Rank-Based Greedy Model Averaging for High-Dimensional Survival Data.

Area of Science:

• Statistics
• Biostatistics
• Machine Learning

Background:

• Model averaging improves prediction accuracy, but existing methods are limited to low-dimensional settings with fully observed data.
• High-dimensional predictors in survival data pose challenges for accurate risk prediction.

Purpose of the Study:

• To propose a novel rank-based greedy (RG) model averaging method for accurate prediction of risk effects in high-dimensional survival data.
• To develop a computationally efficient and robust approach against model misspecification.

Main Methods:

• Utilized transformation models with splitting predictors as working models.
• Employed a smooth concordance index function for deriving candidate predictions and optimal model weights.
• Applied a greedy algorithm tailored for high-dimensional data.

Main Results:

• Derived an asymptotic error bound for optimal weights under mild conditions.
• Demonstrated that weights for correct submodels approach one in probability.
• Showcased robust performance through extensive simulations and real-world data analysis.

Conclusions:

• The proposed rank-based greedy model averaging method offers a flexible, efficient, and robust solution for high-dimensional survival data.
• The approach effectively enhances prediction accuracy without requiring a correct joint model or transformation function estimation.
• Numerical studies confirm the superior performance compared to traditional regularization methods.