Joint Research Conference
June 24-26, 2014
A weighted pseudolikelihood approach for eliminating ascertainment bias and improving quality when utilizing a case-control GWAS to assess genetic association with smoking habits)) We present new statistical analyses of data arising from a case-control study of lung cancer, in which we are interested in identifying Single Nucleotide Polymorphisms (SNPs) that are associated with multiple secondary phenotypes of smoking behavior in the population. To eliminate potential ascertainment bias resulting from the case-control sampling, we incorporate inverse probability weighting into our testing and variable selection procedures, which are also robust to misspecification of the true correlation structure across outcomes. We demonstrate the effectiveness of both procedures through theoretical and empirical analysis, and show that our proposed testing method is more powerful than commonly used alternatives while maintaining the correct type I error rate, and that improperly accounting for case-control sampling can lead to biased results. Furthermore, for variable selection, we show that our proposed weighted Bayesian Information-like Criteria (BIC) approach can out-perform the standard unweighted BIC approach. Because our testing and variable selection approaches borrow strength from the correlation across outcomes, as well as the correlation across SNPs, we enjoy improved quality and greater reproducibility than the more traditional individual-outcome and individual-SNP based approaches. The proposed methods identify several SNPS of biological
interest that are associated with smoking behavior.
This is joint work with Tamar Sofer, Department of Biostatistics, Harvard School of Public Health, David C. Christiani, Departments of Environmental Health and Epidemiology, Harvard School of Public Health Xihong Lin, Department of Biostatistics, Harvard School of Public Health.