<p>Spatial variable selection remains a methodological challenge in epidemiology, especially when predictors are highly correlated and the response variable is bounded. Conventional penalized regressions, such as LASSO and its variants, are common with Gaussian, Poisson, and binomial likelihoods. In epidemiology, incidence proportions constrained between 0 and 1 require flexible models beyond Gaussian assumptions. This study employs a Bayesian hierarchical framework to identify key ecological and sociodemographic determinants of malaria incidence while addressing response variable boundedness, spatial dependence, and multicollinearity. Two regularization priors—the Bayesian LASSO and the global–local horseshoe prior—were compared for variable selection. Spatial effects were modeled using smooth functions of geographic coordinates. Modelled malaria incidence and socio-ecological predictors obtained from the Nigeria 2018 Demographic and Health Survey (NDHS) were used, and model performance was assessed using RMSE, R², WAIC, and posterior uncertainty intervals. Both horseshoe and LASSO priors have proven to be efficient in Bayesian spatial variable selection and prediction of bounded response variables with strongly correlated predictors. However, the horseshoe prior yielded a negligible superior predictive performance, a higher R², and lower RMSE and WAIC values relative to the LASSO prior. This result signifies that both models provided equivalent predictive adequacy and generalization capacity while effectively handling spatial autocorrelation and multicollinearity. The Bayesian spatial variable selection framework employing the adaptive shrinkage improves interpretability, predictive accuracy, and uncertainty quantification for bounded outcomes. This methodological advancement offers a robust, transparent approach to spatial epidemiological inference, providing valuable insights for malaria control and wider disease-mapping applications.</p>