Prison violence dramatically drops with analytics-based risk assessment
SAS helps prisons predict and mitigate violence using data.
in staff assaults
The Indiana Department of Correction uses real-time data and predictive analytics from SAS to reduce violent attacks on staff by 50%
It was just another day at the Indiana Department of Correction (IDOC). Researchers Sarah Schelle and Bret Ellis were chatting by the water cooler when Deputy Commissioner of Operations James Basinger approached them. He appeared visibly shaken. Basinger pulled out his phone and showed them the gruesome picture. Another violent attack on a prison guard. This time, the beating was so severe the guard had to be hospitalized.
“We have to do something about this,” Basinger said.
That is when Schelle and Ellis sprang into action and started doing what they do best – using data to solve complex problems.
The attack that day was not an isolated incident. That spring, the IDOC suffered a spike of brutal assaults against staff members. During an 18-month period, 320 violent assaults were recorded monthly on average across the IDOC’s adult facilities.
“Not only was this surge in violence disturbing, it’s expensive for the state in terms of workers’ compensation and non-productive hours,” Schelle says. “The attacks are costly for our staff, our facilities and the people of Indiana.”
The IDOC’s risk assessment tool clearly wasn’t working. The Indiana Risk Assessment System (IRAS) was specifically designed to predict the likelihood of recidivism – not violence. Furthermore, IRAS uses only static information captured at intake to predict violence. Critically, it fails to incorporate ongoing changes to a prisoner’s risk status during incarceration.
In other words, a low-risk prisoner on day one might become high-risk on day 100. But without a way to track and communicate these changing risk factors, prison staff were left in the dark, and an unexpected assault could be waiting just around the corner.
By identifying where we can interrupt stressors, we can help reduce violence and put these people in a better situation long-term. Sarah Schelle Executive Director of Legislation and Data Science IDOC
Predicting violence using behavioral data
After discussing the problem with operations staff, it became apparent to Schelle and Ellis that a tool designed to detect and communicate when offenders were most volatile would systematically reduce assaults on guards and fellow inmates.
Work began on a new risk assessment tool using predictive analytics from SAS. The tool needed to do three things: incorporate all relevant data sources, more accurately predict violence and clearly communicate this insight to facility staff.
Using data management tools from SAS, the IDOC started aggregating data from various systems including IRAS and its offender information management systems. This introduced dozens of near-real-time risk variables into the equation. Facility staff were involved in a focus group, which helped to hone the variables within the model. Their feedback included interplays between variables included in the first model, which, without their feedback, were operating in isolation within the model. After the inclusion of these items, the model was much more powerful. In total, more than 100 risk variables were tested as modeling inputs.
“We tried to stay away from demographics and stick with factors we can change through rehabilitation and policy,” Schelle says. “Ultimately, we don’t want to ‘scarlet letter’ prisoners, but rather highlight changing risk factors to help facility staff deal with them appropriately.”
The IDOC began testing different predictive models. Using conduct violations as outcome data, Schelle and Ellis compared model types and inputs to pinpoint the most accurate model. Interestingly, many traditional risk indicators such as current age and sentence length proved to be statistically insignificant. Whereas factors like job status and recent contraband history were shown to be more effective predictors of violence.
In the end, a decision tree was chosen as the model of choice. Decision trees use automatic interaction detection, a form of machine learning, to calculate risk scores in an explainable way – an important factor when creating policies, according to Schelle. “If the model shows a person is more violent because they are idle, then we can work on reducing idleness through policy.”
IDOC – Facts & Figures
Communicating risk to facility staff
With its new risk assessment model designed to predict violence rather than recidivism, the IDOC was halfway to its goal. Next, it needed a way to share the insight with prison staff.
The IDOC chose SAS Visual Analytics for this work. Now, every adult facility has access to dashboards containing pertinent risk data. The weekly hotlist, for example, shows which prisoners were recently recategorized as high-risk. This allows mental health, custody and program staff members to discuss these prisoners at weekly meetings and determine the best course of action. This could involve moving the inmate to different housing or assigning an extra guard to handle the inmate.
Also, by knowing which factors caused a prisoner to become high-risk, the IDOC can proactively address these issues to mitigate violence. “Research shows certain stressors cause individuals to act out in violent ways,” Schelle says. “By identifying where we can interrupt those stressors, we can help reduce violence and put these people in a better situation long-term.”
Staff assaults drop by 50%
By many accounts, the new risk assessment tool is working as planned. Compared to IRAS, the new tool is four times more accurate in predicting violence. As a result, staff assaults dropped by 50% during a six-month test period. And inmate-on-inmate assaults dropped by 20%.
Time savings is another huge benefit, according to the IDOC. To rescore prisoners weekly using the IRAS model – which relied on 30-minute interviews at intake to assess risk – it would take staff members 13,500 hours a week to process 27,000 inmates. And the results wouldn’t be nearly as accurate. The SAS model does this work labor-free.
Finally, the IDOC has implemented a number of policies and programs using the new model. For example, a mental health stepdown unit was formed to help integrate prisoners back into the general population. And new education programs have been created to train employees on mental health issues and management techniques.
Although the IDOC is still in the process of rolling out the model across its facilities and training staff on how to best use the information, Schelle says the project has been well-received, especially with frontline staff.
“It gives our line officers peace of mind,” she says. “The tool provides a new level of situational awareness that they previously didn't have. It makes a huge difference in their day-to-day work.”
본 문서에 나오는 결과는 본 문서에 설명된 특정 상황, 비즈니스 모델, 데이터 입력 및 컴퓨팅 환경에 적합하게 되어 있습니다. 각 SAS 고객의 경험은 고유한 것으로, 비즈니스 및 기술적 변수에 따라 달라집니다. 따라서 모든 서술은 비전형적인 것이라는 점을 고려해야 합니다. 실제 절약, 결과 및 성능 특성은 개별 고객의 구성 및 조건에 따라 달라질 수 있습니다. SAS는 모든 고객이 비슷한 결과를 달성할 수 있다고 보증하거나 진술하지 않습니다. SAS 제품과 서비스에 대한 유일한 보증은 해당 제품 및 서비스에 대한 서면 계약의 보증서에 명시되어 있습니다. 본 문서의 어떠한 내용도 추가 보증을 구성하는 것으로 해석될 수 없습니다. 고객은 SAS 소프트웨어의 성공적인 구현에 따라 합의된 계약적 교환 또는 프로젝트 성공 요약의 일환으로 성공 사례를 SAS와 공유했습니다. 브랜드 및 제품 명칭은 각 기업의 상표입니다.