The representativeness heuristic has been used to explain a number of judgemental biases. We have discussed the gambler’s fallacy and insensitiv-ity to sample size as examples of errors of prediction. In diagnosis problems the representativeness heuristic has been held responsible for base-rate neglect (or rather insufficient weight of base rates). This theme has been explored in detail in Chapter 2 of the present volume.
Tversky and Kahneman (1983) further argued that representativeness can lead to the fallacious belief that a combination of one likely and one unlikely event is more likely than the unlikely event taken by itself. This so-called conjunction fallacy has been treated in this book in Chapter 1. For instance, people thought it was likely that Bjorn Borg would win the Wimbledon tennis final, because it looked like a typical thing for a champion like Borg to do. They thought it would be rather unlikely for him to lose the first set of the match. This would be less typical of Borg. The conjunction – losing the first set but winning the match – contains a combination of typical and less typical elements. This conjunction was believed by many participants to have an intermediate probability, rather than the even lower probability that follows logically from the combination of high and low p events.
Text box 9.4 Intuitions about sample size: From samples to population Two teachers want to find out whether there are more male or female students at the university. One of them checks a small class of 15 students, finding 9 (60%) male and 6 female students. The other one studies a larger class of 45, finding 18 (40%) male and 27 female students. What is more probable: (a) There are altogether more male students; (b) there are more female students; (c) there is an equal number of male and female students.
Kahneman and Tversky (1973) also showed that use of the represen-tativeness heuristic could lead to nonregressive predictions. It has been known since the time of Francis Galton that use of imperfect predictors should lead to less extreme predictions. Extremely tall parents will have tall offspring, but since the heights of parents and offspring are not perfectly correlated, we should expect these children to be, on the average, somewhat shorter than their parents; conversely, children of exceptionally short par-ents should be in general taller than their parpar-ents. Filial regression had, in fact, already been observed by Homer, in a passage of the Odyssey: “Few are the sons that are like their father in breed; The most part are worse, scarce any their fathers excel” (Book 2, Verse 277–78, S. O. Andrew’s translation).
Being exclusively concerned with the superior part of the distribution, Homer failed to comment on the complementary fact that inferior fathers often have sons of a more hopeful breed. Evidently, Homer felt sons to be less representative of their illustrious origins than they ought to have been, documenting a very early instance of the representativeness heuristic failing to square with the facts.
It further follows from the concept of statistical regression that, when a measure is not perfectly reliable, we must expect the top scorers on one occasion to be distributed somewhat closer to the mean on the second occa-sion (and vice versa). From the point of view of representativeness, however, a typical top scorer should continue to excel, and an individual scoring in the 75th percentile should remain around the 75th percentile on the second occasion also. A drop in performance would accordingly be attributed to change or some other systematic process, rather than to chance. Tversky and Kahneman (1974) tell the story about a flight instructor who used to praise students who had performed exceptionally well, only to find that, as a rule, they performed worse on the next occasion. Instead of realizing that per-formances are not completely reliable indicators of skill, and thus bound to regress for purely statistical reasons, he felt forced to conclude that praise has a negative rather than the intended positive effect.
A corollary of the problem of non-regressive predictions is that people tend to make the same predictions based on invalid measures as they would do on more reliable and valid ones. So for instance when two groups of participants were asked to predict the grades of hypothetical students based on their relative standing (percentile scores) on (a) a grade point average scale, or (b) a mental concentration test, they produced in both cases almost identical, non-regressive predictions (Kahneman & Tversky, 1973). In other words, they appeared to use the less valid and reliable mental concentration test with the same confidence as a perfectly valid predictor. Extreme predic-tions based on invalid predictors have been described as manifestapredic-tions of an illusion of validity.
The representativeness heuristic has over the years been applied to an increasing range of phenomena in the field of judgement and decision making. It has been proclaimed to be “perhaps, our most basic cognitive
heuristic” (Fiske & Taylor, 1991, p. 384). One of its attractions has been that it seems also to be applicable to expert judgements in a variety of fields.
Another is its link to the area of causality judgements.
Expert judgements
In their very first paper on judgemental biases, Tversky and Kahneman (1971) showed that even scientists with a solid background in statistics place too much confidence in the results of small samples. They presented a ques-tionnaire to a group of mathematical psychologists, asking what kind of advice they would give a PhD student who has just performed two small-scale, inconclusive experiments (one barely significant and the other not).
Many respondents thought it would be a good idea to speculate about the difference between the results (which could have been a statistical artifact).
The majority thought that the experiment should be repeated a third time, again with a small sample (which could not be expected to reach signifi-cance). Despite their theoretical knowledge of sampling distributions and statistical hypothesis testing, these experts seemed to suppose that small samples are highly representative of their populations, apparently believing in a “law of small numbers” (as a proxy for the well-known “law of large numbers” in statistical theory).
