Methods for correcting the accuracy in Mendelian randomization

Abstract

Mendelian randomization (MR) uses genetic variants as instrumental variables (IVs) to investigate the causal relationship between exposure and outcome. It has become widely popular due to its versatile applications in epidemiological research. Its rising popularity is largely driven by the ease of accessing summary-level data from large consortia, making it a cost-effective choice for researchers. In this thesis, we focus on three issues in MR that result in potential bias in causal inference. We first address the “winner’s curse” in MR, which arises from selecting genetic markers based on their significance or ranking. To mitigate this bias, we adapt the bootstrap-based BR-squared method to function with summary-level data. Our findings reveal that the correction methods can effectively reduce bias, albeit with an increase in variability. We then develop a method that accounts for the correlation caused by sample overlap while addressing potential bias from weak instruments. This proposed method yields stable causal estimates, although the standard errors of causal estimates may not be precisely estimated. Lastly, we introduce a novel approach for identifying invalid instrumental variables showing signs of horizontal pleiotropy. We recommend using the bootstrap method to account for the data-driven process of IV selection. Our results indicate that the bootstrap intervals approach the nominal level of coverage rate when the proportion of invalid IVs is less than 50%.