Mendelian Randomization: Using Inherited Genes as a Natural Experiment

The short answer

Mendelian randomization is a way to test whether a biomarker actually causes a disease, rather than just traveling alongside it. It works because the gene variants you inherit are dealt at conception, roughly at random with respect to the diet, income, and habits you later acquire, so a variant that nudges a biomarker up or down mimics a small lifelong experiment. If people who carry the biomarker-raising variants also get more disease, that points toward cause. If they do not, a correlation seen in ordinary studies was probably driven by something else.

Why correlation keeps fooling us

Most of what we know about risk factors comes from observational studies. We measure a biomarker, say a blood lipid or an inflammatory marker, follow people for years, and see who develops disease. The problem is that people with a high biomarker often differ in dozens of other ways. They may exercise less, eat differently, smoke more, or carry an unmeasured illness that raises both the biomarker and the disease. Statisticians call this confounding, and there is also reverse causation, where early disease quietly changes the biomarker before diagnosis. Both can manufacture a convincing association that dissolves the moment you intervene.

Randomized trials solve this by assigning the exposure with a coin flip, which balances the confounders on average. But you cannot randomize a lifetime of high cholesterol, and trials are expensive, short, and sometimes impossible for ethical reasons. That gap is where the genetic idea earns its place.

Genes as a lottery ticket

Here is the mechanism I find most elegant. When egg and sperm form, chromosomes are shuffled and the two copies of each gene are separated so that a child receives one at random from each parent. A variant that happens to raise, say, a person's fasting glucose is therefore distributed across the population largely independently of whether that person will grow up wealthy or poor, active or sedentary. The genotype is fixed from birth and cannot be changed by later behavior, which also rules out reverse causation, since disease in your fifties cannot rewrite the DNA you were born with.

So a variant becomes a kind of natural randomizer. Divide people by whether they carry the biomarker-raising version, compare their disease rates, and you have something that behaves like a small, lifelong, accidental trial. This is the core of Mendelian randomization, named for Gregor Mendel's laws of inheritance. In formal terms the genetic variant serves as an instrumental variable for the biomarker.

Genome-wide association studies, the kind of work that shaped my doctoral training on the genetics of type 2 diabetes at the Lund University Diabetes Centre, are what make this practical. They have mapped thousands of common variants to traits like glucose, insulin secretion, and body fat distribution. Each variant usually moves a biomarker by only a sliver, so researchers often combine many into a single genetic score to get a stronger, cleaner handle on the exposure.

The three assumptions that hold it together

The method is only as good as three conditions, and every serious analysis has to defend them.

Relevance

The genetic instrument must be genuinely and reliably associated with the biomarker. Weak instruments, variants that barely move the exposure, produce noisy and biased answers. This is the easiest assumption to check, because the association with the biomarker is measurable.

Independence

The variant must not be linked to confounders of the biomarker-disease relationship. Inheritance gives us a strong head start here, but it is not automatic. Population structure, where ancestry correlates with both certain variants and certain environments, can smuggle confounding back in. Careful matching of ancestry and sensitivity checks are how analysts guard against it.

Exclusion restriction

The variant must affect the disease only through the biomarker under study, and by no other route. This is the hardest to satisfy and the one that breaks most often. Many genes are pleiotropic, meaning they influence several traits at once. If a variant raises your biomarker and also, through a separate pathway, changes disease risk, the tidy causal story collapses. You cannot fully prove this assumption, so the field has built statistical tools to probe it, comparing many instruments to see whether they tell a consistent story and flagging the outliers that suggest a hidden pathway.

What it can and cannot tell you

Used well, Mendelian randomization has repeatedly clarified debates that observational data left muddy. It has strengthened the case that some lipid fractions causally drive cardiovascular disease while casting doubt on others that merely tracked with risk. In metabolic research it helps separate biomarkers that are drivers of disease from those that are bystanders or consequences, which is exactly the distinction that matters when deciding what to target with a drug.

The limits deserve equal respect. The method estimates the effect of a lifelong, small difference in a biomarker, which is not the same as the effect of a large, late intervention with a medicine. A gene that nudges glucose slightly from birth may not predict what happens when you lower glucose sharply at age sixty. The estimates can be imprecise when variants are weak, and they assume the biomarker's effect is roughly linear across its range, which is not always true. Canalization, the body's tendency to buffer genetic perturbations during development, can also blunt the signal. And a null result does not always mean no effect; it can mean the instrument was too weak to detect one.

I read these studies the way I read any single line of evidence, as one witness rather than a verdict. When a genetic analysis, a randomized trial, and a mechanistic understanding of physiology all point the same way, the conclusion becomes hard to dismiss. When they disagree, the disagreement is usually telling you something worth chasing.

This article is educational and not medical advice; if you have questions about your own risk or biomarkers, please talk with your own clinician.

The appeal of the approach, for me, is honesty about uncertainty. It does not claim to replace experiments. It borrows a fragment of randomness that nature already performed at conception, reads it carefully, and states plainly what that fragment can and cannot support.