Table 1 Mean square error (MSE), false positive rates (FP) by race, and NIJFM scores on held-out (50%) test data with competition cutoff of 0.5 for decision boundary. Bootstrap standard errors reported in parentheses
From: A note on the multiplicative fairness score in the NIJ recidivism forecasting challenge
Model | MSE | FP (black) | FP (white) | NIJFM |
|---|---|---|---|---|
Linear Reg. | 0.192 (0.002) | 0.048 (0.004) | 0.030 (0.003) | 0.793 (0.005) |
Logistic Reg. | 0.192 (0.002) | 0.067 (0.004) | 0.042 (0.004) | 0.787 (0.005) |
Linear Reg. (Trunc.) | 0.192 (0.002) | 0 | 0 | 0.808 (0.002) |
Linear Reg. (Shrink) | 0.192 (0.002) | 0.034 (0.003) | 0.030 (0.003) | 0.804 (0.004) |
Convex Surrogate | 0.193 (0.002) | 0.043 (0.003) | 0.026 (0.003) | 0.794 (0.004) |
BFGS | 0.193 (0.002) | 0.035 (0.003) | 0.021 (0.003) | 0.796 (0.004) |
Linear Reg. (Balanced) | 0.192 (0.002) | 0.048 (0.004) | 0.030 (0.003) | 0.793 (0.005) |
Linear Reg. (Group) | 0.192 (0.002) | 0.045 (0.004) | 0.033 (0.003) | 0.798 (0.005) |
XG Boost | 0.192 (0.002) | 0.045 (0.004) | 0.030 (0.003) | 0.796 (0.005) |
XG Boost (Trunc.) | 0.193 (0.002) | 0 | 0 | 0.807 (0.002) |
XG Boost (Shrink) | 0.192 (0.002) | 0.033 (0.003) | 0.030 (0.003) | 0.804 (0.003) |