Confounding, Mediation, Moderation, and General Considerations in Regression Modeling

Abstract

Although causation cannot be established solely by statistical methods, multivariable modeling can be a useful tool in illuminating hypotheses about causal processes. In the present chapter, we discuss several concepts that are vital to exploiting multivariable models for this purpose. After a brief overview of modeling in general, we present the concept of confounding, including a brief introduction to the graphic representation of causal hypotheses. From graphic models, we move on to a relatively extensive section on mediation, including some specific recommendations on conducting mediation tests, and further discussion of graphs. We then present a short section on testing interactions and subgroup analysis, followed by some final comments on sample size, variable selection, and the preservation of measurement forms.