A radical new approach to measuring recidivism risk

NOTE: This post has been updated as of 4/2 to incorporate additional research.

Researchers at the RAND Corporation have proposed a radical new approach to measuring recidivism risk that raises questions about decades of received truth about the prevalence of reoffending after people leave prison.  At least since the 1990s, the Bureau of Justice Statistics has measured risk of recidivism at the time of a person’s last interaction with the justice system, when the statistical cohort includes many who are frequent participants in the criminal system as well as those for whom it is a one-time affair.  As a result, employers and others tend to interpret background checks as overstating the risk posed by someone who in fact may have been living in the community for years without criminal incident, and is unlikely to become criminally involved again.

In Providing Another Chance: Resetting Recidivism Risk in Criminal Background Checks, Shawn Bushway and his RAND colleagues argue that risk should instead be measured at the time a background check is conducted, after an individual has had an opportunity to demonstrate their ability to reintegrate lawfully as well as their propensity to reoffend.  They label this the “reset principle,” and argue that this more individualized approach to risk assessment promises to improve the predictive value of criminal background checks.  In fact, they propose that it will “strengthen the case that people with convictions can, and usually do, change their ways.”

Coupled with other studies showing that the risk of recidivism depends on a variety of factors (e.g., age at time of offense), this new RAND study suggests that general “time to redemption” research should not be relied upon to predict future behavior of specific individuals.

The premise of the “reset principle” seems reasonable and even self-evident once explained.  And, if models based on the reset principle are developed into viable tools that employers and others can use to assess recidivism risk, these models may offer many with criminal histories a way to demonstrate that they should be offered another chance.  But there is a good reason why this has not already been done: it has been hard to identify data that would account for a sufficiently long period of time that an individual spends free in the community after their last interaction with the criminal justice system to accurately measure risk of reoffending. As Bushway and his colleagues point out, this sort of data collection requires long-term surveillance that implicates issues of privacy.

The large data set from the North Carolina Department of Public Safety allowed the researchers to measure the frequency with which people who have a conviction are reconvicted over a ten-year period, how quickly their likelihood of reoffending declines, and how their risk profiles change.  They reached three highly significant conclusions based on the North Carolina data:

  • The majority of individuals with a conviction do not have a subsequent conviction.
  • A person’s likelihood of being convicted again declines rapidly as more time passes
  • After a sufficient period without a new conviction, even people initially deemed to be at highest risk for recidivism (such as those with a more extensive criminal background) transition to risk levels that appear similar to those initially at the lowest risk.

The RAND researchers caution that “prediction models that work in a specific context are not guaranteed to work in other contexts,” and that “there is no one-size-fits-all solution.” With that caveat, based on their conclusions respecting the North Carolina data, they make a series of recommendations:

  • Policymakers should recognize that, over an extended sampling period, most people who get convicted are not reconvicted. This provides a fact base for policymaking that differs from findings by the Bureau of Justice Statistics that articulate that, in a given cohort of people released from prison (e.g., in a given year), most people experience another conviction.
  • Updates to the Uniform Guidelines on Employee Selection Procedures can validate a new class of models, such as those that satisfy the reset principle, providing employers a more certain defense to challenges to their employment decisions.
  • Policymakers and other decisionmakers should make determinations about risk thresholds that are applied in a particular setting (e.g., an employer deciding how much recidivism risk is appropriate for a given job description) because those thresholds implicate issues of equity and fairness.
  • Data quality can limit the development of successful recidivism risk models, and policymakers should consider creating data infrastructure that supports models that adhere to the reset principle.
  • Policymakers should understand that exploring and stressing models that adhere to the reset principle for bias will be crucial. Model predictions may reflect the unfair systemic biases in the current criminal justice system.
  • Tools that use models that adhere to the reset principle should be developed judiciously and after carefully considering many systemic factors regarding fairness. An adequate assessment of bias should include a comparison to the current state. Even an imperfect tool could provide more opportunities to candidates against whom the current system is biased than the current methods.

This research was sponsored by Arnold Ventures and conducted by the RAND Justice Policy Program.