Adding to an already overwhelming global poverty crisis, the pandemic increased the need for humanitarian aid distribution. In December, the UN appealed for $35 billion to be allotted to the 160 million in need around the world. With limited resources, and some countries cutting assistance, it’s important that the very poorest receive funds first. But need is incredibly hard to assess and usually relies on a combination of survey data that’s only sporadically collected, along with a rather limited well of geospatial data. The targeting that results is often indefinite and overly broad.
Improving that pinpointing is the goal of a collaboration between UC Berkeley’s Center for Effective Global Action and Facebook’s Data for Good, the tech giant’s policy branch. The joint effort is producing extremely granular “micro-estimates” of socioeconomic status, down to 2-kilometer-by-2-kilometer squares, for 135 low- to middle-income countries. Each grid square contains a measure of absolute wealth, or the average wealth of people in that area, in dollar terms, and of relative wealth, compared to other areas in the same country. The new model—called the Relative Wealth Index—will be freely available to nonprofits and governments as they decide how to distribute cash assistance to the developing world. “The more granular you’re targeting, the more likely it is that more benefits will go to the poorer people than wealthier people,” says Joshua Blumenstock, associate professor at UC Berkeley, who’s leading the initiative there.
The current primary way to assess poverty levels is by household surveys: in-person questionnaires conducted on the ground level, funded by USAID and other charities. That forms the “gold standard data set for determining where poor people live,” says Laura McGorman, policy lead for Facebook Data for Good. But these expensive surveys are only conducted about once every decade and only cover a small percentage of households per country: For instance, the most recent iteration only surveyed households in 13.8% of Nigerian wards, the smallest administrative municipalities in that country. That presents two main problems: Data remain scarce and too high-level to be effective.
What’s novel about the Relative Wealth Index approach is that it combines publicly available survey data with nontraditional, predictive data, including high-resolution satellite imagery, topographic maps, mobile network data, and connectivity data, much of which Facebook has collected. “It’s a bit of a kitchen-sink approach,” says Blumenstock, adding that any data that has geospatial markers is helpful for predicting relative wealth. For instance, satellite imagery can show population densities closer or further away from roads and infrastructure, which would suggest greater and less relative wealth, respectively. A region with fewer smartphones and Wi-Fi connectivity would suggest less relative wealth, as would certain cell data trends such as lower frequency and length of calls, and less volume of data used.
Together, both sets of data train a machine learning model that predicts the absolute and relative wealth of each 2.4-km-squared grid cell in 135 countries (there are about 19.1 million). That shows the extreme local variation that wasn’t possible before. “Depending on if you’re in Village A in Pakistan, or City Y in the DRC, you’re going to have individual observations that are very specific,” McGorman says. “Crazy specific, down to the grid tile level.” (The survey data is only available for 56 countries; for the rest, they use just the nontraditional data, so measurements may be slightly less accurate. The data is presented along with accuracy confidence levels.)
This new level of granular data, and the “fine-grain maps” generated, are aimed at policymakers planning cash assistance strategies. They’re already being used by the Nigerian and Togolese governments. In Nigeria, authorities now know rough poverty levels for 100% of the country’s wards. Similarly, in Togo, the data is down to the canton level. In the Togolese case, UC Berkeley has been working with the government and cash relief nonprofit GiveDirectly, to distribute cash to the poorest residents via mobile money.
Berkeley doesn’t necessarily have the capacity to work with each potential partner. The idea is for governments and organizations to be able to use the data “off the shelf” for their local contexts, accessible in raw data formats, or via interactive map visualizations. Institutions such as the World Bank, the IMF, and USAID should also benefit. “The development banks are likely to get the most juice out of this in their foreign assistance approaches,” McGorman says. That’s especially true during the pandemic, when relief is urgent; COVID-19 spurred the acceleration of this project, which has been four years in the making.
There’s one more major, pressing use case: COVID-19 vaccines. In the past, nontraditional data such as population density maps have aided vaccine distribution and awareness. In 2016 and 2017, 3,000 Red Cross volunteers, on the ground in Malawi for a mass education campaign about measles and rubella vaccines, were able to accelerate their work by efficiently filtering out the 97% of terrain that was uninhabited, as indicated by AI-powered maps. Now, as COVID-19 vaccines will likely be in short supply as they roll out to developing countries, the Relative Wealth Index—”not only now showing where people live, but where the poorest of the poor live,” McGorman says—may assist in providing an equitable distribution, ensuring the shots get to the poorest and most afflicted first.