyoak (1984; Flannagan, Fried, & and Henri Zukier for theoretical suggestions. Keith Hol- Holyoak, 1984) shows that people are also yoak, Daniel Kahneman, Jon Krosnick, Darrin Lehman, sensitive to dispersion and distribution shape. and Lee Ross provided thoughtful criticism of an earlier They presented subjects with computer-gen- draft of this article. Sara Freeland provided able editorial help. erated variations on a particular visual theme, Requests for reprints should be sent to Richard E. for example, a pattern of light and dark Nisbett, Institute for Social Research, University of Mich- patches on a grid or a pattern of colored igan, Ann Arbor, Michigan 48106. rectangles. 297 298 RICHARD E. NISBETT AND ZIVA KUNDA

Subjects were more willing to classify a subject to some unique sources of bias. The given variant as an exemplar of a particular following potential errors in the perception type of pattern if the previous exemplars were of social distributions have been identified highly variable on the dimensions composing (cf. Nisbett & Ross, 1980): the type, than if previous exemplars were less 1. People seem likely to be influenced by variable. Subjects were also highly sensitive the availability of exemplars (Tversky & to distribution type. When the dimensions Kahneman, 1973). Because extreme behaviors from which exemplars were constructed were and opinions are likely to be salient (indeed distributed in a normal bell-shaped curve, people with extreme dispositions and opinions subjects readily learned to classify a pattern are likely to take steps to guarantee that they as being of the particular type that in fact will be salient), their overrepresentation in generated it. The probability of assigning a memory may lead to an overestimation of pattern to its parent type was greatest near the variability of social distributions. Indeed, the center of the dimensions underlying the it has been shown that individuals with ex- type and fell off in proportions dictated by treme physical or social characteristics are the shape of the normal curve. more available in memory and are conse- When subjects were instead presented with quently estimated as more frequent than cor- patterns generated by bimodal, U-shaped di- responding numbers of less extreme individ- mensions, their classifications did not track uals (Rothbart, Fulero, Jensen, Howard, & the actual frequencies of the dimension values. Birrell, 1978). Instead, subjects made their guesses about 2. A special case of availability, that is, classification as if they believed the distribu- knowledge about one's own value on social tions of the dimensions were normal. Even dimensions, might produce systematic errors after a very large number of trials exposing in estimates of central tendency. One's per- them to different patterns, subjects failed to ception of the entire distribution might be distribute their guesses in the same fashion shifted toward one's own location. This false as the actual dimensions. Instead, their guesses consensus tendency has already been docu- were distributed in an essentially flat fashion mented for a number of attitudes (e.g. Fields across the dimensions. & Schuman, 1976) and behaviors (Ross, Work on the perception of physical dimen- Greene, & House, 1977). sions thus indicates that people can be quite These investigators and others whom they accurate in assessing central tendency and cite (as well as subsequent studies, e.g., Judd dispersion and that they have a very heavy & Johnson, 1981 and van der Pligt, 1984), bias toward perceiving normality in distri- have shown that people tend to assume that bution shapes. Given the apparent frequency a higher fraction of the population shares of normal distributions in nature, it is possible their own views, or would behave as they that this bias might serve to make perception have, than is the case. The work gives us little more accurate and efficient. Observation of indication, however, of the degree of distortion a very few exemplars would serve to give a that this produces in estimates of central fair idea of central tendency, and, if normality tendency and no indication at all of the of the dimensions is presumed, variance or effects on estimates of dispersion or distri- standard deviation could also be estimated bution shape. with fair accuracy. Thus the classification of 3. There is a well-known tendency of group new objects or events would be maximally members to exaggerate the differences be- efficient in a statistical sense, so long as the tween themselves and out-group members underlying dimensions are in fact distributed (e.g., Campbell, 1967; Child & Doob, 1943; approximately normally. Quattrone, in press; Sherif, Harvey, White, Although it is possible that people are also Hood, & Sherif, 1961; Tajfel, 1970; Vinacke, accurate about the central tendency and dis- 1949). This suggests the possibility that people persion of social distributions and have a might have a bias toward assuming bimodality useful bias toward assuming their normality, for social distributions—"my kind" and "not there are good grounds for suspecting that my kind" (a tendency actually found in one they might not be very accurate. This is study by Judd & Johnson, 1981). because beliefs about social dimensions are Current theory and some evidence therefore SOCIAL DISTRIBUTIONS 299 suggest that people may not be as accurate that beliefs about social distributions might in their perceptions of statistical parameters be quite inaccurate, perhaps due to the kinds for social distributions as in their perceptions of biases discussed earlier. for physical distributions. Additional support Thus it is important to determine how for this expectation is suggested by research accurate people's perceptions of social distri- on inferential errors. Substantial evidence butions are, both because of the "first order" shows that people do not make proper infer- importance of accuracy about social distri- ences when reasoning about problems involv- butions for purposes of social conduct and ing social distributions. For example, Hamill, because of the "second order" importance of Wilson, and Nisbett (1980) showed that peo- accuracy for purposes of inferences about ple may draw essentially the same conclusions causality and meaning. about a population when told that an extreme- The results of an investigation of the ac- appearing case is a highly biased sample of curacy of perception of social distributions the population as when they are told that the also would have educational implications. It case is drawn from the center of the popula- has been shown that instruction in statistics tion distribution. improves people's ability to answer statistical It is possible to explain errors such as these questions about everyday social objects and either in terms of people's inadequate com- events (Fong et al., 1984; Krantz, Fong, & mand of the requisite statistical rules (Kahne- Nisbett, 1984). If it should turn out that man & Tversky, 1972, 1973; Tversky & people have very inaccurate beliefs about Kahneman, 1974) or in terms of their in- social distributions, however, then this would ability to picture the relevant social distri- imply that training should also focus on butions. Recent work shows that faulty beliefs teaching people how to represent social dis- may play a bigger role in producing inferential tributions accurately. errors than originally thought. It appears that people may have some Study 1: Perception of Behavioral and appreciation of statistical rules but that these Attitudinal Dimensions may be applied only to domains that lend themselves to accurate representation of the Method central tendencies and dispersions of the rel- Selecting Dimensions evant distributions (Fong, Krantz, & Nisbett, 1984; Jepson, Krantz, & Nisbett, 1983; In Study 1, we examined the accuracy of subjects' perceptions of the mean, standard deviation, and shape Kunda & Nisbett, 1984; Nisbett et al., 1983). of a number of behavioral and social distributions. To For example, people seem to be able to choose the dimensions, we tacitly denned a space of reason statistically about abilities and about social attitudes and behaviors including public and private sports events but to have more difficulty ones, socially desirable and socially undesirable ones, and reasoning statistically about personality traits seemingly important and trivial ones. From these we chose dimensions haphazardly within the constraint that and social behaviors. we wanted to include a large number of nonnormal This seems to be the case because it is distributions in order to see whether subjects would be easier to perceive the distributions of the biased toward seeing them as normal. former sorts of events than of the latter. The Thus, we tried to include dimensions that would produce bimodal distributions and J distributions with units of measurement for abilities and sports the central tendency far to the right or left of the scale events are usually clear (score on the test, of logical possibilities for the dimension. The resulting number of baskets shot), and they are often set of items is very arbitrary and cannot be considered observed in situations that are relatively fixed, an unbiased sample from any precisely denned universe therefore permitting comparisons among of behaviors and attitudes. It is unclear, of course, people. whether any procedure could approximate such a sample. Traits and most other social behaviors are not so clearly measurable: There are no Subjects and Procedure obvious units for shyness or honesty, for Subjects were 247 University of Michigan students example; and it is rare to observe large num- enrolled in introductory psychology classes. Only 14% of the subjects had had any training in statistics. They bers of people in situations that can be pre- participated in groups of 3 to 8. Sixty-two of the subjects sumed identical, so that their behavior can were asked to indicate the frequency of their actual be directly compared. These results suggest behavior on each of 17 dimensions, and 87 subjects were 300 RICHARD E. NISBETT AND ZIVA KUNDA asked to indicate their attitudes on each of 21 dimensions. married? (c) How do you feel about women being allowed Forty-eight subjects were asked to estimate the distribu- to have an abortion on demand? (d) How do you feel tions of 100 of their fellow University of Michigan about Ronald Reagan's conduct of the Presidency to students on each of the behavioral dimensions (i.e., their date? (e) How do you feel about U.S. intervention to reports of their behavior), and 50 subjects were asked to support the government in El Salvador? (f) How do you estimate the distributions of the attitudinal dimensions. feel about increasing defense spending? (g) How do you feel about the University's plan to eliminate some schools and departments in order to generate money for new Behavioral Items programs? (h) How do you feel about Harold Shapiro's conduct of the University Presidency to date? Subjects were asked 17 questions about behavior. These The objects that subjects were asked to evaluate on 5- were all answered on the scale below: point scales were the following: (a) the actress Jane

