Mapping and Forecasting District-Level Stunting Dynamics in Indonesia Toward SDG Target 2.2: A Hybrid Bayesian-Machine Learning Spatiotemporal Analysis

This study introduces a spatiotemporal framework at the district level in Indonesia to examine and forecast stunting prevalence. The empirical analysis draws on data from 514 districts observed over 2022–2024, with short-term projections extended to 2025–2027 in line with the SDG 2.2 agenda. The modeling methodology is based on a Bayesian spatiotemporal formulation with the SPDE-INLA method. Instead of handling spatial and temporal lags separately, the model simultaneously incorporates them to reflect dependencies that change across both dimensions. This structure facilitates a more flexible representation of underlying risk dynamics. To improve prediction performance, we augment the baseline model with a hybrid component. Specifically, residual variation from the Bayesian specification is further explored using machine learning methods, providing an additional layer of adjustment. Spatial dependence is assessed through three alternative weighting schemes—KNN, Queen contiguity, and distance-based matrices—which are compared prior to selecting the final specification. The empirical specification includes nine key predictors within a semi-parametric framework. Several covariates are allowed to depart from strict linearity by accommodating time-varying effects. Three algorithms were evaluated during the prediction process to determine their abilities to capture the residual structure: XGBoost, Random Forest, and Elastic Net. Spatiotemporal clustering is examined through exceedance probabilities, resulting in the identification of seven unique cluster patterns. The findings consistently indicate that poverty is the main factor influencing stunting dynamics, with evident regional spillovers and temporal variations. Persistent hotspots are primarily located in eastern Indonesia. From a predictive standpoint, the hybrid specification—particularly the variant based on XGBoost—delivers the most stable performance. The forecast results indicate a gradual reduction in stunting prevalence throughout the forecast period. This study establishes persistent geographic inequalities in child nutrition risk and translates them into district-specific intervention priorities, providing decision-support information to further SDG Target 2.2 and its relationships with SDGs 1, 3, 4, and 6.