Simultaneous prediction of multiple dimensions of food security

Program officers are faced with difficult modeling choices when monitoring the state of diverse yet related dimensions of household food security. We compare results of single- and multi-output frameworks, applying both econometric and machine learning models, to jointly predict three interrelated outcomes of food security: undernourishment (access), dietary diversity (utilization), and food consumption stability (stability). Using the 2013–2022 Kyrgyz Integrated Household Survey, we evaluate generalized linear models (GLM), a multi-output generalized structural equation model (GSEM), and single- and multi-output artificial neural networks (ANNs) under both cross-sectional and temporal validation settings. The results show that accounting for non-linearity of relationships in our specifications yields greater predictive gains than modeling outcomes jointly. Among the models evaluated, the simplest model (single outcome GLM) often performs as well as others, and even better in some specifications, while also providing interpretable parameter estimates. However, accounting for non-linearity with our single-outcome ANN often achieves superior performance in low-outcome groups, which is especially relevant for policies aimed at targeting vulnerable households. Adding complexity through joint prediction of the related outcomes improves precision in some cases, but reduces recall, increasing exclusion risk. This trade-off makes such approach less desirable from both equity and implementation perspectives. The findings emphasize that complex patterns tie household socio-economic and demographic characteristics with food security dimensions. These results can guide program officers navigating the trade-offs between prediction performance, interpretability, and implementation; showing that linear models can deliver robust predictions, but specific non-linear, data-driven models might allow better targeting of vulnerable households.