Skip to main content
Category: Effects
Type: Judgment heuristic
Origin: Daniel Kahneman and Amos Tversky, the Hebrew University, Jerusalem, Cognitive Psychology (1972); prediction studies in Psychological Review (1973); the catalog in Science (27 September 1974)
Also known as: Judging probability by resemblance
Quick Answer — Representativeness Heuristic judges how likely a case is by how much the case resembles a type, or how much a sample looks like the process that would produce it. Daniel Kahneman and Amos Tversky set the rule out in 1972. A close resemblance can track a real pattern. It still leaves out how common the type is, how large the sample is, and how reliable the sketch is.

​
What is Representativeness Heuristic?

Representativeness Heuristic is the shortcut of judging a probability by resemblance. The case looks like a type, or the sample looks like the process that would have made it, and that look is used as the probability.
In the representativeness heuristic, probabilities are evaluated by the degree to which A is representative of B, that is, by the degree to which A resembles B.
That sentence is from Amos Tversky and Daniel Kahneman, Science, 27 September 1974. Their picture is a former neighbor’s sketch of Steve: shy, tidy, a need for order, and a passion for detail. People judge the chance he is a librarian by how much he resembles the stereotype of a librarian. The 1974 article does not print a percentage for Steve. It points to a counted study of the same kind of ranking. The 1972 paper splits the rule in two. A sample is representative when it is similar, in its essential features, to the population it is said to come from. It is also representative when it reflects the salient features of the process that generated it. A shirt can look like one person’s uniform. The fit is a real clue. The fit does not say how many such shirts that person owns, or whether today’s small pile is as steady as a month of laundry. Availability Heuristic judges by how easily examples come to mind. Anchoring Effect starts from a number and adjusts. Those are the other two shortcuts in the 1974 article. This one uses resemblance. In many situations the more representative outcome is in fact the more likely one. The misses come from facts that change the probability and do not change the resemblance: how common the type is, how large the sample is, and how much noise is in the evidence.

​
Representativeness Heuristic in 3 Depths

  • Beginner: The cue is a case that looks like a type you already know. Write the type. Then ask what else wears the same look, and how often that type actually shows up.
  • Practitioner: Keep three lines apart. One line is the resemblance. One line is how common the type is before the sketch. One line is how many observations the proportion comes from. The first line does not replace the other two.
  • Advanced: A column of average ranks can track similarity and still not be a probability for any one person. In the Tom W. study below, the correlation of .97 is nine pairs of averages. An empty file and a worthless sketch are also different. With no sketch, the 1973 medians followed the base rate. With a sketch built to fit either profession, both groups landed on 50 percent.

