Flexible yet Sparse Bayesian Survival Models With Time-Varying Coefficients and Unobserved Heterogeneity.

Area of Science:

• Medical research
• Biostatistics
• Statistical modeling

Background:

• Survival analysis is crucial in medical research.
• Existing models face challenges in balancing simplicity and flexibility.
• Simple models have strong assumptions; flexible models require extensive tuning.

Purpose of the Study:

• To present a novel survival model using Bayesian hierarchical shrinkage.
• To automatically determine covariate status (static, time-varying, or excluded).
• To balance model simplicity and flexibility while minimizing tuning and quantifying uncertainty.

Main Methods:

• Bayesian hierarchical shrinkage.
• Automatic covariate selection.
• Efficient Markov chain Monte Carlo (MCMC) sampler.
• Implementation in the R package shrinkDSM.

Main Results:

• The proposed model effectively balances simplicity and flexibility.
• It minimizes the need for manual tuning by researchers.
• The method naturally quantifies uncertainty in covariate effects.
• Demonstrated advantages over existing models via simulations and a clinical dataset.

Conclusions:

• The Bayesian hierarchical shrinkage approach offers an improved survival modeling strategy.
• The shrinkDSM package provides an efficient tool for implementing this method.
• This approach is beneficial for analyzing complex medical datasets, such as adenocarcinoma of the gastroesophageal junction.