False discovery rate controlling procedures for discrete tests

TL;DR

This paper develops more powerful false discovery rate controlling procedures tailored for discrete data, improving upon existing methods and demonstrating their application in pharmacovigilance for early adverse reaction detection.

Contribution

It introduces novel FDR procedures that exploit data discreteness and outperform previous methods, including a new step-down procedure dominating existing ones.

Findings

New procedures have FDR levels closer to the nominal level.

The proposed step-down procedure outperforms previous methods.

Application to pharmacovigilance demonstrates practical utility.

Abstract

Benjamini and Hochberg (1995) proposed the false discovery rate (FDR) as an alternative to the family-wise error rate in multiple testing problems, and proposed a procedure to control the FDR. For discrete data this procedure may be highly conservative. We investigate alternative, more powerful, procedures that exploit the discreteness of the tests and have FDR levels closer in magnitude to the desired nominal level. Moreover, we develop a novel step-down procedure that dominates the step-down procedure of Benjamini and Liu (1999) for discrete data. We consider an application to pharmacovigilance spontaneous reporting systems, that serve for early detection of adverse reactions of marketed drugs.