New algorithm to help resettle refugees, improve their integration

Current resettlement approaches
Stanford notes that in recent years, a record number of people have been displaced as a result of war, persecution, and other human rights violations, surpassing the numbers seen after the Second World War. In 2016 alone, about 65.6 million people were forced to flee their homes, according to the United Nations’ refugee agency.

Often, countries that resettle refugees in their communities do so either somewhat randomly or according to local capacity of hosting communities at the time of refugees’ arrival. In the United States, refugees who have family members at a particular location are directed to join them there. But refugees without preexisting ties are free to be sent to various locations, and current approaches do not match them to locations where the evidence suggests it would be easiest for them to integrate.

“Our motivation was to bring the best of cutting-edge social science to an area of high policy priority that needs innovation but, because of the limited resources and challenges of navigating large numbers, has not been able to innovate from within,” Weinstein said.

The group developed their algorithm based on socioeconomic data from more than 30,000 refugees, aged 18 to 64, placed by a major resettlement agency from 2011 to 2016 in the United States. The data also included where those refugees were resettled, and their eventual employment status.

Based on this data, the team had the algorithm predict employment probability and optimal locations for a group of refugees who arrived toward the end of 2016 and compared those predictions with how these refugees actually fared in their new homes.

The group found that if the algorithm had selected locations for refugees’ resettlement, the average employment rate among those refugees would have been roughly 41 percent higher.

The team went through the same process with data from asylum seekers who had been resettled in Switzerland between 1999 and 2013. They predicted the employment rate would have been 73 percent higher among asylum seekers who arrived in 2013 if they had been assigned to the places the algorithm identified as optimal.

“The employment gains that we’re projecting are quite substantial, and these are gains that could be achieved with almost no additional cost to the governments or resettlement agencies,” said Kirk Bansak, a lead author of the study and a political science Ph.D. student. “By improving an existing process using existing data, our algorithm avoids many of the financial and administrative hurdles that can often impede other policy innovations.”

Promising results, more research needed
The researchers are not advocating for the algorithm to replace the decision-making of resettlement officials.

“Our approach preserves the ability of policy-makers to set their own parameters and priorities,” the researchers wrote. “For instance, in a computer-assisted assignment process, the algorithm might provide several recommendations, and placement officers could use their own discretion to determine the final assignment or override any suggestions.”

Yet in contrast to more expensive policy interventions, such as job or language training for refugees, the results of the algorithm, the code of which is available for free to any organization or government, are promising, the researchers said.

“The fact that we are able to generate such significant gains because of a simple change to the resettlement process is a demonstration of just how important it is to bring data-driven insights to policy-making processes,” Weinstein said.

The group said they still need to confirm the algorithm’s predictions through prospective tests that implement this approach in real time. The research team is now developing a number of pilot programs in partnership with governments and resettlement agencies to test the algorithm’s power.

— Read more in K. Bansak el al., “Improving refugee integration through data-driven algorithmic assignment,” Science 359, no. 6373 (19 January 2018) (DOI: 10.1126/science.aao4408)