Mendelian randomization, lipids and coronary artery disease: trade-offs between study designs and assumptions

Background. Mendelian randomization (MR) studies have been described as naturally occurring randomized controlled trials (RCTs). However, MR often deviates from appropriate RCT design principles and relies heavily on two-sample approaches. We used data from the Million Veteran Program (MVP) to empirically evaluate the impact of study design choices and use of one- versus two-sample MR in a study of lipids and coronary artery disease. Methods. Our MR study included MVP participants of European descent with no history of coronary artery disease or contraindications to low-density lipoprotein cholesterol (LDL-C)-related therapies. We sequentially modified the eligibility criteria, study duration and follow-up to reflect common study design decisions for MR. In all designs, we used one- and two-sample approaches to estimate 10-year risks of coronary artery disease per 39 mg/dL increase in LDL-C or 15.6 mg/dL increase in high-density lipoprotein cholesterol (HDL-C). Results. For LDL-C, one-sample estimates varied across designs (odds ratios from 1.50 [95% CI: 1.34,1.68] to 2.23 [95% CI: 1.93,2.59]) and were most sensitive to the inclusion of prevalent outcome events in the analysis. Odds ratios obtained via two-sample MR were attenuated (1.13 [95% CI: 1.01,1.26] to 1.30 [95% CI: 1.15,1.46]). For HDL-C, we observed inverse or null relationships and estimates were qualitatively similar across all designs (odds ratios from 0.76 [95% CI: 0.68,0.86] to 0.93 [95% CI: 0.65,1.34]). Conclusions. MR estimates can, in practice, be impacted by decisions in study design due to trade-offs between different biases, and investigators should evaluate the sensitivity of their estimates to different design decisions.