Alemayehu GELETU

Correlating High-Resolution Predicted Food System Indicators: A Residual Spatial Decomposition Framework

High-resolution prediction of food system indicators provides new opportunities to examine how food supply, food environment, and nutrition outcomes co-occur across space and time. Using Ethiopia as a case study, we generated spatially continuous 5 × 5 km surfaces for 122 food system indicators from 2010 to 2022 using a stacked model-based geostatistical framework. The framework combines four machine learning algorithms: Random Forest, Generalized Additive Models, LASSO, and XGBoost with a Bayesian Gaussian process meta-learner that integrates predictions while accounting for residual spatio-temporal dependence.
A central challenge in correlating predicted indicator surfaces is that observed associations may be spurious, arising from shared environmental predictors or spatial autocorrelation rather than meaningful relationships between food system components. To address this, we decompose each indicator on the linear predictor scale into a covariate-driven mean component and a structured spatial residual component. Correlating the residual spatial fields provides an analogue to partial correlation, measuring whether indicators co-vary geographically after their covariate-driven signals have been absorbed by their own prediction models. We complement this with global and local Lee’s L statistics to explicitly account for neighborhood structure and identify where associations are strong, weak, or reversed. Together, these methods provide a robust framework for analyzing cross-indicator spatial association in predicted food system surfaces.

Agenda > Tous les événements