ADDRESSING RECIDIVISM INTERVENING TO REDUCE TECHNICAL VIOLATIONS AND

ADDRESSING RECIDIVISM INTERVENING TO REDUCE TECHNICAL VIOLATIONS AND

ADDRESSING RECIDIVISM INTERVENING TO REDUCE TECHNICAL VIOLATIONS AND IMPROVE OUTCOMES FOR EX-OFFENDERS Preliminary Draft - Do Not Cite PROJECT TEAM EVAN ABSHER Kauffman Foundation DANIELLE FULMER City of South Bend KATE

PAWASARAT City of Milwaukee ERIC ROCHE City of Kansas City NOLAN ZAROFF City of Milwaukee LIL SEBASTIAN Team Mascot Applied Data Analytics | Team C2 Preliminary Draft - Do Not Cite 2

RESEARCH QUESTION What factors influence the likelihood that an individual will recidivate within two years for a technical parole or mandatory supervised release (MSR) violation? Applied Data Analytics | Team C2 Preliminary Draft - Do Not Cite 3 PROJECT GOALS Understand the effect of relevant factors on recidivism Develop a prioritization tool to identify those most likely to recidivate within 2 years for a technical or

MSR violation Applied Data Analytics | Team C2 Preliminary Draft - Do Not Cite 4 BACKGROUND: WHY 2 YEARS? 59.5% Recidivate within 2 Years* 100% Percent Releases 80% 60% 40% 20% 0% 0

6 12 18 24 30 36 42 48 54 60

Months Since Release Source: US Bureau of Justice Statistics, Recidivism rates of prisoners released in 2005 from prisons in 30 states * Here recidivate refers specifically to re-arrest, not necessarily re-incarceration Applied Data Analytics | Team C2 Preliminary Draft - Do Not Cite 5 BACKGROUND: WHY TECHNICAL? 20 percent to 31 percent of all admissions to prison in Illinois over the past 10 state fiscal years have been for technical MSR violations Source: Illinois Sentencing Policy Advisory Council, Drivers of the Sentenced Population: MSR Violators (Summer 2013) Applied Data Analytics | Team C2 Preliminary Draft - Do Not Cite

6 BACKGROUND: WHY TECHNICAL? Policy solutions might be simpler and more politically feasible: Is it as simple as better transit access? More flexible parole officer check-ins? Developing an achievable employment plan prior to exit? Applied Data Analytics | Team C2 Preliminary Draft - Do Not Cite 7 DESIGN & APPROACH Use machine learning to predict recidivism risk Target: individuals who exit prison

between 2005 and 2013 Outcome variable: recidivism within 2 years for a technical or MSR violation Applied Data Analytics | Team C2 Preliminary Draft - Do Not Cite 8 DESIGN & APPROACH MACHINE LEARNING PROCESS Prepare and explore data Engineer features Test models and select best

Evaluate results Applied Data Analytics | Team C2 Preliminary Draft - Do Not Cite 9 RELEVANT DATA SETS Illinois Department of Corrections admissions data Illinois Department of Corrections exit data Illinois Department of Employment Security wage record data Applied Data Analytics | Team C2 Preliminary Draft - Do Not Cite 10

EXIT POPULATION CHARACTERISTICS Race / Ethnicity Gender 0.4% 9.2% 29.0% 59.7% 10.8% 90.8% Male Applied Data Analytics | Team C2 Preliminary Draft - Do Not Cite

African American Hispanic/Latino White Other 11 EXIT POPULATION CHARACTERISTICS Number of Exits by Year 40,000 2007 2008 2009 Applied Data Analytics | Team C2

Preliminary Draft - Do Not Cite 28,864 35,462 2006 27,705 34,302 0 28,293 34,878 2005

10,000 28,872 36,764 20,000 37,894 30,000 2010 2011 2012 2013 12

EXIT POPULATION CHARACTERISTICS 26% of Exits were Previously Admitted for a Technical Violation Return Additional Other; 0; 0.4% 9.0% Mittimus; 0.01; 0.9% Direct from Court 33.6% Discharge & Recommitted Technical MSR

