ABSTRACT

Mendelian randomization (MR) studies use genetic variants associated with modifiable exposures as instrumental variables to assess their possible causal relationship with outcomes in observational studies. In recent years, the number of published two-sample MR studies has increased rapidly, partially due to the increasing availability of large-scale genome-wide association studies (GWAS) summary data. These two-sample MR studies often employ the same sample to select relevant genetic variants and construct final causal estimates, leading to biased causal effect estimates due to the winner’s curse phenomenon. On the other hand, the validity of MR analyses critically depends on instrumental variable assumptions, which often be violated in applications due to widespread pleiotropic effects. Also, there are maybe unknown sample overlaps between the two GWAS summary datasets due to the current trends of collecting biobank datasets.
In the first part of this talk, we address the fundamental challenge of the winner’s curse by proposing a novel framework that systematically breaks the winner’s curse and provides an unbiased estimate of the genetic association effect after selection. Built upon the proposed framework, we introduce a novel rerandomized inverse variance weighted (RIVW) estimator that is provably consistent when selection and parameter estimation are conducted on the same sample. In the second part of this talk, we propose a new approach, referred to as Causal Analysis with Rerandomized Estimator (CARE), that corrects winner’s curse bias, weak pleiotropic effects, and unknown sample overlaps. We demonstrate the utilities of CARE through negative control analyses and its applications to identify possible causal risk factors for COVID-19 severity. This is joint work with Jingshen Wang and Xinwei Ma.