However, domain expertise can sometimes counteract some of the more extreme biases due to representativeness thinking. For instance, experience with the ups and downs of the stock market could make the predictions of a professional investor more regressive than those of a novice. Yet even a real-world economic market may be biased by the power of representative pre-dictions, manifested as overconfidence in stocks, firms, or football teams that have a recent history of good performance (Tassoni, 1996). Moreover, risky stocks tend to be undervalued, and safe stocks overvalued, by repre-sentativeness reasoning: Safe stocks come from good companies, and investment in good companies should give good returns, that is, investors assume a match between the company and stock quality, making safe stocks attractive even when they are costly (Shefrin, 2001).
Clinical judgements offer rich possibilities for studying diagnoses as well as predictions. Garb (1996) gave clinical psychologists a case description satisfying the DSM-IIIR criteria for antisocial personality disorder. They were then asked to rate (a) the likelihood for five possible diagnoses, as well as (b) the degree to which the case was similar to the “typical” person with these disorders. Only 27% of the clinicians made the “correct” diagnosis (according to the manual). The correlation between probability judgements and representativeness judgements was extremely high, r = .97, indicating that the clinicians used similarity to a prototype rather than a list of criteria to arrive at a diagnosis.
Causality judgements
John Stuart Mill (1856) observed that people, including philosophers, tend to assume a correspondence between cause and effects. Like begets like.
Large effects prompt us to look for large causes. Good effects are attributed to good causes, whereas disasters and human suffering must be due to evil forces. While this is in general a sound heuristic – large objects make in general louder noises than smaller objects, and nice people often make us feel good – exceptions are not difficult to find (small whistles can be deafen-ing, and nice people can be boring). The similarity between Mill’s cor-respondence principle and the representativeness heuristic has made many investigators think that judgements by representativeness also apply to judgements of causation.
Again, these inferences may go both ways: from known causes to hypo-thetical effects, and from known effects to hypohypo-thetical causes. We may for instance expect an acknowledged expert to be a source of valid and reliable information. Informed people (causes) should produce informative state-ments (effects) matching their level of expertise. Unfortunately, experts can be wrong, particularly outside their field of expertise. Even more risky, we may infer the expertise of the speaker from the confidence and specificity of his or her assertions. It is more impressive for a political commentator to announce that Iraq will be attacked on January 27, than simply that war will break out sooner or later. Unfortunately, the specific prediction will be more easily disconfirmed than the vague one, leading to a “preciseness paradox”
(Teigen, 1990), where the speaker has to choose between being believed (by sounding like an expert) and being correct (by using more general and approximate terms).
If causes correspond to effects, we should expect people to prefer causes whose salient features match the salient features of the events to be explained. Lupfer and Layman (1996) found that people favour religious explanations of uncontrollable events with life-altering outcomes, whereas they prefer naturalistic explanations for controllable events, and events with more mundane consequences. In each case the religious attributions were made in agreement with characteristics believed to be “representative” for supernatural versus natural sources of causality.
Gavanski and Wells (1989) suggested that representative causes also apply to hypothetical, counterfactual outcomes. For instance, when we think how an exceptional outcome could have been prevented, we focus on exceptional antecedents, whereas we change normal outcomes by changing a normal antecedent. Causes, or antecedents, are supposed to match out-comes also in magnitude (Sim & Morris, 1998). If an athlete makes a poor overall performance in a triathlon contest, we will blame the failure on her worst rather than on her average or best exercise, even if they all could, in principle, have been improved.
Representativeness, or similarity reasoning, may play a part in scientific theories as well:
• A stutterer behaves in some respects in a similar way to a nervous per-son, and may indeed be anxious about not being able to communicate.
This has suggested anxiety as an aetiologic factor in some theories about stuttering (Attanasio, Onslow, & Packman, 1998).
• When children show few signs of empathy and social interest, a corres-ponding lack of empathy and interest on the part of their caregivers looks like a plausible cause. Thus childhood autism, with its remarkable impairment of reciprocal social interaction, was for many years believed to be due to inadequate mothering.
In these cases, representativeness reasoning suggested a false lead. But there are probably many more cases where the same line of reasoning provides valuable hints. For instance, violent and abusive adults have themselves often been abused by their parents. Violence breeds violence. This looks like a similarity inference, but it is also a truth.
Representativeness broadly defined
If representativeness applies to all the cases we have listed in this chapter, the original definition (Kahneman & Tversky, 1972, see above) appears too narrow. A more general formulation was suggested by Tversky and Kahneman (1982, p. 85): “Representativeness is a relation between a pro-cess or a model, M, and some instance or event, X, associated with that model,” as in the following four basic cases:
• M is a class and X is a value of a variable defined in this class (X could be the typical income of college professors).
• M is a class and X is an instance of that class (X is regarded to be a
“representative” American writer).
• M is a class and X is a subset of M (X is a “representative” sample of the US population).
• M is a causal system and X is a possible consequence.