​
Origin

Kahneman and Tversky set this rule out in three papers. The 1972 paper names it. The 1973 paper tests prediction. The 1974 article places it beside two other shortcuts. In 1972 they published “Subjective probability: A judgment of representativeness” in Cognitive Psychology, volume 3, pages 430–454. Both authors were at the Hebrew University, Jerusalem. The abstract gives the rule two parts. Subjective probability follows similarity to the parent population. It also follows the salient features of the process that generated the sample. Sample size represents no property of the population. They predicted that size would have little or no effect on judged likelihood. They thank Maya Bar-Hillel for work in all phases of the project. She is not a coauthor of that paper. The same paper describes local representativeness. Essential features of the parent population are represented in the whole sample, and people also expect them in each part. A sample that looks representative in every short stretch has too many alternations and too few clusters. They call that expectancy a belief in the law of small numbers. The phrase is from their 1971 paper in Psychological Bulletin, volume 76, pages 105–110. In the 1972 wording, the law of large numbers is treated as if it applied to small numbers as well.
People predict by representativeness, that is, they select or order outcomes by the degree to which the outcomes represent the essential features of the evidence.
That is the thesis of “On the psychology of prediction,” Psychological Review, volume 80, pages 237–251, July 1973. The byline lists the Hebrew University of Jerusalem and the Oregon Research Institute. The journal records the paper as received on 8 November 1972. The studies use both naive subjects and sophisticated ones. The ranking of outcomes by likelihood coincided with the ranking by representativeness. People predicted rare events, and extreme values, when those outcomes were representative. The paper traces unjustified confidence, which it names the Illusion of Validity, to the same rule. It also traces common mistakes about regression to the mean to the expectation that the outcome should be as extreme as the evidence. One study in that paper is the engineer and lawyer task. A cover story said a panel had interviewed and tested 30 engineers and 70 lawyers, all successful, or the reverse, 70 engineers and 30 lawyers. Five short descriptions were drawn from the hundred. Unless a study in the paper says otherwise, the subjects were paid volunteers recruited through a student paper at the University of Oregon. The low-engineer group had 85 subjects. The high-engineer group had 86. Half the subjects in each group judged the probability that the person was an engineer. Half judged the probability that the person was a lawyer. That change of question had no effect. The medians for the two wordings added to about 100 percent for each description, and the paper reports the engineer side. Subjects were told that experts had been highly accurate, and that a bonus depended on coming close to the experts. The paper does not print the bonus amount. With no description, the medians were 70 percent in the high-engineer group and 30 percent in the low-engineer group. Bayes’ rule is the formal way to combine the starting share with what a description adds. For each description, that rule requires the engineer-to-lawyer odds in the high group to be 5.44 times the odds in the low group. Across the five descriptions, the paper instead averages each subject’s mean, leaving out the empty case. Those averages were 50 percent in the low group and 55 percent in the high group, t = 3.23, df = 169, p < .01. The gap is real, and it is small beside 5.44. The figure’s points sit closer to a line of equal judgments than to the Bayesian curve. A sketch written to fit either profession, called Dick, produced a median of 50 percent in both groups. Dick is 30, married, with no children, able, motivated, and well liked. The paper says the sketch carries no information about his profession. The 1974 article is “Judgment under Uncertainty: Heuristics and Biases,” Science, volume 185, number 4157, pages 1124–1131. Both authors are described there as members of the psychology department at the Hebrew University. The article says people rely on a limited number of heuristic principles. The principles reduce a hard probability question to a simpler judgment. They are quite useful, and they sometimes lead to severe and systematic errors. Representativeness is the first of the three. The article then lists ways the shortcut departs from probability: prior probability, sample size, chance, predictability, the illusion of validity, and regression. Several of those sections point back to the 1972 and 1973 papers. The list is a map. It is not six new trials inside the 1974 article.

​
Key Points

Representativeness Heuristic is doing its job when a resemblance is written down and then set beside the facts it does not contain. It is off the job when the resemblance is used as the probability.
1

Set the resemblance beside the count

Write what the case resembles. Then write how common that type is before the sketch, and how many observations you have. The 1974 sketch of Steve is the picture of the first line only. The counted version is Tom W., below. The psychology graduate students ordered fields by similarity to the sketch. Their likelihood ranks tracked those similarity ranks. The fields’ estimated shares ran the other way for the comparison the paper highlights.
2

Ask which sample is allowed to stray

A proportion can match the population and still be the less stable one. In 1972, Kahneman and Tversky asked about two hospitals. About 45 babies are born each day in the larger one, and about 15 in the smaller one. About half of babies are boys. Over a year, which hospital records more days on which more than 60 percent of the babies are boys? Sampling theory picks the smaller hospital. A large sample is less likely to stray from 50 percent. The subjects were 97 Stanford undergraduates with no background in probability or statistics, run in groups of 5 to 12. Each was paid 1,plus1, plus 1 more if a randomly chosen answer was correct. The item was the first of three, and it had two wordings. For “more than 60 percent,” 12 chose the larger hospital, 10 chose the smaller hospital, and 28 chose “about the same,” within 5 percent. The star in the paper is on the smaller hospital. Those three counts sum to 50. For “less than 60 percent,” 9 chose the larger hospital, 11 chose the smaller, and 25 chose “about the same.” The star is on the larger hospital. A less extreme day belongs in the larger sample. Those counts sum to 45. The paper’s sample sentence says 97. This page does not explain the difference of two.
3

