Abstract: Using Deep Learning to Evaluate the Distribution of Community Development Aid (Society for Social Work and Research 25th Annual Conference - Social Work Science for Social Change)

All live presentations are in Eastern time zone.

Using Deep Learning to Evaluate the Distribution of Community Development Aid

Thursday, January 21, 2021
* noted as presenting author
Woojin Jung, Ph.D., Assistant Professor, Rutgers University, New Brunswick, NJ
The effectiveness of aid distribution is an ongoing problem for policymakers, particularly in fragile and conflict-prone states. Our research focuses on aid distribution in Myanmar, a large and diverse country that has struggled with ethnic conflict. Since 2012, there has been a surge of community development aid. However, a critical question remains: Is aid is being distributed to the areas of most need in a way that reduces inequality?

Building upon the literature, this research explores remote sensing techniques to measure poverty in data-sparse contexts. To the best of our knowledge, the performance poverty estimations has not been tested in the sub-national aid determinant context. Our study of Myanmar examines i) the performance level of different methods of poverty estimation and ii) the extent to which poverty and other development characteristics explain community aid distribution.

This study draws from following sources of data: web-scraped community-driven development projects in 2015 (N=12,504, the most recent georeferenced data available), daytime and nighttime satellite imagery in 2015 (50,000 images with a size of 400 by 400 pixels), the Demographic and Health Survey in 2015-2016 (441 village clusters), and conflict event data in 2010-2019 (n=6,455). We first compare the accuracy of four poverty measures in predicting ground-truth survey data. Using the best poverty estimation in the first step, we perform linear and non-linear regression to investigate the association between village characteristics and aid per capita per village.

Our results show that daytime features perform the best in predicting poverty (4.88) as compared to the analysis of RSG color distribution (0.38), Kriging (0.21), and nighttime-based measures (0.18). In our best model, we use a Convolutional Neural Network, pre-trained on ImageNet, to extract 4,097 features from the satellite images. These features are then trained on the DHS wealth data to predict the DHS wealth index for villages receiving aid. Using the best poverty prediction, the OLS estimator reports that aid is more likely to be disbursed to those villages that are less populous and farther away from fatal conflict events. A 100 km increase in the distance to the nearest fatal events is associated with a 7.8% point increase in aid per capita per village. Aid flows to low-asset villages, but only marginally. When using a non-linear, Random Forest classifier, both distance to conflict and predicted wealth are identified as the two largest variables of importance, each of which explains approximately 10% of the variation in residuals.

The performance of our model, in comparison to the literature on small Africa countries, suggests some promises as well as challenges. A relatively small DHS dataset that covers the large size of Myanmar and typological variation both seem to lower the model’s performance. However, the fact that simple Ridge regression alone could correlate image features with the wealth data is a promising indication. Our study implies that policymakers can enhance aid targeting without much computational complexity by 1) expanding ground truth data (by merging administrative and survey data) and 2) pairing them with a fine-tuned deep learning model.