Bias in error rate estimates in discriminant analysis when stepwise variable selection is employed

Abstract

Little attention has been given to properties of error rate estimates obtained from stepwise-selected discriminant analyses. We present simulation results that describe the effects of several factors on the error estimates' bias. Six estimators are compared; most can give substantially biased estimates when used in conjunction with stepwise discriminant analysis. Holdout methods are recommended because they are nearly unbiased and have desirable stability properties

Keywords:

• linear discrimination
• selection bias