After reading the Prison-Based Chemical Dependency Treatment in Minnesota: An Outcome Evaluation and Substance Abuse Treatment for Adults in the Criminal Justice System articles, discuss the following in your initial post: What are the pros and cons of su

Prison-based chemical dependency treatment in Minnesota: An outcome evaluation Grant Duwe Published online: 25 February 2010 # Springer Science+Business Media B.V. 2010 AbstractUsing a retrospective quasi-experimental design, this study evaluated the effectiveness of prison-based chemical dependency (CD) treatment by examining recidivism outcomes among 1,852 offenders released from Minnesota correctional facilities during 2005. Because recidivism data were collected on the 1,852 offenders through the end of 2008, the average follow-up period was 42 months. To minimize the threat of selection bias, propensity score matching was used to create a comparison group of 926 untreated offenders who were not, for the most part, significantly different from the 926 treated offenders. Results from the Cox regression analyses revealed that participating in prison-based CD treatment significantly reduced the hazard ratio for recidivism by 17–25%. Although dropping out of treatment did not increase the risk of recidivism, completing treatment significantly lowered it by 20–27%. The findings also suggest that long-term treatment programs were not as effective as short- or medium-term programs in reducing the risk of recidivism. The study concludes by discussing the implications of these findings.

KeywordsSubstance abuse.

Chemical dependency.

Drug treatment.

Prison.

Recidivism.

Propensity score matching 1 Introduction The impact of substance use on the criminal justice system is substantial. Research has long shown that alcohol and/or illicit drugs figure prominently in criminal offending. In Marvin Wolfgang’s landmark study on homicide in Philadelphia during the 1950s, he reported that alcohol was consumed by either the victim or the offender in approximately two-thirds of the cases (Wolfgang1958). In a survey of J Exp Criminol (2010) 6:57–81 DOI 10.1007/s11292-010-9090-8 G. Duwe (*) Minnesota Department of Corrections, 1450 Energy Park Drive, Suite 200, St. Paul, MN 55108-5219, USA e-mail: [email protected] nearly 7,000 jail inmates, Karberg and James (2005) found that 33% reported being under the influence of alcohol at the time of the offense. Also, in a recent study of 224 Minnesota sex offenders who recidivated with a sex crime, either the victim or the offender had used alcohol and/or drugs at the time of the offense in at least 31% of the assaults (Duwe et al.2008).

Among state and federal prisoners incarcerated in 2004, Mumola and Karberg (2006) reported that 32% committed their offenses under the influence of drugs, and 56% had used drugs in the month preceding the offense. The highest percentages of drug use were found for drug offenders, followed closely by those incarcerated for property offenses. For example, 44% of drug offenders and 39% of property offenders indicated using drugs at the time of the offense. Moreover, the rate of drug use in the month prior to the offense was 72% for drug offenders and 64% for property offenders.

The use and abuse of substances is linked not only to involvement in criminal activity but also to the growth of the prison population, particularly over the last few decades. Due in part to increased penalties resulting from the War on Drugs, the federal and state prison population has more than doubled in size over the last 20 years (Beck and Gilliard1995; Sabol et al.2007). Drug offenses, moreover, accounted for 53% of all federal prisoners in 2006 and 20% of state inmates in 2005 (Harrison and Beck2006; Sabol et al.2007). Within Minnesota, the percentage of drug offenders in the total inmate population grew from 4% in 1989 to 20% in 2008 (Minnesota Department of Corrections2007b,2008). The percentage of drug offenders, however, represents only a fraction of those who are in need of chemical dependency (CD) treatment. Indeed, approximately 85% of the offenders entering Minnesota state prisons during 2006 were determined to be chemically abusive or dependent (Minnesota Department of Corrections2007a).

Given the relatively high rate of substance abuse and dependency among incarcerated offenders, efforts to reduce their risk of reoffense often include the provision of prison-based CD treatment. Previous evaluations of prison-based CD treatment have concentrated mainlyon programs based on the therapeutic community (TC) model. Originating in England during the late 1940s, the TC model regards chemical dependency as a symptom of an individual’s problems rather than the problem itself (Patenaude and Laufersweiller-Dwyer2002). Viewing substance abuse as a disorder that affects the whole person, the TC model attempts to promote comprehensive pro-social changes by encouraging participants to contribute to their own therapy, as well as that of others, through activities such as therapy, work, education classes, and recreation (Klebe and O’Keefe2004).

Individual and group counseling, encounter groups, peer pressure, role models, and a system of incentives and sanctions often comprise the core of treatment interventions within a TC program (Welsh2002). Moreover, to foster a greater sense of community, participants within a prison setting are housed separately from the rest of the prison population.

Previous studies have evaluated prison-based TC programs for federal prisoners (Pelissier et al.2001) as well as for state prisoners in California (Prendergast et al.

2004; Wexler et al.1999), Delaware (Inciardi et al.1997,2004), New York (Wexler et al.1990), Oregon (Field1985), Pennsylvania (Welsh2007) and Texas (Knight et al.1997,1999). In general, the findings from these studies suggest that prison-based 58G. Duwe treatment can be effective in reducing recidivism and relapse. Indeed, in the most recent meta-analysis of the incarceration-based drug treatment literature, Mitchell et al. (2007) found that treatment significantly decreased subsequent criminal offending and drug use in their review of 66 evaluations. The average treatment effect sizes for recidivism and drug use were odds ratios of 1.37 and 1.28, respectively (Mitchell et al.2007).

The most promising outcome results have been found for offenders who complete prison-based TC programs, especially those who participate in post-release aftercare (Inciardi et al.2004; Mitchell et al.2007; Pearson and Lipton1999). In addition, Wexler et al. (1990) reported that treatment effectiveness is related to the length of time an individual remains in treatment, but only up to a point. As time in the TC program increased, so too did the time until rearrest. Time to rearrest was shorter, however, for offenders who had been in the TC program longer than 12 months.

Despite the positive findings from prior outcome evaluations, most of these studies have been limited in one or more ways. Welsh (2002) notes, for example, that previous evaluations have had small sample sizes, have had faulty research designs, and have devoted too little attention to interactions between inmate characteristics, treatment processes, and treatment outcomes. Moreover, Pelissier and colleagues (2001) identified selection bias as the most significant shortcoming of prior studies on prison-based CD treatment. In evaluations of treatment effective- ness, selection bias refers to differences—both observable and unobservable— between the treated and untreated groups that make it difficult to determine whether the observed effects are due to the treatment itself or to the different group compositions. Therefore, although previous evaluations have found that recidivism rates are generally lower for offenders who participate in treatment, this difference may not necessarily be due to the treatment itself, but rather to other differences between treated and untreated offenders.

In their evaluation of the Federal Bureau of Prison’s Drug Abuse Treatment Program, Pelissier and colleagues (2001) used two methods—the instrumental variable approach and the Heckman selection bias model—to control for selection bias. 1After doing so, Pelissier et al. (2001) still found that, within 3 years of release, 31% of treated male offenders had been rearrested in comparison to 38% of the untreated male offenders, which amounted to a recidivism reduction of 19%.

Although treated female offenders were not significantly less likely to recidivate than untreated female offenders, they were 18% less likely to use drugs in the 36 months following release from prison. Treated male offenders, meanwhile, were 15% less likely to have post-release drug use than untreated male offenders.

1.1 Present study Using a retrospective quasi-experimental design, this study evaluates the effective- ness of CD treatment provided within the Minnesota Department of Corrections 1The instrumental variable approach involves locating a variable that is related to selection into treatment but is unrelated to the outcome variable. The variance from the instrumental variable is then used to estimate the impact of treatment on the outcome measure. The Heckman method, on the other hand, requires that the selection pressures be jointly modeled into the sample and post-release outcome (Pelissier et al.2001).

