How GiveDirectly is finding the poorest people in the world—and sending them cash

The economic effects of COVID-19 have drastically driven up the world’s extreme poverty level. The World Bank estimates that the number of people living on less than $1.90 per day will reach 150 million by 2021. GiveDirectly, a charity that has focused for just under a decade on direct cash transfers to people in poverty around the world, particularly in Africa, has been escalating its pandemic relief efforts—and continually innovating with partners to find groundbreaking ways to target the most in need of money. The charity’s latest innovation is harnessing an algorithm, designed by UC Berkeley, that uses artificial intelligence to identify the poorest individuals in the poorest areas, and transfer cash relief directly to them.

Typically, in order to evaluate whom to send money to, GiveDirectly will use poverty data from national surveys, and enroll all the households in a particular area. If it needs to target more narrowly, it will depend on lists from governments, NGOs, and local organizations; use points-based poverty indexes; or rely on subjective assessments. This new initiative allows the targeting to be faster and more accurate, completely contactless (which is vital during the pandemic), and naturally adapting and evolving as data changes over time.

The algorithm works in two stages, using two distinct data sources. First, it finds the poorest neighborhoods or villages in a certain region, by analyzing high-resolution satellite imagery. The tool identifies those areas from hundreds of poverty markers that distinguish poorer from wealthier places, such as roof material, building density, sizes of farm plots, and paved or unpaved roads.

Once the geography is set, the second stage is finding the poorest individuals within those areas, by analyzing their mobile phone data, provided by Togo’s two principal carriers, Togocel and Moov. It distinguishes between richer and poorer folks using clues for more expensive or cheaper usage, like lengths and frequency of phone calls, number of inbound versus outbound calls, and amount of mobile data used. After the poorest individuals are identified, they will be prompted to enroll via mobile phone, and then instantly paid. “The aim is to pick out as many extremely poor as possible, as quickly as possible,” says Han Sheng Chia, special projects director at GiveDirectly.

The project launched in November in Togo, one of the poorest countries in the world, where it’s estimated that 55% live below poverty line. That number is closer to 81% among the rural population, who are the focus of this pilot. In April, in response to COVID-19, Togo’s government established an innovative cash transfer system called Novissi, spearheaded by Cina Lawson, Togo’s Minister of Postal Affairs and Digital Economy. This allowed the government to send cash relief via mobile to approximately 12% of the population. So, an infrastructure already exists, but they wanted to expand it to more rural areas, where it’s harder to pick out the most in need without the right technology.

Josh Blumenstock, associate professor at the UC Berkeley School of Information, first wrote the paper on mapping poverty using satellite and mobile data in 2017, before meeting Chia to discuss application. Blumenstock stresses that they are not telling the tool what to look for; rather, machine learning enables it to recognize patterns itself. In order to train it on what to look for, they surveyed a large sample of 15,000 citizens across the poorest 100 of Togo’s cantons, on which the government wanted to focus, asking them “a rich set of questions about their socioeconomic status,” including on their income and spending, and whether they’d missed meals in the last week. They paired this with their mobile phone data to help the algorithm find patterns, which it could then scale up to the entire populations of those cantons.

Once identified, eligible citizens get a text notification to enroll; after a few clicks, cash is transferred directly to their phones using mobile money technology. (They can then withdraw cash at local shops.) Approximately $5 million will be delivered in total, sending cash every month for five months, in the sum of $15 for women and $13 for men per month, which they’ve calculated as the figure to cover their “minimum basket of goods” to survive. So far, 30,000 Togolese have been paid, out of a goal of 58,000, or the poorest 10% (it’s limited by funds available from GiveDirectly, which is getting financing largely from TED’s Audacious Project). GiveDirectly chose to donate more to women, Chia says, since they are the primary caregivers, and as a way to incentivize higher participation among a group that’s been historically left out of such social protection programs.

After five months, the teams will analyze how effective the pilot was, and tweak it as necessary. Blumenstock admits it isn’t totally robust yet: For instance, it may be that the algorithm excludes some people through biases, even despite thorough training. It also excludes people who don’t have cellphones. Ninety percent of Togo’s rural population does have one phone per household; the team is also tying registration to SIM cards, not phones, so multiple family members can purchase a cheap SIM and link it to the family phone. The team has also tried to reduce privacy issues as much as possible, by minimizing the data shared and accessed, and through anonymizing and encrypting it. “It’s not as perfect as one would like it to be,” Blumenstock says. “To the extent that we’re sort of stepping on privacy concerns, it really is necessary to make the program work.”

If funding is secured, they can then scale up the project in Togo or in other countries, such as Bangladesh and Nigeria, with whom Chia says they’ve been in talks. He hopes governments and NGOs are able to use the tool in the future, and the system could prove to be a model for “pre-positioning” relief ahead of disasters and humanitarian crises. “We don’t think of this as the only mechanism to provide cash and social protections to needy people,” Blumenstock says. “But it is an effective way to get cash out really quickly, to a lot of people.”