Group differentiated prediction

Abstract

Studies of predictive bias have frequently shown that a prediction equation based on majority group members tends to overpredict the criterion performance of minority group members. Two statistical artifacts that may cause the overprediction finding are reviewed and evaluated using data for black and white students at 30 law schools. It is shown that (1) the degree of overprediction decreases as the predictive accuracy for white students increases, and (2) that overprediction can be caused by the effects of selection on variables not included in the regression model. Use of Heckman’s (1979) procedure to adjust the estimates of the regression parameters was found to essentially eliminate overprediction.