COMPAS Analysis using Aequitas

In 2016, Propublica reported on racial inequality in automated criminal risk assessment algorithms. The report is based off of this analysis. Using a clean version of the COMPAS dataset, found in the Propublica repo, we demostrate the use of the Aequitas bias reporting tool.

Northpointe’s COMPAS (Correctional Offender Management Profiling for Alternative Sanctions) is one of the widest used risk assessment tools, algorithms which are used in the criminal justice system to guide decisions such as how to set bail. The ProPublica dataset represents two years of COMPAS predicitons from Broward County, FL.

Analysis by Ari Anisfeld and Abby Stevens.

import pandas as pd
import seaborn as sns
from aequitas.group import Group
from aequitas.bias import Bias
from aequitas.fairness import Fairness

df = pd.read_csv("../../../examples/data/compas_for_aequitas.csv")
df.head()

	entity_id	score	label_value	race	sex	age_cat
0	1	0.0	0	Other	Male	Greater than 45
1	3	0.0	1	African-American	Male	25 - 45
2	4	0.0	1	African-American	Male	Less than 25
3	5	1.0	0	African-American	Male	Less than 25
4	6	0.0	0	Other	Male	25 - 45

The data was preprocessed for compatibility with Aequitas. Notice the mandatory score and label_value columns along with a set of attribute columns specific to this data set. See documentation for more information about input data. (entity_id is not necessary for this example. As a reserve column name, Aequitas will not treat entity_id as an attribute column.)

Exploratory Analysis

Risk assessment by race

COMPAS produces a risk score that predicts a person’s likelihood of commiting a crime in the next two years. The output is a score between 1 and 10 that maps to low, medium or high. For Aequitas, we collapse this to a binary prediction. A score of 0 indicates a prediction of low risk according to COMPAS, while a 1 indicates high or medium risk.

This is based on ProPublica’s interpretation of Northpointe’s practioner guide.

"According to Northpointe’s practitioners guide, COMPAS “scores in the medium and high range garner more interest from supervision agencies than low scores, as a low score would suggest there is little risk of general recidivism,” so we considered scores any higher than “low” to indicate a risk of recidivism."

In the bar chart below, we see a large difference in how these scores are distributed by race, with a majority of white and Hispanic people predicted as low risk (score = 0) and a majority of black people predicted high and medium risk.

g = sns.countplot(x="race", hue="score", data=df[df.race.isin(['African-American', 'Caucasian', 'Hispanic'])])

Levels of recidivism

This dataset includes information about whether or not the subject recidivated, and so we can directly test the accuracy of the predictions. First, we visualize the recidivsm rates across race.

Following ProPublica, we defined recidivism as a new arrest within two years. (If a person recidivates, label_value = 1). They “based this decision on Northpointe’s practitioners guide, which says that its recidivism score is meant to predict ‘a new misdemeanor or felony offense within two years of the COMPAS administration date.’”

g = sns.countplot(x="race", hue="label_value", data=df[df.race.isin(['African-American', 'Caucasian', 'Hispanic'])])

Putting Aequitas to the task

The graph above shows the base rates for recidivism are higher for black defendants compared to white defendants (.51 vs .39). The prior graph shows that the predictions do not match the base rates.

Practitioners face the challenge of determining whether or not such patterns reflect bias or not. The fact that we have multiple ways to measure bias adds complexity to the decisionmaking process. With Aequitas, we provide a tool that automates the reporting of various fairness metrics to aid in this process.

Applying Aequitas progammatically is a three step process represented by three python classes:

Group(): Define groups

Bias(): Calculate disparities

Fairness(): Assert fairness

Each class builds on the previous one expanding the output DataFrame.

Group()

Aequitas expects a dataframe with predefined columns score, and label_value and treats other columns (with a few exceptions) as attributes against which to test for disparities. In this cases we include race, sex and age_cat.

g = Group()
xtab, _ = g.get_crosstabs(df)

model_id, score_thresholds 1 {'rank_abs': [3317]}
COUNTS::: race
African-American    3696
Asian                 32
Caucasian           2454
Hispanic             637
Native American       18
Other                377
dtype: int64
COUNTS::: sex
Female    1395
Male      5819
dtype: int64
COUNTS::: age_cat
25 - 45            4109
Greater than 45    1576
Less than 25       1529
dtype: int64

The get_crosstabs() command tabulates a confusion matrix for each subgroup and calculates commonly used metrics such as false positive rate and false omission rate. It also provides counts by group and group prevelances.

