Forecasting the prevalence of child acute malnutrition using environmental and conflict conditions as leading indicators

Millions of children worldwide experience acute malnutrition. Forecasts of prevalence that afford sufficient reliability, precision, and advance warning are valuable to facilitate anticipatory action capable of mitigating the extent and downsides of crises. Existing research and resources lack prediction based on statistical analysis with broad cross-national scope and a focus on identifying leading indicators. We model the prevalence of child acute malnutrition at the level of subnational geographic regions (generally first-order administrative divisions), highlighting environmental conditions (precipitation, temperature, vegetation) and lethal and non-lethal conflict activity as main predictors, alongside demographic and geographic characteristics, and involving a temporal vantage point framework that reflects requirements of practical application. Estimations are performed using the random forest machine-learning algorithm, trained on data from 36 countries across mainland Sub-Saharan Africa spanning 2003-2019, including a novel compilation of measurements of prevalence rates drawn from DHS, MICS, and SMART surveys. Our results show strong predictive performance that remains consistent with lead times extending out from one month to 12 months. All the environmental and conflict factors register as important leading indicators. The findings reinforce the potential of relying on model-based approaches to bolster the foundations for humanitarian measures that are better positioned to reduce negative repercussions of food insecurity.