Loss of Power in Two‐Stage Residual‐Outcome Regression Analysis in Genetic Association Studies

Abstract

In the area of genetic epidemiology, investigating the association between genetic risk factors and phenotypic outcomes is an important challenge. Often in such studies, the analysis approach needs to control for potentially confounding factors, such as clinical, demographic, or environmental variables. Multiple linear regression (MLR) is often used for such analysis when the outcome of interest is a quantitative trait. Under this model, the effects of genetic factors are estimated effectively and covariates are adjusted correctly. In large-scale genome-wide studies with a relatively large number of genetic and environmental factors present, some alternative analytic strategies are being used to address the issues of covariate adjustment. One such approach is a two-stage residual-outcome regression analysis (2SR). At stage one, the outcome is regressed on all the covariates. At stage two, the residual-adjusted outcome is then regressed on genetic factors. 2SR has been employed in genetic association studies [Choy et al., 2008], as well as genetic linkage studies [Slager and Iturria, 2003]. Zeegers and colleagues have demonstrated that the 2SR has equal power with other covariate-adjusted methods in genetic linkage studies [Zeegers et al., 2004], but the use of such a two-stage approach in association analyses is less well understood. While there may be some potential advantages with the 2SR approach in both ease of use and computational and data management efficiency, some issues, such as the bias of estimates, power, and type I error, need to be thoroughly investigated.