Comparative Performance Analysis of Bayesian Hierarchical Models Versus Classical Statistical Approaches in Predicting Breast Cancer Treatment Outcomes: Evidence from Kenyan Healthcare Settings

Abstract

Current breast cancer treatment prediction models inadequately quantify uncertainty and fail to account for institutional clustering effects, particularly in resource-constrained healthcare settings. This study compared the performance of Bayesian hierarchical models against classical frequentist approaches for predicting pathological complete response (pCR) in breast cancer patients. We analyzed data from 5,400 breast cancer patients across 12 Kenyan treatment centers. Three progressively complex models were developed: single-level logistic regression (M0), Bayesian empty hierarchical model (M1), and Bayesian hierarchical model with clinical covariates (M2). Performance comparison utilized multiple metrics including Area Under the Curve (AUC), Brier Score, calibration measures, and information criteria. The Bayesian hierarchical model demonstrated superior performance with AUC = 0.837 compared to classical approaches (AUC = 0.752). Bayesian methods showed consistent 2-8 unit improvements in information criteria across all model complexity levels. The hierarchical structure captured 26.5% of outcome variation attributable toinstitutional clustering (ICC = 0.265), which classical models failed to address. Uncertainty quantification through credible intervals provided clinically meaningful prediction confidence assessment. Bayesian hierarchical approaches significantly outperform classical statisticalmethods in breast cancer treatment outcome prediction, particularly in settings with institutional clustering. The explicit uncertainty quantification and superior discrimination make Bayesian