Integrated Dynamic Risk Assessment for Community Supervision (IDRACS)

Predicting Rearrest With An Artificial Intelligence Tool for the National Institute of Justice

In the United States 1.8 million people are incarcerated in jails or prisons. However, more than double that number are serving all or part of their sentence in the community. In 2023, roughly 3.8 million people in the United States were on either probation or parole. Community supervision agencies often use risk assessments to aid the practice of supervising people on probation or parole by categorizing those on supervision based on their risk of rearrest (e.g., low, moderate, high) and aligning the intensity of supervision and services with the risk level. 

Risk assessments are often based on statistical models that correlate certain characteristics with the risk of arrest while on probation or parole. Yet many risk assessment tools only use static measures that are collected once at intake or are infrequently updated. In practice, this can mean that someone’s risk profile looks the same on their first day of supervision as it does after their first or fifth year, regardless of any progress they may have made.

Addressing Risk Classification Challenges with Artificial Intelligence

The National Institute of Justice (NIJ) sought to incorporate Artificial Intelligence (AI) into community corrections. In response to their request for proposal, RTI International, in partnership with the George Department of Community Supervision (DCS) and Applied  Research Services, Inc. (ARS), constructed and implemented the Integrated Dynamic Risk Assessment for Community Supervision (IDRACS) risk classification tool. 

The model development process began by creating a longitudinal dataset from state agency records covering over 160,000 individuals who began supervision in Georgia between 2016 and 2019. The Georgia DCS dataset included traditional static measures collected at intake, such as criminal history, the underlying charge that led to probation or parole, and other fixed characteristics (e.g., age at supervision start, sex, prior criminal history). In addition, they provided detail from case management files, including the supervision conditions, occurrence of technical violations, drug test results (both positive and negative), employment verification, and other measures. This rich dataset that captured changes over time allowed for the development of accurate, tailored models intended to predict the risk of a serious rearrest or revocation for felony or violent misdemeanor charges while on supervision. 

The team explored and compared a range of traditional statistical and advanced machine learning methods, ultimately determining that multi-period logistic regression models, stratified by supervision type, sex, and key time intervals on supervision, provided both predictive accuracy and operational transparency. RTI developed and tested these models, incorporating insights from George DCS research and operations staff. This collaboration yielded several novel features that led to model improvements:

1. RTI developed time-specific models that mirrored Georgia DCS’s supervision practices, with predictions generated from the following time periods:
   1. The first ninety days, when intensive supervision requires multiple check-ins and often includes regular drug testing.

   2. The remainder of the first year on supervision, when check-ins are regular, but drug testing is typically only carried out for cause.

      Beyond the first year of supervision, when check-ins are less common and many data fields are infrequently updated, employing time-specific models ensured that we compared apples-to-apples, so that someone’s relative risk profile was based on how much time they’ve been on supervision.

2. Central to the tool are dynamic factors—like employment, drug testing results, and supervision conditions—that evolve during community supervision. These features allow for someone’s risk to increase or decrease based on their progress on supervision.
3. RTI used feature selection algorithms to test the utility of including someone’s complete criminal history. As a result, models in the first ninety days only use the last five years of criminal history, as these models performed better than those that included lifetime criminal history measures. Furthermore, the models after the first ninety days only use the last two years of criminal history and rely on the dynamic features of an individual’s time on supervision to inform predictions of rearrest.
4. Analysts incorporated race into the model-building process but omitted race from predictions. Including race when developing the models ensures that baked in differences in criminal history do not produce biased predictions. This results in models that do not predict more accurately based on someone’s race.
5. RTI staff compared these traditional inferential models to machine learning algorithms and found that the machine learning models yielded only slight improvements in accuracy but offered limited transparency and posed more obstacles when moving to the implementation stage.

Integration into an Existing System for Efficiency

RTI worked with Georgia DCS IT and operations staff to ensure that nothing was lost in translation when attempting to incorporate models based on analytical datasets into an active case management system. Extensive validation collaboration with Georgia DCS IT and line officers, and iterative model refinement ensured that the final tool could be integrated successfully into the agency's case management system. This offered a dynamic risk assessment that not only improved predictive accuracy over prior models but also allowed for agency feedback. 

Given that supervision levels were set from the extant primarily static risk scores, incorporating automated dynamic and time-specific predictions into Georgia DCS’s CMS allowed for individuals to transition down from high levels of supervision if they showed signs of progress. It also allowed for those who demonstrated acute risk to be transitioned to higher levels of supervision. 

Tapping into Technology to Enhance Supervision and Community Support

The IDRACS tool stands as a testament to the power of interdisciplinary collaboration and evidence-based practice in enhancing public safety and supporting community supervision. Continuing to mindfully leverage the latest AI technologies and innovations in the public safety field while working with agencies to utilize their data will lead to greater efficiency in allocation of agency resources more successful supervision outcomes, and improved public safety as agencies are able to tailor their supervision strategies to those presenting the greatest immediate risk.