Do not make a short run look like the long run

A fair process has a mix in the long run. A short stretch does not have to display that mix. The 1974 article describes a coin tossed for heads or tails. In that article, people regard heads, tails, heads, tails, tails, heads as more likely than heads, heads, heads, tails, tails, tails. They also rank it above heads, heads, heads, heads, tails, heads. The first looks irregular. The second looks patterned. The third has too many heads to look fair. A fair coin gives every specific sequence of that length the same probability. The 1974 article cites the 1972 paper for the judgment. The 1972 paper does not print a count of how many people chose among those three sequences. Statistical Thinking is the habit that asks for the length of the run before it reads the look.
4

Treat an empty file and a blank sketch as different

With no description, the 1973 engineer medians were 70 percent and 30 percent. Those are the stated shares. Dick’s sketch was built to be uninformative, and both groups gave a median of 50 percent. Worthless specific evidence pulled the judgment off the base rate. The five informative sketches left a gap of 50 percent against 55 percent. That gap is not the gap Bayes’ rule required. Before you trust a thin interview, write the judgment you would have made with no interview at all.

​
Applications

Use Representativeness Heuristic as a check on a judgment that is mostly a resemblance. Keep the resemblance. Put the missing count, the sample size, and the empty-file judgment on the same page.

Put the field's share next to the sketch

When a school file or a hiring note “looks like” a field, write the share of people in that field before you rank the person. In the Tom W. study, the base-rate group put humanities and education at 20 percent and computer science at 7 percent. The prediction group still ranked computer science as the more likely field for more than 95 percent of its members.

Write the size of a short streak

A week of quiz scores, a child’s new habit, or a small team’s wins can look like the new normal. Write how many observations the proportion uses. The hospital item is the picture. A day above 60 percent boys is a weaker sign in a ward of about 45 births than in a ward of about 15. The look of the day is the same. The sample is not.

Score the interview you could have skipped

If the notes would fit many roles, treat them as Dick’s sketch. Both base-rate groups in 1973 gave that sketch a median of 50 percent. Write the answer you would give with an empty file. If the two answers differ, the notes have to earn the difference with a feature that actually separates the roles.

Separate a cluster's look from its length

In a public meeting, a clinic huddle, or a news argument, a short cluster can look random, or it can look like a streak. Name the process you think it represents. Then write how many events are in the sample. A sequence that looks fair is not, on that look alone, more probable than a clump of the same length.

