Bias, confounding and interaction: Three fundamental themes in the field of causal inference

Abstract

Causation is the act or agency that produces an effect. In epidemiology assessing the cause of disease is an essential aim. When an epidemiological study observes a statistical association between an exposure and an outcome, a number of possible explanations need to be considered, some of which are alternatives to the existence of a cause–effect relationship. Indeed, the effects of chance (random error), bias (systematic error) or confounding can produce spurious results, leading to wrong conclusions about the existence of a true association. Systematic error, also known as bias, refers to deviations that are not due to chance alone. Estimates affected by bias are inaccurate. Confounding occurs when the observed effect of an exposure on an outcome depends on the role of a third variable (confounder) that is associated with the exposure under study and that is also a cause of the outcome, thereby distorting the estimated magnitudes of the association. Interaction refers to the effect of two risk factors on an outcome. If the effect of the first risk factor varies in the different strata of the study population that is defined by the second risk factor, then there is a biological interaction. This differs from confounding, in which the estimates of the measures of effect are the same in each of the strata defined by the levels of the confounding variable. The variable that determines the difference between strata is defined as the effect modifier.