How we investigated Sweden's Suspicion Machine

Summary

Detailed methodology documentation. Reports six fairness metrics tested: demographic parity, false positive error rate, predictive parity, equal burden (Petersen et al.), ISF supervisory definition, and SIA's internal fairness process. Provides exact ratios: women 1.5x overselected, foreign background 2.5x, no degree 3.31x, below-median income 2.97x. False positive rates: women 1.7x, foreign background 2.4x, no degree 3x, below-median income 3x. Documents that only 166 of 5,520 suspected fraud cases (3%) led to convictions in 2022. Fraud estimate drops from 24% to 6% with threshold adjustment. Lists eight consulting academic experts. GitHub: github.com/Lighthouse-Reports/suspicion_machines_sweden.