Estimation procedures for the false discovery rate: a systematic comparison for microarray data

The microarray technology developed in recent years allows for measuring expression levels of thousands of genes simultaneously. In most microarray experiments the measurements are taken under two experimental conditions. Statistical procedures to identify differentially expressed genes involve a serious multiple comparison problem as we have to carry out as many hypothesis testings as the number of candidate genes in the experiment. If we apply the usual type I error rate α in each testing, then the probability to reject any truly null hypothesis will greatly exceed the intended overall α level. We focus on the recent error control concept of the false discovery rate FDR for which an increasing number of competing estimates as well as algorithms is available. However, there is little comparative evidence. For parametric as well as nonparametric test statistics relevant FDR procedures and typical parameter settings are discussed, including the use of ‘correcting’ constants in the estimation of the pooled variance. An in-depth simulation study is performed aiming at the aforementioned points with respect to sound statistical inference for microarray gene expression data. Finally the famous Hedenfalk data set is analyzed in a similar fashion and conclusions are drawn for practical microarray analysis.