Never Once 2-5 6-11 Once a 2-3 Once 2-3 4-6 Daily a year times times month times a a week times times a year a year month a week a week

The questions asked were the following: (a) How often Fonda, (b) the television anchor Dan Rather, (c) the do you play tennis? (b) How often do you shoot pool? movie Raiders of the Lost Ark, (d) the movie Star Wars, (c) How often do you smoke marijuana? (d) How often (e) Pepperidge Farm cookies, (f) McDonald's hamburgers, do you have any trouble getting to sleep? (e) How often (g) Saudi Arabians, (h) the Moral Majority, (i) Campbell's are you bothered by nervousness? (f) How often are you chicken noodle soup, (j) Ritz crackers, (k) the model troubled by headaches? (g) How often do you feel blue Brooke Shields, (1) former Secretary of State Alexander or down? (h) How often are you bothered by nightmares? Haig, and (m) international oil companies. (i) How often do you attend religious services? (j) How often do you go to concerts? (k) How often do you go to parties? (1) About how often do you get drunk during Results the year? (m) How often do you go to bars or nightclubs? We examined subject accuracy about dis- (n) How often do you drink alcoholic beverages? (o) How often do you go to movies? (p) How often do you get tribution parameters—central tendency and angry with someone? and (q) How often are you late to dispersion—in three complementary ways, class? (a) We calculated average correlation, across items, between a given subject's estimates Attitudinal Items and the actual values. This measure provides an indication of the extent to which people Subjects were asked eight questions about attitudes on are calibrated about the ranking of the items 10-point scales, and to make sure that our findings about on each of these parameters, that is, about accuracy would not be limited to the case of the 10- point scales, with their particular format, subjects were their relative magnitudes, (e.g., a high corre- also asked to evaluate 13 objects on 5-point scales. The lation between subjects' estimates for the two scales are reproduced below: means of different behaviors and the actual

Very Strongly Disapprove Somewhat Slightly Somewhat Approve Strongly Very strongly disapprove disapprove disapprove approve approve strongly disapprove approve

Dislike Dislike somewhat Neither like nor dislike Dislike somewhat Dislike very much very much