Prison-based chemical dependency treatment in Minnesota: An outcome evaluation 59 (MNDOC) by comparing recidivism outcomes between treated and untreated offenders released from prison in 2005. As discussed later in more detail, propensity score matching (PSM) was used to individually match the untreated offenders with those who received CD treatment. Similar to the instrumental variable and Heckman approaches used by Pelissier and colleagues (2001), PSM is a method designed to control for selection bias. More specifically, PSM minimizes the threat of selection bias by creating a comparison group whose probability of entering treatment was similar to that of the treatment group. Although PSM has been used in at least one recent study on community-based CD treatment (Krebs et al.2008), this study is one of the first to use it in a prison-based treatment evaluation.

In addition to PSM, this study attempts to further control for rival causal factors by analyzing the data with Cox regression, which is widely regarded as the most appropriate multivariate statistical technique for recidivism analyses. Moreover, by comparing 926 treated offenders with a matched group of 926 untreated offenders, the sample size used for this study (n= 1,852) is one of the larger prison-based CD treatment studies to date. Finally, to achieve a more complete understanding of the effects of prison-based treatment, multiple treatment and recidivism measures were used.

Despite these strengths, there are several limitations worth noting. First, in measuring the effectiveness of CD treatment, the two most common outcome measures are substance abstention and criminal recidivism. Although abstention is an important and arguably more sensitive measure of CD treatment effectiveness, data on post-release substance use were not available for this study. Therefore, in focusing exclusively on recidivism, this study may not fully capture whether CD programming is effective. Second, in providing a continuum of care from the institution to the community, aftercare programming is often considered a critical component to effective CD treatment. Data on post-release aftercare programming, however, were not available on the offenders examined here. As a result, the differences observed between the treatment and comparison groups (or lack thereof) may be attributable, in part, to differences in the extent to which offenders participated in aftercare programming while in the community.

These limitations notwithstanding, this study attempts to address several questions central to the substance-abuse treatment literature. First, does treatment reduce offender recidivism? Second, what effect does treatment outcome (i.e., drop out or complete) have on reoffending? Finally, what impact does program duration have on recidivism?

In the following section, this study describes the provision of CD treatment within the MNDOC. After discussing the data and methods used in this study, the results from the statistical analyses are presented. This study concludes by discussing the implications of the findings for the prison-based treatment literature.

2 Chemical dependency treatment in the MNDOC Shortly after their admission to prison in Minnesota, offenders undergo a brief (20– 40 min) CD assessment conducted by a licensed assessor. Of the newly admitted offenders who receive a CD assessment, approximately 85% are directed to enter CD 60G. Duwe treatment because they are determined to be chemically abusive or dependent. In making CD diagnoses, which are basedon both self-report and collateral information, CD assessors utilize DSM-IV criteria for substance abuse. Among the criteria for abuse are problems at work or school, not taking care of personal responsibilities, financial problems, engaging in dangerous behavior while intoxi- cated, legal problems, problems at home or in relationships, and continued use despite experiencing problems. The criteria for dependence, meanwhile, include increased tolerance; withdrawal symptoms; greater use than intended over a relatively long period of time, inability to cut down or quit; a lot of time spent acquiring, using, or recovering from use; missing important family, work, or social activities; and knowledge that continued use would exacerbate a serious medical or psychological condition. Although the vast majority of newly admitted offenders are considered to be CD abusive or dependent, not all treatment-directed offenders have the opportunity to participate in prison-based treatment since the number of treatment-directed offenders (nearly 3,000 annually) exceeds the number of treatment beds available (about 1,800 annually).

TheMNDOCcurrentlyusesinformationrelatingtooffenderneedsand recidivism risk in prioritizing inmates for treatment. This information, however, was not routinely considered from 2002–2005, the period of time covered in this study. Rather, among offenders directed to treatment, prioritization decisions were based primarily on the amount of time remaining to serve. Offenders with shorter lengths of time until their release from prison were often selected over those with more time to serve.

During the 2002–2005 period, the MNDOC provided CD programming to both male and female offenders in seven of the 11 state facilities that house adult inmates. Although there are variations among the different programs provided at each facility, all of the CD treatmentoffered by the MNDOC is modeled on TC concepts. Housed separately from the rest of the prison population, offenders admitted to treatment were involved in 15–25 h of programming per week. The CD programs, which maintained a staff-to-inmate ratio of 1:15, emphasized each offender’s personal responsibility for identifying and acknowledging criminal and addictive thinking and behavior. Moreover, the CD programming generally included educational material that addressed the signs and symptoms of CD, the effects of drug use on the body, the effects of chemical use on family and relationships, and the dangers of drug abuse. In addition to completing an autobiography that focused on prior chemical use, program participants completed work relating to relapse prevention.

The MNDOC offered short-term (90 days), medium-term (180 days), and long- term (365 days) CD programming during the 2002–2005 period. The short-term programs, which were primarily psycho-educational with minimal individual counseling, emphasized the relationship between substance-abuse issues and criminal behavior. Participants in these programs were expected to increase their level of active participation as they progressed through the program. The medium- and long-term programs, on the other hand, included education, individual counseling, and group counseling components. Therefore, aside from program duration, the main distinction between the short-term programs and the medium- and long-term programs was that the former contained little emphasis on individual or Prison-based chemical dependency treatment in Minnesota: An outcome evaluation 61 group counseling, primarily due to the relatively short period of time over which to deliver the programming.

In 2006, the MNDOC refocused its CD programs to long-term treatment of at least 6 months or more. The decision to discontinue the short-term programming was due, in part, to evidence which seemed to suggest that short-term programs are not as effective as ones that are longer in duration (Minnesota Office of the Legislative Auditor2006). More specifically, in its report on substance-abuse treatment across the state, the Minnesota Office of the Legislative Auditor found that recidivism rates for short-term program participants were higher than those for offenders who participated in medium- and long-term programs. However, the simple bivariate analyses performed by the Minnesota Office of the Legislative Auditor did not control for factors known to affect recidivism (e.g., criminal history, age at release, institutional disciplinary history, type of offense, etc.). Therefore, rather than demonstrating that short-term treatment is less effective, the higher recidivism rates for short-term participants may simply reflect that they had, in comparison to the medium- and long-term participants, a greater risk of reoffense prior to entering treatment.

3 Data and methodology This study uses a retrospective quasi-experimental design to determine whether CD programming has an impact on recidivism. More specifically, the effectiveness of CD treatment was evaluated by comparing recidivism outcomes between treated offenders and a matched comparison group of untreated offenders who were released from prison in 2005. To ensure that offenders in the comparison group were similar to those in the treatment group, the population for this study consisted only of inmates who received a positive CD assessment (i.e., they were determined to be chemically abusive or dependent) and were directed to enter CD treatment prior to their release from prison. In addition, because valid and reliable CD treatment data were not available prior to 2002, the population from which the treatment and comparison groups were drawn includes only offenders who were admitted to prison after December 31, 2001. As a result, this study does not include offenders with longer sentences who were directed to CD treatment. 2Still, the study captured the vast majority of offenders released in 2005 who were directed to CD treatment given that only 8% of the releasees from 2005 were admitted to prison prior to 2002.

Overall, there were 3,499 offenders directed to CD treatment who were admitted to prison after 2001 and released during 2005. Of these 3,499 offenders, there were 1,164 who participated in CD treatment while in prison. Of the remaining 2,335 offenders, there were 35 who refused to enter CD treatment. Because the 35 treatment refusers did not participate in treatment, these offenders were removed from the study so as not to bias the results from the statistical analyses. Before doing so, however, an attempt was made to remove an additional source of bias by using 2In Minnesota, the sentences for offenders committed to the Commissioner of Corrections consist of two parts: a minimum prison term equal to two-thirds of the total executed sentence, and a supervised release term equal to the remaining one-third.

62G. Duwe PSM to identify a comparison group of offenders from the pool of untreated offenders (n= 2,300) who were not offered treatment, often due to a lack of available treatment beds. The procedures used to address potential bias resulting from treatment refusers are discussed later in this section.

