The Limits of Complex Algorithms in Predicting Recidivism

In the “Centre for Criminology & Sociolegal Studies: Criminological Highlights” published by the University of Toronto, the ability of complex algorithms to predict recidivism or re-offending while on pretrial release is discussed. In a study by Julia Dressel and Hany Farid, the researchers assessed the accuracy of algorithm based prediction system COMPAS.

They looked at data from 7,214 defendants in one county in Florida. They compared the predictions made by COMPAS to two other sources of predictions: (1) Ordinary statistical predictive models and (2) Intuitive predictions by ordinary people.

Dressel and Farid found that COMPAS’s prediction of recidivism (arrest within two years) was better than chance (65.2% accuracy). But its accuracy was not better than predictive models using simple logistic regression models or from their tests on human intuition. The logistic regression models took into account age, sex, charge, and some features of the criminal record. When it came to testing intuition, individuals were presented with summaries of information about the Florida defendants that included sex, age, offence, and three features of their criminal record.

Although the COMPAS model was better than chance, it predicted recidivism in black defendants more commonly than for white defendants when it in fact had not occurred. The commercial software COMPAS disadvantaged black defendants and gave advantages to white defendants. Their tests on human intuition revealed the same racial bias.

 

(Views are my own and do not represent the views of any organization.)

Comments

  1. “Although the COMPAS model was better than chance, it predicted recidivism in black defendants more commonly than for white defendants when it in fact had not occurred. The commercial software COMPAS disadvantaged black defendants and gave advantages to white defendants.” This can prove to be even more dangerous than bias generated from human findings because there are so many educated, not so educated and uneducated who find “techno”-, “machine”-, “computer”-generated to be more intelligent and less fallible than anything human and will hold this result to be truth the whole truth and nothing but the truth especially if supported and propagated by the leaders they choose to follow who use such information to distract from the lack of viable solutions to any problems that may plague their followers.

  2. Criminological Highlights is basically just a newsletter containing non-peer-reviewed summaries. You should at least provide a link to the actual paper, especially because it’s published in an open-access journal: http://advances.sciencemag.org/content/4/1/eaao5580 For those wondering, Science Advances is a peer-reviewed journal.