The questions asked on 10-point scales were the following: (a) How do you feel about people your own age values for the means, would indicate that taking hard drugs? (b) How do you feel about people subjects are knowledgeable about the relative your age living with another person while not being frequency of behaviors of different kinds.) SOCIAL DISTRIBUTIONS 301

(b) We calculated simple average absolute "actual behavior" rather than as "reported deviation of the estimates from the actual actual behavior.") parameter. This measure provides an indica- Although significance levels for analyses of tion of how close the estimates of the average type (b), absolute average deviation, are re- person are to the actual parameters, for the ported, these are descriptive merely, because average dimension, (e.g., small average abso- the various estimates are not independent. lute discrepancies of estimates of behavior The results of these analyses are summarized means from actual behavior means would in Table 1. indicate that subjects are highly accurate about the absolute frequency of various be- Accuracy About Central Tendency haviors.) (c) In order to provide a way of evaluating Behavioral dimensions. As may be seen these absolute deviations and assessing their in the top row of Table 1, subjects' perceptions implications, we also looked at the number of the means of behavioral dimensions were of cases, drawn randomly from actual distri- moderately accurate on the whole. The aver- butions, that was required to equal the ac- age correlation between a given subject's es- curacy of the average subject's guesses about timates of the means of the 17 behavioral parameters as measured by absolute devia- dimensions and the actual means of the tions. This measure indicates the size of the dimensions was .42 ± .06.' The average ab- sample, drawn from the actual distribution, solute error of estimates of the means of that would provide estimates of the parame- behavioral distributions was only 1.19 on the ters that were as accurate as those of the 10-point scale. Because actual means ranged average subject. This measure was based on from 2.66 to 5.95, the accuracy was not due computer-drawn, randomly generated samples to limits on the range of possible error. There of different «s from the actual distribution of was, however, a systematic bias in the subjects' each item. estimates of behavioral distributions toward For each item we drew 10 samples of each overestimating behavior of all kinds. For 13 n. We calculated the mean and standard of the 17 distributions, the average estimate deviation of each such sample and then cal- of the mean was significantly higher than the culated, as we did for subjects' estimates, the actual mean for the dimension.2 absolute deviations of these statistics from Partly because of the overall bias toward the parameters of the actual distributions overestimation of the frequency of behaviors, from which each sample was drawn. These subjects' guesses about means were not ter- were averaged over items for each sample ribly accurate when compared with the results size, and the resulting average absolute devia- of computer-drawn random samples of the tions were compared to those of the subjects' actual distributions. On the average, a sample estimates. Accuracy is estimated by the num- of only 2 cases from behavioral distributions ber of cases in the random samples required came as close to the actual means as did the to equal subjects' discrepancies from actual guesses by a typical subject. values. The more cases required, the more Attitudinal dimensions. Subjects were accurate subjects are. more accurate about the means of attitudinal For each of these analyses, the basic datum dimensions. The average correlation between a given subject's estimate of the means of the was our computation of the estimated param- attitudinal distributions (pooling the results eters, based on subjects' histograms. To avoid for 10-point scales and 5-point scales) and circumlocutions, however, we will refer to the actual means of the distributions was subjects' "estimates of parameters" rather .73 ± .05. The average subject's guess about than to our computation of the parameters the mean of 10-point scale items was only based on subjects' estimates of distributions .89 away from the actual mean. (Actual means across scale points. (Similarly, we asked subjects about their reports of behavior rather ' AH intervals represent 95% confidence intervals for than measuring their behavior, but to avoid the correlations. circumlocutions we will refer to these as 2 All p values reported are based on two-tailed tests. 302 RICHARD E. NISBETT AND ZIVA KUNDA

Table 1 Summary of the Results of Study 1

Average correlation of Average absolute each subject's deviation of Size of randomly drawn estimates with the Overestimation estimates from sample with same actual parameters of parameters parameters absolute deviation

Central tendency Behaviors .42 ± .06 yes 1.19 2 Attitudes .73 + .05 no .89 3.57 (5-point) .48 Dispersion Behaviors .47 ± .05 13% .45 6 Attitudes .26 ± .08 9% .63 10.57 (5-point) .18

ranged from 2.91 to 6.14.) The average sub- from 1.15 to 2.40.) It required a random ject's guess about the mean of the 5-point sample of 6 cases to equal this level of scale items was only .48 away from the actual accuracy. mean. (Actual means ranged from 2.26 to Attitudinal dimensions. The average cor- 4.12). There was no systematic distortion of relation between a given subject's estimates the location of the mean toward either end of the standard deviations of the attitudinal of the scale. Subjects as a group were signif- distributions and their actual standard devia- icantly wrong in their estimates of only 12 of tions was .26 ± .08. Subjects overestimated the 21 means; 5 of these were overestimates the standard deviations of attitudinal dimen- and 7 were underestimates. On average, it sions by 9% on the average. Results of t tests required a random sample of 3.57 cases from contrasting standard deviations of estimated an attitudinal distribution to equal the accu- distributions with standard deviations of ac- racy of the typical subject's guess about the tual distributions yielded nine significant mean. overestimates and only two significant underestimates. The absolute discrepancy of standard de- Accuracy About Dispersion viation estimates from actual standard deviations was .63 for the average subject for 10- Behavioral dimensions. Subjects' estimates point scales (where the actual range was 1.81 about dispersion were also fairly accurate, to 3.19) and .18 for 5-point scales (actual despite a bias toward overestimating variabil- range .68 to 1.26). It required a random ity. The average correlation between the stan- sample of 10.57 cases to equal this level of dard deviation calculated for a given subject's accuracy. estimates of the behavioral distributions and It is striking that it required a substantially their actual standard deviations was .47 ± larger random sample of the actual distribu- .05. On average, subjects overestimated the tion to equal subjects' average accuracy about standard deviation of behavioral distributions standard deviations than to equal subjects' by 13%. average accuracy about means. Because the The results of t tests contrasting standard statistician's calculation of standard deviation deviations of estimated distributions with is substantially more complicated than the standard deviations of actual distributions calculation of a mean, this is surprising on yielded nine significant overestimates and its face. On the other hand, it is distinctly only four significant underestimates. The ab- possible that subjects' rough-and-ready esti- solute discrepancy of standard deviation es- mates of dispersion are actually easier for timates from actual standard deviations was them to compute and less prone to error .45 for the average subject, across dimensions. than their rough-and-ready estimates of (The range of actual standard deviations was means. SOCIAL DISTRIBUTIONS 303

