A Comparison of Mendelian Randomization Methods in The Presence of Pleiotropy

Abstract

Through considering genetic variants as instrumental variables, Mendelian randomization has become a widely used method for inferring causal relationships between exposures and outcomes. However, the robustness of the Mendelian randomization method relies on stringent assumptions, particularly the assumption on absence of horizontal pleiotropy. Although time-invariant genetic factors serve as excellent candidates for estimating causal effects between gene expression and disease outcomes, the pleiotropy is pervasive under such situations. In practice, particularly among highly polygenic phenotypes, pleiotropy is common and present in thousands of genetic variants across many loci. While many recent Mendelian randomization methods effectively mitigate uncorrelated pleiotropy, few methods directly address correlated pleiotropy. In this study, we conduct a comprehensive comparison of six representative Mendelian randomization methods: IVW-random, MR-Egger regression, cML, BWMR. MR-Corr2, and MR-Horse. Among three different scenarios of horizontal pleiotropy, we found MR-Corr2 and cML that explicitly model correlated pleiotropy and use LD-aware structure to produce the most reliable causal estimates as the number of valid IVs is limited. IVW-random and BWMR consistently perform better especially with only uncorrelated horizontal pleiotropy.