How to use Mendelian randomization to anticipate the results of randomized trials

Abstract

This editorial refers to ‘Differential effects of PCSK9 variants on risk of coronary disease and ischaemic stroke’†, by J.C. Hopewell et al., on page 354. A Mendelian randomization study design attempts to introduce a randomization scheme into an observational study to assess whether an observed association between an exposure and an outcome is likely to be causal. These studies use genetic variants that are associated with the exposure of interest as a proxy for higher or lower levels of the exposure.1 Because allocation to genetic variants is approximately random and occurs at conception, this study design should be less susceptible to confounding, reverse causation, and other forms of bias that can limit the validity of observational studies. Perhaps the most intuitive way to explain the concept of Mendelian randomization is by way of analogy with a randomized trial. Indeed, Mendelian randomization studies have been called nature’s randomized trials.2 , 3 For example, numerous genetic variants are associated with lower LDL cholesterol (LDL-C). Each of these variants is inherited approximately randomly at the time of conception in the process sometimes referred to as Mendelian randomization. Therefore, inheriting an LDL-C-lowering allele is analogous to being randomly allocated to an LDL-C-lowering therapy, while inheriting the other allele is analogous to being randomly allocated to usual care. If allocation is indeed random and if the variant under study is associated only with LDL-C, but not with other pleiotropic effects, then the only difference between groups of persons with and without one or more LDL-C-lowering alleles will be their plasma LDL-C level. Therefore, measuring the association between these variants and the risk of cardiovascular disease should provide an unconfounded estimate of the causal effect of lifelong exposure to lower LDL-C on the risk of cardiovascular disease in a manner analogous to a long-term randomized trial. The analogy between a Mendelian randomization study and a randomized trial has created a great deal of interest in using this study design to anticipate the expected results of randomized trials evaluating therapies that target causal exposures. However, the analogy between a Mendelian randomization study and a randomized trial has limitations. Mendelian randomization studies evaluate the effect of genetically determined lifelong changes in an exposure on an outcome, whereas randomized trials evaluate the effect of therapeutically induced short-term changes in the same exposure. The challenge when attempting to use Mendelian randomization to anticipate the results of randomized trials is therefore to translate the causal effect of lifelong changes in an exposure on an outcome to the effect that can be expected to occur in response to short-term therapeutically induced changes in that exposure on the outcome of interest. Unfortunately, the results of a Mendelian randomization study cannot be used to estimate directly the expected effect of a therapy in a short-term trial. This is because the effect of most causal exposures appears to have cumulative effects on the associated outcome over time. For example, lifelong exposure to lower LDL-C is associated with an ∼3-fold greater reduction in the risk of cardiovascular events per unit change in LDL-C as compared with short-term reductions in LDL-C during treatment with a statin.4 Therefore, using the results of a Mendelian randomization study to anticipate directly the results of a randomized trial will in general lead to an overestimate of the expected effect size. Despite this important limitation, however, Mendelian randomization can be used to anticipate the results of randomized trials accurately.5 , 6 This can be accomplished by comparing the effect of a variant of interest with the effect of a reference variant that mimics a therapy that has been evaluated in randomized trials (Figure 1). Algorithm for using Mendelian randomization (MR) studies to the anticipate results of randomized trials. In cardiovascular medicine, the HMGCR gene (which encodes HMG-CoA reductase, the target of statins) is an excellent reference standard. Variants in HMGCR closely mimic the effect of statins which have been evaluated in numerous randomized trials.7 Comparing the effect of a genetic variant with the effect of HMGCR variants can thus be used to anticipate the results of a randomized trial indirectly.5 , 6 For example, common variants in the NPC1L1 gene that encodes Niemann–Pick C1-Like 1, the target of ezetimibe, have nearly identical effects on the risk of cardiovascular disease per unit change in LDL-C as compared with variants in the HMGCR gene.5 This finding implies that treatment with ezetimibe should reduce the risk of cardiovascular events by approximately the same amount as treatment with a statin per unit reduction in LDL-C. This is precisely what the IMPROVE-IT trial reported.8 Similarly, variants in the PCSK9 and HMGCR