Location-Dependent Bayesian Stacking for Nigerian Household Wealth Prediction Under Model Uncertainty

Reliable prediction of household wealth is essential for monitoring socioeconomic inequality and supporting evidence-based economic planning. Geoadditive models provide a flexible framework for modelling household wealth by incorporating linear, nonlinear, and spatial effects. However, prediction commonly relies on selecting a single model, which ignores model uncertainty, or on conventional Bayesian Stacking with global ensemble weights, which may not adequately capture location-specific predictive performance. This study proposed and evaluated a location-dependent Bayesian stacking approach to account for model uncertainty and improve the spatial prediction of household wealth in Nigeria. Conventional Bayesian Stacking was modified by incorporating location-dependent ensemble weights, resulting in Geoadditive Local Stacking (GALS). The proposed approach was evaluated using data from 39,050 respondents in the 2024 Nigeria Demographic and Health Survey. Four competing Gaussian Geoadditive models incorporating socioeconomic and spatial effects were fitted, and predictive performance was compared with conventional Bayesian Stacking and single-model selection using expected log predictive density (ELPD), with higher ELPD indicating better predictive performance. State-level household wealth exhibited strong positive spatial autocorrelation (Moran's I = 0.703, p < 0.001), with mean wealth scores ranging from −1.077 in Kebbi State to 1.316 in Lagos State. GALS achieved the highest predictive performance, with a mean ELPD of −0.8350, compared with −0.8355 for conventional Bayesian Stacking and −0.8373 for the best single model. Location-dependent Bayesian Stacking provides a reliable framework for household wealth prediction under model uncertainty and can support evidence-based socioeconomic planning and policy in Nigeria.