Spatial Bayesian Networks Reveal Complex Drivers of Waste Disposal in Eswatini

Municipal solid waste (MSW) management remains a persistent sustainability chal-lenge in low- and middle-income countries, where uneven service coverage and rapid spatial change produce heterogeneous household disposal behaviours and substantial environmental externalities. This study develops a spatially explicit Bayesian network framework to map and explain six dominant household solid-waste disposal pathways across Eswatini using enumeration areas (EAs; n = 2,326) and nationally consistent census-linked predictors. Separate Tree-Augmented Naïve Bayes (TAN) models were trained for regular collection, irregular collection, open burning, public dumping, backyard pit disposal, and undesignated disposal, integrating socio-demographic, in-frastructural, accessibility, environmental, and neighbourhood-context variables, while explicitly quantifying predictive uncertainty using posterior entropy and Kull-back–Leibler (KL) divergence. Hold-out evaluation (465 test EAs; 1,861 training EAs) shows strong pathway-specific performance, with overall accuracy ranging from 0.497-0.989 across targets and ex-pected-value prediction errors of RMSE = 0.148-0.289 and MAE = 0.141-0.242. Uncer-tainty surfaces reveal low entropy in structurally homogeneous, well-served urban cores and elevated uncertainty in peri-urban transition zones where disposal behav-iours are mixed and services are unreliable. KL divergence highlights a limited subset of EAs where local conditions strongly update national expectations—priority loca-tions for targeted interventions and improved data collection. The framework provides policy-ready, uncertainty-aware evidence to support area-based service planning and sub-national monitoring relevant to SDG 11.6.1 in data-constrained contexts.