Analytics Stories

CHAPTER 58Can an Algorithm Improve Parole Decisions?

In 2019, the American prison system had 2.3 million people in jail (www.prisonpolicy.org/reports/pie2019.html). The United States has 4.4% of the world's population but 22% of the world's prisoners. In 2016, the United States had 4.5 million people on parole or probation. (www.bjs.gov/index.cfm?ty=pbdetail&iid=6188). In 2018, there were 91,600 parole officers (www.bls.gov/ooh/community-and-social-service/probation-officers-and-correctional-treatment-specialists.htm), so on average a parole officer supervises around 50 people on parole or probation.

A 2018 Department of Justice (DOJ) study followed for nine years prisoners who were released in 2005 (www.bjs.gov/content/pub/pdf/18upr9yfup0514.pdf). This, of course, includes prisoners who were released but not paroled. The DOJ study found that 83% of all prisoners released in 2005 committed a crime within nine years of their release. (Eighty-seven percent of blacks were arrested compared to 83% of whites.)

In recent years, sophisticated data science algorithms have been widely applied to aid the U.S. government, cities, and states in their parole decisions. Common inputs to these algorithms include information similar to the information collected by the government on the Post Conviction Risk Assessment (PCRA) (see Jennifer Skeem and Christopher Lowenkamp's article “Risk, Race, and Recidivism: Predictive Bias and Disparate Impact,” Criminology, vol. 54, no. 4, 2016, pages 680–712). ...