3.1 Dependent variable Recidivism, the dependent variable in this study, was defined as a (1) rearrest, (2) felony reconviction or (3) reincarceration for a new sentence. Recidivism data were collected on offenders through December 31, 2008. Considering that offenders from both the treatment and comparison groups were released during 2005, the follow-up time for the offenders examined in this study ranged from 36–48 months.

Data on arrests and convictions were obtained electronically from the Minnesota Bureau of Criminal Apprehension. Reincarceration data were derived from the Correctional Operations Management System (COMS) database maintained by the MNDOC. The main limitation with using these data is that they measure only arrests, convictions, or incarcerations that took place in Minnesota. As a result, the findings presented later likely underestimate the true recidivism rates for the offenders examined here.

To accurately measure the total amount of time offenders were actually at risk to reoffend (i.e.,“street time”), it was necessary to account for supervised release revocations in the recidivism analyses by deducting the amount of time they spent in prison from the time of release to the end of the observation period or to the first recidivism event, whichever came first. Failure to deduct time spent in prison as a supervised release violator would artificially increase the length of the at-risk periods for these offenders. Therefore, the time that an offender spent in prison as a supervised release violator was subtracted from his/her at-risk period, but only if it preceded a rearrest, a reconviction, a reincarceration for a new offense, or if the offender did not recidivate (i.e., no rearrest, reconviction, or reincarceration for a new offense) prior to January 1, 2009.

3.2 Treatment variables Given that the central purpose of this study is to determine whether CD programming has an impact on recidivism, CD treatment is the principal variable of interest. In an effort to achieve a more complete understanding of its potential impact on recidivism, six different treatment measures were used in this study.

The first CD treatment variable compares offenders who entered CD treatment with a comparison group of similar offenders who did not. As such, CD treatment was measured as“1”for offenders who participated in treatment between the time of admission (after 2001) and release (2005) from prison. Offenders who did not participate in CD treatment (the comparison group) were given a value of“0.” Two measures were used to assess the impact of treatment outcome on reoffending. The variable, treatment completer, compares offenders who completed treatment or successfully participated until release (1) with untreated offenders (0).

The treatment dropout variable, on the other hand, compares offenders who quit or were terminated from treatment (1) with untreated offenders (0).

Prison-based chemical dependency treatment in Minnesota: An outcome evaluation 63 Three measures were created to assess the effects of program duration. As noted above, during the 2002–2005 period, the MNDOC had short-term, medium-term, and long-term CD treatment programs. The variable, short-term program, compares short-term participants (1) with untreated offenders (0). The medium-term program variable contrasts medium-term participants (1) with untreated offenders (0), whereas the long-term program variable is dichotomized as long-term participants (1) or as untreated offenders (0).

3.3 Independent variables The independent, or control, variables included in the statistical models were those that were not only available in the COMS database but also those that might theoretically have an impact on whether an offender recidivates. These variables cover the salient factors that are either known or hypothesized to have an impact on recidivism. The following lists these variables and describes how they were created:

Offender Sex: dichotomized as male (1) or female (0).

Offender Race: dichotomized as minority (1) or white (0).

Age at Release: the age of the offender in years at the time of release based on the date of birth and release date.

Prior Felony Convictions: the number of prior felony convictions, excluding the conviction(s) that resulted in the offender’s incarceration.

Metro Area: a rough proxy of urban and rural Minnesota, this variable measures an offender’s county of commitment, dichotomizing it into either metro area (1) or Greater Minnesota (0). The seven counties in the Minneapolis/St. Paul metropolitan area include Anoka, Carver, Dakota, Hennepin, Ramsey, Scott, and Washington. The remaining 80 counties were coded as non-metro area or Greater Minnesota counties.

Offense Type: five dichotomous dummy variables were created to quantify offense type; i.e., the governing offense at the time of release. 3The five variables were person offense (1 = person offense, 0 = non-person offense); property offense (1 = property offense, 0 = non-property offense); drug offense (1 = drug offense, 0 = non-drug offense); felony driving while intoxicated (DWI) offense (1 = DWI offense, 0 = non-DWI offense); and other offense (1 = other offense, 0 = non-other offense). The other offense variable serves as the reference in the statistical analyses.

Length of Stay (LOS): the number of months between prison admission and release dates.

Institutional Discipline: the number of discipline convictions received during the term of imprisonment prior to release.

Dependency Assessment: dichotomized as either (1) chemically dependent or (0) chemically abusive for offenders who received positive chemical dependency assessments at intake. 3The“governing offense”is the crime carrying the sentence on which an offender’s scheduled release date is based. Although offenders may be imprisoned for multiple offenses, each with its own sentence, the governing offense is generally the most serious crime for which an offender is incarcerated.

64G. Duwe Length of Post-Release Supervision: the number of months between an offender’s first release date and the end of post-release supervision; i.e., the sentence expiration or conditional release date, the greater of the two.

Type of Post-Release Supervision: four dichotomous dummy variables were initially created to measure the level of post-release supervision to which offenders were released. The four variables were intensive supervised release (ISR) (1 = ISR, 0 = non-ISR); supervised release (SR) (1 = SR, 0 = non-SR); work release (1 = work release, 0 = non-work release); and discharge (1 = discharge or no supervision, 0 = released to supervision). Discharge is the variable that serves as the reference in the statistical analyses.

Supervised Release Revocations (SRRs): the number of times during an offender’s sentence that s/he returned to prison as a supervised release violator.

4 Propensity score matching PSM is a method that estimates the conditional probability of selection to a particular treatment or group given a vector of observed covariates (Rosenbaum and Rubin 1984). The predicted probability of selection, or propensity score, is typically generated by estimating a logistic regression model in which selection (0 = no selection; 1 = selection) is the dependent variable while the predictor variables consist of those that theoretically have an impact on the selection process. Once estimated, the propensity scores are then used to match individuals who entered treatment with those who did not. Thus, one of the main advantages with using PSM is that it can simultaneously“balance”multiple covariates on the basis of a single composite score. Although there are a number of different matching methods available, this study used a“greedy”matching procedure that utilized a without replacement method in which treated offenders were matched to untreated offenders who had the closest propensity score (i.e.,“nearest neighbor”) within a caliper (i.e., range of propensity scores) of 0.10. 4 In matching untreated offenders with treated offenders on the conditional probability of entering treatment, PSM reduces selection bias by creating a counterfactual estimate of what would have happened to the treated offenders had they not participated in treatment. PSM has several limitations, however, that are worth noting. First, in order to produce unbiased treatment effect estimates, the selection model must contain all of the variables related to the selection process and the outcome variable, and these variables must be measured without error (Berk 2003). Consequently, because propensity scores are based on observed covariates, PSM is not robust against“hidden bias”from unmeasured variables that are associated with both the assignment to treatment and the outcome variable. Second, there must be substantial overlap among propensity scores between the two groups in order for PSM to be effective (Shadish et al.2002); otherwise, the matching process will yield incomplete or inexact matches. Finally, as Rubin (1997) points out, PSM tends to work best with large samples. 4The greedy procedure is a matching algorithm that generates fixed matches. In contrast, optimal matching algorithms produce matches after reconsidering all previously made matches.

Prison-based chemical dependency treatment in Minnesota: An outcome evaluation 65 Although somewhat limited by the data available, an attempt was made to address potential concerns over unobserved bias by including as many theoretically relevant covariates (17) as possible in the propensity score models. More important, however, Rosenbaum bounds sensitivity analyses were conducted to evaluate the extent to which the treatment effects obtained are robust to the possibility of hidden bias. In addition, this study later demonstrates that there was substantial overlap in propensity scores between the treated and untreated offenders. Further, the sample- size limitation was addressed by assembling a relatively large number of cases (n= 3,394) on which to conduct the propensity score analyses.

4.1 Matching treatment refusers and non-refusers In an effort to minimize the bias resulting from treatment refusers, an attempt was made to identify a comparison group of untreated offenders who were not offered treatment in order to remove these offenders from the comparison group pool.

