Criticism of a hierarchical model using Bayes factors

Area of Science:

• Biostatistics
• Medical Statistics
• Health Services Research

Background:

• Heart transplantation is a critical procedure with varying outcomes across hospitals.
• Assessing hospital performance requires robust statistical models to account for patient heterogeneity and surgical factors.
• Hierarchical models are commonly used but require careful validation.

Purpose of the Study:

• To analyze heart transplant outcomes in US hospitals.
• To evaluate the suitability of a Poisson/gamma exchangeable hierarchical model.
• To compare the existing model against alternative models addressing specific concerns.

Main Methods:

• Analysis of heart transplant surgery data from a two-year period in the US.
• Application of a Poisson/gamma exchangeable hierarchical model.
• Development and comparison of alternative models using Bayes factors.
• Sensitivity analysis using graphical displays and Bayes factor plots.

Main Results:

• The study identified concerns regarding the suitability of the initial hierarchical model.
• Alternative models were constructed to address issues like hierarchical structure, outliers, prior hyperparameters, covariates, and exchangeability.
• Bayes factors were employed to quantitatively compare models, aiding in the selection of the most appropriate statistical approach.
• Graphical methods assessed the sensitivity of the analysis to model assumptions and perturbations.

Conclusions:

• The evaluation highlighted potential limitations of the standard hierarchical model for heart transplant data.
• Model comparison using Bayes factors provides a rigorous framework for selecting appropriate statistical models in healthcare research.
• Sensitivity analyses are crucial for understanding the robustness of findings to model choices and data characteristics.