Bayesian Mendelian Randomization methods for Index Trait Bias correction in subsequent trait genome-wide association studies

Index-Trait bias (also called Index-Event Bias) can occur in genetic studies due to conditioning on incident trait, which can bias genetic associations with subsequent traits. We propose the use of two Bayesian Mendelian Randomization (MR) methods (Bayesian weighted MR [BWMR] and [MR-HORSE]) to correct index-trait bias in genome-wide association studies (GWAS) of subsequent traits. We compare these Bayesian MR methods to previously proposed methods for index-trait bias through a simulation study. We observe that BWMR has similar Type I error compared to using an inverse variance weighted MR (MR-IVW), weighted median MR (MR-WMED), and Dudbrige (MR-Egger), but has an inflated Type I error compared to SlopeHunter. MR-HORSE and Slopehunter have similar Type I errors for smaller correlations between incident and subsequent traits, however, MR-HORSE and Slopehunter have better controlled Type I error for a large negative correlation and large positive correlation, respectively. All methods have comparable power across correlations between incident and subsequent traits. We applied the methods to a GWAS of subsequent Acute Ischemic Stroke (AIS) or 3pt Major Adverse Cardiovascular Event (MACE) after an incident AIS event in the Million Veterans Program (MVP) and for Fasting Insulin (FI) adjusted for BMI, and we observed slight differences in the results between the methods. We observed that a single index trait bias correction method is not optimal across all scenarios, therefore applying multiple methods and checking for consistency between the estimates could provide an approach to determine the presence of and correction for index trait bias.