Confounding

Confounding is when a third variable distorts the apparent link between an exposure and an outcome. That third variable is the confounder. It is independently tied to both the exposure and the outcome. By the classic rules, a confounder must be a cause (or stand-in) of the outcome. It must be unevenly spread between the exposed and unexposed groups. And it must not lie on the causal pathway itself. In longevity research, the 'healthy-user bias' is the classic example. People who adopt a preventive habit (statins, caloric restriction, exercise) tend to have healthier baseline behaviors and backgrounds. So an apparent survival benefit in an observational study might reflect those hidden advantages, not the intervention. Adjustment methods can help. These include multivariable regression, propensity-score matching, inverse-probability weighting, and Mendelian randomization. But they rarely erase confounding fully. Unmeasured or poorly measured variables leave 'residual confounding'. Shrank et al. (2011) showed this. Even heavy covariate adjustment could not fully remove healthy-user bias in Medicare drug studies. One useful tool is the 'E-value' (VanderWeele and Ding, 2017). It tells you how strong an unmeasured confounder would have to be to fully explain an observed link. That gives you a yardstick for how solid any observational longevity finding really is.

Last reviewed: June 22, 2026

This definition is educational and is not medical advice, a diagnosis, or treatment. Talk to a doctor about any health decisions. Read our full medical disclaimer