Combining parametric, semi-parametric, and non-parametric survival models with stacked survival models

Area of Science:

• Statistics
• Biostatistics
• Machine Learning

Background:

• Non-parametric survival models offer robustness but can have high variance in small samples.
• Parametric and semi-parametric models have lower variance but are sensitive to misspecification.
• A bias-variance trade-off exists in survival function estimation, particularly with limited data.

Purpose of the Study:

• To introduce and evaluate stacked survival models for estimating conditional survival functions.
• To demonstrate the adaptive balancing of strengths and weaknesses of diverse survival models.
• To compare the performance of stacked survival models against individual models and cross-validation selection.

Main Methods:

• Stacked survival models combine parametric, semi-parametric, and non-parametric models using optimal weighting.
• Prediction error minimization is employed to determine model weights.
• An extensive simulation study across various scenarios was conducted.

Main Results:

• Stacked survival models consistently performed well across a wide range of simulation scenarios.
• The stacked approach adaptively balanced the bias-variance trade-off inherent in survival estimation.
• Stacked survival models achieved performance comparable to or better than models selected via cross-validation.

Conclusions:

• Stacked survival models provide a flexible and high-performing method for conditional survival function estimation.
• This approach effectively leverages the advantages of multiple modeling strategies.
• The methodology was successfully applied to a real-world dataset from a German breast cancer study.