Gene-gene interaction analysis for the survival phenotype based on the standardized residuals from parametric regression models

Most of statistical methods in genome-wide association studies (GWAS) have been developed to identify SNP-SNP interactions using a binary phenotype for case-control studies. However, there is an interest in identifying SNPs and combinations of SNPs that relate to the survival phenotype in many cancer studies. Recently, Gui et al. (2011) proposed a novel method, called Surv-MDR, for detecting gene-gene interactions associated with survival time by modifying the multifactor dimensionality reduction (MDR) method based on the log-rank test. However, the Surv-MDR method needs more-intensive computations and does not allow for a covariate adjustment. Lee et al. (2011) proposed an alternative approach, called Cox-MDR, for detecting gene-gene interactions based on a Cox model by extending generalized multifactor dimensionality reduction (GMDR) to the survival time phenotype. The main idea of Cox-MDR is to use a martingale residual as a score to classify multi-level genotypes as high and low risk groups. Cox-MDR also allows the effects of discrete and quantitative covariates to be easily adjusted in the Cox model and requires much less computation than Surv-MDR. In this paper, we propose to extend the idea of a Cox-MDR to the parametric regression model, called PRM-MDR for identifying gene-gene interactions associated with the survival phenotype. The main idea of PRM-MDR is to use a standardized residual from a parametric regression model as a score to classify multi-level genotypes as high and low risk groups while keeping the rest of the MDR procedure unchanged. We performed a simulation study to compare PRM-MDR with Cox-MDR as well as the Surv-MDR method in detecting two-way interactions. The simulation setting is constructed as similar as that of Gui et al. (2011), so that three methods are comparable. In addition, we compared the power of PRM-MDR with adjusting the covariates versus without adjusting the covariates in two different scenarios. In the first scenario, we - djusted the covariates that potentially have the significant effect on the survival phenotype such as age while we adjusted the covariates that might not have any effect on the survival phenotype such as sex in the second scenario.