Comparison of Variable Selection Methods for High-Dimensional Data in Applications to Genetic Association Studies with Gene-Environment Interactions

• Activity Number: 123
• Type: Contributed
• Date/Time: Monday, August 4, 2014 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Jaejoon Song*+ and Michael Swartz
• Companies: MD Anderson Cancer Center and University of Texas Health Science Center at Houston
• Keywords: Bayesian variable selection ; Boosting ; LASSO ; Gene-environment interaction

Abstract

In genetic association studies of complex diseases, joint consideration of genetic and environmental factors may be desirable, to identify networks of interacting factors that contribute to the disease risk. Several variable selection methods have been proposed for high dimensional variable selection to simultaneously consider all genetic markers, environmental factors and interaction effects, while achieving computational efficiency. Bayesian approaches include the hierarchical approach by Yi et al. (2011) and stochastic search variable selection via spike and slab priors. Non-Bayesian approaches for high dimensional variable selection include component-wise gradient boosting and the least absolute shrinkage and selection operator. In this study, we evaluate these Bayesian and non-Bayesian methods of variable selection in a simulated dataset. Simulated data includes a binary phenotype and interactions between genetic markers and an environmental factor. Bayesian and non-Bayesian approaches are evaluated in a simulation study. Model identification and computational efficiency of the methods are discussed.