Table 2 Percentage of Subjects Who Estimated a Given Distribution Shape for Each Type of Actual Distribution Shape

Actual Estimated distribution shape distribution shape JLeft Normal J Right Bimodal

JLeft 30 44 2 24 Normal 12 64 4 20 J Right 1 38 43 18 Bimodal 9 51 6 34

A statistician can obtain a quite good that were necessarily somewhat arbitrary) to estimate of the standard deviation of a sample be either J Left (that is, central tendency on of 20 cases simply by dividing the range by the extreme left of the scale), J Right, Bi- 4 or 5 (cf. Snedecor, 1946, p. 198). No modal, or Unimodal Symmetric (i.e., resem- comparably easy estimate of the mean of 20 bling a normal distribution).3 Approximately cases exists. In addition, any rough calculation half of the distributions defined by these of the standard deviation relying on the range criteria were coded as normal and a third requires memory only for two values—the each of the remainder as J Left, J Right, and highest and the lowest observed—and these Bimodal. values would tend to be quite memorable. Table 2 shows the cross-tabulation of esti- Comparable accuracy for the mean would mated and actual distribution types, pooled require either remembering many more cases across behavioral and attitudinal dimensions. than that or forgetting just the right ones! Subjects were overwhelmingly accurate at A seeming anomaly in the attitude data identifying normal distributions. Where the should be mentioned. Although the accuracy distribution was nonnormal, subjects' guesses of mean estimates was considerably higher either favored the actual distribution shape than the accuracy of standard deviation esti- or favored a normal distribution, with the mates as measured by correlations, the reverse actual distribution shape coming in a strong was true as measured by the size of the second. These data indicate (a) very substan- randomly drawn sample with an equally large tial accuracy in perception of distribution absolute deviation. We suspect that this was shapes and (b) an even more substantial bias true in part because subjects' estimates of the in favor of seeing distributions as unimodal standard deviations for attitudes in general symmetric. were very close to the true range of standard An interesting view of subjects' accuracy deviations, whereas the actual differences about all distributional parameters, including among standard deviations for attitudes were shape, may be seen in Figures 1 and 2. Figure very small. 1 presents 16 of the 17 actual behavioral This latter fact resulted in extreme trun- distributions and the average of all subjects' cation of range for standard deviations of estimates about the distributions. (One of the attitudes. Thus, although subjects were in three questions about alcohol consumption general very close to the actual values in is not included to save space, but it is not every case, they were not very well calibrated notably different in any respect from the to the slight differences that existed between others.) Figure 2 presents all 8 of the 10- cases.

3 Accuracy About Distribution Shapes For example, to be designated unimodal symmetric, the highest point of a subject's Id-point scale distribution Subjects' guesses about distribution shapes had to be within slots 3-8 and the second highest point had to be adjacent or, if not, then the difference between showed a heavy bias toward assuming nor- the second and third highest points had to be less than mality. All distributions, both actual and 5%. Figures 1 and 2 show how various curves ended up estimated, were denned (by a set of criteria being classified by the system. 304 RICHARD E. NISBETT AND ZIVA KUNDA point attitudinal scale distributions and 8 of havioral and attitudinal distributions. Among the 13, 5-point attitudinal scale distributions other things, this means that when 2 or more (the dropped distributions are not notably people pool their estimates of social distri- different from the retained ones). The curve butions they normally would come closer to shape for each actual distribution is indicated the truth rather than compounding their in- at the top of each distribution. accuracy due to shared biases. The reader is warned that the averaging of We have not yet examined the possibility all estimated distributions allows for the pos- that subjects show a substantial false consen- sibility of greatly overestimating individual sus effect. Note that the overall accuracy subject accuracy. With that caveat, it may be portrayed in Figures 1 and 2 as well as the observed that accuracy about distribution apparent bias toward normality could theo- shapes, at the group level, is remarkable retically be produced entirely by false consen- indeed. Whatever individual biases there are sus. If each subject estimated the distribution in the perception of social distribution pa- by constructing a unimodal distribution with rameters, they tend to cancel out, save for a mode corresponding to his or her own the previously mentioned overestimates of position, all the errors would cancel each the sheer amount of behavior of all kinds other out when pooled together, and the that people engage in and the mild overesti- resulting aggregated picture would be similar mate of the standard deviation of both be- to that portrayed in Figures 1 and 2. Such a