​
Case Study

The counted test of predicting by resemblance is Tom W., in Kahneman and Tversky’s July 1973 paper. Three groups answered different questions. The paper’s default pool, unless a study says otherwise, was paid volunteers recruited through a student paper at the University of Oregon, in group settings. A base-rate group of 69 people estimated the percentage of first-year graduate students in the United States in each of nine fields. The means were 15 percent for business administration, 7 percent for computer science, and 9 percent for engineering. Humanities and education were 20 percent. Law was 9 percent, library science 3 percent, and medicine 8 percent. Physical and life sciences were 12 percent. Social science and social work were 17 percent. A similarity group of 65 people read a personality sketch and ranked how similar Tom W. was to the typical student in each field. Rank 1 would be the closest match. The sketch calls him highly intelligent and short on creativity. He wants order, neat systems, and competence. The writing is dull, with occasional science-fiction flashes. He has little sympathy for other people, does not enjoy them, and is self-centered, with a deep moral sense. The mean similarity ranks were 2.1 for computer science and 2.9 for engineering. Business administration was 3.9, library science 4.2, and physical and life sciences 4.5. Law and medicine were both 5.9. Humanities and education were 7.2. Social science and social work were 8.2. The prediction group was different people. It was 114 graduate students in psychology at three major universities in the United States. They read the same sketch. They were told a psychologist had written it in Tom’s senior year of high school, from projective tests, and that Tom was now a graduate student. They ranked the nine fields by the likelihood that he was in each one. The mean likelihood ranks were 2.5 for computer science and 2.6 for engineering. Business administration and physical and life sciences were both 4.3. Library science was 4.7, law 5.2, and medicine 5.8. Humanities and education were 7.6. Social science and social work were 8.0. The paper correlates the columns of means in its Table 1. The correlation between judged likelihood and similarity is .97. The correlation between judged likelihood and the estimated base rate is -.65. For that second correlation, the ranks were reversed so that a high likelihood counted as a high value. These are correlations across nine fields. They are not correlations computed inside each student. More than 95 percent of the prediction group judged computer science more likely than humanities or education. The base-rate means put those prior odds at about 3 to 1 the other way, 20 percent against 7 percent. The paper says the actual odds are considerably higher. After the ranking, the same graduate students estimated how often several kinds of information would pick the right field on the first try. The median was 23 percent for projective tests. It was 53 percent for high school seniors’ own reports of their interests and plans. They held the tests in low esteem. They still ordered the fields by a sketch that came from those tests. The lesson is narrow. Likelihood order tracked similarity order. It did not track how common the fields were. The evidence they used was evidence they called weak. Stereotype Effect is the broader habit of letting a group picture shape perception and treatment. Here the measured step is narrower. A resemblance rank was used as a likelihood rank. The result does not travel as a hiring rate or a diagnosis rate. The .97 is nine pairs of averages. The people who estimated the field shares were not the people who ranked Tom. Nobody in the prediction group was shown those shares and then asked to ignore them. The design compares judgments across groups. It does not assign a decision aid.

​
Boundaries and Failure Modes

Representativeness Heuristic is the wrong name for a judgment that already uses the base rate, the sample size, and the reliability of the evidence. It is also the wrong name for a questionnaire percentage carried into a hiring meeting or a clinic as if it were that setting’s error rate. The first boundary is the empty file. In the engineer study, no description at all produced medians of 70 percent and 30 percent. The groups used the shares they had been given. Dick’s uninformative sketch produced 50 percent in both groups. The five other sketches produced average means of 50 percent and 55 percent. That difference is statistically significant. It is not the factor of 5.44 that Bayes’ rule required, and the plotted points stay nearer equal judgments than the Bayesian curve. People in this study used the base rate when the file was empty. A sketch with nothing to say about the profession was enough to leave the base rate. Do not describe the result as identical judgments, and do not describe it as a total neglect on every item. The second boundary is export. Do not cite .97, -.65, or “more than 95 percent” as the rate at which hiring, admissions, or diagnosis follows a stereotype. Do not cite 23 percent or 53 percent as a measured hit rate of projective tests or of students’ own reports. Those are medians of what the graduate students guessed about hit rates. Do not cite 12, 10, and 28, or 9, 11, and 25, as the rate at which hospitals, parents, or analysts ignore sample size. They are choices on one item, by Stanford undergraduates, under two wordings. Do not cite the 1974 parentheses of 21, 21, and 53 as the answers to the single question about days above 60 percent. Those parentheses sit beside that wording in the 1974 article. They equal the sums of the two 1972 forms, 12+9, 10+11, and 28+25. The correct hospital is the smaller one for “more than,” and the larger one for “less than.” The third boundary is the word “probable.” In 1983, Tversky and Kahneman, in Psychological Review, volume 90, pages 293–315, showed that a conjunction can be more representative than one of its parts. Linda is described as a 31-year-old philosophy student, outspoken, and concerned with discrimination and social justice. “Bank teller and active in the feminist movement” resembles that sketch more than “bank teller” does. The conjunction cannot be more probable than the broader category, because everyone who is both is also a bank teller. Ralph Hertwig and Gerd Gigerenzer, in the Journal of Behavioral Decision Making, volume 12, pages 275–305 (1999), argue that many readers take “probable” to mean plausible or supported by the story. In their Study 3, 20 students at the University of Munich estimated frequencies for a set of 200 women who fit Linda’s description. None of them violated the conjunction rule. A footnote in that paper puts the median violation rate, across listed frequency-estimation versions of Linda, at 17 percent, and a reviewed median for probability rankings at 87 percent. Those medians mix studies. Study 3 is 20 people and one wording. A fitting detail can raise resemblance. Whether a person treats that rise as a rise in probability depends on the question’s words. A valid sign is still allowed to count. The 1974 article says these shortcuts are quite useful when the resemblance really does track the process. The error is letting the resemblance stand in for the base rate, the sample size, or the reliability of the evidence. In the same 1973 paper, 73 of 108 graduate students were given an observed IQ of 140. They had been told the score mixes a true score and a random error. They still put a 95 percent interval symmetrically around 140. On a separate flight-school question from the authors’ own experience, not one of the graduate students who answered named regression. Instructors had seen good maneuvers followed by worse ones and had blamed the praise. That pair of results belongs with regression to the mean. It is not a second Tom W.

