机器学习与粮食援助影响的异质性：改进瞄准的基础

摘要

We investigate whether in-kind food aid can generate larger nutritional gains through the strategic allocation of different food boxes across households. Using data from a randomized evaluation of a nationwide food aid program in Egypt, we compare a staple-heavy food box with a more expensive nutrition-sensitive food box. We apply machine learning methods to estimate household-level heterogeneity in treatment effects and use these estimates to evaluate allocation rules under budget constraints. We find substantial heterogeneity in impacts on calorie intake, food insecurity experience, and dietary diversity. Observable household characteristics predict this heterogeneity, but the relevant predictors differ across outcomes and food box types. Predicted impacts align closely with baseline deprivation for calorie intake and dietary diversity, but less so for food insecurity experience, particularly under the staple box. Thus, trade-offs between targeting the most deprived and the most impacted households are limited. Instead, the more salient challenge is prioritizing among competing nutritional objectives. Using a multi-armed Qini curve that accounts explicitly for program costs, we show that, under a fixed budget, the estimated optimal policy assigns the nutrition-sensitive food box to most treated households despite its higher cost. The gains from covariate-based allocation are strongest for food insecurity experience and more limited for the other outcomes. Overall, our findings highlight the potential of ML-informed targeting in food aid programs, while also showing that targeting priorities depend on the nutritional objective being pursued.