Abstract
Motivation: Combining P-values from multiple statistical tests is a common exercise in bioinformatics. However, this procedure is non-trivial for dependent P-values. Here we discuss an empirical adaptation of Brown’s Method (an extension of Fisher’s Method) for combining dependent P-values which is appropriate for the correlated data sets found in high-throughput biological experiments.
Results: We show that Fisher’s Method is biased when used on dependent sets of P-values with both simulated data and gene expression data from The Cancer Genome Atlas (TCGA). When applied on the same data sets, the Empirical Brown’s Method provides a better null distribution and a more conservative result.
Availability: The Empirical Brown’s Method is available in Python, R, and MATLAB and can be obtained from https://github.com/IlyaLab/CombiningDependentPvaluesUsingEBM.1