# Here's a subset of the crosstab data
xtab[['attribute_name', 'attribute_value', 'fpr', 'fnr', 'tpr', 'tnr', 'for', 'fdr', 'group_size']].round(2)

	attribute_name	attribute_value	fpr	fnr	tpr	tnr	for	fdr	group_size
0	race	African-American	0.45	0.28	0.72	0.55	0.35	0.37	3696
1	race	Asian	0.09	0.33	0.67	0.91	0.12	0.25	32
2	race	Caucasian	0.23	0.48	0.52	0.77	0.29	0.41	2454
3	race	Hispanic	0.21	0.56	0.44	0.79	0.29	0.46	637
4	race	Native American	0.38	0.10	0.90	0.62	0.17	0.25	18
5	race	Other	0.15	0.68	0.32	0.85	0.30	0.46	377
6	sex	Female	0.32	0.39	0.61	0.68	0.24	0.49	1395
7	sex	Male	0.32	0.37	0.63	0.68	0.33	0.36	5819
8	age_cat	25 - 45	0.33	0.37	0.63	0.67	0.32	0.39	4109
9	age_cat	Greater than 45	0.17	0.57	0.43	0.83	0.24	0.46	1576
10	age_cat	Less than 25	0.54	0.26	0.74	0.46	0.42	0.36	1529

Interpreting the crosstab We see that African-Americans have a false positive rate (fpr) of 45%, while Caucasians have a false positive rate of only 23%. This means that black people are far more likely to be falsely labeled as high-risk than white people. On the otherhand false ommision rates (for) and false discovery rates (fdr) are much closer for those two groups.

Bias()

We calculate disparities as a ratio of a metric for a group of interest compared to a base group. For example, the False Negative Rate Disparity for black defendants vis-a-vis whites is:

Below, we use get_disparity_predefined_groups which allows us to choose reference groups that clarify the output for the practitioner. Aequitas Bias() class includes two additional get disparity functions, which automate base group selection.

b = Bias()
bdf = b.get_disparity_predefined_groups(xtab, {'race':'Caucasian', 'sex':'Male', 'age_cat':'25 - 45'})

get_disparity_predefined_group()

The results are appended to the crosstab dataframe. We see a slice of this below.

bdf.columns

Index(['attribute_name', 'attribute_value', 'k', 'model_id', 'score_threshold',
       'tpr', 'tnr', 'for', 'fdr', 'fpr', 'fnr', 'npv', 'precision', 'pp',
       'pn', 'ppr', 'pprev', 'fp', 'fn', 'tn', 'tp', 'group_label_neg',
       'group_label_pos', 'group_size', 'total_entities', 'prev',
       'ppr_disparity', 'pprev_disparity', 'precision_disparity',
       'fdr_disparity', 'for_disparity', 'fpr_disparity', 'fnr_disparity',
       'tpr_disparity', 'tnr_disparity', 'npv_disparity',
       'ppr_ref_group_value', 'pprev_ref_group_value',
       'precision_ref_group_value', 'fdr_ref_group_value',
       'for_ref_group_value', 'fpr_ref_group_value', 'fnr_ref_group_value',
       'tpr_ref_group_value', 'tnr_ref_group_value', 'npv_ref_group_value'],
      dtype='object')

bdf[['attribute_name', 'attribute_value', 'ppr_disparity', 'pprev_disparity', 'fdr_disparity', 'for_disparity',
       'fpr_disparity', 'fnr_disparity']]

	attribute_name	attribute_value	ppr_disparity	pprev_disparity	fdr_disparity	for_disparity	fpr_disparity	fnr_disparity
0	race	African-American	2.545667	1.690224	0.906085	1.213154	1.912093	0.586416
1	race	Asian	0.009368	0.718384	0.611748	0.433839	0.370749	0.698482
2	race	Caucasian	1.000000	1.000000	1.000000	1.000000	1.000000	1.000000
3	race	Hispanic	0.222482	0.857099	1.120464	1.001616	0.915887	1.165140
4	race	Native American	0.014052	1.915691	0.611748	0.578453	1.598854	0.209544
5	race	Other	0.092506	0.602147	1.115085	1.048203	0.629057	1.417970
6	sex	Female	0.216801	0.904348	1.336425	0.734738	0.990343	1.055810
7	sex	Male	1.000000	1.000000	1.000000	1.000000	1.000000	1.000000
8	age_cat	25 - 45	1.000000	1.000000	1.000000	1.000000	1.000000	1.000000
9	age_cat	Greater than 45	0.204782	0.533914	1.192804	0.746232	0.503031	1.531238
10	age_cat	Less than 25	0.519231	1.395369	0.935673	1.313873	1.621868	0.696781

