Skip to main content
Category: Effects
Type: Judgment bias
Origin: Paul E. Meehl and Albert Rosen, Psychological Bulletin (May 1955); the counted file study in Daniel Kahneman and Amos Tversky, Psychological Review (July 1973); the cab problem as reported by Tversky and Kahneman (15 May 1981)
Also known as: Underweighting the prior
Quick Answer — Base Rate Neglect is the underuse of how common an outcome was, once a specific case, test, or witness report is added. Kahneman and Tversky counted it in 1973. With an empty file, people used the stated share. A sketch that fit either profession pulled both groups to 50 percent. The prior still belongs in the answer. The new evidence has to earn any change.

​
What is Base Rate Neglect?

Base Rate Neglect is the underuse of a prior. The prior is how common the outcome was before the new evidence. A case description, a test result, or a witness report arrives, and the judgment leans on that new piece.
The failure to appreciate the relevance of prior probability in the presence of specific evidence is perhaps one of the most significant departures of intuition from the normative theory of prediction.
That sentence is from Daniel Kahneman and Amos Tversky, Psychological Review, July 1973. A prior is the share you would use if the file were empty. Bayesian Thinking is the formal habit of combining that share with what the new evidence adds. Base Rate Neglect is the miss: the new piece is treated as the answer. A smoke alarm that is usually right about smoke still rings in a kitchen where fires are rare. The ring is about this room. How often fires happen here is a different number. You need both before you call it a fire. The alarm can be a real clue. It does not replace the share of kitchens that are actually burning. Representativeness Heuristic judges by resemblance. The 1973 paper treats the file study below as one result of that rule. A sketch that looks like an engineer is read as the chance he is an engineer. Base Rate Neglect is the narrower miss on this page. The prior is dropped, or nearly dropped, once the case arrives. A later medical quiz, below, is a case Leda Cosmides and John Tooby argue resemblance does not explain. The two names are neighbors. They are not aliases.

​
Base Rate Neglect in 3 Depths

  • Beginner: The cue is a vivid case. Before you read it, write the share you would give with an empty file. That share is the prior. The case has to move you off it for a reason you can point to.
  • Practitioner: Keep three numbers on one line. One is the empty-file share. One is what the new evidence would say if the two outcomes were equally common. One is the answer you are about to give. Landing on the second number, when the first number is far from equal, is the pattern in the studies below.
  • Advanced: A prior that explains the case and a prior that only counts the group are the same arithmetic and not the same judgment. The 50 percent against 55 percent gap below is a real shift. It is not the factor of 5.44 that the formal rule required. A questionnaire percent is not a clinic’s error rate, and it is not a hiring rate.

​
Origin

The warning is older than the student experiments. The counted laboratory pattern begins with a file of engineers and lawyers, and with a cab witness. In May 1955, Paul E. Meehl and Albert Rosen published “Antecedent probability and the efficiency of psychometric signs, patterns, or cutting scores” in Psychological Bulletin, volume 52, pages 194–216. Meehl was at the University of Minnesota. Rosen was at the VA Hospital in Minneapolis and at the same university. Their point is about a test, a sign, or a cutting score. It earns its place when it produces more correct decisions than the base rate alone would produce. They write that base rates were virtually never reported in that clinical literature, so the comparison often could not be made. This paper is a warning about published tests. It is not the engineer study, and it is not a count of hallway answers. Kahneman and Tversky put the prior into a file that subjects could use. “On the psychology of prediction” is 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. Unless a study in the paper says otherwise, the subjects were paid volunteers recruited through a student paper at the University of Oregon, in group settings. The case study below is their engineer and lawyer file. The 1974 article in Science lists insensitivity to prior probability as the first way representativeness departs from probability. That section points back to the 1973 study. It is not a new trial. Maya Bar-Hillel, in Acta Psychologica, volume 44, pages 211–233 (1980), names this laboratory pattern the base-rate fallacy. She explains it by relevance. Information that seems to be about the particular case outranks a rate that seems only to describe the group. This page keeps the name Base Rate Neglect for that underuse. A separate page can take the argument form. Amos Tversky and Kahneman returned to the prior in a report dated 15 May 1981, “Evidential Impact of Base Rates.” Tversky was at Stanford University. Kahneman was at the University of British Columbia. The report was prepared for the Office of Naval Research. It distinguishes a causal base rate from an incidental one. A causal base rate suggests a factor that makes any one case more likely to come out one way. An incidental base rate does not. The report traces the cab problem to their 1972 Oregon Research Institute monograph, “On prediction and judgment.” This page does not print a volume number for that monograph. The report says later work, including Don Lyon and Paul Slovic in Acta Psychologica, volume 40, pages 287–298 (1976), studied versions of the problem. The medians below are the ones the 1981 report itself states.

