Robust inference on effects attributable to mediators: A controlled-direct-effect-based approach for causal effect...

Area of Science:

• Causal inference
• Biostatistics
• Epidemiology

Background:

• Effect decomposition is crucial for investigating causal mechanisms with multiple mediators.
• Traditional causal mediation analysis relies on strong assumptions and overlooks interacting mechanisms.
• Unifying mediation and interaction is vital for comprehensive causal mechanism investigation.

Purpose of the Study:

• To propose a novel measure, the effect attributable to mediators (EAM), for effect decomposition.
• To provide a measure that incorporates both mediation and interaction effects.
• To develop a method with weaker identification assumptions than existing causal mediation analysis.

Main Methods:

• Extended the framework of controlled direct effects to define EAM.
• Developed a semiparametric estimator for EAM, ensuring robustness to model misspecification.
• Validated the asymptotic properties of the proposed estimator.

Main Results:

• EAM quantifies the extent to which an effect can be eliminated by manipulating mediators.
• EAM provides insights into a mediator's involvement through mediation, interaction, or both.
• The proposed semiparametric estimator demonstrates robustness and fully realized asymptotic properties.

Conclusions:

• EAM is a more appropriate measure than path-specific effects for clinical and medical studies.
• The developed method offers weaker identification assumptions compared to causal mediation analysis.
• Applied EAM to demonstrate the elimination of hepatitis C virus infection's effect on mortality by controlling alanine aminotransferase and treating hepatocellular carcinoma.