A1738
Title: Robust sparse Bayesian semiparametric modeling for longitudinal genetic studies
Authors: Jie Ren - Indiana University School of Medicine (United States) [presenting]
Kun Fan - The University of Texas Health Science Center at Houston (United States)
Cen Wu - Kansas State University (United States)
Shuangge Ma - Yale University (United States)
Abstract: In high-dimensional longitudinal interaction studies, performing variable selection and statistical inference is challenging when the assumed interaction model is misspecified under heterogeneous disease phenotypes. Partially motivated by an analysis of the Childhood Asthma Management Program (CAMP) data with high-dimensional single nucleotide polymorphism (SNP) measurements, a novel sparse robust Bayesian mixed model is developed to capture nonlinear longitudinal interactions. The proposed mixed model is robust to outliers and heavy-tailed distributions in the response variable, as well as to the misspecification of nonlinear interaction effects, in terms of both parameter estimation and statistical inference. A Gibbs sampler incorporating spike-and-slab priors is developed to promote exact sparsity in the identification of appropriate forms of main and interaction effects. Simulation results reveal superior performance of the proposed fully robust Bayesian analysis in identification, estimation, and statistical inference compared with alternative methods. In particular, the proposed analysis enables exact statistical inference in misspecified interaction models for the first time by yielding Bayesian credible intervals with nominal coverage probabilities in finite samples. Application of the proposed and alternative methods to the CAMP study sheds novel insight into Asthma etiology.
