Abstract

Paul Meehl (1954/1996) compared statistical with clinical prediction of important social outcomes (e.g., staying out of jail on parole).  He reviewed about 20 studies, in which information available to the statistical prediction rule (SPR) and the clinician were identical. The clinical prediction was never superior.  Later, Sawyer (1966) extended the work to situations in which the clinician had access to information (e.g. interviews) not available to the SPR; there was still no evidence in favor of the clinician.   Reviews involving over a hundred examples by Dawes, Faust, and Meehl (1989) and Grove and Meehl (1996) reinforced the initial findings.   

Most of the models were linear.  What Dawes and Corrigan showed in 1974 was that linear models with ad hoc or even randomly chosen (but in the right direction) coefficients outperformed clinicians.  These models worked because they gave results very close to the optimal SPR. Systems that treat people as if they are “interchangeable” (by using the same weights for all) are superior to those based on an “internal” analysis of particular people, the latter even suspect when variables such as “aggressiveness” are included, because such variables have meaning only with reference to other people.  Treating people as “interchangeable” also satisfies the Kantian categorical imperative.