Abstract
<title>Abstract</title> <p>Robust prediction and well calibrated uncertainty underpin evidence informed decisions in science,engineering and public health. This study is framed within Bayesian hierarchical spatial models builtwith integrated nested Laplace approximation with stochastic partial differential equation (INLASPDE),which propagates uncertainty and allows neighbouring locations to borrow strength so thatunsampled grid cells still receive estimates with credible uncertainty bands. The study focuses on thekey question of which predictors to include in the Bayesian hierarchical models. Two approacheswere compared: (i) Machine-learning (ML) based feature selection—Bayesian Additive RegressionTrees (BART) and Support Vector Regression (SVR). (ii) Stepwise regression based feature selection—backward and forward. Cameroon serves only as an empirical illustration drawn from integratedsurvey sources; the findings concern the methods rather than this specific setting. Results: the modelsbuilt with machine-learning based feature selection led to between 13% to 32% reduction in mean absoluteerror (MAE), 8% - 34% reduction in root mean square error (RMSE), and 61% - 87% reductionin absolute bias (BIAS). Conclusion Even when survey data are sparse, combining machine-learningvariable selection with fully Bayesian spatial modelling yields more accurate small-area estimatesand more clearly quantified uncertainty surfaces when compared with the corresponding stepwiseregression. The workflow can be adapted to other contexts within small-area estimation.</p>