Mendelian randomization: letting genes do the randomizing

Mendelian randomization tests whether an association between a measurable trait, a hormone level, a cholesterol reading, and a disease is actually causal, by using a genetic variant linked to that trait as a stand-in for a randomly assigned treatment. Because which version of a gene a person inherits is set at conception, effectively at random with respect to the rest of their life, a genetic variant cannot be caused by later confounders like diet, smoking, or income the way an observed trait can. If the variant that raises a trait also lowers disease risk, that is evidence the trait itself matters; if it does not, the original association was probably driven by something else.

The best-known demonstration concerns HDL cholesterol, long associated in observational studies with lower heart attack risk. A 2012 study in The Lancet used two genetic instruments, a variant in the endothelial lipase gene and a score built from 14 HDL-raising variants, to test whether that association held up. In ordinary observational data, each one-standard-deviation rise in HDL cholesterol was associated with an odds ratio of 0.62 for myocardial infarction, a large apparent protective effect. But among carriers of the HDL-raising genetic variants, whose HDL cholesterol was higher for life, there was no such reduction: the genetic-score estimate for the same one-standard-deviation rise was an odds ratio of 0.93, statistically indistinguishable from no effect. LDL cholesterol, tested the same way as a built-in comparison, showed the opposite pattern: its genetic and observational estimates agreed closely.

Two traps follow. First, the method depends on the genetic variant affecting disease only through the trait being studied, not through some other pathway, an assumption researchers have to argue for case by case rather than simply assume. Second, a null mendelian randomization result does not prove a trait is irrelevant to disease everywhere; it shows that the particular biological route by which the variant raises the trait did not lower risk, which is not automatically the same finding as every possible drug or lifestyle change that raises the same trait.

So when a headline reports that a supplement, drug, or lifestyle change should work because it raises some biomarker linked to better health in observational studies, ask whether that link has been checked against a genetic instrument. A trait that tracks with better outcomes in cohort studies but shows no effect in a mendelian randomization analysis, as HDL cholesterol did, is a strong signal that raising it directly may not help. Pair a biomarker claim with a look at whether it has been tested this way, the kind of scrutiny a randomized trial would otherwise have to supply.