Abstract

Epidemiologic analyses often attempt to decompose the effect of an exposure on an outcome into its effect via a number of different pathways. For example, the effect of heavy alcohol consumption on systolic blood pressure (SBP) may be separated into an effect via body mass index (BMI), an effect via the liver enzyme gamma-glutamyl transpeptidase (GGT), an effect via both BMI and GGT , and an effect via other pathways (not through BMI or GGT ) – often called the direct effect. Much progress has been made, mainly due to contributions from the field of causal inference, in understanding the precise nature of estimands that capture these sorts of effects, the assumptions under which they can be identified from data, and statistical estimation methods for doing so. However, the focus in the causal inference literature has been mostly on the decomposition of an effect around and through a single mediator, or a set of mediators considered en bloc, hence the two components: a direct and an indirect effect.

In this talk we describe novel, counterfactually-defined path-specific effects that permit the decomposition of the total effect of an exposure on an outcome into a sum of numerous path-specific effects through many mediators, where the mediators are permitted to have a causal effect on each other. We show that there are many ways in which this decomposition can be done, discuss the strong structural and modelling assumptions under which the effects can be estimated, together with a sensitivity analysis approach when a particular subset of the assumptions cannot be justified. Illustrating these ideas using data on alcohol consumption, SBP , BMI and GGT from the Izhevsk Family Study, we focus on the ambitious nature of multiple mediation analyses, giving some practical guidance on how progress can be made.