Transformation and Multivariate Methods for Improving Power in Genome-Wide Association Studies

Abstract

This dissertation proposes methods for improving power while maintaining valid inference in genome-wide association studies (GWAS) of common, complex, quantitative traits. Chapter 1 proposes association tests that incorporate the rank-based inverse normal transformation (INT). In the direct approach, the phenotype itself is transformed, whereas in the indirect approach, phenotypic residuals are transformed. Since neither test is uniformly most powerful, these approaches are combined into an adaptive INT-based omnibus test (O-INT). Chapter 2 proposes Surrogate Phenotype Regression Analysis (Spray) for leveraging information from surrogate outcome to improve inference on a partially missing target outcome. The target and surrogate outcomes are jointly modeled within a bivariate normal regression framework. Estimation in the presence of bilateral missingness is performed using the expectation conditional maximization algorithm, and a Wald test for the total effect of genotype on the target outcome is derived. Chapter 3 proposes Synthetic Surrogate Analysis (SSA) for extending Spray to the setting of multiple candidate surrogates. Rather than directly modeling the target outcome together with multiple surrogates, the candidate surrogates are instead combined into a univariate summary measure, termed the synthetic surrogate, which is jointly analyzed with the target outcome using the bivariate framework. Combining multiple surrogates to predict the target outcome improves power while maintaining computational tractability. All proposed estimation and inference procedures have been implemented in publicly available R packages.