The role of estimation in Mendelian randomization: should Mendelian randomization investigations provide estimates?

Abstract

Mendelian randomization (MR) makes causal claims by treating genetic variation in an analogous way to randomization in a clinical trial. MR investigations can be viewed as analogous to a randomized encouragement design, in that genetic variants do not determine the precise level of an exposure, but increase liability to it. As such, an MR estimate typically does not represent an achievable or well-defined causal effect in terms of the exposure, as it reflects the impact of a life-long shift in the trajectory of the exposure, which likely differs between individuals. We advocate for MR investigations to be performed to assess evidence for a causal hypothesis, rather than to estimate a well-defined causal quantity. MR estimates are useful to combine evidence across genetic variants, to assess the validity of variants as instruments, to provide confidence intervals, and to compare estimates across outcomes. However, numerical estimates from MR should not be over-interpreted. The value of an MR investigation is not to quantify the magnitude of effect for a well-defined intervention in the exposure. Instead, it provides a distinct source of evidence to increase or decrease confidence in a causal hypothesis, which can be triangulated with evidence from other sources.