Study 1
Davis and colleagues (2021) used a quasi-experimental design with propensity score matching to assess the effect of the Milwaukee Diversion program on convictions, rearrest, and time to first rearrest at the 2-year follow-up.
All defendants in the sample were arrested in 2013. Defendants were screened for program eligibility based on their Level of Services Inventory Short Version (LSI-R:SV) scores, as described above in the Program Description. Since all defendants (including those held at suburban facilities and those who were arrested on weekends, neither of whom were considered for inclusion in the program) were administered the LSI-R:SV, it was possible to identify a pool of comparison cases with similar scores to those defendants who were diverted into the program.
Demographic, criminal history, instant case outcome (identified as either the arrest that triggered entry into the diversion program or, for comparison defendants, the first arrest within the specified timeframe), and 2-year rearrest data were obtained for both diversion and potential comparison defendants from the Milwaukee District Attorney’s Office. The Milwaukee Administrative Office of the Courts provided LSI-R:SV risk assessment data—including both raw scores and risk classifications—for both diversion and comparison defendants. These scores were used to select potential comparisons for the diversion program before the propensity score calculation.
Propensity score matching was then used to statistically equalize treatment and comparison groups on an array of demographic, criminal history, and instant case variables. For all instant cases and priors, the charge type and severity were identified first, then the group differences in the means or frequencies on the entire array of demographic, prior arrest, current charge, and other instant case variables were calculated. Next, a backward stepwise logistic regression was performed. The predicted probability of treatment group membership in the final step of this regression—the propensity score—was used to match each treatment case to its closest comparison case, using a straightforward one-to-one matching algorithm.
After matching, the diversion treatment group (n = 139) was an average of 29.6 years old, and the comparison group (n = 139) averaged 27.0 years old. The sample was predominately male (71 percent of both the treatment and comparison groups), Black (53 percent of the treatment group, 53 percent of the comparison group), and White (40 percent of the treatment group, 40 percent of the comparison group). The treatment group had an average of 0.50 prior arrests, compared with 0.60 prior arrests for the comparison group. The most common instant case charge categories were drug possession (26 percent of both the treatment and comparison groups), theft (21 percent of the treatment group, 25 percent of the comparison group), and other charge (29 percent of the treatment group, and 22 percent of the comparison group). There were no statistically significant differences between the matched treatment and comparison groups on any measured characteristics.
To facilitate cross-site comparisons, odds ratios were computed for each outcome as an estimate of effect size. Survival analyses (Cox regression) were conducted to supplement the findings on the dichotomous rearrest outcomes. No subgroup analysis was conducted.
