Title: "Empirical Bayes Method for Variable Selection with Applications to Genome-Wide Association Studies"
Speaker: Min Zhang, Department of Statistics, Purdue University
Place: LILLY G126; September 28, 2010, Tuesday, 4:30pm

Abstract

Available high-throughput technologies make it necessary to select important predictors out of a large number of candidates while exploiting their complicated relationships. Bayesian variable selection methods combined with Markov chain Monte Carlo algorithms have been explored for such tasks due to easy implementation. However, Bayesian variable selection methods usually choose convenient values of hyperparameters, which may comprise the data analysis. They are also challenged by exponentially growing number of variables, even though current computers are becoming more powerful. Here we propose an iterated conditional modes/medians (ICM/M) algorithm which will be employed to implement an empirical Bayes variable selection. First, iterated conditional modes are employed to optimize values of the hyperparameters so as to implement the empirical Bayes method. Second, iterated conditional medians are used to estimate the model coefficients and therefore implement the variable selection function. Our simulation study suggests fast computation and superior performance of the proposed method. Even in the simple case of variable selection, it outperforms LASSO and its variants. The developed method has been applied to data from genome-wide association studies.

This is a joint work with Vitara Pungpapong (Ph.D. candidate, Statistics)

Associated Reading:

Li and Li. 2008. Network-constrained regularization and variable selection for analysis of genomic data. Bioinformatics. 24(9):1175-1182.



Click here for a full schedule of BIOINFORMATICS SEMINARS, past and present.