​
Key Points

Base Rate Neglect is on the page when a prior was available, the file was not empty, and the answer moved toward the new evidence. It is off the page when the file was empty, or when the prior itself explains the case.
1

Write the empty-file answer first

The prior is the answer you would give with no sketch, no test, and no witness. In the 1973 file study, that answer matched the stated share. The medians were 70 percent and 30 percent. In the 1981 cab problem, 85 percent of the cabs are green and 15 percent are blue. When the witness sentence was left out, the report says almost all subjects gave .15. Statistical Thinking starts from that share. The case has not earned a different number yet.
2

Make a blank sketch show its work

Worthless specific evidence is not the same as no evidence. Dick’s sketch was built to fit either profession. He is 30, married, with no children, able, motivated, and well liked. Both groups gave a median of 50 percent. The paper says the sketch carries no information about his profession. The formal answer is still the group’s share, 70 percent or 30 percent. A detail that would fit many roles has to be scored as blank. If your answer moved, name the feature that earned the move.
3

Do not read the witness as the posterior

A hit-and-run cab is seen at night. A witness says the cab was blue and is right about each color 80 percent of the time. The Bayesian answer in the 1981 report is .41. True blue reports are 80 percent of the 15 percent, which is 12 parts. False blue reports are 20 percent of the 85 percent, which is 17 parts. Twelve parts out of 29 is .41. The cab is still more likely green. Across slightly different versions, the report says the median and modal answer is typically .80, the witness’s hit rate. It does not print one sample size for that “typically.” In the engineer file, the shift from a low share to a high share was 50 percent against 55 percent. Bayes’ rule required the high group’s engineer-to-lawyer odds to be 5.44 times the low group’s, for each description. The small gap is not that factor.
4

Ask whether the prior explains the case

The same arithmetic is not the same story. When the 1981 report replaced the fleet share with a causal share, the two companies were about equal in size, and 85 percent of cab accidents involved green cabs. The Bayesian answer is still .41. The answers were highly variable, and the median was .60, between the witness’s .80 and .41. The base rate was no longer ignored. It was not the formal answer either. A version that replaced the witness with a less specific fact, an intercom installed in 80 percent of green cabs and 20 percent of blue cabs, had a median of .48. The report, citing Bar-Hillel, says that fact did not replace the base rate.

​
Applications

Use Base Rate Neglect as a check whenever a case arrives with a share you could have written first. Keep the case. Put the empty-file answer on the same page.

Write the empty file before the sketch

For a course file, a hiring note, or a promotion packet, write the share of people in that role before you read the sketch. In 1973 the empty file produced 70 percent and 30 percent. Dick’s uninformative sketch produced 50 percent in both groups. If the notes would fit either candidate, keep the share. A feature has to separate the roles before it moves the number.

Count the false alarms in this group

A positive test is not the chance of the condition. Write how common the condition is in the group being tested, and how many false alarms the error rate makes in that group. The hallway quiz below is the picture. It is one question, asked of 60 people at four teaching hospitals. It is not the rate at which those hospitals misread charts.

Put the witness next to the fleet

A review, a referee, or a neighbor can be right 80 percent of the time and still point at a rare type. Write the hit rate and the share of that type before you adopt the call. In the cab problem the hit rate is .80 and the blue share is .15. The report’s Bayesian answer is .41. The typical median it describes is .80. Those are questionnaire answers, not a court’s error rate.

Ask how the sample was picked

A family story, a class list, or a stack of interviews can carry a percent that was chosen, not a percent that explains the case. Tversky and Kahneman, summarizing Icek Ajzen’s 1977 experiments, report a larger move when the percent explained the outcome. For an exam, a high base rate of success raised the judged chance by .34 under the causal wording, and by .12 under the incidental wording, each compared with the low base rate. Both differences were significant. For a course choice, the causal means were .65 and .36. The incidental difference was .025 and was not significant. Ask who picked the sample.

