Household-level spatial modelling of child height-for-age in Northern Province, Rwanda: comparison of geographically...

Childhood stunting remains a major public health concern in low- and middle-income countries, with persistent sub-national disparities. This study examined spatially varying associations between Height-for-Age z-scores (HAZ) and selected child, maternal, and household factors, and compared Geographically Weighted Regression (GWR), Multi-scale GWR (MGWR) and Geographically Neural Network Weighted Regression (GNNWR) in their ability to capture spatial heterogeneity and non-linearity. We analysed data from a cross-sectional survey conducted in December 2021 in the Northern Province of Rwanda. The survey covered 615 households, with analyses performed for 601 children aged 1-36 months. After imputation, multicollinearity screening, and feature selection, HAZ was modelled using Ordinary Least Squares (OLS), GWR, MGWR and GNNWR. Model performance was assessed using coefficient of determination (R2), Root Mean Square Error (RMSE), Akaike Information Criterion/AIC Corrected (AIC/AICc), and Moran's I of residuals. Overall stunting prevalence was 27.1%. Spatially varying associations were observed for child age, sex, birthweight, underweight status, selected childcare practices, maternal support, and household living conditions. GNNWR achieved the best training performance (R2 = 0.66; RMSE=0.74) followed by MGWR (R2 =0.51; RMSE=0.88) and GWR (R2 =0.43; RMSE=0.95). Validation performance declined for all models, although GNNWR retained a higher validation (R2=0.28) with negligible residual spatial autocorrelation (Moran's I = -0.005). GNNWR provided the strongest overall fit among the compared models and highlighted local spatial variation in associations with HAZ. However, the modest validation performance indicates that the findings should be interpreted cautiously. The workflow offers a portable framework for analysing DHS-like geocoded household datasets, but broader applicability should be assessed using spatial cross-validation, and sensitivity analyses.