Confounding

A confounder is a variable associated with both the exposure and the outcome, that is not a step on the causal pathway between them. It distorts the apparent exposure–outcome association — manufacturing, masking or exaggerating an effect. It is controlled by design (randomisation, restriction, matching) or by analysis (stratification, multivariable adjustment, propensity scores). Crucially, adjustment can only touch confounders you have measured.

A triangle. The confounder sits above and points down to both the exposure and the outcome, while the exposure–outcome link you actually care about runs along the bottom. Because the confounder is tied to both ends, an association along the bottom can appear even when the exposure does nothing — the confounder is doing the work.

Ask three questions of any candidate confounder: is it associated with the exposure? is it an independent risk factor for the outcome? is it off the causal pathway (not a mediator that the exposure causes)? Only if all three are yes is it a true confounder you must control. Adjusting for it should move the effect estimate; if it doesn’t, it wasn’t confounding the result.

Confounding is why observational studies can never prove causation outright: there may always be an unmeasured factor explaining the association. Randomisation is special precisely because it balances confounders you never even thought to measure. Every “adjusted for age, sex and smoking” analysis is an admission that residual confounding might remain.

• Confounder = linked to exposure AND outcome, off the causal path
• Design control: randomise / restrict / match
• Analysis control: stratify / adjust / propensity
• Adjustment fixes measured confounders only

HRT and coronary heart disease — large observational studies (notably the Nurses’ Health Study) found post-menopausal women on hormone replacement therapy had roughly one-third the coronary heart disease risk of non-users, and HRT was recommended for cardiac prevention. When the Women’s Health Initiative RCT (JAMA 2002) randomised >16,000 women, HRT actually increased early coronary events. The observational “benefit” was largely healthy-user confounding: women who took HRT were healthier and better-resourced to begin with. No amount of multivariable adjustment in the observational data rescued the answer — only randomisation balanced the unmeasured confounders and reversed the conclusion.

• Adjusting for a mediator — a variable on the causal pathway (caused by the exposure) is not a confounder; adjusting for it wrongly hides the real effect.
• Assuming “fully adjusted” means causal — residual and unmeasured confounding always remain in observational data.
• Confusing confounding (a nuisance to remove) with effect modification (a real difference to report).