Causal mediation analysis with one or multiple mediators: A comparative study.

Abstract

Mediation analysis decomposes the causal effect of a treatment on an outcome into an indirect effect, mediated through intermediate variables, and a direct effect, operating through other mechanisms. However, mediation analysis is challenging due to the need to accurately adjust for confounders of the treatment, mediators, and outcomes, which may involve complex nonlinear relationships. Machine learning offers a promising solution by accommodating flexible function forms to account for confounders. It can be integrated into various estimators, resulting in a complex landscape for the practitioner. We evaluate parametric and nonparametric implementations of classical and more recent estimators, providing a thorough assessment of direct and indirect effect estimation in causal mediation analysis for binary, continuous, and multidimensional mediators. Through a comprehensive benchmark using simulated data, we demonstrate that advanced statistical approaches, such as the multiply-robust and double-machine-learning estimators, perform well across most simulated settings and real-world data. As an application example, we analyze hypertension, a factor known to influence cognitive functions, to determine if this effect is mediated by changes in brain morphology, using the U.K. Biobank brain imaging cohort. Our findings indicate that for hypertension, a substantial portion of the effect is mediated by alterations in brain structure. This work provides guidance to the practitioner from the formulation of a valid causal mediation problem, from the verification of identification assumptions to the choice of an appropriate estimator. (PsycInfo Database Record (c) 2026 APA, all rights reserved).