Estimation of separable direct and indirect effects in a continuous-time illness-death model

Area of Science:

• Causal Inference
• Survival Analysis
• Biostatistics

Background:

• Understanding the causal pathways of exposure on time-to-event outcomes is crucial in medical research.
• Traditional mediation analysis methods like natural direct and indirect effects face limitations with time-dependent processes and competing risks.
• The illness-death process model is a standard framework for analyzing time-to-event data with intermediate states.

Purpose of the Study:

• To propose and define novel separable direct and indirect effects for analyzing baseline exposure effects in a continuous-time illness-death process.
• To extend existing causal mediation frameworks to situations where the mediator is truncated by the terminal event.
• To develop statistically sound methods for estimating these effects and assess their theoretical properties.

Main Methods:

• Utilized the concept of separable (interventionist) effects to define direct and indirect causal pathways.
• Derived conditions for the identifiability of these effects, addressing potential structural assumptions.
• Developed plug-in estimators and multiply robust, asymptotically efficient estimators based on efficient influence functions.

Main Results:

• Successfully defined meaningful mediation targets for separable direct and indirect effects, even when the mediating event is truncated.
• Established conditions for identifiability and discussed the validity of underlying structural assumptions.
• Demonstrated the performance of the proposed estimators through a simulation study and application to real-world registry data.

Conclusions:

• The proposed separable effects provide a robust framework for causal mediation analysis in continuous-time illness-death processes.
• The developed estimators offer reliable tools for quantifying direct and indirect effects in complex survival data.
• This methodology enhances our ability to understand exposure effects in longitudinal health studies.