AN ALGORITHM CHOOSES WHERE REFUGEES START

Refugees often struggle to find work in their host country. In twenty European countries, the paper notes, they are about 12% less likely to be employed than comparable migrants, a gap that lasts ten to fifteen years. The first months and years after arrival matter most, and so does where a refugee is placed. Yet in many countries, including Switzerland, placement follows administrative quotas, with officials having little information on where each person would do best.

A recommendation, not a decision

In 2018, Kirk Bansak, Jens Hainmueller and colleagues proposed algorithmic refugee matching. A machine-learning model, trained on past arrivals, predicts how likely each refugee is to find work in each possible location. An optimisation step then recommends placements that maximise expected employment while respecting all quotas. A human officer keeps the final decision.

Until now, the benefits had been estimated only by replaying past data. This study, led by Bansak (University of California, Berkeley), Hainmueller (Stanford University) and Dominik Hangartner (ETH Zurich), reports what the authors believe is the first randomised controlled trial of the approach, run with Switzerland’s State Secretariat for Migration (SEM), using an early version of their tool, GeoMatch.

A blind lottery in a real administration

From January 2020 to June 2023, every eligible case — a family or a single adult who had received refugee status or provisional protection, and who could be placed in any canton — was assigned by computer to one of two groups, 50/50:

• algorithm: the canton GeoMatch predicted would maximise employment;
• control: a canton drawn at random among the available slots, as in the usual procedure.

Placement officers saw the same screen either way, and neither they nor the refugees knew which group a case was in. Each group had its own identical copy of the quotas by canton and by nationality. That design rules out a simple trick: the algorithm could not send more people to cantons with strong job markets. It could only improve the fit between a person and a place.

The model knew very little: sex, age, marital status, household size, arrival date, nationality and native language. Education, work history and language skills are not recorded in the register it used. Officers followed the displayed recommendation about 97% of the time.

The final sample: 2,000 cases, 2,156 adults, mainly from Afghanistan (54%), Turkey (26%) and Syria (9%).

The results

The main outcome, fixed in a pre-registered plan, is the share of months employed during the first three years. In the control group, adults worked on average 22.3% of those months.

• All cohorts (2020–2023): +2.2 percentage points for the algorithm group, about 10% more, with a confidence interval running from +0.05 to +4.3 points.
• Post-Covid cohorts (2022–2023): +3.9 points, about 17% more. The trial began weeks before the pandemic, and the model had been trained on pre-pandemic data; effects were indistinguishable from zero for the 2020 and 2021 cohorts. These comparisons by cohort were not pre-registered.
• Growing with time: 36 months after placement, the share of households with at least one adult in work was 5.2 points higher (about 11%).
• More stable jobs: a job lasting at least six months, +5.6 points; an open-ended contract at three years, +4.5 points.

Cheap, by design

Counting only public finances — benefits no longer paid, plus contributions and taxes — the authors value each extra month of employment at about 1,950 Swiss francs. With an assumed running cost of 50 francs per adult, they estimate about 1,300 francs of net benefit per adult, a ratio of roughly 27 to 1. They compare the effect at three years with that of programmes adding hundreds of hours of language training, while stressing that the algorithm is not a substitute for such courses.

They call the trial a conservative test: an early, frozen version of the tool, few predictors, coarse geography, strict quotas, and only half the caseload available to optimise. Whether newer, less constrained versions would do better remains to be tested.

Who built it, who tested it

The authors developed the GeoMatch method themselves. They declare no competing interests and state that they received no payments from the Swiss government for developing or operating the tool; the work was funded by Stanford University, foundations and philanthropies including Google.org, and a European research grant. Beyond refugees, they see the trial as rare field evidence that prediction tools can improve high-stakes public decisions when they support, rather than replace, the people who make them — and as a reminder that such tools must be retrained when the world changes, as it did in 2020.