A Restricted Empirical Bayes Approach to Detecting Genetic Association

• Activity Number: 299
• Type: Contributed
• Date/Time: Tuesday, August 2, 2011 : 8:30 AM to 10:20 PM
• Sponsor: Biometrics Section
• Author(s): Zhenyu Yang*+ and David R. Bickel
• Companies: University of Ottawa and University of Ottawa
• Keywords: multiple comparison procedure ; multiple testing ; empirical Bayes ; local false discovery rate ; GWAS ; constrained likelihood

Abstract

In microarray data analysis, measurements of the expression of thousands of genes enable simultaneously testing thousands of null hypotheses, where each null hypothesis says a particular gene is not differentially expressed across the conditions studied. Similarly, in a genome-wide association study, measurements of the genotypes across hundreds of thousands of DNA sites enable simultaneously testing hundreds of thousands of null hypotheses, where each null hypothesis says a particular site is not associated with a trait. At the hypothesis testing level, the main difference between GWAS data and microarray data is not the number of hypotheses tested but rather is the fact that p, the proportion of false null hypotheses, is much smaller in the case of genetic association data. (Leading geneticists put p for GWASs within an order of magnitude of 1 in 100,000.) We find that the standard false discovery rate (FDR) methods tend to identify many times as many sites as associated with disease than is biologically plausible in light of the smallness of p. By construction, a recent restricted-parameter method of estimating the local FDR (arXiv:1104.0341) yields much more tenable results.