Comparison of Stopping Rules in Forward Stepwise Discriminant Analysis

Abstract

Abstract Criteria based on conditional and estimated unconditional probabilities of correct classification are employed to compare alternative stopping rules that can be used with the forward stepwise selection method in the two-group multivariate normal classification problem. Based on Monte Carlo studies of 48 sampling situations, it is found that F to enter (.10 ≤ α ≤ .25) and a rule based on the maximum estimated unconditional probability often perform better than the strict use of all variables. Although the relative gains in classification are not large, the reductions in the numbers of variables to be used may be substantial.