Fall Seminar Series - November 17, 2005
University of Minnesota
School of Statistics
College of Liberal Arts

Some Recent Work on Multiple Comparisons: Non-discovery Rate and Stratified False Discovery Control

Radu Craiu
Department of Statistics
University of Toronto

Thursday, November 17, 2005
3:30 PM, 115 Ford Hall
Minneapolis, East Bank Campus
Social at 3:00 PM, 300 Ford Hall

Abstract

The multiplicity problem has become increasingly important in many scientific studies. The control of false discovery rate (FDR) represents a different point of view than the classical control of the familywise error rate (FWER). In this talk, I will present our recent work on two FDR-related problems. I will first discuss a new quantity of interest, non-discovery rate (NDR) which is the fraction of non-rejections among the cases in which the null hypotheses are false. I will describe the property and utility of NDR and use it to address several issues including a). type II error rate and power in the context of multiple hypothesis testing. b). trade-off relationship between FDR and NDR. c). choice of an appropriate FDR level. Secondly, I will investigate the performance of false discovery control after a set of p-values has been stratified based on auxilliary information. Under the fixed rejection region framework, I will demonstrate that the aggregated FDR is a weighted average of the stratified FDR. Under the fixed FDR framework, I will provide the condition under which the expected total number of correct rejections from the stratified FDR approach is greater than that from the aggregated FDR approach. For both problems, I will illustrate the ideas with analyses of microarray gene-expression data, pedigree data and genome-wide association data.

This is joint work with Lei Sun from the Departments of Public Health Sciences and Statistics, University of Toronto.