compas: Criminal Offenders Screened in Florida

In fairml: Fair Models in Machine Learning

Criminal Offenders Screened in Florida

Description

A collection of criminal offenders screened in Florida (US) during 2013-14.

Usage

data(compas)

Format

The data contains 5855 observations and the following variables:

• age, a continuous variable containing the age (in years) of the person;

• juv_fel_count, a continuous variable containing the number of juvenile felonies;

• decile_score, a continuous variable, the decile of the COMPAS score;

• juv_misd_count, a continuous variable containing the number of juvenile misdemeanors;

• juv_other_count, a continuous variable containing the number of prior juvenile convictions that are not considered either felonies or misdemeanors;

• v_decile_score, a continuous variable containing the predicted decile of the COMPAS score;

• priors_count, a continuous variable containing the number of prior crimes committed;

• sex, a factor with levels "Female" and "Male";

• two_year_recid, a factor with two levels "Yes" and "No" (if the person has recidivated within two years);

• race, a factor encoding the race of the person;

• c_jail_in, a numeric variable containing the date in which the person entered jail (normalized between 0 and 1);

• c_jail_out, a numeric variable containing the date in which the person was released from jail (normalized between 0 and 1);

• c_offense_date, a numeric variable containing the date the offense was committed;

• screening_date, a numeric variable containing the date in which the person was screened (normalized between 0 and 1);

• in_custody, a numeric variable containing the date in which the person was placed in custody (normalized between 0 and 1);

• out_custody, a numeric variable containing the date in which the person was released from custody (normalized between 0 and 1);

Note

The data set has been pre-processed as in Komiyama et al. (2018), with the following exceptions:

• the race variable has not been reduced to a binary factor with levels "African-American" and "not African-American";

• the variables type_of_assessment, v_type_of_assessment have been dropped from the analysis because they take the same value for all observations;

• variables like c_jail_in and c_jail_out that encode dates have been jointly rescaled to preserve the temporal ordering of events.

In that paper, two_year_recid is the response variable, sex and race are the sensitive attributes and the remaining variables are used as predictors.

References

Angwin J, Larson J, Mattu S, Kirchner L (2016). "Machine Bias: Theres Software Used Around the Country to Predict Future Criminals."
https://www.propublica.org

Examples

data(compas)

# convert the response back to a numeric variable.
compas$two_year_recid = as.numeric(compas$two_year_recid) - 1

# short-hand variable names.
r = compas[, "two_year_recid"]
s = compas[, c("sex", "race")]
p = compas[, setdiff(names(compas), c("two_year_recid", "sex", "race"))]

m = nclm(response = r, sensitive = s, predictors = p, unfairness = 0.05)
summary(m)

m = frrm(response = r, sensitive = s, predictors = p, unfairness = 0.05)
summary(m)

fairml documentation built on May 31, 2023, 6:02 p.m.