Comparison between the SSVS and the LASSO for Genome Wide Association Studies

Abstract

Due to the multiple loci control nature of complex phenotypes, there is great interest to test markers simultaneously instead of one by one. In this paper, we compare three model selection methods for genome wide association studies using simulations: the Stochastic Search Variable Selection (SSVS), the Least Absolute Shrinkage and Selection Operator (LASSO) and the Elastic Net. We also apply the three methods to identify genetic variants that are associated with daunorubicin-induced cytotoxicity. The simulation studies were performed by using the genotype data of 60 unrelated individuals from the CEU population in the Hapmap project. For the cytotoxicity data, we used 3,967,790 markers across the whole genome for 56 unrelated individuals from the CEU population. Using Sure Independence Screening as the pre-screening procedure, the SSVS gives a small model while the LASSO gives an intermediate sized model and the Elastic Net provides a large model. The three models share many common markers although the model sizes are different. The model sizes are subject to various cutoffs and parameters. The SSVS outperforms the LASSO and the Elastic Net in simulation studies. We also demonstrate the ability of the SSVS, the LASSO, and the Elastic Net to handle the situation when the number of markers is larger than the number of samples.