Risk-Averse Placement Optimization in Refugee Resettlement

The intentional use of analytical modeling for refugee resettlement is an emerging area of research. One primary stream uses machine learning to estimate future refugee employment likelihoods based on past refugee placement and outcomes data, that are then integrated as match quality scores within an optimization framework to identify optimal refugee-community matches. While these scores serve as necessary point estimates, inherent uncertainty exists with respect to match quality score estimation|uncertainty that can lead to suboptimal resettlement outcomes, thereby putting vulnerable refugees at a disadvantage when it comes to accessing scarce employment opportunities. We mitigate this uncertainty by explicitly factoring in risk into the placement optimization, including a new form from the perspective of individual refugee families. We differentiate our family-level risk measures from classical measures that emphasize risk in the aggregate form and demonstrate the effectiveness in risk reduction while retaining most of the total expected employment. We further explore fairness considerations, advocating that the most vulnerable of refugees are at least as well off as others. Computational results on actual refugee data demonstrate the effectiveness of the proposed optimization framework that embeds family-level risk aversion, generating optimal placements with less employment volatility. Key managerial insights highlight important factors available to improve operational decision making for risk-averse refugee resettlement.