Propensity scores were computed for the 35 treatment refusers and the 2,300 untreated offenders by estimating a logistic regression model in which the dependent variable was refusal of treatment (i.e., the 35 treatment refusers were assigned a value of“1”, while the 2,300 untreated offenders in the comparison group pool received a value of“0”). The predictors were the 17 control variables described earlier. After obtaining propensity scores on the 2,335 offenders, a greedy matching procedure was used to match 35 untreated offenders not offered treatment with the 35 treatment refusers.

Of the 1,199 offenders who received a treatment offer, there were 35 who refused, resulting in a refusal rate of 3%. 5If a similar refusal rate is assumed among the 2,300 offenders not offered treatment, then approximately 70 of the untreated offenders would have refused a treatment offer. As a result, it was necessary to remove an additional 35 untreated offenders who were not offered treatment. Accordingly, after removing the 35 untreated offenders who were matched to the treatment refusers, a second logistic regression model was estimated to generate propensity scores on the 35 offenders who refused treatment and the remaining 2,265 who did not receive a treatment offer. A greedy matching procedure was then used, once again, to match 35 untreated offenders without a treatment offer with the 35 treatment refusers.

Along with the 35 treatment refusers, the 70 matched offenders not offered treatment were removed from the remaining analyses. In doing so, the number of untreated offenders in the comparison group pool was reduced by 105 from 2,335 to 2,230.

4.2 Matching treated and untreated offenders Similar to the approach described above with treatment refusers, propensity scores were calculated for the 1,164 treated offenders and the 2,230 untreated offenders by estimating a logistic regression model in which the dependent variable was participation in prison-based treatment (i.e., the 1,164 group offenders were assigned a value of“1”, while the 2,230 offenders in the comparison group pool received a 5The 1,199 offenders include the 1,164 who participated in treatment and the 35 who refused to enter treatment.

66G. Duwe value of“0”). The predictors were the 17 control variables used in the statistical analyses (see Table1). As shown in Fig.1, there was substantial overlap in propensity scores between the treated and untreated offenders, even though the difference in mean propensity score was statistically significant at the .01 level (see Table2).

After obtaining propensity scores for the 3,394 offenders, a greedy matching procedure was used to match the untreated offenders with the treated offenders.

Because the matching process is often a trade-off between the size of the bias reduction and the proportion of cases that can be matched (DiPrete and Gangl2004), matches were not obtained for all of the treated offenders. However, in using a relatively narrow caliper of 0.10, matches were found for 926 treatment participants, which accounts for 80% of the total number of treated offenders (n= 1,164).

Table2presents the covariate and propensity score means for both groups prior to matching (“total”) and after matching (“matched”). In addition to tests of statistical significance (“t-testp-value”), Table2provides a measure (“Bias”) developed by Table 1Logistic regression model for assignment to treatment Predictors Coefficient Standard error Male–0.315* 0.134 Minority–0.288** 0.085 Age at release (years)–0.002 0.005 Metro 0.003 0.084 Prior felonies–0.023 0.013 Offense type Person offenders–0.027 0.138 Property offenders 0.027 0.139 Drug offenders–0.008 0.136 DWI offenders 2.051** 0.338 Assessed as dependent 0.535** 0.081 Institutional discipline–0.046** 0.012 Length of stay (months) 0.056** 0.004 Length of supervision (months)–0.013** 0.003 Supervision type ISR 1.542** 0.253 Supervised release 2.143** 0.236 Work release 1.814** 0.260 SR revocations 0.056 0.062 Constant–2.795 0.330 n3,394 Log-likelihood 3805.104 Nagelkerke R 2 0.210 **p< .01 *p< .05 Prison-based chemical dependency treatment in Minnesota: An outcome evaluation 67 Rosenbaum and Rubin (1985) that quantifies the amount of bias between the treatment and control Bias¼ 100 Xt Xc ffiffiffiffiffiffiffiffiffiffiffiffiffiffi S2 tþS2 c ðÞ 2 q samples (i.e., standardized mean difference between samples), where XtandS 2 t represent the sample mean and variance for the treated offenders and XcandS 2 c represent the sample mean and variance for the untreated offenders. If the value of this statistic exceeds 20, the covariate is considered to be unbalanced (Rosenbaum and Rubin1985). As shown in Table2, the matching procedure reduced the bias in propensity scores between treated and untreated offenders by 96%. Whereas the p-value was 0.00 in the unmatched sample, it was 0.40 in the matched sample. In the unmatched sample, there were three covariates that were significantly imbalanced (i.e., the bias values exceeded 20). However, in the matched sample, covariate balance was achieved insofar as there were no covariates with bias values greater than 20. The average reduction in bias for the 17 covariates was 46%.

4.3 Matching for treatment outcome and program duration As noted above, this study also examines the effects of treatment outcome and program duration on recidivism. Because untreated and treated offenders were matched individually, it is possible to estimate the effects of treatment outcome by Fig. 1Distribution of propensity scores by treatment assignment 68G. Duwe Table 2Propensity score matching and covariate balance for treatment Variable Sample Treated meanUntreated meanBias (%) Bias reductiont-test p-value Propensity score Total 0.44 0.29 74.28–95.74% 0.00 Matched 0.40 0.40 3.17 0.40 Male Total 89.60% 90.72% 3.02 13.69% 0.30 Matched 89.85% 88.55% 3.44 0.37 Minority Total 40.81% 50.36% 15.77–85.36% 0.00 Matched 43.52% 44.92% 2.31 0.54 Age at release (years) Total 33.55 32.97 5.12–68.51% 0.08 Matched 33.44 33.26 1.61 0.67 Metro Total 49.74% 52.87% 5.11–93.10% 0.08 Matched 51.30% 51.51% 0.35 0.93 Prior felony Total 2.45 2.51 1.62–90.42% 0.58 Matched 2.55 2.55 0.16 0.97 Person offenders Total 27.41% 34.84% 13.30–95.61% 0.00 Matched 28.62% 28.94% 0.58 0.88 Property offenders Total 24.66% 24.84% 0.35 304.00% 0.91 Matched 24.62% 25.38% 1.43 0.71 Drug offenders Total 30.41% 27.85% 4.59–29.31% 0.12 Matched 30.24% 32.07% 3.24 0.39 DWI offenders Total 5.24% 0.81% 19.13–35.48% 0.00 Matched 4.21% 1.51% 12.34 0.00 Other offenders Total 12.29% 11.66% 1.58–65.91% 0.59 Matched 12.31% 12.10% 0.54 0.89 Assessed as dependent Total 63.66% 51.66% 20.10–75.85% 0.00 Matched 58.75% 61.66% 4.85 0.20 Institutional discipline Total 2.36 2.86 9.61–66.84% 0.00 Matched 2.50 2.66 3.19 0.40 Length of stay (months) Total 17.46 11.55 47.86–98.46% 0.00 Matched 16.29 16.19 0.74 0.86 Length of supervision (months) Total 18.95 17.60 4.14 58.72% 0.25 Matched 18.60 17.06 6.56 0.47 Intensive supervised release Total 18.30% 25.38% 14.33–86.42% 0.08 Matched 21.38% 20.41% 1.95 0.61 Supervised release Total 64.95% 46.86% 30.47–94.03% 0.00 Matched 62.10% 63.17% 1.82 0.63 Work release Total 14.86% 12.51% 5.52–86.21% 0.06 Matched 14.15% 13.82% 0.76 0.84 Discharge Total 1.89% 15.25% 46.23–97.53% 0.00 Matched 2.38% 2.59% 1.14 0.77 Supervised release revocations Total 0.42 0.39 3.75–96.73% 0.01 Matched 0.48 0.48 0.12 0.98 Total treatedn= 1,164 Total untreatedn= 2,230 Matched treatedn= 926 Matched untreatedn= 926 Prison-based chemical dependency treatment in Minnesota: An outcome evaluation 69 separately comparing completers and dropouts with their untreated counterparts in the comparison group. Likewise, the effects of program duration can be analyzed by separately comparing short-, medium-, and long-term program participants with their matched pairs of untreated offenders. Yet, using the matched pairs produced by the propensity score model for treatment participation could yield biased estimates of the effects for treatment outcome and program duration considering that the initial match between treated and untreated offenders was based on a different measure of treatment (participation). 6 To address this issue, separate propensity score models were estimated for each of the five additional measures of treatment: (1) treatment completers, (2) treatment dropouts, (3) short-term participants, (4) medium-term participants, and (5) long- term participants. Specifically, five logistic regression models were estimated in which the 17 aforementioned predictors were regressed against dependent variables that contrasted the untreated offenders (n= 2,230) with the treatment completers (n= 843), treatment dropouts (n= 321), short-term participants (n= 671), medium-term participants (n= 393), and long-term participants (n= 100). After obtaining propen- sity scores from the five logistic regression models, untreated offenders were then matched—using a caliper of 0.10—with treated offenders for each of the five treatment measures. The matching process yielded match rates of 84% (708 of 843) for treatment completers, 96% (306 of 321) for treatment dropouts, 90% (606 of 671) for short-term participants, 90% (352 of 393) for medium-term participants, and 98% (98 of 100) for long-term participants. Comparisons between the matched pairs for the five treatment measures, which are not presented here but can be obtained from the author on request, revealed that all propensity score and covariate means had bias values less than 20.

