Research shows that machine learning can improve the way we place refugees in the United States.
While the Beltway argues over whether President Donald Trump called Haiti a “shithole” or a “shithouse”—or whether the outburst in question is simply more Fake News—the global refugee crisis continues more or less unabated.
In 2017, the number of forcibly displaced persons around the world soared past 65 million, a figure that includes both refugees—those living outside their home countries—and the internally displaced. Every day, according to the United Nations High Commissioner for Refugees, more than 28,000 people are forced to flee their homes.
A recently published study in Science offers a 21st-century solution to an age-old question: where best to settle refugees upon their arrival in a new country.
Jens Hainmueller and his colleagues at Stanford University’s Immigration Policy Lab used machine learning and reams of historical data to design an algorithm that optimizes a refugee’s chance of finding employment in her new home.
“Refugees are among the world’s most vulnerable populations,” the authors write. “After experiencing war, violence, and years of living in overcrowded refugee camps, refugees arrive in a new country with few resources and must acclimate to an unfamiliar local language, economy, and culture. Refugees frequently remain economically marginalized, with low levels of employment in the years following their arrival.”
Hainmueller and his colleagues built an algorithm that studied historical settlement data—including characteristics such as country of origin, language skills, sex, and age—and “learned” where particular newcomers fared best. The researchers trained their model on data from 33,000 working-age refugees from 2011 to 2016. They then tested it on 900 refugees who arrived at the end of 2016.
The results were impressive. “The median refugee’s predicted probability of employment in the United States more than doubled, increasing from approximately 25% to 50%,” the authors write. “Our optimized assignment increased the probability of finding employment across the entire distribution of refugees, including those who were least likely and most likely to find work.”
On average, 34 percent of refugees in the test group found work within 90 days of settling in the U.S. Using an algorithm to guide the settlement process, the researchers predict, that figure would climb to 48 percent—a 41 percent improvement.
A job isn’t a silver bullet, but it’s a start. In addition to economic stability, it provides community and self-sufficiency, even purpose. But not every city or town offers a refugee the same odds of finding work.
THE SYRIAN MIGRANT CRISIS YOU’VE NEVER HEARD OF: AS MILLIONS OF SYRIANS ARE FORCED TO FLEE THEIR HOME COUNTRY, THE DESCENDANTS OF EARLIER MIGRANTS ENJOY A LIFE OF CULTURAL AND ECONOMIC ASSIMILATION.
“A particular place could be an onramp or an obstacle,” Hainmueller tells Pacific Standard. Denver, for instance, might have a thriving Afghan community, making it easier for refugee from Kabul to find work there, while Los Angeles has no such network. Or a town in Alabama might have a meatpacking plant that readily employs young people, but which rarely hires older workers. “Many of these factors matter,” Hainmueller says. “The key is, they matter differently in different locations.”
On an anecdotal level, this seems obvious. In Switzerland, for example, African refugees from Francophone countries are 40 percent more likely to find work when they’re settled in the French-speaking part of the country rather than the German-speaking part. But trying to catalog these synergies in an exhaustive way is a complex process of trial and error, hideously inefficient for a human to tackle.
“Think about it as a matching problem,” Hainmueller says. “If you have three families and three locations, you can do it with a paper and pencil. In the real world, there’s thousands of refugee families and hundreds of locations. It’s basically impossible for a case work or placement officer, using an Excel sheet, to come up with the optimal matching.”
(Photo: Spencer Platt/Getty Images)
At present, refugee settlement in this country is a basically random process. If a refugee doesn’t have pre-existing ties—family, friends, or an official sponsor—they are assigned to one of nine voluntary resettlement agencies in a weekly lottery. From there, they’re typically sent to whichever local office has the most capacity at the moment. According to the authors, “Placement officers make assignment decisions prior to refugees’ arrival and without interviewing the refugees.”
Hainmueller envisions a future where, rather than replacing them altogether, computers simply help their human counterparts make better decisions. For instance, an algorithm could suggest the three most optimal settlement locations for a refugee family, based on its particular characteristics, but a case worker, drawing on her experience and professional judgment, would make the final call.
Because machine-learning algorithms are flexible, they improve with time. “Say Afghans no longer do well in Denver because their networks stop being effective or the labor market is saturated,” Hainmueller says. “Once the data comes in, the algorithm will figure out, ‘Oh, Afghans are no longer doing well there.’ But they might start to do well in another place. There’s immediate feedback built into the system that’s absent from the process right now.”
“The world is facing the most significant refugee crisis in the post-war period,” Hainmueller says. “The U.S. has been a leader in refugee resettlement, but that leadership had waned somewhat recently. Refugee policies, like immigration policies generally, are dominated by ideology rather than sounds evidence. We haven’t seen a lot of innovations in this space. Cash assistance, language instruction, training programs: These turn out to be very expensive and difficult to scale.”
“The nice thing is, from a policy perspective, this doesn’t really cost you anything more,” he continues. “It’s just a smarter way of doing the allocation. Rather than doing it in a haphazard, quasi-random fashion, as we’re doing it right now, we might as well do it in a more data-driven way, where we send people to the places they’re more likely to succeed.”