A surrogate ℓ₀ sparse Cox's regression with applications to sparse high-dimensional massive sample size time-to-event...

Area of Science:

• Biostatistics
• Computational Statistics
• Survival Analysis

Background:

• Sparse high-dimensional massive sample size (sHDMSS) time-to-event data pose significant computational challenges for existing survival regression methods.
• Current software often fails to operate effectively with such large and complex datasets.

Purpose of the Study:

• To develop a scalable ℓ0-based sparse Cox regression tool specifically designed for sHDMSS time-to-event data.
• To extend the broken adaptive ridge (BAR) methodology to the Cox model for improved performance on large datasets.

Main Methods:

• Extension of the ℓ0-based broken adaptive ridge (BAR) methodology to the Cox proportional hazards model.
• Utilizing reweighted ℓ2-penalized regression iteratively.
• Leveraging high-performance implementations of ℓ2-penalized regression for scalability.

Main Results:

• The developed BAR Cox regression estimator demonstrates selection consistency.
• The method provides oracle performance for parameter estimation.
• It exhibits a grouping property for highly correlated covariates.
• An R package implementing the BAR method is developed and validated.

Conclusions:

• The BAR Cox regression method provides a scalable and effective solution for analyzing sHDMSS time-to-event data.
• The method is rigorously validated through simulations and a real-world application on a large trauma registry dataset.