Mallow’s Cp for Selecting Best Performing Logistic Regression Subsets

Abstract

Mallow’s Cp is used herein to select maximally accurate subsets of predictor variables in a logistic regression. Across a wide variety of data sets, an examination of the cross-validated prediction accuracy, posited as the ultimate criterion for model performance, contrasts the leave-one-out performance of Mallow’s Cp selections with the accuracy afforded by optimal subsets. Losses in accuracies ranged from no loss in several data sets up to a maximum of 10%. The performance of Cp selected subsets can be viewed as promising. It is posited that one should also consider parsimony and the richness of multiple optimal models.