Semiparametric causal mediation analysis of cluster-randomized trials for indirect and spillover effects

Area of Science:

• Statistics
• Epidemiology
• Biostatistics

Background:

• Cluster-randomized trials (CRTs) are increasingly used in public health and social sciences.
• Causal mediation analysis in CRTs is challenging due to data dependency and limited methods.
• Existing methods often rely on restrictive parametric assumptions.

Purpose of the Study:

• To develop novel, robust statistical methods for causal mediation analysis in CRTs.
• To address key mediation effect estimands: natural indirect effect, individual mediation effect, and spillover mediation effect.
• To provide semiparametric efficient estimators for improved causal inference in clustered settings.

Main Methods:

• Development of a formal semiparametric efficiency theory for mediation in CRTs.
• Derivation of efficient influence functions (EIFs) for multiple mediation estimands.
• Implementation of doubly-robust estimators using parametric models and data-adaptive machine learning for nuisance functions.

Main Results:

• Proposed doubly-robust methods offer semiparametric efficiency for estimating mediation effects in CRTs.
• Simulation studies confirm the good finite-sample performance of the new estimators.
• The methods are illustrated through a reanalysis of a real-world CRT dataset.

Conclusions:

• The developed semiparametric methods provide a robust framework for causal mediation analysis in CRTs.
• These methods relax parametric assumptions, offering more reliable effect estimation.
• The findings advance the methodology for understanding complex causal pathways in clustered experimental designs.