​
Case Study

The counted file is the engineer and lawyer study in Kahneman and Tversky’s July 1973 paper. A cover story said a panel had interviewed and tested 30 engineers and 70 lawyers, all successful in their fields, or the reverse, 70 engineers and 30 lawyers. Thumbnail descriptions had been written. Subjects would see five, chosen at random from the hundred. For each one they gave the probability that the person was an engineer, on a scale from 0 to 100. The low-engineer group had 85 subjects. The high-engineer group had 86. Unless a study says otherwise, this paper’s subjects were paid volunteers recruited through a student paper at the University of Oregon. The same task, the cover story said, had been done by experts who were highly accurate. A bonus depended on coming close to those experts. The paper does not print the bonus amount. In each group, half the subjects 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. The paper reports the engineer side. One printed sketch is Jack. He is 45, married, with four children. He is conservative, careful, and ambitious. He shows no interest in political and social issues. His free time goes to home carpentry, sailing, and mathematical puzzles. The paper does not print a separate median for Jack in the text. After the five descriptions came the empty file. Suppose you are given no information about a person drawn at random. What is the probability he is an engineer? The medians were 70 percent in the high group and 30 percent in the low group. Those are the stated shares. Bayes’ rule, in odds form, is how the paper scores the rest. The likelihood ratio is whatever the sketch adds. It is the same sketch in both groups, so it cancels when the two groups are compared. For every description, the engineer-to-lawyer odds in the high group should be 5.44 times the odds in the low group. The figure plots each description’s median under the high prior against its median under the low prior. Only the empty file falls on the Bayesian curve. The other five points fall close to the line of equal judgments. Every one of those points is closer to that line than to the Bayesian curve. The effect of the prior is small, and it is statistically significant. For each subject the paper averages the five descriptions, leaving out the empty file. The average of those means was 50 percent in the low group and 55 percent in the high group, t = 3.23, df = 169, p < .01. The paper calls the effect minimal. It also says the prior was largely ignored once individuating information was available. Do not call the judgments identical. Do not call the 5-point gap the factor of 5.44. Dick shows how little the sketch had to say. He is 30, married, with no children. He has high ability and high motivation, promises to be successful, and is well liked. The description was built to be uninformative about his profession. The median was 50 percent in both groups. Dick is one of the five descriptions inside the 50 and 55 averages. He is not a sixth case with his own average.
Evidently, people respond differently when given no specific evidence and when given worthless evidence. When no specific evidence is given, the prior probabilities are properly utilized; when worthless specific evidence is given, prior probabilities are ignored.
That is the paper’s contrast between the empty file and Dick. Stereotype Effect is the broader habit of letting a group picture shape how a person is seen. Here the measured step is the prior. It was used when the file was empty. A sketch with nothing to say about the profession was enough to leave it. The result does not travel as a hiring rate. The subjects were not choosing employees. They were not shown a workplace base rate and then followed for a year. The 70, the 30, the 50, and the 55 are medians and averages on a questionnaire. A bonus was offered for matching unnamed experts. The design compares two stated shares. It does not assign a decision aid.