5 Analysis In analyzing recidivism, survival analysis models are preferable in that they utilize time-dependent data, which are important in determining not only whether offenders recidivate but also when they recidivate. As a result, this study uses a Cox regression model, which uses both“time”and“status”variables in estimating the impact of the independent variables on recidivism. For the analyses presented here, the“time” variable measures the amount of time from the date of release until the date of first rearrest, reconviction, reincarceration, or December 31, 2008, for those who did not recidivate. The“status”variable, meanwhile, measures whether an offender reoffended (rearrest, reconviction, or reincarceration for a new crime) during the period in which s/he was at risk to recidivate. In the analyses presented below, Cox regression models were estimated for each of the three recidivism measures for all 6It is worth noting that results from Cox regression models analyzing treatment outcome and program duration based on matches from the treatment participation propensity score model were similar to those reported in this study. That is, completing treatment significantly reduced recidivism, whereas dropping out of treatment had no effect. Similarly, for program duration, short-term programs significantly decreased recidivism, while long-term programs did not have a statistically significant impact. Medium- term programs significantly reduced rearrest and reconviction, but did not have a statistically significant effect on reincarceration.

70G. Duwe six treatment variables (participation, completer, dropout, short-term, medium-term, and long-term).

6 Results Compared to the untreated offenders, those who received treatment had lower rates of reoffending for all three recidivism measures. As shown in Table3, which breaks down recidivism rates by treatment participation, outcome, and program type, offenders who completed treatment or successfully participated until their release had lower reoffense rates than treatment dropouts for all three recidivism measures.

In addition, offenders who participated in medium-term programs had the lowest recidivism rates, followed by those who entered long-term programs.

These findings suggest that: (1) prison-based treatment may have an impact on recidivism, (2) completing treatment may significantly lower the risk of recidivism, and (3) medium- and long-term programs may be more effective at reducing recidivism than short-term programs. It is possible, however, that the observed recidivism differences between treated and untreated offenders, treatment com- pleters and dropouts, and short-term andother treatment participants are due to other factors such as time at risk, prior criminal history, discipline history, or post- release supervision. To statistically control for the impact of these other factors on reoffending, Cox regression models were estimated for each of the three recidivism variables across all six treatment measures (participation, completers, dropouts, short-term, medium-term, and long-term).

6.1 The impact of chemical dependency treatment on recidivism 6.1.1 Treatment participation The results in Table4indicate that, controlling for the effects of the other independent variables in the statistical model, participation in a prison-based CD treatment program significantly reduced the hazard ratio for all three recidivism measures (rearrest, Table 3Recidivism rates by treatment participation, outcome, and program length Rearrest Reconviction Reincarcerationn Untreated offenders 63.5 39.5 29.6 926 Treated offenders 59.8 33.7 23.8 926 Treatment outcome Treatment completers 57.1 29.8 20.6 650 Treatment dropouts 66.3 42.8 31.2 276 Length of program Short-term treatment 67.1 36.8 25.6 562 Medium-term treatment 46.7 27.5 20.3 291 Long-term treatment 56.2 34.2 23.3 73 Prison-based chemical dependency treatment in Minnesota: An outcome evaluation 71 reconviction, and reincarceration for a new offense). Put another way, treated offenders recidivated less often and more slowly than untreated offenders; as a result, those who participated in treatment survived longer in the community without committing a new offense. In particular, CD treatment decreased the hazard by 17% for rearrest, 21% for reconvictions, and 25% for reincarcerations for a new crime.

The results also showed that the hazard ratio was significantly greater for males (all three measures), minorities (all three measures), younger offenders (all three measures), offenders with a metro-area county of commitment (reconviction and reincarceration), offenders with prior felony convictions (all three measures), DWI offenders (all three measures), offenders with institutional discipline convictions (all three measures), offenders with supervised release revocations (reconviction and reincarceration), and offenders with shorter lengths of stay in prison (rearrest and reconviction) and time under post-release supervision (all three measures). The risk (hazard) was significantly less, however, for offenders released to intensive supervised release (reconviction and reincarceration) and work release (reconviction and reincarceration).

Table 4Cox regression models for treatment participation Variables Rearrest Reconviction Reincarceration Hazard ratio SE Hazard ratio SE Hazard ratio SE Chemical dependency treatment 0.828** 0.060 0.792** 0.077 0.746** 0.091 Male 1.448** 0.104 1.665** 0.148 1.964** 0.185 Minority 1.276** 0.064 1.273** 0.083 1.350** 0.098 Age at release (years) 0.981** 0.004 0.981** 0.005 0.982** 0.006 Metro 1.118 0.064 1.378** 0.084 1.321** 0.100 Prior felonies 1.083** 0.008 1.088** 0.009 1.100** 0.009 Offense type Person offenders 0.896 0.103 1.034 0.131 0.984 0.153 Property offenders 1.058 0.099 1.121 0.125 1.107 0.144 Drug offenders 0.930 0.102 0.804 0.134 0.783 0.159 DWI offenders 2.400** 0.265 2.436** 0.346 4.003** 0.412 Assessed as dependent 1.034 0.062 1.064 0.081 1.006 0.095 Institutional discipline 1.038** 0.008 1.024* 0.010 1.035** 0.011 Length of stay (months) 0.983** 0.003 0.988** 0.004 0.992 0.005 Length of supervision (months) 0.979** 0.003 0.982** 0.004 0.975** 0.006 Supervision type Intensive supervised release 0.697 0.192 0.586* 0.229 0.530* 0.264 Supervised release 0.860 0.170 0.734 0.199 0.718 0.226 Work release 0.741 0.195 0.571* 0.238 0.518* 0.280 Supervised release revocations 0.919 0.049 1.193** 0.056 1.152* 0.065 n1,852 1,852 1,852 **p< .01 *p< .05 72G. Duwe The results for the control variables were, for the most part, similar across all six measures of treatment (participation, completer, dropout, short-term, medium-term, and long-term). As such, the ensuing discussion of the results presented in Tables5, 6,7,8will focus strictly on the effects found for the other five treatment measures.

6.2 Treatment outcome As shown in Table5, which analyzes the impact of treatment outcome on reoffending, dropping out of treatment—either quitting or being terminated—did not have a statistically significant effect on any of the three recidivism measures.

Completing treatment, however, had a significant impact on all three types of recidivism, reducing the hazard by 22% for rearrest, 20% for reconviction, and 27% for reincarceration.

