Outline:

This document has two quick regression tables to check whether our tired-eyes feature contains any signal. As a reminder, we had re-collected that label after changing the description.

Here is how we now prompt the Mturk-raters:

Tired eyes: does this person appear to have eyes that are strained or tired? At low values, this person may have eye bags, have visibly tired looking eyes, be squinting, have black circles or puffiness around their eyes, droopy eyelids, or dark sunken eyes. At high values, this person may have alert looking eyes, smooth skin around their eyes, and no puffiness or eye bags.

Overall: This doesn’t seem to have worked very well, and we appear to have no real significant in the full sample here !

Checking label distribution:

Note This label has an average inter-rater correlation of only about 10% which is (as far as I know) quite low. So we may need to iterate on the label-gathering process for this !

Regressing tired-eyes on algorithmic predictions:
Correlation between human-labeled novel feature and algorithms prediction
Dependent variable:
ML-Face Prediction
(1) (2) (3) (4)
Tired-eyes -0.003** -0.0003 -0.001 -0.002**
(0.001) (0.001) (0.001) (0.001)
Well-Groomed -0.016***
(0.001)
Age 0.001*** 0.0003*** 0.0002***
(0.0001) (0.0001) (0.0001)
Sex-Male 0.115*** 0.113*** 0.115***
(0.002) (0.003) (0.002)
Sex-Unknown 0.055 0.057 0.044
(0.099) (0.098) (0.097)
Race-Black -0.018*** -0.017*** -0.016***
(0.004) (0.004) (0.004)
Race-Unknown 0.007 0.013* 0.017**
(0.007) (0.007) (0.007)
Race-Asian -0.023** -0.021* -0.016
(0.012) (0.011) (0.011)
Race-Indian 0.006 0.013 0.017
(0.024) (0.024) (0.024)
Skin-tone -0.044*** -0.041*** -0.044***
(0.006) (0.006) (0.006)
Attractiveness -0.006*** 0.0005
(0.002) (0.002)
Competence -0.009*** -0.006***
(0.002) (0.002)
Dominance 0.004*** 0.004***
(0.001) (0.001)
Trustworthiness -0.005*** -0.003*
(0.002) (0.002)
Constant 0.266*** 0.179*** 0.247*** 0.288***
(0.006) (0.008) (0.011) (0.011)
Observations 9,598 9,598 9,598 9,598
Adjusted R2 0.0004 0.204 0.220 0.234
Note: p<0.1; p<0.05; p<0.01
Regressing tired-eyes on detention outcome:
Does fat-faced predict judge decisions
Dependent variable:
Judge Detain Decision
(1) (2) (3) (4) (5)
Tired-eyes -0.006 -0.001 -0.003 -0.003 -0.002
(0.004) (0.004) (0.004) (0.004) (0.004)
Well-Groomed -0.014*** -0.005
(0.005) (0.005)
ML-Face 0.620***
(0.044)
Age -0.001** -0.001*** -0.001*** -0.001***
(0.0004) (0.0004) (0.0004) (0.0004)
Sex-Male 0.098*** 0.094*** 0.095*** 0.024**
(0.011) (0.011) (0.011) (0.012)
Sex-Unknown -0.160 -0.162 -0.174 -0.201
(0.421) (0.420) (0.420) (0.415)
Race-Black -0.065*** -0.062*** -0.062*** -0.052***
(0.016) (0.016) (0.016) (0.015)
Race-Unknown -0.002 0.007 0.011 0.001
(0.031) (0.031) (0.031) (0.031)
Race-Asian -0.083* -0.077 -0.072 -0.062
(0.049) (0.049) (0.049) (0.048)
Race-Indian 0.051 0.064 0.068 0.058
(0.102) (0.102) (0.102) (0.101)
Skin-tone -0.106*** -0.101*** -0.103*** -0.076***
(0.025) (0.025) (0.025) (0.025)
Attractiveness -0.005 0.0002 -0.0001
(0.007) (0.007) (0.007)
Competence -0.021*** -0.018** -0.015**
(0.007) (0.007) (0.007)
Dominance 0.010* 0.009* 0.007
(0.005) (0.005) (0.005)
Trustworthiness -0.013* -0.012* -0.010
(0.007) (0.007) (0.007)
Constant 0.263*** 0.282*** 0.407*** 0.444*** 0.266***
(0.024) (0.035) (0.045) (0.047) (0.048)
Naive-AUC 0.51 0.571 0.586 0.588 0.633
Observations 9,598 9,598 9,598 9,598 9,598
Adjusted R2 0.0001 0.012 0.016 0.017 0.037
Note: p<0.1; p<0.05; p<0.01