Violator 26.0% MSR Violator, New Sentence Return Additional Mittimus 30.1% Other Applied Data Analytics | Team C2 Preliminary Draft - Do Not Cite 13 EXIT POPULATION

CHARACTERISTICS Average days in prison during most recent stay: 445 Average number of stays: 1.69 between 2013 and 2015 Average days out before return: 574 Applied Data Analytics | Team C2 Preliminary Draft - Do Not Cite 14 TARGET POPULATION CHARACTERISTICS Recidivate within 2 Years Recidivate for Technical Violation 21.2% 39.9%

60.1% 78.8% Recidivate No Recidivate Applied Data Analytics | Team C2 Preliminary Draft - Do Not Cite Technical Other 15 MACHINE LEARNING What is machine learning? A method in which a computer program learns from experience (data) with respect to a task and a performance measure

Source: Rayid Ghani presentation, 2017 Applied Data Analytics | Team C2 Preliminary Draft - Do Not Cite 16 MACHINE LEARNING Policy relevance: 1. 2. 3. 4. Description Detection Prediction Behavior change Source: Rayid Ghani presentation, 2017

Applied Data Analytics | Team C2 Preliminary Draft - Do Not Cite 17 MACHINE LEARNING Magic Loop: Loops data through series of machine learning models Tests different parameters that can be adjusted for each model Applied Data Analytics | Team C2 Preliminary Draft - Do Not Cite 18 MACHINE LEARNING Machine Learning Methods Tried:

Decision Tree Extra Trees Gaussian Naive Bays Gradient Boosting K-Nearest Neighbor Applied Data Analytics | Team C2 Preliminary Draft - Do Not Cite

Logistical Regression Random Forest Support Vector Machines And others 19 MACHINE LEARNING Understanding Random Forests Ensemble machine learning method Best suited for classification problems Constructs many decision trees and gives most predicted outcome overall Assumes representative population sample Applied Data Analytics | Team C2 Preliminary Draft - Do Not Cite 20

MACHINE LEARNING Understanding Random Forests Number of features chosen at each split in the tree randomly selected Creating many trees allows us to combine them into a model that is not over-fitted Applied Data Analytics | Team C2 Preliminary Draft - Do Not Cite 21 MACHINE LEARNING Our Random Forest Model

Impute values for missing data 75 features included Split into training and testing sets Train model and apply to test data Set threshold at 10% for evaluation Applied Data Analytics | Team C2 Preliminary Draft - Do Not Cite 22 PRECISION/RECALL CURVES AT VARYING POPULATIONS Applied Data Analytics | Team C2 Preliminary Draft - Do Not Cite 23

MODEL RESULTS Baseline: 21% of sample has a violation within 2 years of exit For the 10% riskiest population: Accuracy: 81% Precision: 0.61 Recall: 0.19 Applied Data Analytics | Team C2 Preliminary Draft - Do Not Cite 24 MODEL RESULTS Most important classifiers (in order)

Gender (male) Marital status Race/ethnicity Veteran status Sex offender status Applied Data Analytics | Team C2 Preliminary Draft - Do Not Cite Employment plan Habitual offender Gang member

Drug use 25 LIMITATIONS (OF THE TEAM) Little or no prior knowledge of Illinois prison system Little or no prior knowledge of machine learning method Limited time to perform analysis Possible changes in how variables are recorded / measured over study period Applied Data Analytics | Team C2 Preliminary Draft - Do Not Cite 26 LIMITATIONS (OF OUR ASSUMPTIONS) Every individual / story is unique

Even technical violations are complex Dont incorporate results of interventions already explored by IL DOC Predictive model at exit cannot include subsequent relevant factors (e.g., successful employment after release) Applied Data Analytics | Team C2 Preliminary Draft - Do Not Cite 27 LIMITATIONS (OF THE ANALYSIS) Review from subject-matter expert Clarify scope and impact of data entry errors, changes in variable coding over time or institution Fine-tune model performance Improve model based on specific intervention threshold, e.g., 5% most at-risk v. 20% most at-risk

