Confounding and Effect Modulation

Overview

An important feature of a regression model is its ability to include multiple covariates and thereby statistically adjust for possible imbalances in the observed data before making statistical inferences. This process of adjustment has been given various names in different fields of study. In traditional statistical publications, it is sometimes called the analysis of covariance, while in clinical and epidemiologic studies it may be called control for confounding. Interactions between covariates may also be included in the model and regarded as effect modifiers in the sense that the effect on the outcome differs according to the level of the moderator variable. When an outcome is correlated with a study variable but the relationship disappears when adjusted by a third variable, the third variable is often called a mediating variable or mediator. In an epidemiological study of the strength of the association between smoking status and lung cancer, the relationship may be affected by other variables such as the drinking habits, extent of exposure to tobacco smoke, or age of the subject, or other personal or environmental conditions. Variables other than smoking that affect the relationship of smoking and lung cancer are often described as modulating variables. Modulating variables are further classified as confounding variables, effect moderators, effect modifiers, or mediating variables according to their finer properties. Some terms and interpretations used to distinguish different types of modulation are based on statistical definitions and may thus be measured and tested objectively. In other instances, judgments regarding causality will be required, thus introducing concepts not readily amenable to statistical analysis.

Definitions of terms such as mediation, moderation, and confounding have been questioned because of their implied dependence on the unquantifiable concept of causality. This entry illustrates these through simple statistical modeling and figures, and provides examples of the roles of confounding, mediating, and moderating variables. The first example illustrates the role of “helplessness” as a moderating variable where patients with [Page 176]a “low” helplessness index could have decreasing depression even with an increasing “swollen joint count,” whereas patients with a “high” helplessness index have the opposite relationship (increasing depression with an increasing swollen joint count). The second example illustrates the “mediating” role of pain and comorbidities on the association observed between body mass index (BMI) and total unhealthy days (TUD). The association vanishes when the mediating variables are adjusted in the model. The analyses for each of the examples are adjusted for various confounding variables.

...