Instrumental Variables: Estimating a Cause When You Cannot Measure Every Confounder

An instrumental variable is a factor that nudges the exposure you care about, has no other route to the outcome, and shares no common cause with it. When such a factor exists, you can estimate the causal effect of the exposure even when confounders are unmeasured, because the instrument creates variation in exposure that is, in effect, random. The catch is that the method rests on assumptions that are partly untestable, so the credibility of the answer depends less on the arithmetic than on whether those assumptions hold. Genetics can supply strong instruments, which is why Mendelian randomization has become a workhorse of modern epidemiology.

The problem instruments are built to solve

Most non-randomized comparisons in medicine are haunted by confounding. If people who drink more coffee also smoke more and sleep less, then a raw association between coffee and heart disease blends the effect of coffee with everything that travels alongside it. Standard adjustment handles the confounders you can name and measure. It does nothing for the ones you cannot, and those are often the ones that matter.

A randomized trial dissolves this problem by assigning exposure at random, so that on average the exposed and unexposed groups differ only in the exposure. It is the gold standard precisely because it balances the unmeasured along with the measured. But trials are expensive, sometimes unethical, and often impossible for exposures like lifelong cholesterol level or body weight. The instrumental variable idea asks a pointed question. Is there something in nature or in the structure of a system that assigns exposure in a way that mimics randomization, at least partially?

What makes an instrument valid

An instrument earns its name by satisfying three conditions, and the second and third are where the real work lies.

Relevance

The instrument must actually move the exposure. This is the one assumption you can check directly, by measuring how strongly the instrument predicts the exposure. Weak instruments, ones that shift exposure only slightly, produce unstable and biased estimates, which is why analysts report the strength of the first stage rather than assuming it.

Independence

The instrument must share no common cause with the outcome. It should be unrelated to the confounders that plague the exposure-outcome relationship. If your candidate instrument is itself entangled with socioeconomic status or baseline health, it is not clean, and the estimate inherits the same confounding you were trying to escape.

The exclusion restriction

This is the demanding one. The instrument must affect the outcome only through the exposure, by no other path. If it reaches the outcome through some side channel, the method quietly attributes that side effect to the exposure and hands you a biased number. This restriction cannot be proven from the data alone. You argue for it from mechanism and subject knowledge, and a thoughtful analysis spends most of its honesty budget here.

When all three hold, the logic is elegant. The instrument induces variation in exposure that is unlinked to confounders, so comparing outcomes across levels of the instrument isolates the exposure effect. What you recover is a ratio: how much the outcome moves per unit of instrument, divided by how much the exposure moves per unit of instrument.

A clean example

Consider a study of whether a medication reduces a bad outcome, where sicker patients are more likely to receive the drug, so a naive comparison is confounded by severity. Suppose some clinics adopt the medication early and others late, and a patient clinic is essentially a matter of geography rather than prognosis. Distance to an early-adopting clinic can then act as an instrument. It shifts the probability of receiving the drug (relevance), plausibly bears no relation to a patient underlying severity (independence), and should touch the outcome only by changing whether the drug is prescribed, not through any direct biological path (exclusion). The difference in outcomes across geography, scaled by the difference in treatment, then estimates the drug effect. No one had to measure severity. That is the whole appeal.

This design carries a limitation worth naming. The effect you estimate applies only to the people whose exposure the instrument actually changed, those who took the drug because their clinic adopted it early and would not have otherwise. That is a local effect, not necessarily the average across everyone, and honest reporting says so.

Genes as instruments

Nature runs a version of this experiment at conception. When egg and sperm form, genetic variants are allocated in a way that is close to random with respect to the environment and behavior a person later encounters. A variant that raises lifelong LDL cholesterol, for instance, is not systematically paired with diet, income, or smoking. This is the foundation of Mendelian randomization, which uses a genetic variant as an instrument for the exposure it influences.

The mapping to the three assumptions is direct. Relevance means the variant reliably shifts the exposure, say a cholesterol level or a body weight. Independence draws on the near-random allocation of alleles across a population. Exclusion requires that the variant affect the outcome only through that exposure, which is where the biology has to be interrogated. The chief threat is pleiotropy, a single gene influencing several traits at once, opening a back path from variant to outcome that bypasses the exposure of record. Much of the methodological progress in this field has been about detecting and correcting for that back path.

My own doctoral work at the Lund University Diabetes Centre sat squarely in the genetics of type 2 diabetes, and that vantage point is what makes the promise and the fragility of genetic instruments feel concrete to me. A variant can be a beautifully clean natural experiment for one trait and a confounded mess for another, and only mechanism tells you which. Used carefully, genetic instruments have helped clarify whether biomarkers are causes or merely bystanders.

Reading an instrumental variable claim

When you meet one of these studies, ask three plain questions. Is the instrument strong enough to move the exposure? Is there a credible argument, grounded in mechanism, that it has no back door to the outcome? And whose effect is being estimated? A method that manufactures a natural experiment is powerful because it reaches past the confounders you failed to measure, yet that power is borrowed against assumptions you can only partly verify. The right posture is neither dismissal nor faith. It is scrutiny of the assumptions, in the open.

This article is educational and not medical advice; for decisions about your own health, please talk with your own clinician.