Abstract:
|
Variance component (VC) based score tests are a popular and powerful class of methods for assessing interactions between groups of genomic markers and environmental exposures. However, many existing methods tend to give inflated type I errors due to challenges in estimating the null model which contains main effects for the genomic markers. Least squares based estimation fails especially when the number of genetic markers increases, and alternative ridge regression gives biased estimates due to the penalty function. To overcome these difficulties, we propose an improved VC based score test that estimates the main effects under the null hypothesis using biascorrected ridge regression. To construct the test statistic, we adapt the classical kernel association test under a linear kernel but use a novel empirical corrected projection matrices corresponding to the biascorrected ridge regression. We also give pvalues from asymptotic mixture chi squared distribution. Simulations and real data analysis demonstrate that proposed biascorrected interaction tests improve the type I error control compared with current methods, while still maintaining power.
|