Vocational and Educational Programs: Claire Perry April 2014
by user
Comments
Transcript
Vocational and Educational Programs: Claire Perry April 2014
Vocational and Educational Programs: Impacts on Recidivism Claire Perry * Haverford College April 2014 This paper looks at the relationship between educational and vocational program participation and recidivism. Using a cohort of individuals released from state prison in 1994 across five states and tracked for three years, this paper takes into consideration both rearrest and re-confinement. It finds that vocational programs in particular have significant reductions in both re-arrest and re-confinement; results are primarily driven by programs in Illinois and New York. This paper builds off the extensive literature on the topic by accounting for varying levels of program completion, including different types of recidivism, controlling for state-to-state variation, and looking at both education and vocation programs. In addition to predicting recidivism based on program participation, this study seeks to control for motivation of people choosing to participate in educational and vocational programs using instrumental variable analysis and looks at time until recidivism using proportional hazard models. First, I would like to thank my adviser, Professor Anne Preston, whose guidance, support, and expertise have been invaluable throughout this process. Additional thanks go to Professor Ball for instrumental help with data manipulation, Kim Minor for logistical support, Norm Medeiros for research assistance, and Matt Durose, from the Bureau of Justice Statistics, for help re-creating variables in the original BJS report. * Table of Contents Introduction ........................................................................................................................................................................... 1 Background ............................................................................................................................................................................ 4 Literature Review ................................................................................................................................................................ 7 Data......................................................................................................................................................................................... 15 Methods ................................................................................................................................................................................ 16 Results ................................................................................................................................................................................... 27 Conclusion ........................................................................................................................................................................... 48 References ........................................................................................................................................................................... 55 Appendices .......................................................................................................................................................................... 61 Perry 1 Introduction The continued expansion of America’s prison system has refocused debate on the cost of incarceration, judicial fairness, and the effectiveness of programs designed to rehabilitate incarcerated individuals and reduce recidivism. Many studies indicate that educational and vocational programs for people in prison can reduce recidivism and thus better the lives of formerly incarcerated individuals2 and decrease the burden on the criminal justice system (Davis et al. 2013; Pew 2011; Wilson et al. 2000). A 2011 report by the Pew Center on the States notes that “State corrections spending, driven almost entirely by prison expenditures, has quadrupled over the past two decades, making it the second fastest growing area of state budgets, trailing only Medicaid. Total state spending on corrections today is more than $50 billion a year” (Pew 2011). This increase is due, in part, to increases in prison populations. In the period from 1987 to 2007, the national prison population nearly tripled. Today, one in 100 Americans are behind bars in federal prison, state prison, or local jail (Warren et al. 2008). With so many incarcerated Americans and so much taxpayer money going toward prisons, people are asking the question: are our prisons working? One way to measure success of the prison system is by looking at recidivism. A commonly accepted definition of the recidivism rate is the proportion of individuals released from prison that are rearrested, reconvicted, or re-incarcerated within three years of their initial release (Pew 2011). If prisons are meant both to protect society from criminals and to “correct” individuals for successful re-entry to society, then recidivism indicates a fault on both counts. Despite massive increases in corrections spending, the recidivism rate has remained fairly constant since 1983 at around 40% (Pew 2011; Hughes et al. 2002). This consistency means that, despite evidence of There is discussion as to which tern is most appropriate to refer to individuals who were formerly incarcerated. Though not without its problems, “formerly incarcerate individuals” is used throughout this paper as opposed to ex-inmate, exprisoner, ex-offender, prisoner, or returning citizen. For more information about recent legal changes in Philadelphia see Cherri 2013. 