Developing the Risk Instrument

Sidebar to the article Reconsidering the Project Greenlight Intervention: Why Thinking About Risk Matters by James A. Wilson and Christine Zozula

We developed our risk instrument from data on individual attributes that have been strongly associated with criminal recidivism.[1] Our data include numerous measures of criminal history such as current offense; numbers of misdemeanor and felony arrests and convictions; bench warrants; and database indicators for drugs, weapons and firearm offenses. Standard demographic data such as age, race/ethnicity and educational level, as well as some information on substance use were contained in the data files originally provided by the New York State Department of Correctional Services. We estimated a number of different models for constructing a risk scale. In doing so, we paid special attention to the literature on the predictors of offender recidivism, but we also tested all of the variables available to us and considered their potential meaning for respondent outcomes.

Following Gottfredson and Snyder, we used logistic regression to obtain unstandardized coefficients for variables that predicted new arrests.[2] Variables that were statistically associated included prior parole revocations, prior felony arrests, bench warrant indicators, substance use measures, release age and borough of release.[3] We included borough of release because it could potentially indicate opportunities and networks available to individuals recently released from prison. Given the lack of dynamic risk predictors (i.e., predictors amenable to change, such as antisocial attitudes and peer associations, substance use, poor social control/impulsivity, family environment, and education/employment), geographic location may be the next best thing because it suggests neighborhood characteristics such as employment opportunities, living arrangements and exposure to pro-social peers. Once our scale was constructed, we defined three risk levels for sample size reasons, but rather than simply dividing the scale into thirds, we selected the bottom 30 percent as "low risk," the top 30 percent as "high risk," and the middle 40 percent as "medium risk."[4]

Chart: Assessing the Risk Scale: Rearrest by Risk Level of Study Participants
View larger image and text description

Figure 1 is a very basic illustration of how well our risk instrument discriminates between those classified as low, medium and high risk. A 27 percentage point difference distinguishes the difference between low and medium risk, and a 22.3 percentage point difference exists between medium and high risk. The degree of discrimination across risk categories for any arrest is statistically significant. To our knowledge, no one has established specific criteria for what constitutes low-, medium- and high-risk offenders, but we believe our scale represents, to a reasonable degree, these conceptual categories.


[1] See, e.g., Andrews, D.A., Ivan Zinger, Robert D. Hoge, James Bonta, Paul Gendreau, and Francis T. Cullen, "Does Correctional Treatment Work? A Clinically Relevant and Psychologically Informed Meta-Analysis," Criminology 28 (1990): 369–404. Andrews, D.A., and James Bonta, The Psychology of Criminal Conduct, 4th Edition, Cincinnati, OH: Anderson Publishing (2006). Gendreau, Paul, Tracy Little, and Claire Goggin, "A Meta-Analysis of the Predictors of Adult Offender Recidivism: What Works!" Criminology 34 (1996): 575–608.

[2] Gottfredson, Don M., and Howard M. Snyder, The Mathematics of Risk Classification: Changing Data into Valid Instruments for Juvenile Courts (pdf, 48 pages), OJJDP Report, Washington, D.C.: Office of Juvenile Justice and Delinquency Prevention, July 2005, NCJ 209158.

[3] Educational level and race/ethnicity were not included because they were not predictive in any of the models tested and because the use of race/ethnicity variables in such scales raises ethical concerns.

[4] As one might expect, whether we divided our risk levels into thirds, quartiles or some other grouping made little difference. We ultimately decided on the 30-40-30 distribution in order to capture those who were at slightly lower and slightly higher risk, but in practical terms, other divisions did not yield different results.

Date Created: November 3, 2011