Bayesian Variable Selection in High Dimensional Survival Time Cancer Genomic Datasets using Nonlocal Priors
share
Variable selection in high dimensional cancer genomic studies has become very popular in the past decade, due to the interest in discovering significant genes pertinent to a specific cancer type. Censored survival data is the main data structure in such studies and performing variable selection for such data type requires certain methodology. With recent developments in computational power, Bayesian methods have become more attractive in the context of variable selection. In this article we introduce a new Bayesian variable selection approach that exploits a mixture of a point mass at zero and an inverse moment prior which is a non-local prior density on the Cox proportional hazard model coefficients. Our method utilizes parallel computing structure and takes advantage of a stochastic search based method to explore the model space and to circumvent the computationally expensive MCMC procedure. It then reports highest posterior probability model, median probability model and posterior inclusion probability for each covariate in the design matrix. Bayesian model averaging is also exploited for predictive power measurements. The proposed algorithm provides improved performance in identifying true models by reducing estimation and prediction error in simulation studies as well as real genomic datasets. This algorithm is implemented in an R package named BVSNLP.
READ FULL TEXT