Have Trouble Getting to Sleep Get Drunk

Come Late to Class

Go to Concerts Smoke Marijuana Play Tennis Go to Parties Figure 1. Actual and estimated distributions for behavioral items. Letters at the top of each distribution denote coded shape of actual distribution: N = Normal, JL = J Left, JR = J Right, B = Bimodal. SOCIAL DISTRIBUTIONS 305 heuristic could also produce the apparent want to contaminate estimates about distri- bias toward normality, because it is true even butions with reports of own behavior and for the nonnormal distributions that the po- attitude, or vice versa.) In Study 2, we asked sitions of a large proportion of subjects are both sets of questions of the same subjects. located in the middle range of the scale; if This made it possible to examine the degree these subjects used their own position as the of false consensus in estimates. The previous sole basis for judging central tendency they literature on false consensus does not allow would estimate the distribution to be normal. for a very good estimate of the degree of But such a strategy could not possibly produce error because in most studies the data were any accuracy about dispersion—one's own dichotomized, and it was shown merely that position on a distribution provides no infor- subjects tended to overestimate the number mation about the variability of the distribu- of people who behaved in the rough way they tion nor does one's own position on different did or who held the approximate opinion distributions provide any information about they did. their relative variability. Examination of false consensus bias can Study 2: False Consensus in the Perception only be achieved by asking the same subjects of Social Distributions about their estimates of the population dis- Method tributions and about their own locations in One hundred three University of Michigan students the distributions. (In Study 1, we did not enrolled in introductory psychology classes were presented

JR JL-

University's Budget Cuts Unlwslty President's Conduct

U.S. Intervention In El Selvedor Increeslng Defense Spending

International Oil Compsnles

Moral Majority CsmpbeirsSoup Ssudl Arabians Figure 2. Actual and estimated distributions for attitudinal items. Letters at the top of each distribution denote coded shape of actual distribution: N = Normal, JL = J Left, JR = J Right, B = bimodal. 306 RICHARD E. NISBETT AND ZIVA KUNDA

Table 3 Summary of the Results of Study 2

Central tendency Behaviors .55 ± .04 yes .95 2 Attitudes .56 ± .10 no 1.18 2.36 (5-point) .60 Dispersion Behaviors .44 ± .04 9% .43 7 Attitudes .25 ± .06 15% .63 7.29 (5-point) .26 with the behavioral items in Study 1 and were asked were closer to the actual mean than to the both to indicate their own position on the scale and to subject's own level at p < .01 (though it should distribute 100 students enrolled in introductory psychology over the 10 scale points. (The reference population be noted that these are descriptive only be- differs from that in Study 1, where subjects were asked cause of lack of independence). The average about University of Michigan students. The match between discrepancy between the subject's guess about actual population and the description of the population the mean and the actual mean was .95, given to subjects was thus closer in Study 2 than in Study whereas the average discrepancy between the 1.) Another 120 introductory students were presented with the attitudinal items in Study 1 and were asked subject's own value and the actual mean was both to indicate their own position on the scale and to 1.95;tfll8) = 12.95, p<.001. distribute 100 introductory students over the 10 (or 5) Attitudes. The average correlation between scale points. For both behavioral and attitudinal items, subjects' own attitudes and the mean of their half the subjects answered for themselves first and half gave estimates about the distributions of other subjects' estimated distributions was .32 (p < .001; reports first. Procedure was otherwise identical to that in the correlation was .34 when subjects indi- Study 1. cated their own attitude first and .28 when Results they gave their estimates about the population first). For all but one of the attitudinal di- As may be seen in Table 3, the major mensions, the average subject estimated the analyses produced results virtually identical mean of the distribution to be closer to its to those of Study 1. Significance levels tended actual mean than to the subject's own attitude. to be higher, however, because the N was All but seven of the estimated means were greater. closer to the actual mean than to the subject's own attitude at p < .01. The comparison of False Consensus and Accuracy about Means the overall average discrepancies, which were .80 for estimate versus actual and 1.17 for There was a decided false consensus effect, own versus actual, was highly significant, but it was not of great magnitude, either for f(98) = 9.95, p<. 001. behaviors or for attitudes. Subjects' estimates of means are thus influ- Behaviors. The average correlation, across enced by their own values, but this influence dimensions, between subjects' own behavioral is not great and their estimates of the means level and the mean of their estimated distri- are substantially closer to the actual mean butions was .28 (p < .01; the correlation was than to their own level. Thus, subject's own .23 when subjects indicated their own re- value is not the exclusive, or even the major, sponse first and .34 when they gave estimates contribution to accuracy about means. about the distribution first). For all of the behavioral dimensions the average subject estimated the mean of the distribution to be False Consensus and Distribution Shape closer to its actual mean than to the subject's There was a decided relation between sub- own level: All but one of the estimated means jects' own positions and the shape of their SOCIAL DISTRIBUTIONS 307

