An adaptive statistical method for the discriminant problem

This paper proposes an adaptive statistical method for the discriminant problem. The method selects Fisher’s linear or Smith’s quadratic discriminant function or the nearest neighbor method for use on the holdout samplle, depending upon which method minimizes the sum of overall accuracy and balance on the estimation sample. A simulation study examines the two group discriminant problem with variables generated from both bivariate normal and nonnormal distributions. The resulting misclassification rates indicate that the adaptive method is an effective alternative to existing statistical and linear programming methods.