Bayesian Hierarchical Spatiotemporal Models Outperform Spatial Autoregressive Models in Predictive Stability: Evidence from Mississippi County-Level Premature Mortality, 2015–2025 County Health Rankings Releases

County-level premature mortality in high-disparity settings exhibits strong spatial clustering and occasional abrupt shocks, yet few studies have compared the predictive performance of frequentist spatial autoregressive (SAR) models with Bayesian hierar-chical spatiotemporal alternatives under rigorous out-of-sample validation. Using 11 years of Mississippi county-level years of potential life lost (YPLL) data from County Health Rankings & Roadmaps (2015–2025 releases; underlying deaths ≈2012–2023; 82 counties, 902 observations), we contrasted a panel SAR model with year fixed effects against a Bayesian hierarchical model combining a BYM-2 spatial prior and an AR(1) temporal process. Both models used identical Queen-contiguity weights and seven standardized covariates. In leave-one-year-out cross-validation, the Bayesian model achieved a mean out-of-sample RMSE of 1,491 versus 1,690 for SAR (11.8% improve-ment) and generalized more reliably during the COVID-19 mortality surge (RMSE 2,184 vs. 2,559). Only the injury-death rate retained a credible positive association in the Bayesian model; the remaining six covariates — all significant in SAR — were shrunk toward zero. Both models left substantial residual spatial autocorrelation (Moran’s I ≈ 0.39–0.45), but Bayesian shrinkage and partial pooling yielded markedly more stable forecast. These findings demonstrate that Bayesian hierarchical spatio-temporal models provide more robust and policy-relevant predictions than SAR mod-els in sparse, high-disparity settings.