Dynamic Placement in Refugee Resettlement

TL;DR

The paper tackles how the placement of resettled refugees within host-country communities affects employment outcomes. It develops two dynamic allocation frameworks—two-stage stochastic programming and a Walrasian-equilibrium-inspired approach using shadow prices—to optimize job prospects across batches of arrivals, including non-unit family sizes and batching. On HIAS data from 2014–2019, the methods reach over 98% of hindsight-optimal employment, far surpassing greedy baselines around 90%, and are implemented in Annie™ Moore with a human-in-the-loop interface for robust real-world deployment. The work also analyzes uncertainty in total arrivals, proposes robust trajectory-based priors for future arrivals, and demonstrates practical performance with batching and non-unit cases, making the approach scalable for large-scale refugee resettlement operations.

Abstract

Employment outcomes of resettled refugees depend strongly on where they are placed inside the host country. Each week, a resettlement agency is assigned a batch of refugees by the United States government. The agency must place these refugees in its local affiliates, while respecting the affiliates' yearly capacities. We develop an allocation system that suggests where to place an incoming refugee, in order to improve total employment success. Our algorithm is based on two-stage stochastic programming and achieves over 98 percent of the hindsight-optimal employment, compared to under 90 percent of current greedy-like approaches. This dramatic improvement persists even when we incorporate a vast array of practical features of the refugee resettlement process including indivisible families, batching, and uncertainty with respect to the number of future arrivals. Our algorithm is now part of the Annie MOORE optimization software used by a leading American refugee resettlement agency.

Figures (24)

• Figure 1: Total employment obtained by different algorithms, assuming that cases are split into multiple cases of size $1$. Capacities are the final capacities of the fiscal year. For the potential algorithms, total employment is averaged over 10 random runs. The numbers in the bars denote the absolute total employment; the bar height indicates the proportion of the optimum total employment in hindsight.
• Figure 2: Evolution of the per-refugee match score in order of arrival, for fiscal years 2016 and 2019 in the experiment of \ref{['fig:unit']} (split cases, final capacities). Consecutive match scores are smoothed using triangle smoothing with width 500.
• Figure 3: Remaining priced capacity at the time of arrival of different refugees, for fiscal years 2016 and 2019 in the experiment of \ref{['fig:unit']} (split cases, final capacities).
• Figure 4: Total employment, where cases are not split and arrive in batches. Capacities are the final fiscal year capacities. In contrast to \ref{['fig:unit']}, cases are treated as indivisible, cases arrive in batches, and the batching variants of greedy and the potential algorithms are used. For the potential algorithms, the mean employment across 50 random runs is shown.
• Figure 5: Distribution of the total employment obtained by instantiating PMB with different potential methods and different $k$, in the experiment of \ref{['fig:full']} (whole cases, batching, final capacities) and over 50 random runs per algorithm.