CMStatistics 2022: Start Registration
View Submission - CMStatistics
B1351
Title: A robust whole-genome Mendelian randomization approach for improved estimation and inference of causal effects Authors:  Haoyu Zhang - National Cancer Institute (United States) [presenting]
Zhonghua Liu - Columbia University (United States)
Xihong Lin - Havard University (United States)
Abstract: Mendelian randomization (MR) uses genetic variants as instrumental variables (IV) to assess causal associations between modifiable risk factors and diseases. The method only requires summary-level statistics from genome-wide association studies (GWAS), reducing logistical load and making it widely used. MR makes several strong assumptions, which can be violated in practice and lead to biased estimates. For example, MR estimates can be biased due to weak IVs, pleiotropic effects, or sample overlaps. To address these issues, we develop a whole-genome MR method (WMR) that accounts for linkage disequilibrium (LD) across genetic variants. We assume pleiotropic effects follow distributions with mean 0 and a fixed variance, and estimate the variance of pleiotropiceffects using LD-score regression. Our simulation analyses mimic real LD patterns using 1000 Genomes Project European haplotype data. We simulate different proportions of causal SNPs, levels of pleiotropic effects, and sample overlap proportions. We find that WMR is robust to weak IVs and different levels of pleiotropic effects. Meanwhile, WMR provides a smaller empirical standard error than alternative approaches since WMR uses more correlated IVs. We apply the proposed methods along with other approaches to many traits, including BMI, lipids-related traits, cardiovascular disease, breast cancer, etc. The standard error of WMR estimates is smaller than alternative approaches, consistent with simulations.