| Sumario: | In-kind food aid programs remain prominent world-wide. Targeting in these programs is complex due to potential distortions in consumption. This paper advances the literature by moving beyond poverty-based targeting to address nutritional objectives. Using data from a randomized controlled trial (RCT), we apply machine learning (ML) techniques to analyze heterogeneity in impacts across nutritional outcomes, aiming to inform targeting based on observable characteristics. We find that such characteristics significantly predict heterogeneity in treatment effects, though relevant predictors differ by outcome and treatment type. Building on recent literature advocating for balancing of deprivation and expected impact, we show that, in our context, the trade-off between targeting the most impacted versus the most deprived households is limited. Instead, the main challenge is prioritizing among competing nutritional objectives. Our findings indicate that ML methods can inform outcome-specific targeting criteria, though these criteria vary across outcomes and are imperfectly correlated.
|
|---|