6.3 Program duration As shown earlier in Table3, offenders who entered medium-term programs had the lowest recidivism rates, whereas short-term participants had the highest rates. The results presented in Tables6,7,8, however, show that both the short- and medium- term programs had statistically significant effects on all three recidivism measures.

In contrast, long-term programs did not have a statistically significant impact on any type of recidivism. The hazard ratio for short-term participants was, relative to their untreated counterparts, 18% lower for rearrest, 18% lower for reconviction, and 24% lower for reincarceration. In addition, compared to their untreated matched pairs, the hazard ratio for medium-term participants was 32% lower for rearrest, 28% lower for reconviction, and 30% lower for reincarceration.

Given that medium-term participants had the lowest recidivism rates, it is perhaps not that surprising to find that medium-term programming had a statistically significant effect on all three recidivism measures. Interestingly, however, the results suggest that short-term programming was more effective than long-term program- ming even though the latter had lower recidivism rates. Although short-term participants had the highest rates of reoffense, they also had more prior felony convictions, shorter lengths of stay in prison, shorter post-release supervision periods, and they were less likely to be released to supervision—all factors that significantly increased the risk of recidivism. Yet, after controlling for the effects of these and other factors such as time at risk, it was participation in the short-term programs—as opposed to the long-term programs—that had a statistically significant effect on all three recidivism measures.

6.4 Sensitivity analyses 6.4.1 Rosenbaum bounds Although the results suggest that prison-based CD treatment reduces recidivism, PSM controlled only for bias among the observed covariates. As a result, the possibility exists that unobserved selection bias may account for the significant treatment effects. Hidden bias can occur when two offenders with the same observed Prison-based chemical dependency treatment in Minnesota: An outcome evaluation 73 Table 5Cox regression models for treatment outcome Variables Treatment completer Treatment dropout Rearrest Reconviction Reincarceration Rearrest Reconviction Reincarceration Hazard ratio SE Hazard ratio SE Hazard ratio SE Hazard ratio SE Hazard ratio SE Hazard ratio SE Treatment outcome Complete 0.783** 0.069 0.800* 0.093 0.730** 0.113 Drop out1.022 0.100 1.067 0.130 0.882 0.148 Male 1.344* 0.116 1.349 0.162 1.699* 0.212 1.220 0.185 1.360 0.253 1.810 0.306 Minority 1.427** 0.075 1.398** 0.101 1.557** 0.122 1.117 0.110 1.135 0.143 1.365 0.163 Age at release (years) 0.982** 0.004 0.984** 0.006 0.990 0.007 0.976** 0.006 0.972** 0.008 0.962** 0.010 Metro 1.069 0.075 1.311** 0.100 1.325* 0.122 1.115 0.106 1.347* 0.140 1.063 0.157 Prior felonies 1.069** 0.010 1.081** 0.010 1.091** 0.011 1.077** 0.015 1.090** 0.018 1.113** 0.020 Offense type Person offenders 0.857 0.126 0.944 0.163 0.861 0.196 0.847 0.178 1.093 0.231 1.034 0.270 74G. Duwe Variables Treatment completer Treatment dropout Rearrest Reconviction Reincarceration Rearrest Reconviction Reincarceration Hazard ratio SE Hazard ratio SE Hazard ratio SE Hazard ratio SE Hazard ratio SE Hazard ratio SE Property offenders 1.082 0.119 1.098 0.153 1.193 0.179 0.987 0.175 1.076 0.230 1.118 0.266 Drug offenders 0.842 0.121 0.665* 0.162 0.633* 0.198 0.971 0.199 0.967 0.270 0.888 0.315 DWI offenders 1.684 0.324 1.600 0.460 1.785 0.606 3.554** 0.430 3.519* 0.557 6.487** 0.681 Assessed as dependent 0.957 0.072 1.006 0.098 1.026 0.118 1.079 0.106 1.270 0.140 1.207 0.157 Institutional discipline 1.036* 0.015 1.028 0.019 1.039 0.023 1.017* 0.008 1.021* 0.010 1.026* 0.011 Length of stay (months) 0.980** 0.004 0.987* 0.006 0.989 0.007 0.980** 0.005 0.981** 0.007 0.987 0.008 Length of supervision (months) 0.982** 0.003 0.983* 0.005 0.976** 0.007 0.980* 0.006 0.982* 0.008 0.976* 0.011 Supervision type Intensive supervised release 1.292 0.347 1.023 0.454 1.053 0.509 1.200 0.281 0.703 0.339 0.439* 0.397 Supervised release 1.652 0.324 1.513 0.420 1.386 0.464 1.209 0.254 0.929 0.305 0.724 0.354 Work release 1.372 0.338 1.203 0.441 0.965 0.497 0.437 0.579 0.466 0.669 0.497 0.697 Supervised release revocations 0.930 0.060 1.218** 0.070 1.274** 0.081 0.891 0.081 1.288** 0.092 1.268* 0.104 n1,416 1,416 1,416 612 612 612 **p< .01 *p< .05 Prison-based chemical dependency treatment in Minnesota: An outcome evaluation 75 covariates have different chances of receiving treatment due to an unobserved covariate. If this unobserved covariate is related to the outcome (recidivism) affected by treatment, then the failure to account for this hidden bias can alter conclusions drawn about the effects of treatment.

The sensitivity of the results to hidden bias was tested by using a method developed by Rosenbaum (2002) that calculates a bound on how large an effect an unobserved covariate would need to have on the treatment selection process in order to reverse inferences drawn about the effects of treatment. The Rosenbaum bounds sensitivity analysis produces a test statistic, gamma, that measures the threshold at which an unobserved covariate would cause the estimated treatment effect to no longer be statistically significant (i.e.,p> .05). More specifically, the closer the gamma value is to 1, the stronger the possibility that the effect can be explained away by an unobserved covariate. Therefore, an estimated treatment effect with a Table 6Cox regression models for program duration: first rearrest Variables Short-term Medium-term Long-term Hazard ratio SE Hazard ratio SE Hazard ratio SE Program duration Short-term treatment 0.821** 0.070 Medium-term treatment 0.683** 0.107 Long-term treatment 1.052 0.227 Male 1.396** 0.128 2.531* 0.425 1.669 0.294 Minority 1.281** 0.077 1.355* 0.113 1.617* 0.227 Age at release (years) 0.976** 0.004 0.986* 0.007 0.961** 0.013 Metro 1.245** 0.076 1.080 0.113 1.015 0.221 Prior felonies 1.075** 0.010 1.087** 0.018 1.148** 0.033 Offense type Person offenders 0.909 0.127 0.885 0.165 1.193 0.358 Property offenders 1.024 0.117 1.264 0.194 1.629 0.356 Drug offenders 0.881 0.125 0.933 0.165 1.191 0.358 DWI offenders 1.708 0.385 2.489** 0.332 2.079 0.563 Assessed as dependent 0.954 0.072 1.023 0.112 0.891 0.237 Institutional discipline 1.019 0.010 1.033* 0.013 1.021 0.025 Length of stay (months) 0.982** 0.004 0.989* 0.005 0.973** 0.011 Length of supervision (months) 0.989** 0.004 0.979** 0.004 0.989 0.008 Supervision type Intensive supervised release 1.257 0.244 0.477* 0.330 0.969 0.818 Supervised release 1.423 0.211 0.492* 0.317 1.533 0.775 Work release 1.164 0.247 0.463* 0.336 0.780 0.896 Supervised release revocations 0.922 0.062 0.976 0.080 0.684* 0.171 n1,212 704 196 **p< .01 *p< .05 76G. Duwe gamma value of 1.5, for example, would be more sensitive to hidden bias than an effect with a gamma value of 2.0.

It is important to emphasize, however, that the Rosenbaum bounds method is limited in two important ways. First, the sensitivity analysis does not indicate whether unobserved bias exists. Rather, it simply identifies how large the hidden bias would need to be to nullify the estimated treatment effect. Second, as DiPrete and Gangl (2004) point out, the Rosenbaum bounds method is a“worst-case” scenario to the extent that it assumes the hypothetical unobserved covariate is an almost perfect predictor of the outcome variable (recidivism).

