Confounding and Interaction

Confounding occurs when the subpopulations of the source population being compared would have different outcomes over the risk period under study, even if they were subject to the same treatment or exposure. In this situation, the exposed and non-exposed groups are not comparable or exchangeable, in that the compared groups would have had different outcomes even if they had received the same treatment or exposure. Interaction is a highly ambiguous term: In statistics, it usually refers to the need for a product of variables in a regression model. In epidemiology, however, it usually means that, in some sense, the causal effect of exposure on the outcome varies with some other factor; that is, to estimate the effect of exposure, we must first know the level of the other factor. This causal conceptualization of interaction subsumes two distinct concepts: effect-measure modification (statistical interaction for an effect measure) and biological interaction. Effect-measure modification explicitly refers to a situation in which the measure under consideration is indeed a measure of the effect of exposure, and can be identified (estimated) from epidemiological data if confounding and measurement biases can be fully controlled. In contrast, biological interaction instead refers to interdependence of the mechanisms of action, and as such is not fully identified (estimable) from epidemiological data.