Analytics for Refugee Resettlement

Ahani, Narges

Etd

Analytics for Refugee Resettlement

Public

Every year tens of thousands of refugees are resettled to dozens of host countries. While there is growing empirical evidence that the initial placement of refugee families profoundly affects their lifetime outcomes such as employment, there have been few attempts to optimize resettlement decisions. This dissertation is about the use of analytics in refugee resettlement. We leverage machine learning, matching theory, integer optimization, stochastic programming, risk modeling, and interactive visualization to improve the refugee resettlement decision-making process and benefit involved stakeholders including refugees, communities, and resettlement agencies. First, recognizing that specific synergies exist between refugee characteristics and resettlement communities that affect refugee employment outcomes, we use machine learning to train predictive models on past refugee placement and outcome data to estimate employment likelihood of refugees in communities. These estimated values are then used as refugee-community match quality scores in an integer optimization model for optimal matching of to-be-arriving refugees into the network of communities. We implemented our analytical approaches into an innovative, interactive refugee resettlement decision support software, Annie™ Moore, that assists HIAS, one of the nine U.S. resettlement agencies, with matching refugees to their initial placements by providing optimized, data-informed recommendations. Our software suggests optimal placements while giving substantial autonomy to the resettlement staff to fine-tune recommended matches through interactive visualization, thereby streamlining their resettlement operations. Back-testing indicates that Annie™ can improve short-run employment outcomes by 22%–38%. Second, we consider the dynamic nature of refugee resettlement. While refugees arrive weekly, capacities are assigned to communities on an annual basis. By only allowing resettlement staff to manually set weekly capacities, this may result in consuming annual capacities in an overly greedy or conservative manner. In other words, allocating the weekly batches of arriving refugees with arbitrary weekly community capacity settings leads to sub-optimal total annual employment, as each batch of arrivals is allocated by separately maximizing the expected employment of this batch and without considering future arrivals. While the optimized value for total annual expected employment can only be realized when perfect information is available for all future arrivals, to better approximate this hindsight optimal employment we introduce a dynamic allocation system based on two-stage stochastic programming to improve employment outcomes. Our algorithmic approach places refugees to communities using not only employment probabilities, but also estimates of the value of the remaining slots of capacity for each community, leveraging this critical information on whether each slot is more useful for placing the current refugee or a yet-unknown refugee arriving later in the year. This algorithm is able to achieve over 98 percent of the hindsight-optimal employment compared to less than 90 percent for existing myopic approaches. We incorporated our dynamic placement algorithm into Annie™. Third, we consider a new extension to our earlier optimization model to account for risk. We recognize that inherent error exists with respect to the estimation of employment probabilities. This results in uncertainty with respect to expected optimized outcomes for refugees, that is, employment likelihoods for optimized refugee-community placements. Directly related to this uncertainty, we introduce the concept of risk in refugee resettlement optimization. Although numerous studies exist on risk in the context of optimization, the related literature largely interprets risk as expressed by uncertainty around total expected objective function value. Considering that the expected outcome of each refugee family—the employment likelihood at its optimal placement—is just as important as the total expected employment from maximizing all refugee-community placements, we provide an alternative family-level definition of risk that properly accounts for vulnerability of refugees and is useful in the context of refugee resettlement. We seek to mitigate this alternative definition of risk from an optimization point of view. Our modeling approach explicitly incorporates family-level risk into the formulation by accounting for both the total expected outcomes as well as risk related to placement outcome uncertainty for refugee families. To hedge against the risk, we weigh the trade-off in lower expected outcomes associated with less risk, generating optimal solutions that satisfy risk-averse decision makers. Optimizing this trade-off presents significant modeling and computational challenges. Multiple optimization models are proposed and analyzed, and specific measures are developed to quantify the change in risk. We discuss on the functionality of these models and provide experimental results to illustrate their performance. Our results show that risk-averse optimization models can alleviate much of the risk while retaining much of the total expected employment.

Creator