Model selection methods and their application in genome-wide association studies

Abstract

As the data where the number of features greatly exceeds the sample size frequently appear in many areas, high dimensional model selection problem has become common and imminent. We propose the generalized tournament approach cum EBIC for generalized linear models with both main effects and interactions. In screening step, main effects and interactions are screened in consecutive stages until the dimension of features is reduced to a desirable level. In final selection, the modified SCAD method combined with EBIC is developed to select causal features. We show that the modified SCAD method guarantees finite parameter estimates in case of separation. To cater the demand for statistical methods to identify causal SNPs with interaction structure, we apply the generalized tournament procedure in genome-wide association studies. In simulated data, we demonstrate that the generalized tournament approach cum EBIC enjoys higher positive selection rate and lower false discovery rate than interaction-based multiple testing.