useR! 2024

The simultaneous analysis of a large number of statistical tests is ubiquitous in applications. The false discovery rate (FDR) is a popular error rate for avoiding Type-I error inflation, controlled by the famous BH procedure, which is implemented in R’s stats package. As this procedure was designed for continuous test statistics, it is known to be conservative for discrete data. More efficient methods that still guarantee FDR control in the latter setting have been proposed by Döhler, Durand & Roquain (2018). In our talk, we present the R package DiscreteFDR which provides efficient implementations of these procedures. It can be applied as an off-the-shelf tool for commonly used discrete tests such as Fisher’s exact test, or for any arbitrary discrete test by using information on its p-value distribution. After a brief introduction to the statistical background we focus on the implementation, which relies heavily on Rcpp. With large numbers of tests, some tweaks are required to lower computing and RAM requirements. The results are output in an S3 class, for which print, summary and plot methods are available. Finally, we will demonstrate the usage of the package with real data.