“In summary, a relation of representativeness can be defined for (1) a value and a distribution, (2) an instance and a category, (3) a sample and a population, (4) an effect and a cause. In all four cases, representativeness expresses the degree of correspondence between X and M” (Tversky &
Kahneman, 1982, p. 87). This correspondence can be based on statistical beliefs (as in 1), causal beliefs (as in 4), and perceived similarity (as in 2 and 3). When this correspondence has been empirically established, for example by asking people to judge which of two events, X1 or X2, is more representa-tive of M, we would expect probability judgements to be influenced by the representativeness relation. If X1 is regarded as more representative than X2, it will appear to be more likely.
CRITICISMS
The concept of a representativeness heuristic, as well as the biases it was supposed to explain, have often been challenged. Some of the main criticisms are summarized below.
Conceptual vagueness
Representativeness is a very broad concept, applicable to a number of situations. This generality makes it both imprecise and difficult to falsify.
Gigerenzer, the strongest critic of the heuristics-and-biases programme, is not impressed by terms like representativeness, availability, and anchoring:
“These one-word labels at once explain too little and too much: too little, because the underlying processes are left unspecified, and too much, because, with sufficient imagination, one of them can be fit to almost any empirical result post hoc” (Gigerenzer. Todd, & the ABC Research Group, 1999, p. 28). This is a serious criticism if we expect a full-fledged theory capable of modelling and predicting human judgements with a high degree of accuracy. However, representativeness was originally proposed as a more descriptive term, capable of elucidating some general characteristics of human reasoning under uncertainty. The concluding section of the pres-ent chapter prespres-ents some recpres-ent speculations about the nature of the
“underlying processes”.
Biases can disappear
Not all studies show equally strong effects of representativeness. Moreover, in all studies there will be a substantial number of individual participants who appear less susceptible to representativeness reasoning. For instance in the random sequence experiment, many participants will say (correctly) that all sequences have the same probability of occurrence.
Such differences can be attributed to a variety of sources. One is situ-ational transparency. A concrete situation, in which procedures and mech-anisms are clearly visible, will increase the chances of a normative response.
Within-subjects studies, in which participants are asked directly to compare the alternatives, will typically yield more normative answers than between-subjects designs, in which the focal variables are more disguised. A group in a between-subjects design who are only shown the sequence HTTHTH will probably characterize it as a more likely than participants in another group who are asked to characterize the sequence HHHTTT, whereas individual participants who are asked to compare both sequences may “know” that they are equally likely.
People’s use of heuristics is also influenced by their degree of statistical sophistication, ability differences (Stanovich & West, 2000), and more gen-erally whether the task is conceived as a problem that should be solved by mathematical reasoning or simply by “gut feelings”.
Probabilities versus frequencies
Problems can sometimes be made more concrete and transparent by trans-lating probabilities into frequencies. Some evidence suggests, indeed, that people reason more normatively with natural frequency formats (Gigerenzer, 1991; see Chapter 3). But despite claims to the contrary, the judgement
“illusions” do not disappear. In several of the original demonstrations (including those presented in the first section of the present chapter) partici-pants were in fact asked about frequencies.
Even so, it has been suggested that the representativeness heuristic is espe-cially well suited for unique events, whereas the availability heuristic (see Chapter 8) is more applicable to frequentistic probabilities (Jones, Jones, &
Frisch, 1995). Frequency theorists, who believe that probabilities can only be meaningfully assigned to repeated events, have argued that probability judgements by representativeness cannot be given a mathematical interpret-ation, but invoke instead a credibility or plausibility concept (Hertwig &
Gigerenzer, 1999).
Biases are not errors
Some critics have argued that when people appear biased, it is not because they commit errors of judgement, but because the norms do not apply.
People may have been asked ambiguous questions, where a particular answer will appear incorrect given a literal interpretation of the task, but justified given a more pragmatic interpretation. For instance, a question about the likelihood of the HTTHTH sequence may be interpreted as a question about a sequence of “this type” (with alternating Hs and Ts) rather than about exactly this sequence. Conjunction tasks and base-rate tasks have similarly been given pragmatic interpretations that make “conjunction errors” and “base-rate neglect” less fallacious than they originally appeared.
A problem with this criticism is that it is typically raised post hoc (when the results are known) and often assumes that the participants are able to draw very fine distinctions in their interpretation of questions. Indeed, the participants are sometimes attributed a more sophisticated grasp of probability theory than the experimenters.
Alternative explanations
Not all the judgement “illusions” that have been attributed to the represen-tativeness heuristic may, in fact, be due to it. The conjunction fallacy may in some cases be due to a misplaced averaging rule, or judgements of surprise, as discussed in Chapter 1. Base-rate neglect, as discussed in Chapter 2, could sometimes be due to inversion errors (Villejoubert & Mandel, 2002). Simi-larly, the gambler’s fallacy may be due to more magical “balancing beliefs”, in addition to similarity judgements (Joram & Read, 1996). Finally, when
middle numbers in a lottery are preferred to extreme numbers (Teigen, 1983), it could be due to representativeness, but it could also signify a pref-erence for small errors over large ones (with ticket No. 1, one could be very wide of the mark).