The results from the sensitivity analyses reveal that the estimated treatment effects are not particularly robust to hidden bias. With a gamma value of 1.05, the rearrest findings are the most sensitive to the possibility of hidden bias, followed by reconviction (gamma = 1.08) and reincarceration (gamma = 1.10). These results Table 7Cox regression models for program duration: first reconviction Variables Short-term Medium-Ttrm Long-term Hazard ratio SE Hazard ratio SE Hazard ratio SE Program duration Short-term treatment 0.820* 0.093 Medium-term treatment 0.725* 0.143 Long-term treatment 0.994 0.302 Male 1.492* 0.184 1.614 0.604 1.205 0.382 Minority 1.238* 0.100 1.406* 0.153 1.262 0.286 Age at release (years) 0.980** 0.006 0.982 0.010 0.967 0.018 Metro 1.453** 0.100 1.191 0.155 1.006 0.283 Prior felonies 1.078** 0.011 1.144** 0.022 1.226** 0.045 Offense type Person offenders 0.949 0.166 0.921 0.209 2.335 0.528 Property offenders 1.056 0.151 0.755 0.257 1.550 0.520 Drug offenders 0.790 0.167 0.659 0.219 2.155 0.539 DWI offenders 1.896 0.503 2.555* 0.434 5.648* 0.819 Assessed as dependent 1.021 0.096 0.898 0.153 1.132 0.326 Institutional discipline 1.010 0.014 1.043** 0.015 0.993 0.033 Length of stay (months) 0.989* 0.006 0.987* 0.007 0.992 0.013 Length of supervision (months) 0.988* 0.006 0.982** 0.006 0.980 0.011 Supervision type Intensive supervised release 0.787 0.312 0.651 0.418 0.849 0.857 Supervised release 1.118 0.262 0.763 0.394 0.865 0.815 Work release 0.810 0.317 0.678 0.427 0.159 1.311 Supervised release revocations 1.311** 0.072 1.209* 0.087 0.933 0.201 n1,212 704 196 **p< .01 *p< .05 Prison-based chemical dependency treatment in Minnesota: An outcome evaluation 77 suggest that if an unobserved covariate that almost perfectly predicted rearrest differed between matched pairs of treated and untreated offenders by a factor of 1.05 or more, it would be sufficient to undermine the conclusions regarding the treatment effect. To put this statistic in perspective, institutional discipline would be a hidden bias equivalent in that, as shown earlier in Table1, it had a comparable impact on the treatment selection process (b =–0.046). Therefore, if an unobserved covariate existed that perfectly predicted rearrest and had an impact on the treatment selection process similar to institutional discipline, it would be sufficient to invalidate the treatment effect for rearrest. Still, it isworth reiterating,however, that the Rosenbaum bounds method is a“worst-case”scenario. Although existing research has identified a number of factors that are significantly associated with recidivism, none have yet to be shown to be a nearly perfect predictor of reoffending, which is what the Rosenbaum bounds approach assumes.

Table 8Cox regression models for program duration: First reincarceration Variables Short-term Medium-term Long-term Hazard ratio SE Hazard ratio SE Hazard ratio SE Program duration Short-term treatment 0.760* 0.111 Medium-term treatment 0.705* 0.173 Long-term treatment 0.841 0.373 Male 2.093** 0.254 3.033 1.024 1.656 0.475 Minority 1.330* 0.120 1.484* 0.185 1.174 0.340 Age at release (years) 0.978** 0.007 0.981 0.012 0.977 0.021 Metro 1.481** 0.120 1.065 0.188 1.030 0.333 Prior felonies 1.092** 0.011 1.187** 0.025 1.203** 0.051 Offense type Person offenders 0.981 0.197 0.974 0.253 2.329 0.658 Property offenders 1.218 0.175 0.719 0.303 1.586 0.655 Drug offenders 0.786 0.203 0.710 0.266 2.235 0.669 DWI offenders 3.881* 0.601 3.610* 0.514 15.800* 1.224 Assessed as dependent 0.980 0.114 0.893 0.186 0.866 0.380 Institutional discipline 1.007 0.016 1.055** 0.016 1.009 0.036 Length of stay (months) 0.999 0.007 0.987 0.008 0.991 0.016 Length of supervision (months) 0.980* 0.008 0.980** 0.008 0.957* 0.020 Supervision type Intensive supervised release 0.596 0.346 0.508 0.466 1.136 0.933 Supervised release 0.808 0.278 0.683 0.430 0.770 0.891 Work release 0.579 0.360 0.478 0.485 0.284 1.381 Supervised release revocations 1.299** 0.080 1.222* 0.100 0.785 0.250 n1,212 704 196 **p< .01 *p< .05 78G. Duwe 7 Conclusion This study is limited by the absence of data on post-treatment substance use and participation in post-release aftercare programming. Despite these limitations, however, the results are consistent with previous findings showing that prison- based CD treatment significantly reduces offender recidivism. Still, the size of the treatment effect was relatively modest. For example, entering treatment lowered the hazard ratio by 17–25% across all three types of recidivism. These results translate into odds ratios of 1.17 for rearrest, 1.28 for reconviction, and 1.35 for reincarceration (Lösel and Schmucker2005), which can, in turn, be converted into Cohen’sd-values of 0.09 for rearrest, 0.14 for reconviction, and 0.17 for reincarceration (Sánchez-Meca et al.2003). In their meta-analysis of incarceration- based drug treatment studies, Mitchell et al. (2007) reported a treatment effect odds ratio of 1.37, which was based primarily on rearrest as a measure of recidivism. The rearrest odds ratio (1.17) for the treatment effect observed in this evaluation is therefore quite a bit lower than what Mitchell et al. (2007) found among drug treatment studies in general. Moreover, the Cohen’sd-values for all three recidivism measures were under 0.20, which is indicative of a small effect size (Cohen1988).

The findings also indicated that dropping out of treatment did not have a significant effect on recidivism, while completing treatment lowered the risk of reoffending from 20–27%. Consistent with previous research (Wexler et al.1990), the results suggest that more treatment is not always better. That is, increased treatment time appeared to lower the risk of recidivism, but only up to a point.

Although short-term (90 days) and medium-term (180 days) programs had a statistically significant impact on all three recidivism measures, no statistically significant effects were found for long-term (365 days) programming.

The results regarding program duration have implications not only for the MNDOC but also for the prison treatment literature in general. Recall that the MNDOC discontinued its short-term programming in 2006, a decision that was based, in part, on evidence which seemed to suggest that better recidivism outcomes were associated with longer program durations. This evidence, however, consisted primarily of simple recidivism comparisons similar to those presented in Table3.

Yet, as this study has shown, controlling for rival causal factors is critical in determining whether a program (or type of program) has an impact on the outcome measure.

This study suggests that short-term programs can be an effective form of treatment, which is an important consideration given that the MNDOC has, over the last several years, had a growing influx of offenders admitted to prison as either probation or supervised release violators (Minnesota Department of Corrections 2007b). Because these offenders tend to have relatively short lengths of stay in prison (an average of 8 months), developing (or reinstituting) a treatment program for these offenders, even if it is short in duration, may yield a benefit in terms of reduced recidivism.

The growing number of probation and supervised release violators admitted to prison is not unique to Minnesota, however. Probation and parole violators have figured prominently in the dramatic growth in the state and federal prison systems, and are projected to have a sizeable impact on future prison populations (JFA Prison-based chemical dependency treatment in Minnesota: An outcome evaluation 79 Associates2007). Therefore, implementing short-term treatment programs for offenders with shorter lengths of stay (e.g., probation and parole violators) may produce a modest recidivism reduction and, in so doing, help limit the growth of prison populations.