Bias in the data (imputation, structural, etc.) Applied Data Analytics | Team C2 Preliminary Draft - Do Not Cite 28 NEXT STEPS Effect of parole officer on parolee outcomes Effect of transit availability and accessibility to jobs Test model for collinearity, validate data using temporal methods, increase number of features Applied Data Analytics | Team C2 Preliminary Draft - Do Not Cite

29 P e rc e n t o f P ris o n E x its A s s o c ia te d w ith E m p lo y m e n t NEXT STEPS Employment Rates After Prison Exit 30 25 20 15 10 5 0 Q1 Q2 Q3

Q4 Q5 Q6 Q7 Q8 Quarter following Prison Exit No Recidivism within 2 years Recidivism within 2 years for technical violation Recidivism within 2 years Source: Illinois Dept. of Corrections; Illinois Dept. of Employment Security Applied Data Analytics | Team C2 Preliminary Draft - Do Not Cite

30 NEXT STEPS Imprisoning someone is costly; reducing recidivism lowers costs But targeting criminal violators is complex and politically sensitive Intervening in technical violations could be more feasible, low-cost way of addressing incarceration rates Applied Data Analytics | Team C2 Preliminary Draft - Do Not Cite 31 NEXT STEPS [W]eve seen time and time again that

mathematical models can sift through data to locate people who are likely to face great challenges, whether from crime, poverty, or education. Its up to society whether to use that intelligence to reject and punish them or to reach out to them with the resources they need. - Cathy ONeil, Weapons of Math Destruction Applied Data Analytics | Team C2 Preliminary Draft - Do Not Cite 32 QUESTIONS? AND THANK YOU! Applied Data Analytics | Team C2 Preliminary Draft - Do Not Cite 33

Recently Viewed Presentations

  • Communication - Centers for Disease Control and Prevention

    Communication - Centers for Disease Control and Prevention

    Communication Skills for Building Rapport During Contact Investigation Interviewing. Explain to the participants that this section will focus on basic communication skills that can be applied to any health care worker/case encounter; however, the main goal will be to the...
  • Stress Responses & Gene Expression

    Stress Responses & Gene Expression

    Stress Responses & Gene Expression plants must adapt to stresses because of their sedentary lifestyle Fig. 22.2, Buchanan et al. * * Some have common features: e.g., #2-5 can all create water shortages for plant cells.
  • Concentration of Solutions - Romona Olton

    Concentration of Solutions - Romona Olton

    Solutions are homogeneous mixtures consisting of two or more components. The major component of a solution is known as the solvent and the minor component is known as the solute. ... QuestionConverting from g dm-3 to mol dm-3 (1) Some...
  • Chapter 7 Energy Storage What are the main

    Chapter 7 Energy Storage What are the main

    Depending on their group number, elements have different oxidation states, based on their preference to gain electrons (non-metals) or lose electrons (metals). Galvanic Cells: A Laboratory Model A zinc-copper galvanic cell, containing a salt bridge separator between anode and cathode...
  • How to make a quote sandwich - Wheeler Web

    How to make a quote sandwich - Wheeler Web

    This example was a very simple version of a quote sandwich. The context, quote, and analysis can, and often will be, more than one sentence each. The main purpose of quoting is to support your analysis.
  • Panovníci Českých zemí

    Panovníci Českých zemí

    upravil a hudbu dodal Kozoroh - květen 2010. vypracoval: Lukáš Salivar Prosinec 2003 Hudbu dodal: Květen 2010
  • Safety Plan ALL SOULS UU CONGREGATION All Souls

    Safety Plan ALL SOULS UU CONGREGATION All Souls

    An Observer is a Soul who has volunteered to watch for signs of problems and is trained to communicate with others to handle situations as peacefully as possible. An observer will have a communication device with a headset and microphone...
  • Financial Liaison Meeting Presented by the AU Business

    Financial Liaison Meeting Presented by the AU Business

    Negotiated Fringe Benefit Rate. New rates will go into effect October 1, 2011. For fiscal year 2011 carryover, benefit budgets will be treated as salary and wage budgets are currently