Statistics Seminar: Mengjie Bian – Investigating the Winner’s Curse in Mendelian Randomization
Feb 7, 2023
3:30PM to 5:00PM
Date/Time
Date(s) - 07/02/2023
3:30 pm - 5:00 pm
Speaker: Mengjie Bian, McMaster University
Title: Investigatingthe Winner’s Curse in Mendelian Randomization
Abstract:
MendelianRandomization (MR) uses genetic markers in an instrumental variable analysis toestimate the causal effect of an exposure on an outcome. The genetic markersare typically selected as the most significant associations with the exposurein a genome-wide associationstudy (GWAS). Selection might be based on a test statistic exceeding athreshold or on ranking of test statistics. In either case, the associationestimates for these markers are likely to be biased away from 0 because of thisselection. This phenomenon is known as the “winner’s curse”. This bias in the association estimates withexposure may result in the estimate of causal effect in MR being biased. We are interested in investigating, andcorrecting for, the impact of winner’s curse in MR. Several methods have been developed forcorrecting the winner’s curse bias in the GWAS setting. However, there has been little research aboutcorrecting the winner’s curse in MR studies. We performed simulation studies toassess the potential bias from winner’s curse and applied the existing methodsto correct the genetic marker-exposure association which are then used incausal estimation. We also adapted an existing bootstrap method to work in thecommon situation when only summary level data is available and show our adaptedmethod works comparably to the original method using individual level data. The results of the simulation study show thatwinner’s curse can bias causal estimates from MR. Some of the correction methods can reduce thebias but this improvement comes at the expense of increasing the variability. Confidence intervals based on correctedmethods are therefore wider and the power to detect a causal effect is notsubstantially affected by the corrections. An application of the methods to real datashows similar results.
Date/Time: Tuesday February 7, 2023, 3:30 – 5:00
Location: MDCL 1115