To Help The Poorest Of The Poor, First You Have To Find Them
In the quest to help the poor, it's difficult to know whose needs are the greatest. Without clear data, it's tough to know who to help first.
The traditional way to look for the poorest of the poor is with household surveys. They are the primary source of data for policy decisions, but they have drawbacks.
"Household surveys are expensive, and the coverage is not great," says Stefano Ermon, an assistant professor of computer science at Stanford University.
So Ermon is heading up a new project to create a method of gathering poverty data using satellite images — which eventually could help policymakers make more informed decisions.
The use of satellite images isn't new. It's the emphasis on daytime pictures that's innovative.
Nighttime pictures can reveal some information about consumption, the "use of goods and services by households." The theory is, if people could afford a high level of consumption, then people could afford to use a lot of electricity. So the nighttime satellite image of the region would be bright. On the other hand, if the region is dark, then people in the region presumably cannot afford to use much electricity. This would mean a low level of consumption in the region.
The problem with nighttime images is that "they don't show much gradation in the poorest parts," Ermon says. In other words, these images don't distinguish poor regions from very poor regions — which is where some policymakers may be aiming their limited resources.
So Ermon and his team developed a computer program to analyze daytime satellite images — just as available as nighttime images but with more visual details. Ermon says his new model goes beyond looking at just differences in electricity use, also analyzing roads, houses and features of the terrain.
Nobody, not even Ermon, completely understands how the new model extracts economic data from geographic information. "The model is extremely complicated [and] we don't know exactly what the model is looking for in the images," Ermon says.
All we know for now is that the model works pretty well.
Indeed, a paper Ermon and his team published in Science this summer states that their program can "identify image features that can explain up to 75 percent of the variation in local level economic outcomes." The paper states that Ermon's new method outperforms nighttime imagery analysis with a "81.2 percent increase."
"What the Stanford team has shown is that daylight beats night lights," wrote Justin Sandefur, a senior fellow at the Center for Global Development, in a post on the center's blog. "This is a cool technique that I'm sure will (and should) get a lot of use among researchers, and will hopefully catalyze further refinements to the approach."
The computer program analyzes satellite images alongside the survey data from five countries where household surveys have been conducted recently — Malawai, Nigeria, Rwanda, Tanzania and Uganda. The program picks out connections between visual features of images and numerical data from the survey.
Sandefur, who helped design the Tanzania survey the Ermon group used, supports the group's work. But he cautions that it's not ready to act as the sole basis for policy decisions. His quantitative analysis of the group's work finds that the group "gets the poverty status right 60 percent of the time," which is not much better than random guessing.
"The accuracy is just not nearly refined enough to target a social program," Sandefur says. "I could guarantee you urban legends would spread about what would make you look poor from space."
So the day of replacing door-to-door surveys with observations from the outer space has yet to come.
Nonetheless, the World Bank is receptive to the idea of incorporating the data.
"We see satellite images as a complement [to household surveys], not a substitute," says David Newhouse, a senior economist at the World Bank. "Putting the two together would make a lot of sense."
Then again, the point of the Ermon group's research is to shed light on regions with little to no survey data.
"If it's that information-poor, then we might as well use this," says Jonathan Drake, an imagery analyst at the American Association of the Advancement of Science. "I think it shows considerable promise."
"We've received a large number of request for data from both academic, government and non-governmental organizations," Ermon says. While not everyone thinks Ermon's project provides accurate enough data for policymaking purposes, Ermon says his method does "better than existing methods." His group plans to work on the project, investigating the "practical benefits of using [their] new poverty data."
Copyright 2020 NPR. To see more, visit https://www.npr.org.