Subspace False Discovery Rate

• Activity Number: 617
• Type: Contributed
• Date/Time: Thursday, August 4, 2011 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Jeffrey S. Spence*+ and Patrick S. Carmack and Thomas C. Ferree
• Companies: The University of Texas Southwestern Medical Center at Dallas and University of Central Arkansas and The University of Texas Southwestern Medical Center at Dallas
• Keywords: false positive errors ; multiple testing ; fMRI ; EEG power spectra ; microarray

Abstract

The false discovery rate (FDR), as originally formulated by Benjamini and Hochberg (BH 1995), has proven effective at limiting false positives in multiple testing. In more extreme cases, however, the power of the BH procedure declines considerably as a function of the number of tests and the fraction of NULL tests. It is common in gene expression studies and functional brain imaging studies, for example, to yield over 100,000 tests among which 99% are NULL. We propose a method that utilizes the BH procedure in subsets of the set of all tests, such that in the union, the FDR is controlled. Subspace FDR (sFDR) does not require any additional assumptions, yet can dramatically improve statistical power. We show, by calculation and simulation, the conditions under which sFDR performs optimally. In addition, we present a data-driven estimation method for the number and configuration of subsets that maximize power-95% of maximum power in best cases-without exceeding the nominal FDR. Finally, we present examples of sFDR in fMRI, EEG power spectra and microarrays, spanning a variety of spatial settings in which signal can be viewed as deriving naturally from different subspaces.