The consequences of adjustment, correction and selection in genome-wide association studies used for two-sample Mendelian randomization

Abstract

Introduction: Genome-wide association studies (GWASs) often adjust for covariates, correct for medication use, or select on medication users. If these summary statistics are used in two-sample Mendelian randomization analyses, estimates may be biased. We used simulations to investigate how GWAS adjustment, correction and selection affects these estimates and performed an analysis in UK Biobank to provide an empirical example. Methods: We simulated six GWASs: no adjustment for a covariate, correction for medication use, or selection on medication users; adjustment only; selection only; correction only; both adjustment and selection; and both adjustment and correction. We then ran two-sample Mendelian randomization analyses using these GWASs to evaluate bias. We also performed equivalent GWASs using empirical data from 306,560 participants in UK Biobank with systolic blood pressure as the exposure and body mass index as the covariate and ran two-sample Mendelian randomization with coronary heart disease as the outcome. Results: The simulation showed that estimates from GWASs with selection can produce biased two-sample Mendelian randomization estimates. Yet, we observed relatively little difference between empirical estimates of the effect of systolic blood pressure on coronary artery disease across the six scenarios. Conclusions: Given the potential for bias from using GWASs with selection on Mendelian randomization estimates demonstrated in our simulation, careful consideration before using this approach is warranted. However, based on our empirical results, using adjusted, corrected or selected GWASs is unlikely to make a large difference to two-sample Mendelian randomization estimates in practice.