Representativeness Heuristic

REPRESENTATIVENESS

Many probabilistic questions people face involve judging the likelihood that an object belongs to a class, an event originates from a process, or a process will generate an event. In answering such questions, people commonly rely on the representativeness heuristic: probabilities are evaluated by the degree to which A is representative of B, that is, by how much A resembles B—hence it is sometimes called the similarity heuristic. For example, if an instance is highly similar to a stereotype, the probability that it comes from that class is judged to be high; if it is dissimilar, the probability is judged low. This heuristic is a case of attribute substitution: a difficult question (“How likely is it?”) is replaced by an easier one (“How similar is it?”), which is answered rapidly and automatically by intuitive, System 1 thinking. Deliberate, rule-governed System 2 reasoning can override this substitution, but often fails to do so unless special effort is applied.

Representativeness belongs to a broader family of prototype heuristics, in which the properties of a prototypical exemplar dominate global judgments about an entire set. The heuristic often produces useful impressions—people who act friendly usually are friendly, very tall and thin professional athletes are more likely to be basketball players, and PhDs are more likely to read The New York Times than those with only a high school education. However, because similarity ignores factors that should affect probability, relying on representativeness alone leads to systematic and severe errors, sometimes called the “sins” of representativeness.

Strong evidence for attribute substitution comes from studies in which separate groups of participants rank outcomes by similarity and by probability. When given a personality sketch of a fictitious graduate student, Tom W.—described as high in intelligence but lacking creativity, needing order and clarity, writing in a dull and mechanical way, and having little sympathy for others—one group ranked nine fields of graduate specialization by how much Tom W. “resembles a typical graduate student,” while another group ranked the fields by how likely he was to be in each. The correlation between the average representativeness and probability rankings was nearly perfect (.97). Similarly, for the famous Linda problem—where Linda is described as a single, outspoken, philosophy-major activist—separate rankings of outcomes by representativeness and by probability yielded a .99 correlation. In that problem, most participants ranked “bank teller and active in the feminist movement” as more probable than “bank teller,” violating a basic logical rule because a conjunction cannot be more likely than one of its constituents. This conjunction error is a direct product of representativeness: Linda more closely resembles a feminist bank teller than a bank teller, so similarity overrules logic. Even those who know the correct answer often feel the intuitive pull; as Stephen Jay Gould reflected, “I know [the right answer], yet a little homunculus in my head continues to jump up and down, shouting at me ‘but she can’t just be a bank teller; read the description!’”

---

Insensitivity to prior probability of outcomes (base-rate neglect)

Prior probability, or base-rate frequency, is a factor that has no effect on representativeness but should greatly affect probability judgments. For instance, when assessing whether Steve—described as “very shy and withdrawn, invariably helpful, but with little interest in people… a meek and tidy soul, with a need for order and structure and a passion for detail”—is more likely a librarian or a farmer, people judge the probability by similarity to the stereotypes of those occupations. Because the description is more representative of a librarian, people rank librarian as more probable, even though there are many more farmers in the population. Base-rate frequency does not alter similarity, so it is neglected.

The Tom W. study illustrates the neglect of base rates starkly. Participants in a third group estimated the proportion of graduate students enrolled in each of the nine fields, yielding base-rate estimates that varied widely (from 3% for Library Science to 20% for Humanities and Education). The description of Tom W. was deliberately constructed to be representative of smaller fields (e.g., library science, computer science) and unrepresentative of larger ones. As intended, the correlation between representativeness and base rates was strongly negative (–.65). Yet the probability rankings almost perfectly mirrored representativeness and completely failed to regress toward the base rates. This is the hallmark of base-rate neglect: even when the personality description is of little diagnostic value, probability judgments stay anchored to similarity rather than shifting toward the base-rate frequencies.

People do use base rates correctly when no individuating information is available. In the absence of a personality sketch, they will estimate the probability that an unknown person is an engineer as .7 when told that a group is 70% engineers and 30% lawyers, and as .3 when the proportions are reversed. But when given a description—even a completely uninformative one such as “Dick is a 30-year-old man… of high ability and high motivation… well liked by his colleagues”—people judge the probability to be .5 regardless of the stated base rates. When no evidence is given, prior probabilities are properly used; when worthless evidence is given, they are ignored.

