Methodologies for Robust and Efficient Genetic Association Analysis

Abstract

Genetic association studies (GAS) aim to identify disease susceptibility genes by investigating the statistical association of a disease with genetic markers in the human genome. Statistical approaches to this end can be divided into two classes: one is more robust to bias due to unmeasured factors/inadequate modeling assumptions but less powerful for detecting the association, and the other is more powerful but less robust. Therefore, as in statistical applications in other scientific areas, we encounter the robustness-efficiency trade-off in GAS. In this talk, I will present two methodologies we recently developed for a better balance between robustness and efficiency in GAS. The first methodology focuses on two popular study designs in GAS: the (population-based) case-control and the (family-based) case-parents designs. It is known that the case-control design is very convenient to implement and hence allows for a large sample size to gain power, but less robust to the "population stratification" (PS) bias, a confounding bias due to unmeasured population structures; the family design is robust to the PS bias but subject to power loss due to the potential difficulty in collecting family controls (parents). We propose a two-stage design/analysis that can gain power from the population-based case-control study performed at stage one, while still maintaining the full robust property of the case-parents design, which is performed at stage two of the proposed design. The second methodology focuses on haplotype-based case-control studies. Haplotypes, the specific combinations of alleles on the same chromosomes, serve as informative genetic units for GAS but can only be indirectly inferred through locus-specific genotypes. Some modeling assumptions such as Hardy-Weinberg equilibrium and gene-environment independence are thus required for the haplotype-based association analysis to be implemented with the incompletely observed data. The existing “prospective likelihood” approach is usually unbiased even if the modeling assumptions are violated but less efficient; the “retrospective likelihood” is more efficient but may be seriously biased when the assumptions fail. We propose a penalized likelihood approach to achieve better bias-efficiency trade-off than the two types of methods, with the penalization given by the LASSO or Ridge penalty. Real data and simulation studies illustrate both the robustness and efficiency of the proposed method.