The Impact of Static Count Significance Threshold Selection on False Discovery Rate Estimation for Gene Expression Data

• Activity Number: 617
• Type: Contributed
• Date/Time: Thursday, August 4, 2011 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Nicholas Bradley Larson*+ and Dan Nettleton
• Companies: Iowa State University and Iowa State University
• Keywords: FDR ; correlation ; microarray ; bias

Abstract

The false discovery rate (FDR) is a convenient measure of statistical significance when conducting multiple simultaneous hypothesis tests. This procedure is of particular use in gene expression data analysis, where investigators wish to discover differentially expressed genes. If investigators determine their significance threshold a priori by universally selecting a particular number of top differentially expressed genes, however, the corresponding estimates of FDR may be misleading. We identify a negatively correlated relationship between the estimates of the FDR and the true proportion of Type I errors via simulations using experimental microarray data under a variety of estimators and conditions. We argue that this phenomenon is due to the correlated nature of the gene expression values, which results in a variable null p-value distribution. We discuss how this leads to FDR estimates being liberally biased when publication bias is taken into account, and caution against the use of a static gene count significance threshold.