Bayesian Modeling for Environmental Association and Gene-Environment Interaction Under Complex Epidemiologic Study Designs.

Abstract

There has been substantial interest in characterizing the joint effects of genetic and environmental factors on disease risk. The first two projects in this dissertation focus on developing novel statistical methods for characterizing effects of environmental exposures on health outcomes under complex study designs. The other two projects evolve around the theme of meta-analysis of gene-environment interactions (GEI) and illustrate how environmental heterogeneity and varying degrees of gene-environment association across studies could influence the characteristics of several meta-analysis approaches. New data-adaptive shrinkage approaches for meta-analysis of interaction effects are proposed.

In project 1, a point source distance-odds model for characterizing exposure effects on disease sub-types under a matched case-control design is considered, where Bayesian methods appear to have advantages in terms of estimation stability, precision and interpretation over frequentist alternatives. Bayesian analysis of data collected under case-crossover designs is considered in project 2. Bayesian equivalence between case-crossover and time-series analysis is first studied, and a full likelihood-based approach that makes less restrictive assumptions on the disease risk model is subsequently introduced. The proposed methods are applied to an asthma study, where risk of asthma events is found to be associated with distance to major roads (project 1) and ambient levels of fine particulate matters (project 2).

In project 3, the problem of meta-analysis of GEI is considered. An adaptively-weighted-estimator that combines meta-analysis and meta-regression estimates of the interaction parameter is proposed. This approach has reduced mean squared error across a spectrum of scenarios. In project 4, the work of project 3 is extended to case-control studies. This project focuses on exploiting gene-environment independence assumption in a study-specific way while conducting meta-analysis of GEI. An empirical Bayes inferential framework that can incorporate uncertainty around the independence assumption by borrowing strength across studies is developed. The proposed class of shrinkage estimators outperforms simpler alternatives that do not incorporate study-specific heterogeneity around the independence assumption. The methods are illustrated by using data investigating type 2 diabetes. Meta-analysis of interactions between SNP on the FTO gene and body mass index reveal some suggestive evidence of interaction in this example.