Using Artificial Intelligence/machine learning to evaluate the distribution of community development aid across Myanmar

Achieving global poverty alleviation goals requires a systematic allocation of resources, particularly at the subnational level. However, assessing the pro-poor nature of development efforts is challenging without community-level poverty data. In the context of Myanmar, our study presents granular methods to estimate poverty, examine targeting, and predict aid distribution based on village-specific attributes. We evaluate multiple poverty estimation methods, leveraging both daytime and nighttime satellite imagery along with geofeatures. Daytime image features, when processed with convolutional neural networks (CNN), provide the most accurate poverty estimates. Using this refined poverty metric, we evaluate the targeting error and deploy machine learning (ML) techniques to predict the block grant size each village receives for community development. Findings show that a majority of beneficiary villages have predicted wealth above the median, resulting in high targeting errors. While impoverished villages tend to receive more grant aid per capita, wealth is not a primary factor. Instead, village capacity and state/ethnicity attributes hold more sway. The study highlights the need for an increased poverty-centric approach in community-based interventions and calls for more transparent aid allocation practice in Myanmar with potential implications for other conflict-prone countries.