Interpreting the disparity ratios The differences in False Positive Rates, noted above, are clarified using the disparity ratio (fpr_disparity). Black people are falsely identified as being high or medium risks 1.9 times the rate for white people. As seen above, False Discovery Rates have much less disparity (fdr_disparity). False discovery is the fraction of false postives over predicted positive in a group. COMPAS is calibrated to have these balanced across groups.

Notice that reference groups have disparity = 1 by design.

Fairness()

Finally, the Fairness() class provides three functions that provide a high level summary. Using FPR disparity as an example and the default fairness threshold, we have:

The Fairness().get_group_value_fairness() function builds on the previous dataframe.

f = Fairness()
fdf = f.get_group_value_fairness(bdf)

We can assess fairness at various levels of detail.

gaf = f.get_group_attribute_fairness(fdf)
gaf

	model_id	score_threshold	attribute_name	Statistical Parity	Impact Parity	FDR Parity	FPR Parity	FOR Parity	FNR Parity	TypeI Parity	TypeII Parity	Unsupervised Fairness	Supervised Fairness
0	1	binary 0/1	age_cat	False	False	True	False	False	False	False	False	False	False
1	1	binary 0/1	race	False	False	False	False	False	False	False	False	False	False
2	1	binary 0/1	sex	False	True	False	True	False	True	False	False	False	False

fdf[['attribute_name', 'attribute_value','Statistical Parity',
       'Impact Parity', 'FDR Parity', 'FPR Parity', 'FOR Parity', 'FNR Parity',
       'TypeI Parity', 'TypeII Parity', 'Unsupervised Fairness',
       'Supervised Fairness']]

	attribute_name	attribute_value	Statistical Parity	Impact Parity	FDR Parity	FPR Parity	FOR Parity	FNR Parity	TypeI Parity	TypeII Parity	Unsupervised Fairness	Supervised Fairness
0	race	African-American	False	False	True	False	True	False	False	False	False	False
1	race	Asian	False	False	False	False	False	False	False	False	False	False
2	race	Caucasian	True	True	True	True	True	True	True	True	True	True
3	race	Hispanic	False	True	True	True	True	True	True	True	False	True
4	race	Native American	False	False	False	False	False	False	False	False	False	False
5	race	Other	False	False	True	False	True	False	False	False	False	False
6	sex	Female	False	True	False	True	False	True	False	False	False	False
7	sex	Male	True	True	True	True	True	True	True	True	True	True
8	age_cat	25 - 45	True	True	True	True	True	True	True	True	True	True
9	age_cat	Greater than 45	False	False	True	False	False	False	False	False	False	False
10	age_cat	Less than 25	False	False	True	False	False	False	False	False	False	False

Interpreting parity In this case, our base groups are Caucasian for race, Male for gender, and 25-45 for age_cat. By construction, the base group has supervised fairness. (The disparity ratio is 1). Relative to the base groups, the COMPAS predictions only provide supervised fairness to one group, Hispanic.

Above, the African-American false omission and false discovery are within the bounds of fairness. This result is expected because COMPAS is calibrated. (Given calibration, it is surprising that Asian and Native American rates are so low. This may be a matter of having few observations for these groups.)

On the other hand, African-Americans are roughly twice as likely to have false positives and 40 percent less likely to false negatives. In real terms, 44.8% of African-Americans who did not recidivate were marked high or medium risk (with potential for associated penalties), compared with 23.4% of Caucasian non-reoffenders. This is unfair and is marked False below.

These findings mark an inherent trade-off between FPR Fairness, FNR Fairness and calibration, which is present in any decision system where base rates are not equal. See Chouldechova (2017). Aequitas helps bring this trade-off to the forefront with clear metrics and asks system designers to make a reasoned decision based on their use case.