Identification of effect modifiers using a stratified Mendelian randomization algorithmic framework

Abstract

Mendelian randomization (MR) is a technique which uses genetic data to uncover causal relationships between variables. With the growing availability of large-scale biobank data, there is increasing interest in elucidating nuances in these relationships using MR. Stratified MR techniques such as doubly-ranked MR (DRMR) and residual stratification MR have been developed to identify nonlinearity in causal relationships. These methods calculate causal estimates within strata of the exposure adjusted to mitigate the impact of collider bias. However, their application to scenarios using a stratifying variable other than the exposure to identify the presence of effect modifiers has been limited. The reliable identification of effect modifiers is key to identifying subgroups of patients differentially affected by risk and protective factors. In this study, we present a stratified MR algorithm capable of identifying effect modifiers of causal relationships using adapted forms of DRMR and residual stratification MR. Through simulations, the algorithm was found to be robust at handling nonlinear relationships and forms of collider bias, accommodating both binary and continuous outcomes. Application of the stratified MR algorithm to 1,715 exposure-stratifying variable-outcome combinations identified two Bonferroni significant effect modifiers of causal relationships in the UK Biobank. The causal effect of body mass index on type 2 diabetes mellitus was attenuated with age, while the effect of LDL cholesterol on coronary artery disease was exacerbated with increased serum urate. Overall, we introduce a tool for detecting effect modifiers of causal relationships, and present two cases with clinical implications for personalized risk assessment of cardiometabolic diseases.