Methods: Data were obtained from administrative databases of a southeastern state Department of Correction. The study dataset included 2,681 prisoners who were eligible for SOT, released from prison between1999-2003, and were tracked until 2008. Because of limited program funding, not all prisoners who volunteer for SOT receive treatment. Covariate-control survival analysis models were tested to assess the effects of volunteerism (Aim 1) and treatment effects (Aim 2) on recidivism rates. Then we used propensity score matching methods - nearest neighbor and generalized boosted regression - to balance samples on factors contributing to selection bias. Several models were tested until a balanced sample was achieved that also retained the maximum sample size. Using Cox proportional hazards for post-matching analysis, we examined recidivism outcomes for participants (n=264) that volunteered for treatment, of whom some completed (n=132) and some were not enrolled (n=132). We then examined a matched sample (n=470) of untreated volunteers (n=235) and nonvolunteers (n=235).
Results: Results from post-matching analysis indicated that volunteerism had no impact on recidivism (HR=1.03; p=.791), and SOT participants had lower recidivism rates than untreated participants (HR=.50; p=.006). Comparisons of propensity score models to covariate-control models revealed changes in magnitude, direction, and significance of some variables. For example, in the volunteerism sample, magnitude of effects differed for volunteerism, personality disorders, and rapist, child molester, and post-release supervision status. Statistical significance changed for personality disorder, child offender, and cognitive-behavioral program variables. Moreover, although not statistically significant, the direction of volunteerism effects changed. Further, propensity score models with balanced and matched samples varied in detecting treatment effects on recidivism depending on the sample size retained. Small adjustments to caliper sizes increased statistical power to detect effects.
Implications: This study has important social work implications. Robust statistical methods can be used to overcome some limitations of administrative data. The findings suggest prison-based SOT is effective at reducing recidivism. However, in the absence of effective interventions, prisoners' motivation to change may not be sufficient to ensure post-release behavior changes. This investigation demonstrates the influence of selection bias on parameter estimates when such bias is not accounted for with rebalanced samples. Finally, researchers should not rely on the sophistication of propensity score methods alone. The importance of power remains and multiple methods should be tested to maximize sample retention.