Stacked survival models for residual lifetime data.

Area of Science:

• Biostatistics
• Epidemiology
• Survival Analysis

Background:

• Accurately modeling disease progression is challenging when onset is insidious and exact timing is unknown.
• Prevalent cohort studies often rely on residual lifetimes measured from a defined screening date to circumvent onset uncertainty.
• Existing methods like nonparametric maximum likelihood estimation (NPMLE) can produce wide confidence intervals, while parametric methods risk model misspecification.

Purpose of the Study:

• To develop a robust survival function estimator for prevalent cohorts with insidious disease onset.
• To address the limitations of NPMLE (wide confidence intervals) and parametric methods (sensitivity to misspecification).
• To propose a stacking procedure that combines multiple survival estimators for improved accuracy and robustness.

Main Methods:

• Utilized right-censored residual lifetime data from a prevalent cohort.
• Proposed a stacking procedure combining nonparametric and parametric survival function estimators.
• Optimized stacking weights by minimizing the Brier Score loss function to enhance estimator performance.

Main Results:

• The proposed stacking procedure offers an alternative to traditional estimators when dealing with uncertain disease onset times.
• This method aims to provide a more robust estimation of the survival function compared to individual parametric or nonparametric approaches.
• Optimal weights derived from Brier Score minimization balance the strengths of different survival models.

Conclusions:

• The stacking procedure effectively overcomes the non-robustness associated with model misspecification in survival analysis.
• This approach enhances the reliability of survival function estimation in prevalent cohorts, particularly for diseases with insidious progression.
• The method provides a flexible framework for integrating diverse survival estimation techniques for improved clinical and epidemiological insights.