confounding

The Catalogue of Bias, maintained by the Centre for Evidence-Based Medicine at the University of Oxford, defines confounding as a distortion that modifies an association between an exposure and an outcome because a factor is independently associated with both. Those two independent links are the test. Age is a confounder of almost everything in health research because older people differ both in what they are exposed to and in what happens to them, so any comparison that ignores age is partly a comparison of ages.

The distortion runs in either direction. It can invent an association that is not there, and it can hide one that is. The catalogue's worked example is hormone replacement therapy: observational studies appeared to show a cardiovascular benefit, and when socioeconomic status and education were adjusted for, that benefit went away. The women taking the therapy had been different from the women not taking it in ways that also affected their hearts.

Researchers have two places to deal with it. In the design, randomisation is described by the catalogue as the best way to reduce the risk, because assigning people by chance breaks the link between the factor and the exposure. In the analysis, stratification, statistical adjustment and propensity scores attempt the same correction after the fact. Neither closes the question. Adjustment only reaches the factors somebody thought to measure, and the catalogue is explicit that there will always be the possibility of unknown confounders that cannot be taken into account. What is left over is called residual confounding, and it is the standing reason an observational study cannot settle cause on its own.