Theses Master's

Actuarial Injustice: Discrimination in Crime Prediction Software

Ciccolini, Julie

The actuarial justice movement has propagated an unprecedented increase in the use of crime prediction software in the criminal justice system. Specifically, two forms of crime prediction software - predictive policing and risk assessment instruments – are now informing high-stakes police and judicial decisions that have direct consequences on individual’s civil rights. While advocates claim that the software can alleviate human biases in the system, critics believe it may actually exacerbate them. Due to the conflicting definitions of fairness across legal, technical, and statistical disciplines, there has been no consensus on the software’s potential for discrimination.

In order to demonstrate how discrimination can manifest in crime prediction software, I examined a risk assessment instrument designed to predict pretrial felony rearrest for racial discrimination. The instrument is currently used in New York City and to date, has never been independently reviewed. I found that while the instrument demonstrates acceptable predictive validity for all racial subgroups, black defendants receive significantly higher scores on average than white defendants. Although there were small effect sizes, these differences may transcend into discrimination via disparate impact. Most noteworthy, I discovered that only one of the eight predictor variables in the model - whether or not a defendant had any prior arrests - was significantly predictive of future re-arrest. In fact, a redesigned model that predicts rearrest based solely on a defendant’s number of prior arrests performed just as well as the original model.

These findings indicate that crime prediction software that utilizes police-generated data to predict police-dependent outcomes is ultimately predicting police activity, not crime. I proffer that this problem is related to the outcome variable at hand and cannot be sufficiently minimized by data manipulation. Therefore, the police and judicial biases that have always plagued America’s criminal justice system will be paralleled in crime prediction software.


  • thumnail for Thesis Paper_Julie Ciccolini_Academic Commons.pdf Thesis Paper_Julie Ciccolini_Academic Commons.pdf application/pdf 633 KB Download File

More About This Work

Academic Units
Institute for the Study of Human Rights
Thesis Advisors
Harcourt, Bernard E.
M.A., Columbia University
Published Here
June 12, 2020