Bias-reduced methods in Genetic Epidemiology

点击翻译按钮获取中文摘要

My area of research is biostatistical methods in genetic epidemiology. The proposed research is motivated by investigations of the genetic basis of complex human traits, which has proven to be more subtle than initially thought. Past studies of associations between common diseases and millions of common genetic variants  across the genome have uncovered many genes that play a part in disease onset and progression, but these findings are just the tip of the iceberg. Recent studies that aim to delve deeper into the complex genetic architecture of common diseases require robust statistical methods able to cope with rare genetic variants, find interactions between genes and environmental exposures, and measure the extent to which a gene acts directly to increase disease risk, or acts indirectly, through an intermediate disease state. My proposed research can be summarized as follows. Penalized likelihood methods for case-control data: To study rare variant associations we require regression methods for predictor variables that are predominantly zeros. The focus of my work for rare variants is on penalized likelihood methods for inference of genetic effects from case-control data. Robust inference of GxE from case-parent trios: In the case-parent trio design genetic variants are measured on affected children (cases) and their parents. Environmental exposures on the child may also be collected. The case-parent trio design gives correct inference of genetic effects, even in studies that pool data across multiple ethic groups. However, Shi et al. (2011) have shown that showed that inference of GxE from case-parent trio data can be biased when the genetic locus being analyzed (the test locus) is not causal, but is correlated with a causal locus and this correlation varies from one ethnic group to another. I am developing methods for inference of GxE from case-parent trio data that are robust to such population stratification bias. Robust methods for mediation analysis: Mediators can be thought of as intermediate disease states on the causal path between a gene and a disease. For example, the effect of a gene on coronary heart disease may be mediated by lipid levels. Mediation analysis decomposes the exposure effect into estimated direct and indirect components. However, such effect estimates are prone to bias in the presence of unmeasured confounding variables, such as population stratification. Building on my work on robust inference of GxE, I plan to develop methods for mediation analysis that are robust to population stratification bias.