Imputation Strategies for the Estimation of Natural Direct and Indirect Effects

Abstract Mediation analysis is widely adopted to infer causal mechanism by disentangling indirect or mediated effects of an exposure on an outcome through given intermediaries, from the remaining direct effect. Traditional approaches build on standard regression models for the outcome and mediator, but easily result in difficult-to-interpret or difficult-to-report results when some of these models involve non-linearities. In this article, we overcome this via a general class of so-called natural effect models, which directly parameterize the (natural) direct and indirect effects of interest. We propose flexible estimation strategies for the direct and indirect effect parameters indexing these models, that are easy to perform with standard statistical software: one based on regression mean imputation and one based on doubly robust imputation. We give a theoretical discussion of the properties of these estimation strategies. We moreover assess their finite-sample performance through a simulation study, and through the analysis of the WHO-LARES study on the association between residence in a damp and moldy dwelling and the risk of depression.