Table 4 Percentage of Subjects Whose Position Corresponds to J-Left, Normal, or J-Right Curves, Estimating the Curve to be Each of the Possible Curve Shapes, for Curves That are in fact J Left, Normal, and J Right* Subjects' estimates Actual Subjects' curve n position JLeft Normal J Right Bimodal JLeft 196 JL 42 34 3 21 199 N 18 50 3 29 9 JR 0 44 12 44 Normal 232 JL 33 37 1 29 1602 N 7 69 2 21 73 JR 8 27 29 36 J Right 25 JL 0 16 8 76 409 N 5 43 32 20 250 JR 10 8 69 14 * Actual bimodal curves were not included in this analysis because a subject's position cannot correspond to a bimodal curve. estimated distributions, suggesting that own tions were in the central range favored normal position played a significant, though not a curves. And a similar pattern is found for J- singular role in determining subjects' esti- Right curves, where subjects whose own po- mates of shape. To examine this relation we sitions were at the extreme right overwhelm- pooled together all the J-Left curves, all the ingly favored J-Right curves in their estimates, Normal curves, and all the J-Right curves whereas subjects whose own positions were for attitudes and behaviors. For each curve in the central range favored normal curves. shape we identified the subjects whose own Further, for each actual curve shape the positions were located at the extreme left of highest accuracy rate (i.e., percentage of cor- the distribution, the subjects whose own po- rect identifications) is found for subjects whose sitions were in the middle range, and the own positions corresponded to the actual subjects whose own positions were at the distribution shape. extreme right of the distribution. Thus there is no question that subjects' If subjects were relying entirely on their own positions influenced their guesses about own position (estimating the curve to be distribution shape. But it is equally clear that unimodal with a mode located at their own false consensus could not be the only mech- position), then these three groups of subjects anism influencing subjects' guesses. First, as would be expected to estimate the curve noted earlier, if false consensus were the only shapes, respectively, as J Left, Normal, and J mechanism, no subject would estimate the Right, and no subject would be expected to distribution to be bimodal. Yet, on average, estimate the curves as Bimodal. Table 4 27% of the guesses are bimodal. presents, for curves that are J Left, Normal, More strikingly, in most cases the majority or J Right, the percentage of subjects whose of subjects do not believe the distribution own positions were in each of the three range shape to correspond to their own positions. categories who estimated the curves to be The only two exceptions to this are for cases each of the possible shapes. where subjects' own positions do in fact It may be seen first in Table 4 that own correspond to the actual curve shape. Further, position clearly played a role in estimates of it may be seen that subjects appear to know curve shape. For example, in the first and when their own position is extremely different second rows of Table 4, for J-Left curves, from that of the modal person. Thus, for J- subjects whose own positions were at the Right curves, not one of the subjects whose extreme left favored J-Left curves in their own positions were in the extreme left range estimates, whereas subjects whose own posi- made the serious error of estimating the 308 RICHARD E. NISBETT AND ZIVA KUNDA curve to be J Left, and for J-Left curves only is not great—a little more than a 10% over- 12% of the subjects whose own positions were estimate of standard deviation on average in the extreme right estimated the curves to across studies and across types of social di- be J Right. mensions—but the overestimation was quite The overall extent of false consensus may systematic. The most natural way to under- be gauged from the following comparison. stand this error is in terms of the availability The average probability of estimating the heuristic. Extreme behaviors and attitudes distribution shape as corresponding to one's are more salient than nonextreme ones, en- own position when one's own position does tirely aside from the fact that people who correspond to the actual distribution shape behave in extreme ways or who hold extreme is 60%. But the average probability of esti- attitudes often take steps to ensure that their mating the distribution shape as correspond- behavior or attitudes are highly salient. When ing to one's own position when one's own people try to estimate dispersion, the extreme position does not correspond to the actual examples will be more available than they shape is only 28%. The difference between are frequent.4 the two percentages reflects the impact of 2. Subjects overestimated the means of forces other than knowledge of one's own behavioral distributions. The extent of the position that affect subjects' estimates. overestimation was not great, but it was Another important finding in Table 4 is highly reliable. It is not clear why people that the apparent bias toward normality ap- should overestimate the sheer frequency of pears to have been due in part to false behavior of most kinds. Perhaps the expla- consensus. Subjects whose own positions were nation is similar to the explanation for over- not in the normal range show only a slight estimation of dispersion. Behavior is more preference for normal curves. These data salient than nonbehavior, thus when one suggest that the tendency to view distributions comes to estimate the frequency of behavior as normal is a rather weak default for subjects for people on some dimension the behavers who are not themselves located fairly near are more salient than the nonbehavers and the center of the distribution. contribute disproportionately to the outcome. There is of course no comparable phenome- Discussion non for attitudes, because it is both extremes In general, we find subjects' estimates about of the dimension that are salient rather than social distribution parameters to be surpris- just one, and we in fact found no systematic ingly close to the actual values. They were tendency to displace the means of attitude not overwhelmingly accurate about any single scales in one direction or the other. This parameter, as evaluated by any given type of explanation is consistent with the literature analysis, and they showed several types of on the feature-positive effect, which docu- systematic bias. The errors, however^ were ments people's tendency to ignore or underuse not very large and the biases underlying the nonoccurrences when making judgments and errors do not seem to be of very great mag- solving problems. This effect has been dem- nitude. We begin with a review of the biases onstrated for a diverse range of domains and then return to an assessment of the likely including, for example, animal and human real world accuracy that people might be learning (Jenkins & Sainsbury, 1970; Sains- expected to display. bury, 1983), judgment of contingency (Nisbett & Ross, 1980; Ward & Jenkins, 1965), and Biases in the Perception of self-perception (Fazio, Sherman, & Herr, Social Distributions 1982). The results provide strong evidence for 4 An alternative explanation for the overestimation of four different sources of bias in estimates of dispersion should be noted. This is the trivializing pos- distributional parameters for social dimen- sibility that our method of presenting the subjects with highly dispersed scale values encouraged them to use sions. those extreme values, whereas had they spontaneously 1. Subjects overestimated the degree of generated, say, a range, they might not have overestimated dispersion. The extent of the overestimation dispersion. SOCIAL DISTRIBUTIONS 309

