Table 2 NIJFM scores and false positive rates on held-out (50%) test data with cutoffs of \(c=0.3\) and \(c=0.7\) for decision boundary. Bootstrap standard errors reported in parentheses
From: A note on the multiplicative fairness score in the NIJ recidivism forecasting challenge
Model | FP black | FP white | NIJFM | FP Black | FP white | NIJFM |
|---|---|---|---|---|---|---|
| Â | (\(c=0.3\)) | (\(c=0.3\)) | (\(c=0.3\)) | (\(c=0.7\)) | (\(c=0.7\)) | (\(c=0.7\)) |
Linear Reg. | 0.485 (0.009) | 0.354 (0.009) | 0.702 (0.010) | 0 | 0 | 0.808 (0.002) |
Logistic Reg. | 0.430 (0.008) | 0.303 (0.009) | 0.706 (0.010) | 0.003 (0.001) | 0.001 (0.001) | 0.807 (0.002) |
Linear Reg. (Trunc.) | 0 | 0 | 0.798 (0.002) | 0 | 0 | 0.808 (0.002) |
Linear Reg. (Shrink) | 0.369 (0.008) | 0.354 (0.009) | 0.795 (0.008) | 0 | 0 | 0.808 (0.002) |
Convex Surrogate | 0.478 (0.009) | 0.377 (0.009) | 0.725 (0.010) | 0 | 0 | 0.808 (0.002) |
BFGS | 0.434 (0.008) | 0.428 (0.010) | 0.795 (0.007) | 0 | 0 | 0.808 (0.002) |
Linear Reg. (Balanced) | 0.487 (0.008) | 0.354 (0.009) | 0.701 (0.010) | 0 | 0 | 0.808 (0.002) |
Linear Reg. (Group) | 0.485 (0.008) | 0.358 (0.009) | 0.705 (0.010) | 0 | 0 | 0.808 (0.002) |
XG Boost | 0.472 (0.008) | 0.377 (0.009) | 0.731 (0.010) | 0 | 0 | 0.808 (0.002) |
XG Boost (Trunc.) | 0 | 0 | 0.798 (0.002) | 0 | 0 | 0.808 (0.002) |
XG Boost (Shrink) | 0.386 (0.008) | 0.377 (0.009) | 0.798 (0.007) | 0 | 0 | 0.808 (0.002) |