Bayesian Spatiotemporal Modelling

Neonatal mortality had the best predictive accuracy (R²=0.99), while under-5 mortality had the weakest fit (R²=0.60). The study used a hierarchical Bayesian framework estimated via Integrated Nested Laplace Approximation in R 4.5.0. Spatial structure was modelled using Besag-York-Mollie priors, and temporal dependence used a first-order random walk. Four nested models were tested, progressively adding environmental, healthcare and demographic variable domains, with model selection based on DIC and WAIC. Island countries were linked to their nearest land neighbour so every country had at least one spatial neighbour and the model could be estimated without singularities. Penalised complexity priors shrank spatial and temporal random effects toward zero unless the data supported heterogeneity. The white-box Bayesian design allows policymakers to trace how covariates, spatial patterns and temporal trends contribute to risk estimates, supporting evidence-based planning.