Nils Olberg | Searcharxiv

arXiv2022

Enabling Trade-offs in Machine Learning-based Matching for Refugee Resettlement

The Swiss State Secretariat for Migration recently announced a pilot project for a machine learning-based assignment process for refugee resettlement. This approach has the potential to substantially increase the overall employment rate of refugees in Switzerland. However, the currently proposed method ignores families' preferences. In this paper, we build on this prior work and propose two matching mechanisms that additionally take families' preferences over locations into account. The first mechanism is strategyproof while the second is not but achieves higher family welfare. Importantly, we parameterize both mechanisms, giving placement officers precise control how to trade off family welfare against overall employment success. Preliminary simulations on synthetic data show that both mechanisms can significantly increase family welfare even with only a small loss on the overall employment rate of refugees.

cs.GT↗

arXiv2021

Search and Matching for Adoption from Foster Care

To find families for the more than 70,000 children in need of adoptive placements, most United States child welfare agencies have employed a family-driven search approach in which prospective families respond to announcements made by the agency. However, some agencies have switched to a caseworker-driven search approach in which the caseworker directly contacts families recommended for a child. We introduce a novel search-and-matching model that captures the key features of the adoption process and compare family-driven with caseworker-driven search in a game-theoretical framework. Under either approach, the equilibria are generated by threshold strategies and form a lattice structure. Our main theoretical finding then shows that no family-driven equilibrium can Pareto dominate any caseworker-driven outcome, whereas it is possible that each caseworker-driven equilibrium Pareto dominates every equilibrium attainable under family-driven search. We also find that, within our model, when families are sufficiently impatient, caseworker-driven search is better for all children. We numerically illustrate that most agents are better off under caseworker-driven search across a wide range of parameter values. Finally, we present an empirical study of an agency that switched to caseworker-driven search, finding a three-year adoption probability that outperformed a statewide benchmark by 44.9%, along with a statistically significant 54% higher adoption hazard rate.

cs.GT↗