Estimating restricted mean treatment effects with stacked survival models

Area of Science:

• Biostatistics
• Survival Analysis
• Epidemiology

Background:

• Restricted mean survival time (RMST) differences are clinically relevant for comparing treatment effects in observational studies.
• Observational data often presents confounding variable imbalances between groups, complicating accurate RMST difference estimation.
• Existing methods for covariate-adjusted RMST differences rely on accurate estimation of the covariate-adjusted survival distribution.

Purpose of the Study:

• To propose and evaluate stacked survival models for robustly estimating covariate-adjusted restricted mean survival time differences.
• To assess the performance of stacked survival models compared to traditional methods, particularly under varying data conditions.
• To demonstrate the utility of the proposed method using real-world post-lung transplant survival data.

Main Methods:

• Utilizing stacked survival models, which combine multiple survival models (parametric, semi-parametric, non-parametric) via weighted averaging to minimize prediction error.
• Estimating the covariate-adjusted survival distribution using the stacked model approach.
• Marginalizing the estimated covariate-adjusted survival distribution over the covariate distribution to obtain the RMST difference.

Main Results:

• Simulation studies indicate that improved covariate-adjusted survival distribution estimation generally leads to better mean squared error for the RMST difference.
• The proposed stacked survival model estimator performs comparably to Cox regression when the proportional hazards assumption holds.
• The stacked survival model estimator significantly outperforms Cox regression when the proportional hazards assumption is violated.

Conclusions:

• Stacked survival models offer a robust method for estimating covariate-adjusted restricted mean survival time differences from observational data.
• This approach provides a reliable way to handle confounding and improve treatment effect estimation in survival analysis.
• The method's flexibility and performance make it valuable for analyzing complex survival data, as shown in the lung transplant center analysis.