As part of our “Research Bites” series, in which we ask data science researchers to spend five minutes telling us in their own words about their work, and opportunities for practical applicability in the context of sustainable development or humanitarian action, today we hear from researchers Kush R. Varshney from IBM Thomas J. Watson Research Center, Brian Abelson from Enigma and Joy Sun from GiveDirectly.
Targeting Direct Cash Transfers to the Extremely Poor
The issue of how to target aid and relief to those that need it most remains of key concern for NGOs and development practitioners. This research used publicly available satellite images of Uganda and Kenya to analyze the roofing materials in villages – using them as a proxy for relative poverty. Information regarding whether a village has mostly thatched or metal roofs was used to target cash transfers, optimizing the efficiency of aid relief. Researchers created a poverty heat-map that could be used by GiveDirectly, an NGO, to better target resources where they are needed most. By aggregating poverty estimates at the village level, researchers produce a ranked list from which top villages are included in GiveDirectly’s planned distribution of cash transfers. The automated approach increases village selection efficiency significantly.
The full research paper can be accessed here.
1. Tell us about your research paper in two sentences
Brian and I worked with Joy and the rest of the GiveDirectly team to develop an automated way to select the poorest villages in a part of rural Kenya to receive unconditional cash transfers. Since village-level poverty data does not exist in Kenya, we relied on the observation that households with thatched roofs are poorer than households with metal roofs and that roof type can be differentiated in satellite images. We used image processing and machine learning techniques to estimate the proportion of thatched roofs in each village from satellite images of the region.
2. Why do your findings matter?
GiveDirectly wants to transfer donations to recipients as efficiently as possible—to date, more than 90¢ of each donated dollar has reached a recipient in Kenya or Uganda. Our work aims to improve on that (already high) efficiency.
3. How could this research be put into practice?
GiveDirectly piloted the approach we came up with in February to select 50 villages in Kenya to receive unconditional cash transfers totaling $4 million. The data-driven approach saved approximately 100 person-hours of manual effort. In the future, our algorithm needs to be tuned to account for the fact that not all structures with roofs are households. In this part of the world, it is common for kitchens to be separate structures and for there to be sleeping houses that dependent sons live in once they reach puberty, both of which should not be counted separately from the main house. Quantifications of poverty level are not only useful for GiveDirectly, but for many other NGOs as well.
4. Why did you select this topic to research?
Both Brian and I volunteer with DataKind to provide pro bono data science services to NGOs. Our DataKind colleagues were the ones who connected us to GiveDirectly, assembling the team of expert volunteers for the project and providing support throughout. The project was conducted under the auspices of DataKind’s flagship DataCorps program, which partners top data scientists with social organizations to collaborate on cleaning, analyzing and visualizing data for social good. You can read more about the Datakind collaboration here.
The researchers will be on hand to answer questions in the comments section below, so we invite practitioners from the development or humanitarian sectors to join the discussion!
Kush R. Varshney @krvarshney
Brian Abelson @brianabelson
Joy Sun @joysun