Checking Correlations

We see no significant correlation with arrest_final_outcome which is a bit worrying. The new tired label is significantly correlated with ml-face (p-hat-cnn below) at the 1% level.

Regression Table 1: Tired & Well-Groomed on ML-Face

Q: How much variation in the CNN prediction is explained by tired ?

Regressing new labels on ML-Face
Dependent variable:
ML-Face Prediction
(1) (2) (3) (4) (5)
Well-Groomed -0.017*** -0.017*** -0.015***
(0.001) (0.001) (0.001)
Tired -0.006*** -0.002* -0.004*** -0.002**
(0.001) (0.001) (0.001) (0.001)
Male 0.116*** 0.117***
(0.003) (0.002)
Unknown Gender 0.061 0.048
(0.098) (0.097)
Age 0.0003*** 0.0002**
(0.0001) (0.0001)
Black -0.003 -0.002
(0.003) (0.003)
Unknown Race 0.015** 0.019***
(0.007) (0.007)
Asian -0.017 -0.011
(0.011) (0.011)
Indian 0.018 0.022
(0.024) (0.024)
Skin-Tone -0.177*** -0.190***
(0.043) (0.043)
Attractiveness -0.005*** 0.0003
(0.002) (0.002)
Competence -0.009*** -0.006***
(0.002) (0.002)
Dominance 0.004*** 0.004***
(0.001) (0.001)
Trustworthiness -0.004*** -0.003
(0.002) (0.002)
Constant 0.334*** 0.277*** 0.342*** 0.228*** 0.258***
(0.005) (0.005) (0.007) (0.009) (0.009)
Observations 9,603 9,603 9,603 9,603 9,603
Adjusted R2 0.025 0.002 0.025 0.217 0.231
Note: p<0.1; p<0.05; p<0.01

Regression Table 2: Tired & Well-Groomed on Detention-outcome

Q: Does tired explain some significant variation in the final detention outcome (like well-groomed does) ?

Regressing new labels and covariates on detention
Dependent variable:
Judge Detain Decision
(1) (2) (3) (4) (5) (6)
ML-Face 0.695*** 0.627***
(0.038) (0.044)
Well-Groomed -0.020*** -0.020*** -0.004
(0.004) (0.004) (0.005)
Tired -0.003 0.002 0.001 0.003
(0.005) (0.005) (0.005) (0.005)
Male 0.099*** 0.027**
(0.011) (0.012)
Unknown Gender -0.167 -0.208
(0.420) (0.416)
Age -0.001 -0.001**
(0.0005) (0.0005)
Black -0.022* -0.020*
(0.011) (0.011)
Unknown Race 0.017 0.008
(0.031) (0.031)
Asian -0.066 -0.054
(0.049) (0.048)
Indian 0.074 0.064
(0.102) (0.101)
Skin-tone -0.314* -0.207
(0.185) (0.183)
Attractiveness -0.002 0.002
(0.007) (0.007)
Competence -0.022*** -0.015**
(0.007) (0.007)
Dominance 0.010* 0.007
(0.005) (0.005)
Trustworthiness -0.014* -0.010
(0.007) (0.007)
Constant 0.058*** 0.329*** 0.244*** 0.322*** 0.294*** 0.168***
(0.011) (0.021) (0.020) (0.026) (0.040) (0.043)
Naive-AUC 0.624 0.531 0.498 0.531 0.579
Observations 9,603 9,603 9,603 9,603 9,603 9,603
Adjusted R2 0.033 0.002 -0.0001 0.002 0.014 0.035
Note: p<0.1; p<0.05; p<0.01

Regression Table 3: Tired on Re-Arrest

Q: Does tired explain re-arrest risk ?

Regressing new labels and covariates on Re-Arrest
Dependent variable:
Re-Arrest Outcome
(1) (2) (3) (4) (5)
Structured Risk Prediction 0.968*** 0.940***
(0.032) (0.035)
Well-Groomed -0.023*** -0.021*** -0.017***
(0.005) (0.005) (0.005)
Tired -0.017*** -0.012** -0.018***
(0.005) (0.005) (0.005)
Male -0.007
(0.011)
Unknown Gender -0.134
(0.431)
Age -0.001**
(0.0005)
Black 0.001
(0.012)
Unknown Race -0.0002
(0.032)
Asian 0.035
(0.050)
Indian -0.056
(0.105)
Skin-tone -0.438**
(0.190)
Attractiveness 0.014*
(0.007)
Competence -0.007
(0.008)
Dominance -0.002
(0.005)
Trustworthiness -0.015**
(0.007)
Constant -0.006 0.396*** 0.352*** 0.433*** 0.245***
(0.011) (0.022) (0.022) (0.028) (0.045)
Naive-AUC 0.685 0.532 0.521 0.536 0.692
Observations 9,603 9,603 9,603 9,603 9,603
Adjusted R2 0.085 0.003 0.001 0.003 0.089
Note: p<0.1; p<0.05; p<0.01