Spatial extended hazard model with application to prostate cancer survival

Area of Science:

• Biostatistics
• Spatial Statistics
• Survival Analysis

Background:

• Standard survival models often struggle with high-dimensional, spatially correlated data.
• Existing methods may not adequately capture complex spatial dependencies in survival outcomes.
• Accurate modeling of areal survival data is crucial in various fields, including epidemiology and public health.

Purpose of the Study:

• To develop a flexible Bayesian semiparametric approach for the extended hazard model.
• To generalize this approach for high-dimensional, spatially grouped data, accommodating county-level spatial correlation.
• To provide efficient methods for model fitting and hypothesis testing, including comparisons with existing models.

Main Methods:

• A Bayesian semiparametric extended hazard model is proposed.
• Normal transformation model and intrinsic conditionally autoregressive prior are used for spatial correlation.
• Efficient Markov chain Monte Carlo algorithms are developed for large, censored survival datasets.
• Bayes factors are employed for per-variable tests of proportional hazards, accelerated failure time, and accelerated hazards.

Main Results:

• The developed Bayesian approach effectively handles high-dimensional, spatially grouped survival data.
• The new methods demonstrate superior performance compared to standard additive Cox models with spatial frailties.
• Efficient algorithms facilitate the analysis of very large and highly censored areal survival data.
• Per-variable tests allow for interpretable model reduction, considering spatial effects.

Conclusions:

• The proposed Bayesian semiparametric extended hazard model offers a powerful and flexible tool for spatial survival analysis.
• This approach provides significant improvements in model fit and interpretability for complex, spatially correlated survival data.
• The efficient computational algorithms make it applicable to large-scale, real-world datasets.