SE · 27 November 2024
Sweden's Suspicion Machine: the Social Insurance Agency's parental benefit risk model
Image: Lighthouse Reports, from the original investigation. All rights reserved by the owner.
- Reported by
- Lighthouse Reports, Svenska Dagbladet
- Story language
- EN · SV
- Methods
- FOI / records request
- Data analysis
In Sweden the Social Insurance Agency refused almost everything for three years, and the investigation went ahead anyway by going through the agency's overseer. It showed that a risk model for parental benefit selected women more than 1.5 times as often as men, and people with a foreign background about 2.5 times as often. Amnesty International called for the system to be discontinued, and in 2025 the agency took it out of use.
lighthousereports.comSweden's Suspicion Machine: the Social Insurance Agency's parental benefit risk modelA machine-learning risk profile used by Försäkringskassan (the Swedish Social Insurance Agency) to select applicants for targeted checks on temporary parental benefit (VAB), which compensates parents for time off to care for a sick child.Sweden · Reported by Lighthouse Reports, Svenska Dagbladet (opens in a new tab)How they investigated
- Methodology: Sweden's Suspicion Machine: the Social Insurance Agency's parental benefit risk model (opens in a new tab)
- Reusable code: Sweden's Suspicion Machine: the Social Insurance Agency's parental benefit risk model (opens in a new tab)
FOI / records request
The agency rejected the team's first request in October 2021 and nearly all of the dozens that followed over three years. The team went to court twice. The way through was the agency's independent supervisory authority (ISF), whose 2018 report had analysed the model: an FOI request to ISF for the underlying dataset succeeded, even though the agency had refused the same data. A separate request produced a document describing the agency's own fairness procedure.
Data analysis
The dataset covered 6,129 people selected for investigation in 2017 (1,047 at random and 5,082 by the model) and whether a mistake was found, split by gender, foreign background, income and education. With eight academic reviewers, the team tested six statistical fairness definitions with bootstrapped significance tests, tested the agency's own fairness procedure against the data, and examined how the agency estimates fraud. Data and code are published.
Read alongside
NL · 2023
Suspicion Machines: Rotterdam's welfare fraud risk scoreSweden has outcome data but the model stayed a black box. Rotterdam is the reverse.
FR · 2023
France's Digital Inquisition: the family benefits fund's risk scoreBoth investigations analysed a benefits-fraud model from released material. In France the code came through access requests and a CADA ruling; in Sweden the data came through the agency's overseer.
UK · 2024
The DWP model that flags Universal Credit advance requests for fraudBoth are benefits-fraud models where the main evidence came from the agency's own fairness analysis. In the UK the department published its numbers after pressure; in Sweden the team ran the tests itself.