Although this study suggests that prison-based CD treatment and, more narrowly, short-term programs can be effective, more evaluations of prison-based programs are needed. Due to the many variations among state and federal correctional populations, it is unlikely that a single study—regardless of how rigorous the design—can conclusively determine whether prison-based treatment works. Rather, by quantitatively reviewing evaluations from multiple jurisdictions, meta-analyses could help better identify what works best for whom under which circumstances. In order to do so, however, the meta-analyses need to be based on an accumulation of rigorous evaluations that effectively control for threats to validity, not least selection bias.

AcknowledgementsThe views expressed in this study are not necessarily those of the Minnesota Department of Corrections. The author wishes to thank the Editor and the three anonymous reviewers for their helpful comments on an earlier draft of this manuscript.

References Beck, A. J., & Gilliard, D. K. (1995).Prisoners in 1994. Washington, D.C: U.S. Department of Justice, Bureau of Justice Statistics.

Berk, R. A. (2003).Regression analysis: A constructive critique. Newbury Park, CA: Sage.

Cohen, J. (1988).Statistical power analysis for the behavioral sciences(2nd ed.). Hillsdale, NJ: Lawrence Earlbaum Associates.

DiPrete, T. A., & Gangl, M. (2004). Assessing bias in the estimation of causal effects: Rosenbaum bounds on matching estimators and instrumental variables estimation with imperfect instruments.Sociological Methodology, 34, 271–310.

Duwe, G., Donnay, W., & Tewksbury, R. (2008). Does residential proximity matter? A geographic analysis of sex offense recidivism.Criminal Justice and Behavior, 35, 484–504.

Field, G. (1985). The cornerstone program: a client outcome study.Federal Probation, 48,50–55.

Harrison, P. M., & Beck, A. J. (2006).Prisoners in 2005. Washington, D.C: U.S. Department of Justice, Bureau of Justice Statistics.

Inciardi, J. A., Martin, S. S., Butzin, C. A., Hooper, R. M., & Harrison, L. D. (1997). An effective model of prison-based treatment for drug-involved offenders.Journal of Drug Issues, 27, 261–278.

Inciardi, J. A., Martin, S. S., Butzin, & Butzin, C. A. (2004). Five-year outcomes of therapeutic community treatment of drug-involved offenders after release from prison.Crime & Delinquency, 50,88–107.

JFA Associates. (2007).Public safety, public spending: Forecasting America’s prison population, 2007– 2011. Washington, D.C: Pew Charitable Trusts.

Karberg, J. C., & James, D. J. (2005).Substance dependence, abuse, and treatment of jail inmates, 2002.

Washington, DC: U.S. Department of Justice, Bureau of Justice Statistics.

Klebe, K., & O’Keefe, M. (2004).Outcome evaluation of the crossroads to freedom and peer I therapeutic communities. Washington, DC: National Institute of Justice.

Knight, K., Simpson, D. D., Chatham, L. R., & Camacho, L. M. (1997). An assessment of prison-based drug treatment: Texas’in-prison therapeutic community program.Journal of Offender Rehabilitation, 24,75–100.

Knight, K., Simpson, D., & Hiller, M. (1999). Three-year reincarceration outcomes in prison therapeutic community treatment in Texas.The Prison Journal, 79, 337–51.

Krebs, C. P., Strom, K. J., Koetse, W. H., & Lattimore, P. K. (2008). The impact of residential and nonresidential drug treatment on recidivism among drug-involved probationers.Crime & Delinquency, Online First.

80G. Duwe Lösel, F., & Schmucker, M. (2005). The effectiveness of treatment for sexual offenders: a comprehensive meta-analysis.Journal of Experimental Criminology, 1,117–146.

Minnesota Department of Corrections (2007a).Chemical dependency treatment services in prison.

St. Paul, MN.

Minnesota Department of Corrections (2007b).Minnesota prison population projections—fiscal year 2007 report. St. Paul, MN.

Minnesota Department of Corrections (2008).Drug offenders in prison.St. Paul, MN.

Minnesota Office of the Legislative Auditor (2006).Substance abuse treatment. St. Paul, MN.

Mitchell, O., Wilson, D. B., & MacKenzie, D. L. (2007). Does incarceration-based drug treatment reduce recidivism? A meta-analytic synthesis of the literature.The Journal of Experimental Criminology, 3, 353–375.

Mumola, C. J., & Karberg, J. C. (2006).Drug use and dependence, state and federal prisoners, 2004.

Washington, DC: U.S. Department of Justice, Bureau of Justice Statistics.

Patenaude, A. L., & Laufersweiller-Dwyer, D. L. (2002).Arkansas comprehensive substance abuse treatment program: Process evaluation of the modified therapeutic community (Tucker Unit).

Washington, DC: National Institute of Justice.

Pearson, F. S., & Lipton, D. S. (1999). A meta-analytic review of the effectiveness of corrections-based treatments for drug abuse.The Prison Journal, 79, 384–410.

Pellisier, B., Rhodes, W., Saylor, W., Gaes, G., Camp, S. D., Vanyur, S. D., et al. (2001).TRIAD drug treatment evaluation project final report of three-year outcomes. Washington, DC: Federal Bureau of Prisons, Office of Evaluation and Research.

Prendergast, M. L., Elizabeth, E. A., Wexler, H. K., Melnick, G., & Cao, Y. (2004). Amity prison-based therapeutic community: 5-year outcomes.The Prison Journal, 84,36–60.

Rosenbaum, P. (2002).Observational studies(2nd ed.). New York: Springer Verlag.

Rosenbaum, P. R., & Rubin, D. B. (1984). Reducing bias in observational studies using subclassification on the propensity score.Journal of the American Statistical Association, 79, 516–524.

Rosenbaum, P. R., & Rubin, D. B. (1985). Constructing a control group using multivariate matched sampling methods that incorporate the propensity score.The American Statistician, 39,33–38.

Rubin, D. B. (1997). Estimating causal effects from large data sets using propensity scores.Annals of Internal Medicine, 127, 757–763.

Sabol, W. J., Couture, H., & Harrison, P. M. (2007).Prisoners in 2006. Washington, D.C: U.S.

Department of Justice, Bureau of Justice Statistics.

Shadish, W. R., Cook, T. D., & Campbell, D. T. (2002).Experimental and quasi-experimental designs for generalized causal inference. Boston: Houghton Mifflin.

Sánchez-Meca, J., Marín-Martínez, F., & Chacón-Moscoso, S. (2003). Effect-size indices for dichotomized outcomes in meta-analysis.Psychological Methods, 8, 448–467.

Welsh, W. (2002).Evaluation of prison based drug treatment in Pennsylvania: A research collaboration between the Pennsylvania Department of Corrections and the Center for Public Policy at Temple University, final report. Washington, DC: National Institute of Justice.

Welsh, W. (2007). A multisite evaluation of prison-based therapeutic community drug treatment.Criminal Justice and Behavior, 34, 1481–1498.

Wexler, H. K., Falkin, G. P., & Lipton, D. S. (1990). Outcome evaluation of a prison therapeutic community for substance abuse treatment.Criminal Justice and Behavior, 17,71–92.

Wexler, H., Melnick, G., Lowe, L., & Peters, J. (1999). Three-year reincarceration outcomes for Amity in- prison therapeutic community and aftercare in California.The Prison Journal, 79, 321–36.

Wolfgang, M. (1958).The patterns of criminal homicide. Philadelphia, PA: University of Pennsylvania.

Grant Duweis Director of Research and Evaluation for the Minnesota Department of Corrections. His publications have appeared inCriminology,Sexual Abuse: A Journal of Research and Treatment,Criminal Justice and Behavior,Crime & Delinquency,Justice Quarterly,Homicide Studies,andWe s t e r n Criminology Review. The author of the book,Mass Murder in the United States: A History,he holds a PhD in Criminology and Criminal Justice from Florida State University. Prison-based chemical dependency treatment in Minnesota: An outcome evaluation 81