Scalable algorithms for semiparametric accelerated failure time models in high dimensions.

Area of Science:

• Statistics
• Biostatistics
• Machine Learning

Background:

• Semiparametric accelerated failure time (AFT) models offer an alternative to Cox proportional hazards models when hazard ratios are not constant.
• Fitting AFT models using rank-based criteria presents computational challenges in high-dimensional settings due to non-differentiability.

Purpose of the Study:

• To propose a novel algorithm for fitting semiparametric AFT models using a penalized rank-based loss function.
• To develop a new criterion for tuning parameter selection that addresses issues with cross-validation for censored data.

Main Methods:

• An alternating direction method of multipliers (ADMM) algorithm was developed to minimize a penalized rank-based loss function.
• The algorithm accommodates various penalties and scales efficiently with the number of subjects and predictors.
• A new tuning parameter selection criterion was introduced to improve model fitting with censored responses.

Main Results:

• The proposed ADMM algorithm demonstrates significant speed improvements over existing methods, particularly for high-dimensional data.
• Penalized rank-based estimators show superior performance compared to penalized weighted least squares estimators in simulation studies.
• Rank-based estimators for semiparametric AFT models are competitive with proportional hazards models in high-dimensional settings, unlike weighted least squares estimators.

Conclusions:

• The developed algorithm provides an efficient and scalable solution for fitting semiparametric AFT models in high-dimensional scenarios.
• Penalized rank-based loss functions offer a robust alternative to weighted least squares for AFT model estimation.
• The findings suggest rank-based AFT models are a valuable tool for analyzing complex survival data, including cancer datasets.