Illusory Correlation: Seeing a Link the Numbers Do Not Support

Illusory correlation in psychology is the perception of a relationship between two variables that the data do not contain. It is an error about covariation, the judgment of whether two things go together, which is why it sits behind so many confident beliefs about what goes with what: a group and a kind of behavior, a treatment and a recovery, a car wash and a rainstorm.

Judging a relationship properly requires four counts. The times both things happened, the times the first happened without the second, the times the second happened without the first, and the times neither happened. People attend heavily to the first of those four cells and barely at all to the last, and an impression built from co-occurrences alone will find a relationship in almost any pair of common events. Rates matter rather than totals, and totals are what memory supplies.

The group version adds distinctiveness. When one group is smaller and one kind of behavior is rarer, the pairing of the two is the least frequent combination available, and it stands out. Those items get encoded more thoroughly and retrieved more easily, so a later frequency estimate or trait rating leans toward them. This is the distinctiveness-based account that Hamilton and Gifford proposed in 1976, following Chapman's earlier work on distinctive word pairs.

That account is not the only one, and the 1994 replication complicated it. McConnell, Sherman and Hamilton varied where in the sequence the rare pairings appeared, so that in one condition they were not rare at the moment they were read, and the illusory correlation appeared just as strongly in all three conditions. Distinctiveness at the time of encoding was not necessary; what mattered was that the items were ultimately rare when the judgment was made. Klaus Fiedler offered a different framing in 2000, a sampling account in which biases of this kind are traced to the information samples a judge has to work from, the sample the environment supplies and the sample the mind then draws from it, and to the constraints on both.

What is agreed is the finding. Equal ratios in unequal groups produce unequal judgments, and the smaller group gets the worse of it.

Loren Chapman gave the effect its name in 1967 in the Journal of Verbal Learning and Verbal Behavior. He showed people pairs of words drawn from two arrays and then asked how often words from one array had appeared with words from the other. Pairs that were strongly associated in meaning were reported as having occurred together more often than they had, and so were pairs made distinctive by their appearance, such as an unusually long word among short ones, even though every pairing had been presented equally often. Chapman and Jean Chapman extended the point to clinical judgment in the same period, showing how the same error would make practitioners confident in diagnostic signs that do not work.

David Hamilton and Robert Gifford took the paradigm to groups in 1976, in "Illusory correlation in interpersonal perception: A cognitive basis of stereotypic judgments" in the Journal of Experimental Social Psychology. Participants read 39 statements, each naming a person, a group and a behavior. Twenty-six statements described members of Group A and thirteen described members of Group B, so Group B was the minority, and both groups had the same ratio of desirable to undesirable behaviors, nine to four. Afterwards participants estimated how many of each group's behaviors had been undesirable, assigned behaviors back to groups, and rated the groups on trait scales. They overestimated the frequency of undesirable behavior by the smaller group and evaluated that group less favorably, with no prior information about either group and no actual difference between them.

Allen McConnell, Steven Sherman and David Hamilton replicated and tested the explanation in 1994 in the Journal of Personality and Social Psychology. Their 102 participants read 36 statements in one of three orders, and in the balanced order, the one the demo above follows, the 34 participants estimated the undesirable share at 54 percent for the smaller group and 39 percent for the larger, against a true value of 33 percent for both, and rated the smaller group at 4.97 against 6.53 for likability. The effect appeared in all three orders, which is the result that sent the explanation back for revision.

The pattern appears wherever an impression of covariation is built from remembered cases. The rows below are illustrations, not findings from the studies cited on this page.

The counters are arithmetic, because the bias lives in what gets remembered and a tally kept on paper does not.

• Draw the two-by-two table and fill in all four cells, including the one for the times neither thing happened. That cell is the one most often left uncollected, and often the one that dissolves the relationship.
• Use rates, not counts. Twelve incidents in a group of 400 and six in a group of 50 point the opposite way from the raw numbers.
• Ask for the base rate before forming the impression. How common is the behavior overall, and how large is each group?
• Count the non-events. The nights with no emergency, the patients with no side effect, the hires who worked out: these are evidence, and they tend to leave far less of a trace in memory than the events do.
• Record as you go. A tally kept at the time cannot be reweighted later by what was distinctive.
• Be most careful where the group is small and the event is rare, because that is exactly the combination the research says will mislead you, and it is also the combination where a wrong conclusion does the most damage to people.

Illusory correlation is specifically an error about a relationship between two variables, which separates it from its neighbors.

• It is not the clustering illusion, which is seeing structure in a single stream of random data, such as a streak in coin flips. Illusory correlation needs two variables and a judgment about how they covary.
• It is not stereotyping, which is the application of group-level expectations to an individual. The 1976 work was offered as one account of where such expectations can come from, since the effect appears in people who have no prior view of either group.
• It is not the base rate fallacy, which is the neglect of a known prior probability when a specific case is judged. In the illusory correlation experiments nobody is given the base rates; the error is in estimating them from what was read.