Kirk Bansak - Algorithmic Assignment of Refugees and Outcome Choice Under Distribution Shift

The algorithmic geographic assignment of refugees and asylum seekers is an area of algorithmic decision-making that has seen an increasing amount of research in recent years and has been implemented on a limited basis in the United States and Switzerland. The basic idea behind this approach is to use data on prior refugee/asylum-seeker arrivals to generate machine learning models that can then be used (along with optimal assignment methods) to match families to those locations within a host country where they are most likely to thrive, according to specific metrics of integration success. In considering and implementing algorithmic refugee assignment, policymakers must choose their metric of success, such as a measure of employment a certain number of years after arrival. This paper considers the problem of building the underlying models in the presence of distribution shift. In particular, it considers how shorter-term versions of the outcome of interest can be leveraged to generate better models. The paper formalizes, proposes, and compares several alternative strategies for this purpose, and establishes conditions under which the different strategies are superior to one another. The paper also presents empirical evidence from real-world data on asylum seekers in the Netherlands. This application demonstrates that the problem posed in this study is not merely a theoretical novelty, but rather a real-world practical issue for consideration in implementations of refugee/asylum-seeker algorithmic assignment (and more broadly by designers of algorithmic decision-making systems in general).