​
Boundaries and Failure Modes

Base Rate Neglect is the wrong name when the file was empty and the stated share was used. It is also the wrong name for a questionnaire percent carried into a clinic or a hiring meeting as that setting’s error rate. The first boundary is the empty file, and the size of the gap. In the engineer study, no description produced medians of 70 percent and 30 percent. Dick produced 50 percent in both groups. Across the five descriptions, null excluded, the averages were 50 percent and 55 percent. That difference is significant. It is not 5.44. The plotted points stay nearer equal judgments than the Bayesian curve. People in this study used the prior when they had no case. A blank sketch was enough to leave it. The phrase to avoid is “they ignored the base rate on every item,” and the phrase to avoid is “the two groups gave the same answer.” The second boundary is the kind of prior. In the 1981 report, a causal accident share moved the cab median from the typical .80 to .60. The answers were highly variable, and .60 is not .41. The intercom version’s median was .48. The report also shows the prior being used on the wrong question. When subjects were told that a witness had identified the color, but not which color, and were asked the chance the witness said blue, the median and modal answer was .15. The correct answer in the report is .29, because a witness who is wrong 20 percent of the time will say blue for some green cabs. The prior was available. It was aimed at the witness’s report as if the witness were perfect. Icek Ajzen’s 1977 experiments, as summarized in that report, found the same split. A percent that explained the exam or the course moved judgments more than a percent an investigator had picked. Those figures are the report’s summary. This page does not re-read Ajzen’s tables, and it does not treat the Yale exam in the cover story as a real course. The third boundary is export, including the medical quiz. In 1978, Ward Casscells, Arno Schoenberger, and Thomas B. Graboys asked a question in 67 consecutive hallway encounters at four Harvard Medical School teaching hospitals. The people asked were 20 house officers, 20 fourth-year medical students, and 20 attending physicians. The question stated a prevalence of 1/1000 and a false positive rate of 5 percent. It asked the chance that a person with a positive result has the disease, if you know nothing about that person’s symptoms or signs. Leda Cosmides and John Tooby, Cognition, volume 58, pages 1–73 (1996), report that 18 percent of that sample answered 2 percent, and 45 percent answered 95 percent. The 2 percent is the usual Bayesian reading. It treats the test as catching the one case in 1,000, which the printed question does not say. About 5 percent of the other 999 people would then be false alarms, about 50, and 1 true positive among about 51 positives is about 2 percent. Cosmides and Tooby gave the same printed question to Stanford undergraduates, 25 paid volunteers in that condition, average age 19.6. Twelve percent answered 2 percent. Fifty-six percent answered 95 percent. Another 12 percent answered with the base rate alone. A second condition changed several things at once. It stated a true-positive rate of 100 percent, defined the false-positive rate as 50 of every 1,000 healthy people, specified a random sample of 1,000, and asked for a frequency. Fifty-six percent of those 25 students answered 2 percent. The paper reports that comparison with the 12 percent as Z = 3.28, p = .0005. A third condition added three counting questions, and 76 percent of 25 students answered 2 percent. One student in the second condition, and none in the third, answered 95 percent. Do not merge the Stanford 12 percent with the Harvard 18 percent. Do not cite 45 percent, 56 percent, or 76 percent as a rate of misdiagnosis. Cosmides and Tooby argue that resemblance does not account for this quiz. They cite Tversky and Kahneman’s 1982 chapter, page 154. The lines they quote from that page say that educated respondents often fail to appreciate the outcome base rate, and that Meehl and Rosen’s warning applies to physicians as well. This page does not treat that citation as a new experiment. The engineer file and the hallway question are not one mechanism with one error rate.

​
Common Misconceptions

Three confusions travel with the prior. A blank sketch gets treated as information. A small gap gets treated as no gap, or as the full Bayesian gap. A questionnaire percent gets treated as a rate in a clinic or an office.
No description at all produced medians of 70 percent and 30 percent. Dick’s sketch was built to fit either profession, and both groups gave 50 percent. The paper’s contrast is between no evidence and worthless evidence. A detail that does not separate the outcomes is not a reason to leave the share. The five-description averages, 50 percent and 55 percent, include Dick. They are not a separate average for five sketches besides him.
The low group’s average was 50 percent and the high group’s was 55 percent, t = 3.23, df = 169, p < .01. The gap is real. Bayes’ rule required a factor of 5.44 in the odds, for each description, and the points sit closer to equal judgments than to that curve. “Largely ignored” is the paper’s phrase. “Identical,” and “off by a factor of 5.44,” are not.
The engineer numbers are a questionnaire about fictional professionals. The cab’s typical .80 is the report’s summary across slightly different versions, with no single sample size printed for that sentence. The causal median of .60 is not the Bayesian .41. The Harvard 18 percent and 45 percent, and the Stanford 12, 56, and 76 percent, are answers to one quiz. The Stanford cells have 25 students each. None of these figures is a count of missed diagnoses, overturned verdicts, or bad hires.
Base Rate Neglect sits next to a resemblance rule, a memory rule, and the formal way of combining a prior with evidence. The links below are neighbors. They are not other names for this miss.

Representativeness Heuristic

Judges a probability by resemblance. The 1973 paper ties the engineer file to that rule. Cosmides and Tooby argue that this resemblance rule does not explain the medical quiz.

Bayesian Thinking

Combines the prior with what the evidence adds. The factor of 5.44, and the cab answer of .41, are that combination. This page is what the judgments did instead.

Statistical Thinking

Asks what the share was before the case, and how the sample was drawn. Those are the lines this miss leaves blank once a sketch or a witness arrives.

Availability Heuristic

Judges frequency by how easily examples come to mind. A vivid case can be easy to recall and still be a rare type. Ease is a different operation from dropping a stated share.

Stereotype Effect

A group picture shapes how a person is seen and treated. The engineer file reads a professional picture as a probability and leaves the stated share.

Confirmation Bias

Favors evidence that fits a belief already held. A prior is a share, not a favored story. Neglect is underusing that share when the case arrives.

​
One-Line Takeaway

Write the empty-file share first, and make the new evidence earn any change.