3. Subjects skewed the central tendencies with which they are familiar. They know of distributions in the direction of their own what major kinds of distributions are possible level on the distribution, for both behaviors and they know which of these are most likely. and attitudes. The average distortion of these For all distribution types examined, they parameters due to false consensus was not show a strong tendency to guess the correct great, however. Subjects' estimates of distri- type. In addition, their estimates of the central bution shapes were also influenced to some tendency and dispersion of distributions are extent by their own positions. The mechanism fairly well-correlated with the actual param- underlying these false consensus effects is eters. For estimates of the means for distri- again most probably the availability heuristic. butions, a random sample of 2 to 4 cases One always has an estimate for at least one's from actual distributions would be required own level on a given dimension. Only under to exceed the accuracy of the average subject's special conditions, such as having unequivocal guesses; for standard deviations a random knowledge of a number of people's positions, sample of 6 to 10 cases would be required. or holding a theory that one is unusual on a Our data thus make it clear that the errors particular dimension, does it seem likely that people make in inductive reasoning about knowledge of own position would fail to have social events are not due to any fundamental an effect on the estimate for the distribution incapacity to characterize the central tenden- as a whole. cies, dispersions, and shape of social distri- 4. Subjects showed a weak bias toward butions. characterizing distributions as normal. It is Note also that the reference population is our guess that this tendency would work not one that would be expected to display toward producing accuracy in most circum- subjects' abilities for parameter estimation to stances. It is of course impossible to know, their best advantage. The University of Mich- because of unanswerable sampling questions igan undergraduate body is a large and highly and scaling questions, whether normal distri- diverse population. Had we studied a small butions are the rule for social events. If they college population or had we asked our sub- are, the bias is clearly useful. People would jects to estimate parameters for some subset be well-served by using their point estimates of the Michigan population with which they (knowledge of their own levels on particular were personally extremely familiar, such as dimensions and those of their friends) as people in their major field or fellow athletes inputs into a formula assuming normality. or sorority members, there is every reason to This typically would produce a characteriza- assume we would have found more accuracy tion of the distribution that is accurate and than we did (cf. Judd & Johnson, 1981). that is expressible with great economy in On the other hand, it should be clear that terms of psychological parameters corre- although our studies do not display subjects' sponding to mean and standard deviation. parameter estimation abilities to their best advantage, they probably also do not display Accuracy in the Perception of them to their worst advantage or even at their Social Distributions ecologically typical level. There is good reason to suspect, for example, that people would The biases noted would not serve, of them- be far less accurate about dispersion for groups selves, to produce serious inaccuracy in the of which they are not themselves members. estimation of social distributions. And one Work by Quattrone (in press; Quattrone & of them, the false consensus bias, is a very Jones, 1980) indicates that people substan- useful heuristic to employ in the absence of tially underestimate dispersion for outgroups. other knowledge: For most of the distributions What our studies show is that when subjects we examined, the majority of people would are explicitly requested to generate social be right to assume that most other people's distributions for a large and diverse group of stances are not very different from their own. which they are members, they can do so with Indeed, the major lesson that we draw substantial accuracy. This does not mean that from our data is that people are skilled in people spontaneously recall or construct social estimating social distributions of populations distributions for most of the social judgment 310 RICHARD E. NISBETT AND ZIVA KUNDA problems that require them. On the contrary, own position but also on additional accurate there is evidence in the work by Nisbett, knowledge. For most of the dimensions we Krantz, Jepson, and Kunda (1983) that the examined, subjects could be expected to have distributions of social dimensions are often reasonably accurate memories for the location not salient and that it is only when subjects of at least several other people. Some of these are specifically cued to the necessity of con- memories could produce more error than structing them that they do so. Without such accuracy, if they are more memorable than cuing, subjects sometimes made inferences as they are typical. But in general, memory of if they believed that the relevant population actual cases could be expected to promote was either uniform or essentially infinitely accuracy. variable. For the purposes of education, then, In addition to memories of actual cases it might be profitable to emphasize the ne- and knowledge about the specific item at cessity of considering social distributions for hand, subjects undoubtedly have theories various classes of problems. The present work about the kind of item in question and about indicates that college students, at least, will social behavior and attitudes in general. For not need much help in constructing such example, although subjects might have little distributions when they understand the need knowledge about the actual attitudes of many for doing so. of their fellow students toward increasing defense spending, they might view this issue Constructing Social Distributions as the kind of issue heavily saturated with right-left ideological orientation and might How do people construct social distribu- have reasonably accurate estimates of the tions? Our attempts to answer this question percentage of the student population occu- will be necessarily speculative, inasmuch as pying various locations on this continuum. our studies were not designed to address it. And, at the most general level, people seem Everyone knows his or her own location to have tacit beliefs about the likely distri- on each of the dimensions we examined, and bution shapes for social dimensions, just as this could serve as a beginning reference they do for physical dimensions. Given point in estimating distribution parameters. knowledge about one's own position, reason- Indeed, it is highly likely that our subjects ably accurate memory for a few cases, and a did use this as a reference point, given that good guess about the shape of the underlying there was a false consensus effect for estima- distribution, subjects could generate quite tion of the mean and of the distribution accurate estimates of the mean and standard shape. deviation of the distribution. Our data also indicate strongly, however, These considerations make it seem less that false consensus could not possibly be the mysterious why subjects were as accurate as only factor influencing the construction of they were in their guesses about various such distributions. False consensus cannot parameters of social dimensions. But it should account for the considerable accuracy about be remembered that they are purely post dispersion, because one's own position does hoc. Indeed, as should be clear by now, we not provide any relevant information. And were quite surprised by the accuracy subjects our data demonstrate that false consensus displayed. Prior to conducting the research, cannot fully account for the accuracy about we thought only of mechanisms that might central tendencies and distribution shapes produce error. either. Estimated means were closer, on average, to the actual values than to subjects' own positions, and subjects estimated the References distribution shapes as corresponding to the Campbell, D. T. (1967). Stereotypes and the perception shapes implied by their own positions much of group differences. American Psychologist, 22, 817- more when there was a genuine correspon- 829. Child, I. L., & Doob, L. W. (1943). Factors determining dence than when there was not. national stereotypes. Journal of Social Psychology, 17, We therefore believe that when constructing 203-219. distributions, subjects rely not only on their Fazio, R. H., Sherman, S. J., & Herr, P. M. (1982). The SOCIAL DISTRIBUTIONS 311

