Abstract

Available high-throughput biotechnologies make it necessary to select important candidates out of massive biomarkers while exploiting their complicated relationship structures. Bayesian variable selection methods can be carried out with the renowned Markov chain Monte Carlo (MCMC) for implementation. Though implementing MCMC is typically easy, it is notoriously slow to converge especially with exponentially growing numbers of biomarkers and complex hidden structures among them. In spite of choosing values for hyperparameters at convenience, empirical Bayes methods can be employed to estimate hyperparameters from data. Here we propose an iterated conditional modes/medians (ICM/M) algorithm to implement an empirical Bayes variable selection in regression models. Our simulation studies suggest fast computation and superior performance of the proposed method. The developed algorithm has also been applied to real genome-wide data.