What a Moonshot to End Extreme Poverty Would Cost

One of the most persistent questions in the field of global development is how much it would cost to end extreme poverty. It sounds like one of those problems economists could kick around forever, the equivalent of calculating the number of angels dancing on pinheads. But a couple of years ago, Paul Niehaus, an economist at the University of California, San Diego, had a notion that it could actually be done.

“I feel like if we can put people on the moon, we can do this, too,” he says.

The best strategy, Niehaus thought, was to give cash directly to the people who needed it. The official definition of extreme poverty is living on less than $2.15 per person per day. “It’s the simplest possible thing,” Niehaus says. “It’s literally saying that if someone is living on $1.50 and we want to get them to $2.15, they need another 65 cents to get there.”

Niehaus is the co-founder and director of GiveDirectly, a nonprofit that organizes direct cash transfers to people in need. Since it was founded in 2009, GiveDirectly has distributed $1 billion to 2 million people in seven African countries and the U.S. Its guiding principle is that people are capable of choosing for themselves what they need to improve their lives without interference from aid organizations. Niehaus believed this approach could work on a global scale.

To calculate how much it would cost to end extreme poverty, all Niehaus would have to do was get the income data for everyone in the world living below the poverty line, figure out how much it would cost to get each individual to the $2.15 threshold, and then add all those numbers together. Simple!

The only problem was that there was no way of knowing the exact incomes of the 700 million people living in extreme poverty. Most poverty gap estimates are based on household surveys that cover tiny percentages of a country’s population — between 0.001% and 0.01%. Self-reported income levels are unreliable, and there are other factors that determine wealth and need: for instance, livestock holdings, the material a house is made of, whether that house has improvements such as solar panels, and whether it’s located in a place with good infrastructure and access to resources.

Niehaus still believed it was possible to solve the problem. But he needed help with the numbers. “It required some optimization, some new math, some new statistical learning that I didn’t know how to do,” he says.

He discussed the problem with Joshua Blumenstock of the School of Information at the University of California, Berkeley, and Stanford Graduate School of Business economics professor Susan Athey, PhD ’95. Athey suggested that her colleague Stefan Wager, an associate professor of operations, information, and technology, who had experience with optimization learning problems, might know what to do. Wager consulted with Roshni Sahoo, a graduate student in computer science at Stanford whose research is about developing statistical methodology to address challenges in economic development and public health. And, says Wager, “Roshni solved it.”

Closing the Gap

Sahoo, Wager, Niehaus, and Blumenstock, along with Leo Selker of UC Berkeley, recently published their results in a National Bureau of Economic Research working paper called “What Would It Cost to End Extreme Poverty?”

According to Sahoo’s calculations, reducing the world’s poverty rate to 1% would cost 0.3% of the global GDP, roughly $318 billion per year. That may sound like a lot, but, as the researchers point out, the world spends seven times that on alcoholic beverages annually. In contrast, another popular proposal, providing universal basic income at the poverty line, would cost $895 billion per year.

To get those numbers, Sahoo began with household survey data from 23 countries that account for half of the world’s poorest people. That data could be used to estimate the aggregate poverty gap, the total amount of cash needed to lift each household above the poverty line for a year — if the researchers knew everyone’s exact income. Yet Sahoo decided to go beyond poverty gap estimation by using a method called policy learning to estimate the cost of a more realistic transfer program operating without perfect information on recipients’ income.

“The big question is how we can use the household survey data in order to learn ways of allocating cash to individuals that still provide poverty reduction guarantees,” she says. “And that’s where the optimization and statistical learning come in.” Her work built on a statistical method developed by Athey and Wager that incorporates observational data and real-world constraints to determine who should benefit from a certain policy. To scale this approach to 23 countries, Sahoo built software that takes household survey data and produces cost estimates for a variety of cash-transfer policies. This engineering solution made it possible to compare different approaches, ranging from universal basic income to targeted transfer programs.

Sahoo’s calculations took into account not only household income but also standards of living. A family that lives in a house with a tin roof, for example, would be considered better off than a family with a thatched roof, even if they bring in the same amount of money. Part of the goal of eradicating poverty is giving people the resources to improve their lives, not just continue to exist at the same level. So if Sahoo’s model were applied to the two hypothetical families, the family with the thatched roof would receive more money than the family with the tin roof because they have a lower standard of living.

A Moral Perspective

“There’s integration of a moral perspective on this,” Niehaus explains. “To me, that’s one of the most exciting aspects of research like this, that there is this integration of the cutting-edge data science with an explicit representation of what’s ethically important.”

“I’ve been working in the space of large-scale data-driven policy learning for a while,” Wager says. “Sometimes you get the impression that the main application area for these methods is ad targeting. But our results here highlight how, given access to high enough quality data, policy learning methods can also help make a difference in public-interest settings.”

The researchers emphasize that their paper is still a work in progress; they want to add more countries to the study and calculate the macroeconomic effects of deploying these policies at a larger scale. Yet they’ve already started to share the results, including with a few of Niehaus and Blumenstock’s acquaintances in Washington. While the Trump administration has slashed U.S. funding for international development, the researchers see scope for a renewed commitment in the future.

Niehaus also finds it satisfying to have a definitive answer whenever people ask him how much they should donate to alleviate extreme poverty: 0.3% of their annual income. For a typical American earning $45,000 per year, that amounts to $135.

“The reaction I’ve had so far,” Niehaus says, “is ‘My gosh, that’s so little, I had no idea.’ And maybe a little excitement and horror that that’s all it would take and we’re not doing it.”