Return to search

Analysis of racial bias in Northpointe's COMPAS algorithm

archives@tulane.edu / This study will evaluate the effectiveness and validity of an algorithm that is widely used in
the criminal justice system. In the year 2000, Northpointe introduced a risk assessment algorithm,
called Correctional Offender Management Profiling for Alternative Sanctions, often referred to as
COMPAS, in attempt to optimize the pretrial and sentencing stages of the process. COMPAS
computes scores for defendants and these scores assist judges in deciding jail terms, sentencing,
and probation. Some argue that such algorithms are more accurate and less biased than humans,
but COMPAS has been highly criticized for perpetuating the systemic racial bias that is currently
present in the criminal justice system.
This analysis will test the logical bases of the algorithm and its predictive ability. The results
of this analysis cast doubt on the usefulness of the COMPAS algorithm if fairness is the goal of the
criminal justice system.
A logistic regression analysis will show that the two most important variables for predicting
recidivism are a defendant’s age and the number of prior offenses committed. A linear model using
only these two predictors will be shown to yield better predictive accuracy than COMPAS. The
analysis will also show that COMPAS scores were unevenly distributed between different racial
groups. In general, African Americans were more likely than Caucasians to be given higher scores
regardless of their recidivism rates. Further, more frequently than Caucasians, African Americans
who were assigned lower scores did not recidivate. / 1 / Adrienne Brackey

  1. tulane:92018
Identiferoai:union.ndltd.org:TULANE/oai:http://digitallibrary.tulane.edu/:tulane_92018
Date January 2019
ContributorsBrackey, Adrienne (author), Cortez, Ricardo (Thesis advisor), School of Science & Engineering Mathematics (Degree granting institution)
PublisherTulane University
Source SetsTulane University
LanguageEnglish
Detected LanguageEnglish
TypeText
Formatelectronic, pages:  45
RightsNo embargo, Copyright is in accordance with U.S. Copyright law.

Page generated in 0.0024 seconds