On causal mediation analysis with a survival outcome

Area of Science:

• Epidemiology
• Biostatistics
• Causal Inference

Background:

• Decomposing total effects into direct and indirect pathways is crucial for understanding causal mechanisms.
• Mediation analysis in time-to-event data requires robust estimation strategies.

Purpose of the Study:

• To propose a theory and develop estimators for natural direct and indirect effects in semiparametric survival models.
• To enhance mediation analysis by addressing confounding and ensuring robustness.

Main Methods:

• Utilizing marginal structural Cox proportional hazards and additive hazards models.
• Developing multiply robust estimators that allow flexible working models for confounding adjustment.
• Introducing semiparametric sensitivity analysis for mediator ignorability.

Main Results:

• New estimators for natural direct and indirect effects are provided for both Cox and additive hazards models.
• The multiply robust approach enhances consistency by requiring only a subset of working models to be correct.
• Sensitivity analysis techniques are developed to assess violations of mediator ignorability.

Conclusions:

• The proposed methods offer robust and flexible tools for mediation analysis in time-to-event outcomes.
• The multiply robust nature of the estimators simplifies inference by reducing the need to identify correct working models.
• The developed sensitivity analysis aids in evaluating the plausibility of causal findings under potential unmeasured confounding.