Bayesian Variable Selection for High-Dimensional Genomic Data with Survival Response

• Activity Number: 623 - Bayesian Variable Selection
• Type: Contributed
• Date/Time: Thursday, August 3, 2017 : 8:30 AM to 10:20 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Amir Nikooienejad* and Valen Johnson
• Companies: Texas A&M University and Texas A&M University
• Keywords: Survival Data ; Bayesian Variable Selection ; Non-local prior ; High Dimensional Data ; Genomic and Genetics

Abstract

There has been a lot of interest in analyzing high dimensional genomics data specially in cancer studies in the past decade. The response variable for majority of such datasets are of survival type. In this article we introduce a new Bayesian approach that exploits a mixture of non-local prior densities and point masses on the Cox proportional hazard model coefficients, to tackle the problem of variable selection. The literature in this context lacks a powerful Bayesian tool and the proposed method can be considered as one. Our method provides improved performance in identifying true models and reducing estimation and prediction error in simulation studies. When applied to real genomic datasets, the proposed model produces predictions with high accuracy while using few explanatory variables. Moreover, we utilize a novel stochastic search based procedure to bypass MCMC and find the highest posterior probability model faster.