Conceptual issues concerning mediation, interventions and composition

Abstract

Concepts concerning mediation in the causal inference literature are reviewed. Notions of direct and indirect effects from a counterfactual approach to mediation are compared with those arising from the standard regression approach to mediation of Baron and Kenny (1986), commonly utilized in the social science literature. It is shown that concepts of direct and indirect effect from causal inference generalize those described by Baron and Kenny and that under appropriate identification assumptions these more general direct and indirect effects from causal inference can be estimated using regression even when there are interactions between the primary exposure of interest and the mediator. A number of conceptual issues are discussed concerning the interpretation of identification conditions for mediation, the notion of counterfactuals based on hypothetical interventions and the so called consistency and composition assumptions.

The Role of Perceived Peer Norms in the Relationship Between Media Violence Exposure and Adolescents' Aggression

“…Although bootstrapping methods have been used as solution for analyses using non-normal variables, these do not fully solve problems with highly skewed variables such as aggressive behavior (B. Muthén, 2011), while Poisson regression models are difficult to interpret in the context of longitudinal moderated-mediation analysis (VanderWeele & Vansteelandt, 2009). Our solution to address the data skewness was to create a change score for aggressive behavior between wave 1 and wave 2.…”

The Causal Mediation Formula—A Guide to the Assessment of Pathways and Mechanisms

“…In other words, the regression coefficient R Y Z·X would no longer equal γ and the difference R Y X − R ZX R Y Z·X would no longer equal α. This follows from the fact that "controlling" or "adjusting" for Z in the analysis (by including Z in the regression equation) does not physically disable the paths going through Z; it merely matches samples with equal Z values, and thus induces spurious correlations among other factors in the analysis (see Pearl 1998;Cole and Hernán 2002;VanderWeele and Vansteelandt 2009). Such correlations cannot be detected by statistical means and, so, regardless of whether the error terms are independent, the difference-incoefficients and product-of-coefficients methods always yield the same (biased) result.…”

The Causal Mediation Formula - A Guide to the Assessment of Pathways and Mechanisms

“…In other words, the regression coefficient R Y Z·X would no longer equal γ and the difference R Y X − R ZX R Y Z·X would no longer equal α. This follows from the fact that "controlling" or "adjusting" for Z in the analysis (by including Z in the regression equation) does not physically disable the paths going through Z; it merely matches samples with equal Z values, and thus induces spurious correlations among other factors in the analysis (see Pearl, 1998;Cole & Hernán, 2002;VanderWeele & Vansteelandt, 2009).…”