feature-positive effect in the self-perception process: Nisbett, R. E., & Ross, L. (1980). Human inference: Does not doing matter as much as doing? Journal of Strategies and shortcomings of social judgment. Engle- Personality and Social Psychology, 42, 404-411. wood Cliffs, NJ.: Prentice-Hall. Fields, J. M., & Schuman, H. (1976). Public beliefs about Peterson, C. R., & Beach, L. R. (1967). Man as an the beliefs of the public. Public Opinion Quarterly, 40, intuitive statistician. Psychological Bulletin, 68, 29-46. 427-448. Posner, M. I., & Keele, S. W. (1968). On the genesis of Flannagan, M., Fried, L. S., & Holyoak, K. J. (1984). abstract ideas. Journal of Experimental Psychology, Induction of the distributional structure of perceptual 77, 353-363. categories. Unpublished manuscript, University of Quattrone, G. A. (in press). On the perception of a Michigan. group's variability. The psychology of intergroup rela- Fong, G. X, Krantz, D. H., & Nisbett, R. E. (1984). tions, Vol. II. Improving inductive reasoning through the application Quattrone, G. A., & Jones, E. E. (1980). The perception of statistical heuristics to everyday problems. Unpub- of variability within in-groups and out-groups: Impli- lished manuscript, University of Michigan. cations for the law of large numbers. Journal of Fried, L. S., & Holyoak, K. J. (1984). Induction of Personality and Social Psychology, 38, 141-152. category distributions: A framework for classification Ross, L., Greene, D., & House, P. (1977). The false learning. Journal of Experimental Psychology: Learning, consensus phenomenon: An attributional bias in self- Memory, and Cognition, 10, 234-257. perception and social perception processes. Journal of Hamill, R., Wilson, T. D., & Nisbett, R. E. (1980). Experimental Social Psychology, 13, 279-301. Insensitivity to sample bias: Generalizing from atypical Rothbart, M., Fulero, S., Jensen, C., Howard, J., & cases. Journal of Personality and Social Psychology, Birrell, P. (1978). From individual to group impressions: 39, 578-589. Availability heuristics in stereotype formation. Journal Jenkins, H. M., & Sainsbury, R. S. (1970). Discrimination of Experimental Social Psychology, 14, 237-255. learning with the distinctive feature on positive or Sainsbury, R. S. (1983). Discrimination learning using negative trials. In D. Mostofsky (Ed.), Attention: Con- positive or negative cues. Canadian Journal of Psy- temporary theory and analysis. New York: Appleton- chology, 27, 46-57. Century Crofts. Sherif, M., Harvey, O., White, B., Hood, W., & Sherif, Jepson, C, Krantz, D. H., Nisbett, R. E. (1983). Inductive C. (1961). Intergroup conflict and cooperation: The reasoning: Competence or skill? Behavioral and Brain robber's cave experiment. Norman, OK: Normal In- Sciences, 6, 494-501. stitute of Group Relations, University of Oklahoma. Judd, C. M., & Johnson, J. P. (1981). Attitudes, polar- Snedecor, G. W. (1946). Statistical methods. Ames, Iowa: ization, and diagnosticity: Exploring the effect of affect. The Iowa State College Press. Journal of Personality and Social Psychology, 41, 26- Tajfel, H. (1970). Experiments in intergroup discrimi- 36. nation. Scientific American, 223, 96-102. Kahneman, D., & Tversky, A. (1972). Subjective proba- Thagard, P., & Nisbett, R. E. (1982). Variability and bility: A judgment of representativeness. Cognitive confirmation. Philosophical Studies, 47(2), 188-202. Psychology, 3, 430-454. Tversky, A., & Kahneman, D. (1973). Availability: A Kahneman, D., & Tversky, A. (1973). On the psychology heuristic for judging frequency and probability. Cog- of prediction. Psychological Review, 80, 237-251. nitive Psychology, 5, 207-232. Kelley, H. H. (1967). Attribution theory in social psy- Tversky, A., & Kahneman, D. (1974). Judgment under chology. In D. Levine (Ed.), Nebraska Symposium on uncertainty: Heuristics and biases. Science, 185, 1124- Motivation (Vol. 15, pp. 192-240). Lincoln: University 1131. of Nebraska Press. van der Pligt, J. (1984). Attributions, false consensus, Krantz, D. H., Fong, G. X, & Nisbett, R. E. (1984). and valence: Two field studies. Journal of Personality Formal training improves the application of statistical and Social Psychology, 46, 57-68. heuristics to everyday problems. Unpublished manu- Vinacke, W. E. (1949). Stereotyping among national- script, Bell Laboratories, Murray Hill, NJ. racial groups in Hawaii: A study in ethnocentrism. Kunda, Z., & Nisbett, R. E. (1984). The psychometrics Journal of Social Psychology, 30, 265-291. of everyday life. Unpublished manuscript, University Ward, W. D., & Jenkins, H. M. (1965). The display of of Michigan. information and the judgment of contingency. Canadian Nisbett, R. E., Krantz, D., Jepson, C., & Kunda, Z. Journal of Psyhology, 19, 231-241. (1983). The use of statistical heuristics in everyday inductive reasoning. Psychological Review, 90, 339- 363. Received May 21, 1984