Description:

Grant Izmirlian, Ph.D.

NCI DCP Biometry Research Group

Presentation Title:

False Discovery Type Procedures: caveats to reproducibility–its all about dispersion

Abstract:

The Benjamini-Hochberg false discovery rate (BH-FDR) procedure is a step-up multiple testing procedure which guarantees control of the false discovery rate (FDR). The FDR is the expected false discovery proportion (FDP) which is an unobserved random ratio. The procedure is typically applied in --omics studies in which a large number of markers is to be thresholded yielding a shorter list of markers for further investigation. In the planning and design of these studies, the average power, which is the expected true positive proportion (TPP) is used. Specifically, a sample size is determined that offers specified average power to control the FDR level given design parameters such as sample size, effect size and prior probability that a given marker is a true signal. In my talk I will demonstrate first by presenting some theoretical results and then with a large simulation study, that both the FDR and average power can be very unreliable due to non-negligible dispersion in the distributions of the FDP and TPP. I will then present alternative procedures for controlling the tail probability of the FDP, and a concept of power which is based upon the tail probability of the TPP. The R function, ‘pwrFDR’, calculates all notions of power under all procedures for controlling the distribution of the FDP. Furthermore, based upon the theoretical results it determines a recommended FDP procedure. We will have time after the talk to demonstrate usage of the package.

For Zoom link please contact Lucia Ontiveros: lucia.ontiveros@northwestern.edu