This neglect is fueled by the fact that once a specific description is provided, what you see is all there is (WYSIATI); the associative machinery creates a vivid impression that overwhelms abstract statistical facts. The result is an excessive willingness to predict unlikely, low-base-rate outcomes. For example, when you see a person reading The New York Times on the New York subway, representativeness suggests the person has a PhD, yet the far larger number of non-college graduates on the subway makes that a poor bet. Similarly, a shy poetry lover is more likely to be a business student than a Chinese literature student, simply because of the vastly larger base of business students—even if every Chinese literature student matches the description.

The failure to attend to base rates is not always absolute, but it often requires deliberate cognitive effort. Experiments show that many people do integrate base-rate information when it is explicitly provided, though the individual case information is almost always weighted more heavily. Instructing people to “think like a statistician” increases use of base rates, while instructing them to “think like a clinician” lowers it. A particularly striking demonstration of the role of mental effort comes from a study with Harvard undergraduates: when told to frown during the task—an expression that increases System 2 vigilance and reduces intuitive overreliance—participants showed some sensitivity to base rates. Those who puffed out their cheeks (an emotionally neutral expression) replicated the typical neglect, relying exclusively on representativeness. This indicates that base-rate neglect is often a product of System 2 laziness rather than ignorance; the mind “knows” that base rates are relevant but applies that knowledge only when it invests special effort.

When the quality of evidence is untrustworthy or scanty, the correct procedure is to keep probability estimates very close to the base rate. For example, in the Tom W problem—where the personality sketch is admitted to be of low validity—the proper response is to slightly reduce the initially high probabilities of well-populated fields and slightly raise the low probabilities of rare specialties, staying near the base-rate anchoring point. This discipline requires active self-monitoring and is not easily executed, because WYSIATI makes it difficult to treat worthless information as equivalent to no information.

---

Insensitivity to sample size

To evaluate the probability of a sample result, people typically rely on the representativeness heuristic. They assess the likelihood of a sample statistic—say, the average height in a random sample of ten men—by its similarity to the corresponding population parameter. Because similarity does not depend on sample size, the judged probability of a sample statistic becomes essentially independent of how many observations are included. When subjects estimated the distribution of sample means for heights, they assigned the same probability of obtaining an average greater than 6 feet whether the sample was 10, 100, or 1,000 men. Emphasizing sample size in the problem did not eliminate this error. Consider the classic hospital problem:

A certain town is served by two hospitals. In the larger hospital about 45 babies are born each day, and in the smaller hospital about 15 babies are born each day. About 50% of all babies are boys, but the exact percentage varies from day to day. Over one year, which hospital recorded more days on which more than 60% of the babies born were boys?

Most participants judged the probability to be about the same for both hospitals, because the event “more than 60% boys” is described by the same statistic and thus appears equally representative of the general population. Sampling theory, however, dictates that the smaller hospital will have far more such days, because a small sample is more likely to deviate from the population mean. This fundamental statistical principle is not part of people’s intuitive repertoire.

The same insensitivity appears in judgments of posterior probability. Given an urn that is either ⅔ red and ⅓ white, or the opposite, people find a sample of 4 red and 1 white balls more convincing evidence for the hypothesis that red is predominant than a sample of 12 red and 8 white, because 80% red seems more representative than 60% red. The correct posterior odds, assuming equal prior probabilities, are 8 to 1 for the small sample and 16 to 1 for the large sample, meaning the larger sample provides stronger evidence despite a less extreme proportion. Intuitive estimates are dominated by the sample proportion, with little regard for sample size, and posterior odds estimates are typically far less extreme than the correct values—a pattern labeled “conservatism.”

---

Misconceptions of chance

People expect that even short sequences generated by a random process will represent the essential characteristics of that process. In coin tosses, sequences like H-T-H-T-T-H are seen as more likely than H-H-H-T-T-T, which does not appear random, or H-H-H-H-T-H, which does not represent the fairness of the coin. This belief in local representativeness leads to an expectation of alternation and a corresponding overestimate of alternations and underestimation of runs.

The gambler’s fallacy—the belief that after a long run of red on a roulette wheel, black is “due”—is a direct result. People view chance as a self-correcting process in which deviations in one direction are compensated by opposite deviations, when in fact deviations are merely diluted as the sequence grows. This misconception also extends to experienced researchers, who are subject to a “law of small numbers,” believing that even small samples will be highly representative and yield statistically significant results. The term originated from a survey of 84 participants at the 1969 meetings of the Mathematical Psychology Society and the American Psychological Association, where even statistically sophisticated respondents expected the law of large numbers to apply to small numbers, revealing a broad bias. Consequently, researchers put too much faith in the replicability of findings from small samples, leading to underpowered studies and overinterpretation.

