Examining the evidence for Mendelian randomization homogeneity assumption violation using instrument association with exposure variance

Background Estimation of the average causal effect using instrumental variable (IV) analyses requires homogeneity of instrument-exposure and/or exposure-outcome relationships. Previous research explored the validity of homogeneity assumptions by testing IV-exposure interaction effects using a set of effect modifiers. However, this approach requires that modifiers are known and measured but evidence for interaction may also be observed through IV association with exposure variance without knowledge of the modifier.

Methods We explored the utility of testing for IV-exposure variance effects as evidence against homogeneity through simulation. We also evaluated the approach of removing IVs from Mendelian randomization (MR) analyses that show strong association with exposure variance (hence are likely to have heterogeneous effects). Our methodology was applied to evaluate homogeneity assumptions of LDL, urate and glucose on cardiovascular disease, gout, and type 2 diabetes, respectively.

Results Under simulation, interaction of IV-exposure and exposure-outcome effects by a single modifier led to bias of the estimated average causal effect (ACE) which could be partially assessed by testing for IV-exposure variance effects. Bias of the ACE attenuated after removing instruments with strong exposure variance effects. In applied analyses, we found no strong evidence of bias from the ACE.

Conclusions We find no strong evidence against estimating the ACE for LDL, urate and glucose on cardiovascular disease, gout, and type 2 diabetes. These approaches could be used in future MR analyses to gain improved understanding of the causal estimand.

Key messages

• Homogeneity of the instrument-exposure and/or exposure-outcome effect is necessary to estimate the average causal effect which is important for developing health interventions

• Partial evidence against the homogeneity assumption can be obtained from testing for the instrument-exposure variance effect which may suggest the presence of effect modification

• This evidence can be used in two ways: i) as a falsification approach to determine if the homogeneity assumption may be violated. ii) to remove genetic instruments from Mendelian randomization analyses providing an estimate that is closer to the average causal effect

• After removing instruments with exposure variance effects, the Mendelian randomization effect of LDL, urate and glucose on coronary heart disease, gout, and type 2 diabetes, respectively showed little difference suggesting no strong evidence against the average causal effect