CMStatistics 2020: Start Registration
View Submission - CMStatistics
Title: Accounting for Winner's curse, weak instrument bias and pleiotropy in Mendelian randomization studies Authors:  Jack Bowden - University of Exeter (United Kingdom) [presenting]
Abstract: Mendelian randomization (MR) is the science of augmenting the analysis of observational data with genetic information to uncover the causal mechanisms of disease. In an MR analysis, single nucleotide polymorphisms (SNPs) are assumed to be an instrumental variable (IV) for a given exposure trait. In the MR field, research has focused on the development of methods to guard against incorrect inferences when using SNPs that exert direct (or pleiotropic) effects on the outcome, not through the exposure. So far, relatively little attention has been given to the issue of bias due to `Winner's curse', induced when the same data are used to select a small number of SNPs as IVs from a much larger candidate pool based on a p-value threshold. A simple way of removing Winner's curse is to use separate data sets for SNP discovery and MR model fitting, but this is statistically inefficient and is also more vulnerable to weak instrument bias. We describe a method for combining the SNP-discovery into the MR analysis to deliver efficient and unbiased estimates of a causal effect, whilst simultaneously accounting for weak instrument bias and pleiotropy. Our approach is based on the method of uniform minimum variance conditionally unbiased estimation (UMVCUE), which has been used as a technique for bias adjustment in genome-wide association studies and multi-arm multi-stage trials.