---

Insensitivity to predictability

Numerical predictions, such as the future profit of a company or the long-term success of a teacher, are often made by representativeness. A very favorable description leads to a prediction of very high profit because that outcome seems most representative; a mediocre description leads to a mediocre prediction. The degree to which the description is favorable is unaffected by how reliable it is or how much it actually permits accurate prediction. Thus, people make predictions that are insensitive to the expected accuracy.

Normative statistical reasoning dictates that when predictability is nil, the same value (e.g., the mean) should be predicted in all cases, and as predictability increases, the range of predictions should widen. In one study, subjects evaluated the quality of a student teacher’s practice lesson in percentile scores and, separately, predicted the same teacher’s standing five years later. The judgments were identical: predictions of a remote criterion were just as extreme as evaluations of the current evidence, ignoring the massive uncertainty inherent in such long-term forecasting. Even though the participants were aware of the limited predictability, their predictions mirrored their evaluation of the input. Similarly, when predicting a student’s GPA based on a test score at the 90th percentile, people will forecast a GPA just as far above average whether the predictor is a highly diagnostic measure of mental concentration, a less diagnostic measure of sense of humor, or even the GPA itself. They fail to adjust their predictions for the limited accuracy of the predictor—a neglect of regression toward the mean.

---

The illusion of validity

When people predict by selecting the outcome most representative of the input, their confidence depends primarily on the degree of representation—the quality of the match between the predicted outcome and the input—with little regard for factors that limit accuracy. A good fit between a description and a stereotype (e.g., a shy, orderly personality and the librarian stereotype) generates great confidence, even if the description is scanty, unreliable, or outdated. This unwarranted confidence is the illusion of validity.

The internal consistency of input information is a major driver of confidence. People feel more certain when predicting the final grade point average of a student whose first-year record consists entirely of B’s than one whose record contains a mix of A’s and C’s. Highly consistent patterns often result from redundant or correlated variables. Yet, in statistics, predictions based on independent inputs are more accurate than those based on redundant ones, given the same individual validities. Redundancy decreases accuracy even as it increases confidence, leading people to be especially confident in predictions that are quite likely to be wrong. This illusion helps explain why many clinicians maintain faith in the unstructured interview despite extensive evidence of its fallibility.

---

Misconceptions of regression

When two measurements are imperfectly correlated, extreme values tend to be followed by values closer to the mean, a phenomenon known as regression toward the mean. For example, children who score highest on one version of an aptitude test will, on average, do somewhat worse on a second equivalent version, and those who score lowest will tend to improve. Despite many natural encounters with regression—in the heights of fathers and sons, the performance of stocks, or consecutive athletic achievements—people do not intuitively expect it. They tend to seek causal explanations when regression occurs, rather than recognizing it as a statistical inevitability.

This blindness arises because regression conflicts with the representativeness heuristic: if the prediction is to be maximally representative of the input, the outcome should be as extreme as the evidence. Because representativeness does not account for regression, people overlook or misinterpret it. A classic illustration is flight instructors’ observation that praising an exceptionally smooth landing is typically followed by a worse landing, while harsh criticism after a rough landing is typically followed by improvement. The instructors concluded that punishment is effective and reward is detrimental. In reality, performance fluctuates with chance errors, so extreme performances naturally regress to the mean regardless of feedback. Because instructors praised after good landings and criticized after poor ones, they fell prey to a pernicious illusion. A similar dynamic pervades social interaction: rewards are usually given when performance is good, and punishment when performance is poor, so by regression alone, behavior will most frequently improve after punishment and deteriorate after reward. People are generally unaware of this contingency, leading to an overestimation of the effectiveness of punishment and an underestimation of reward.

The influence of representativeness extends beyond these specific errors to shape everyday causal beliefs. The principle of “like goes with like” often leads people to accept dubious cause–effect links simply because the cause resembles the effect, and to resist valid explanations that violate this resemblance. A striking historical example is the initial skepticism toward Walter Reed’s discovery that malaria is carried by mosquitoes: a tiny mosquito seemed utterly unrepresentative of the devastating disease it causes, violating the expectation that big effects should have big causes. The very term representativeness has entered everyday discourse precisely because it captures a common lexical void—people frequently invoke it to describe cases where an assessment of similarity guided or misguided a causal inference. The heuristic is thus both a powerful mental shortcut and a pervasive source of error, operating not only in formal probabilistic judgments but also in the routine fabric of how we explain the world around us.