Causal inference with a mediated proportional hazards regression model

Area of Science:

• Biostatistics
• Epidemiology
• Causal Inference

Background:

• Causal mediation analysis with survival data is crucial for understanding complex health relationships.
• Existing methods by VanderWeele (2011) rely on the rare outcome assumption for proportional hazards models.
• This limitation hinders analysis when outcomes are common.

Purpose of the Study:

• To extend VanderWeele's causal mediation analysis for survival data beyond the rare outcome assumption.
• To develop and validate novel approaches for estimating natural direct and indirect effects in proportional hazards models.
• To provide practical methods applicable to non-rare time-to-event outcomes.

Main Methods:

• Developed two novel approaches to estimate natural direct and indirect effects without the rare outcome assumption.
• Utilized numerical integration with cumulative baseline hazard estimation via Breslow method (Cox model) or a piecewise constant hazard model.
• Employed simulation studies for method comparison and applied to the ASSESS-AKI Consortium data.

Main Results:

• Successfully extended causal mediation analysis to non-rare outcomes in survival data.
• The proposed methods provide accurate estimation of natural direct and indirect effects at specific time points.
• Demonstrated the utility of the approaches using both simulated and real-world data.

Conclusions:

• The developed methods overcome the limitations of previous approaches for causal mediation in survival analysis.
• These techniques offer robust tools for researchers studying time-to-event data with common outcomes.
• Facilitates a deeper understanding of direct and indirect effects in various health-related research.