A Bayesian Model Averaging Approach to Gene-Environment Interaction in a GWAS

• Activity Number: 441
• Type: Invited
• Date/Time: Wednesday, August 3, 2011 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Dalin Li and David V. Conti*+
• Companies: University of Southern California and University of Southern California
• Keywords: Gene-Environment Interaction ; Genetic Association ; Bayesian Model Averaging

Abstract

Scans for gene(G)-environment(E) interaction have been mostly neglected in current GWAS. The conventional approach, the case-control analysis, suffers from low power. The alternative case-only analysis can be more powerful but violation of the assumption of independence can greatly bias the results. We propose a model averaging approach for GxE interaction leveraging the fact that most SNPs in a GWAS are not associated with either the outcome nor the environmental factor. When there is no G-E association or main effect of G, removing the corresponding terms in the log-linear model reduces the variance of the interaction estimate. Since there is uncertainty, the method averages over the sub-models with and without the G-E or main effect G terms. A final estimate of effect is the average over the models weighted by the posterior probability of each model. In a single-marker analysis this approach can be more powerful than the case-only approach when the GxE independence assumption holds. When the assumption is violated, the Type I error rate is more robust than the case-only analysis. In genome-wide simulations, this approach has better overall performance.