Bridged parametric survival models: General paradigm and speed improvements

Area of Science:

• Biostatistics
• Survival Analysis
• Health Informatics

Background:

• Large-scale biomedical data from electronic health records (EHRs) necessitate scalable survival models.
• Clustering in EHR data increases computational complexity for survival analyses.
• Existing models struggle to unify marginal and conditional survival perspectives.

Purpose of the Study:

• To develop fast, scalable bridged-survival models within a unified framework.
• To enable transparent estimation and relation of conditional hazard ratios, marginal hazard ratios, conditional acceleration factors, and marginal acceleration factors.
• To address computational challenges in analyzing clustered survival data.

Main Methods:

• Formulation of a Weibull parametric frailty likelihood for clustered survival times.
• Utilizing a nonlinear mixed model with positive stable frailties and Gaussian quadrature.
• Development of a novel closed-form integrated likelihood to reduce computational burden.

Main Results:

• Significant reduction in computational time: 12x faster in R (parfm) and 2400x faster in SAS.
• Preservation of the proportional hazards assumption for the marginal model.
• A unified framework allowing estimation of all four perspective-parameterization combinations.

Conclusions:

• The proposed bridged-survival models offer a computationally efficient and transparent approach for EHR data analysis.
• The Static-Stirling closed-form likelihood enhances the utility of bridged-survival models.
• Provides valuable insights into subject-specific and population-level survival effects.