Multilevel small-area childhood stunting risk estimation: Insights from spatial ensemble learning, agro-ecological and environmentally remotely sensed indicators

Small area childhood stunting risk estimations remain a critical tool for shaping the surveillance policies of such a public health concern. With classical statistical methods, stunting risk metrics are mainly reported at a national level, which inhibits fine-scale insights at a more localized level. To this end, we implemented a novel multilevel small area estimation (SAE) approach for effectively explore the scale effect and non-stationarity of cross-sectional stunting prevalence data, by measuring their nonlinear interactions through the lens of satellite-derived indicators, environmental and agroecological factors using a predictive spatial ensemble learning and explainable artificial intelligence framework in the northern province of Rwanda. We found a wide spatial variability with 27.1% prevalence in childhood stunting with considerable heterogeneity across regions. Random forest consistently outperformed other base learners, including spatial ensemble learning, with an average F1 score and Matthews Correlation Coefficient of 0.8 and 0.65, respectively. Compared to the observed stunting prevalence, the model respectively resulted in 28.2%, 30.7%, 28.2% and 30.2% stunting risk at 1 km 2 and 5 km 2 grid resolutions, village and sector levels evidencing its performance to handle the effect of spatial scale, except the spatial ensemble learning high discriminability between stunting and non-stunting risk probabilities. Our study underscores the influence of livestock intensification and enhanced soil fertility, suggesting the adoption of an integrated farming framework combining the increase of livestock diversity and homestead crops and adapting to climate shocks to improve child health. Our results highlight the need to consider a multilevel framework when planning localized childhood stunting interventions to address persistent stunting risk disparities in the region. • This study conducted multilevel small-area childhood stunting risk estimation in Rwanda. • The scale effect and the non-stationarity of the relationship between stunting risk and environmental and agroecological factors. • Random forest consistently outperformed other base machine learning learners. • Stunting risk was associated with livestock density, water access, and soil fertility. • The high stunting risk presented a spatially heterogeneous pattern across the study area