Towards Better FDR Procedures for Discrete Test Statistics

Abstract

The false discovery rate (FDR) has been a widely used error measure in situations where a large number of tests are conducted simultaneously. Most methods that control the FDR at a prespeci…ed level, or estimate the FDR of a multiple testing procedure (FDR procedures), were essentially developed for continuous test statistics. As such, their performances need to be carefully assessed when applied to discrete test statistics. We review some popular FDR procedures, point out a key reason for their excessive conservativeness when applied to discrete p-values, and suggest an improvement for these methods for such p-values.