Abstract
Two-sample Mendelian randomization (MR) is increasingly used to strengthen causal inference using observational data. This method allows the use of freely accessible summary association results from genome-wide association studies (GWAS) for a range of traits. Some GWAS studies adjust for heritable covariables in an attempt to estimate direct effects of genetic variants on the trait of interest. One, both or neither of the genetic instrumental variables (IVs)-exposure association or genetic IVs-outcome association may have been adjusted for heritable covariables (referred to as GWAS covariables). However, it is unclear how this may affect two-sample MR analysis. We evaluated this in a simulation study comprising different scenarios that could motivate covariable adjustment in a GWAS. Our results indicate that the impact of covariable adjustment is highly dependent on the underlying causal structure. In the absence of residual confounding between exposure and covariable, between exposure and outcome, and between covariable and outcome, using covariable-adjusted summary associations for two-sample MR may eliminate bias due to horizontal pleiotropy. However, the presence of residual confounding (especially between the covariable and the outcome) leads to bias upon covariable adjustment, even in the absence of horizontal pleiotropy. Bias was lower when the true causal effect of the exposure on the outcome was zero compared to a non-zero causal effect. In an analysis using real data from the Genetic Investigation of ANthropometric Traits (GIANT) consortium and UK Biobank, the direction of the causal effect estimate of waist circumference on blood pressure changed upon adjustment of waist circumference for body mass index. Our findings indicate that using covariable-adjusted summary associations in MR should generally be avoided. When that is not possible, careful consideration of the causal relationships underlying the data (including potentially unmeasured confounders) is required to direct sensitivity analyses and interpret results with appropriate caution.