​
Common Misconceptions

Three confusions travel with the shortcut. Resemblance gets treated as the probability. A short sample gets treated as a portrait of the whole process. A questionnaire count gets treated as a rate in the world.
Resemblance is one input. In the Tom W. study it was used as the likelihood rank. More than 95 percent of the prediction group placed computer science above humanities and education. The separate base-rate group had put those fields at about 7 percent and 20 percent. The prediction group also said projective tests would pick the right field about 23 percent of the time. A close match to a picture is not yet a probability. The picture still has to be common enough, and the evidence still has to be reliable enough, for the match to carry the conclusion.
On the “more than 60 percent” wording, 28 of the 50 printed choices were “about the same.” Sampling theory picks the smaller hospital for that wording, and 10 students chose it. On the “less than 60 percent” wording, the correct choice is the larger hospital, and 9 students chose it, against 25 who said “about the same.” The day’s percentage can resemble the population equally in both wards. The chance of a day that far from 50 percent is higher where fewer babies are born. The 1974 counts of 21, 21, and 53 are not this split.
The .97 is the correlation, across nine fields, between two columns of mean ranks. One column is similarity, from 65 Oregon volunteers. The other is likelihood, from 114 psychology graduate students. The -.65 compares that likelihood column with the base-rate column from 69 other volunteers. No student in the prediction group produced all three judgments. The design is not a count of job offers, admissions letters, or clinical calls. Base Rate Neglect takes up the engineer study’s 50 percent against 55 percent. Those two numbers are not a workplace error rate either.
Representativeness Heuristic sits next to shortcuts that use memory, a starting number, a group picture, or a felt pattern. The links below are neighbors. They are not other names for this rule.

Availability Heuristic

Judges frequency by how easily examples come to mind. The 1974 article lists it beside this rule. Ease of recall is a different operation from resemblance.

Anchoring Effect

Starts from a given number and adjusts. It is the third shortcut in the 1974 article. A resemblance rank does not need a numeric anchor.

Regression to the Mean

Extreme evidence is followed by a less extreme result when the measure is noisy. The 1973 paper says representativeness makes that fact hard to expect, because the outcome is supposed to match the evidence.

Statistical Thinking

Asks what varies, how large the sample is, and what the base share was. Those are the facts this shortcut leaves out when it stops at resemblance.

Stereotype Effect

A group picture shapes how a person is seen and treated. This rule is the further step of reading that picture as a probability.

Illusory Correlation

A pairing that fits a theory is felt as a real link. Local representativeness is about a short sequence looking like its process. The felt link is a neighbor, not the same result.

​
One-Line Takeaway

Write the resemblance, then the base share, the sample size, and the judgment an empty file would have given.