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

Area of Science:

• Causal inference
• Statistical modeling
• Machine learning applications

Background:

• Causal mediation analysis estimates indirect effects through mediators and direct effects.
• Accurately adjusting for confounders in mediation analysis is complex, especially with nonlinear relationships.
• Machine learning offers flexible function forms to address confounding in mediation analysis.

Purpose of the Study:

• To evaluate parametric and nonparametric estimators for causal mediation analysis with various mediator types.
• To benchmark advanced statistical approaches like multiply-robust and double-machine-learning estimators.
• To provide guidance on formulating mediation problems, verifying assumptions, and selecting estimators.

Main Methods:

• Comprehensive benchmark using simulated data to assess direct and indirect effect estimation.
• Evaluation of classical and recent estimators, including multiply-robust and double-machine-learning methods.
• Application to U.K. Biobank data analyzing hypertension's effect on cognitive function mediated by brain morphology.

Main Results:

• Advanced statistical approaches, particularly multiply-robust and double-machine-learning estimators, performed well across diverse simulated settings.
• The analysis of hypertension in the U.K. Biobank cohort indicated significant mediation of its effect on cognitive function by brain structure alterations.
• The study provides a thorough assessment of direct and indirect effect estimation for binary, continuous, and multidimensional mediators.

Conclusions:

• Machine learning-integrated estimators enhance causal mediation analysis by flexibly handling confounders.
• Multiply-robust and double-machine-learning methods are recommended for robust direct and indirect effect estimation.
• Hypertension's impact on cognition is substantially mediated by changes in brain structure, highlighting the utility of advanced mediation analysis.