2 Perry 2 effective programs, little has been done to alter dramatically prison recidivism. Given that it costs $78.95 on average per day 3 to keep an individual in prison, more than 20 times the cost of a day on probation, (Warren et al. 2009) the rate at which people return to prison presents a huge financial cost to society (Davis et al. 2013). One way to deal with this issue is by trying to better prepare individuals for life outside of prison through education and vocational courses. Such programs are in high demand4,5 as they help normalize the prison experience, teach essential skills for employment, keep people occupied and engaged, and allow some individuals to earn “good time” in order to reduce their sentences. Investing now saves money later and makes it more likely that people will re-enter society as successful, tax paying citizens (Davis et al. 2013). As seen below in Table 1 the percentage of state prisons offering different types of educational programing has increased slightly from 1995 to 2000. Yet, as seen in Table 2, the percentage of incarcerated individuals participating in these programs has steadily decreased for all program types from 1991 to 2004. TABLE 1: PROGRAMS OFFERED IN STATE AND FEDERAL CORRECTIONAL FACILITIES (PERCENTAGES) 6 Federal State 1995 2000 1995 2000 With Education Programs Adult Basic Education Secondary/GED Special Education Vocational Training 96.8 85.6 92.8 32.8 68.0 91.7 89.3 90.5 54.8 85.7 86.5 73.5 78.5 33.1 52.4 90.3 79.6 82.8 39.2 55.2 College Study Release Life Skills/ Community Adjustment 63.2 8.8 80.0 73.8 6.0 89.3 30.5 10.9 65.4 26.4 7.6 68.3 In 2008 dollars As part of the Serious and Violent Offender Reentry Initiative (SVORI), a Federal initiative to help States reduce recidivism, the Urban Institute and RTI International interviewed then soon-to-be-released incarcerated people. Of their requested reentry needs, 94% requested more education, 82% job training, and 80% post-release employment. 5 Crayton & Neusteter suggest that several states, including CA and MD reported having significant waitlists for educational programs 6 Sources: Stephan & Karberg (2003); Stephan (1997); Crayton & Neuster (2006) 3 4 Perry 3 TABLE 2: PROGRAMS OFFERED IN STATE AND FEDERAL CORRECTIONAL FACILITIES (PERCENTAGES)7 Federal State Basic 1991 10.4 1997 1.9 2004 1.5 1991 5.3 1997 3.1 2004 2.1 Secondary/GED Vocational Training College English as a Second Language Life Skills/ Community Adjustment 27.3 29.4 18.9 - 23.0 31.0 12.9 5.7 - 21.1 30.8 10.1 2.3 28.6 27.3 31.2 13.9 - 23.4 32.2 9.9 1.2 - 19.2 27.0 7.3 1.0 14.3 Given that resources are clearly limited, it begs the question: which programs - vocational training, general education, or a combination - are best at reducing recidivism of people incarcerated in state prisons? In this paper, I take an empirical approach to evaluate the relative effectiveness of vocational training, educational programs, or a combination of the two in reducing recidivism to state prison. I use probit regression analysis to study the largest national dataset on this topic, Recidivism of Prisoners Released in 1994 (ICPSR3355) by the Bureau of Justice Statistics. This dataset is arguably the most comprehensive study of recidivism at the individual level available in the United States. Like some other papers on the topic I consider the difference between educational and vocational programs, however I account for state-to-state variation, look at different levels of participation, attempt to control for selection bias, and consider two measures of recidivism. To deepen this work, I use a proportional hazards model to see if the rate of recidivism overtime is impacted by their involvement in educational or vocational programs. I begin with a brief background of recidivism and discuss the existing literature regarding recidivism and educational and vocational programs in prisons. I provide basic information on the 7 Sources: Harlow (2003); BJS (2007); Crayton & Neuster (2006) Perry 4 data set used in this paper, ICPSR3355, and outline the methods for analysis along with references to specific literature using these statistical forms. Results are given and analyzed; then I conclude. Background Recidivism As stated previously, the recidivism rate is the proportion of individuals released from prison that are rearrested, reconvicted, or re-incarcerated within three years of their initial release (Pew 2011). Recidivism is a popular metric, and while most recent studies use the previous definition, there is still room for interpretation. For example, within that definition there are four distinct ways to quantify recidivism, as noted by the Bureau of Justice Statistics (BJS): Rearrests for a new offense Reconviction for a new crime Resentenced to prison or jail for a new crime Resentenced to prison or jail for a new crime or for a technical violation, also known as reconfinement 8 Officially, the above are measured within a 1,096 day period after release.9 For the purpose of this analysis, I focus on the first and last definitions as the first indicates, to some extent, return to criminal behavior, and the last indicates reengagement with the criminal justice system either through a new crime or a technical violation. While re-arrest is subject to many biases (one can be arrested without actually being guilty), it is a readily available indicator and the first step in return to the criminal justice system. Few analyses look at the difference in a new criminal conviction versus a technical violation as it relates to educational and vocational programs. While both indicate a violation of the law, the latter is more likely to be a function of state laws, intensity of parole sentences, and punishments for technical violations, rather than individual motivation (Pew 8 9 For example, a technical violation may be failing a drug test or not complying with parole guidelines. 1,096 days is considered to be more accurate than simply “3 years” due to the different number of days in each month. Perry 5 2011). For simplicity’s sake the final definition will be used as returning to prison or jail indicates a failed attempt at “correction” and indicates the greatest cost to both society and the individual. While recidivism is a popular metric, it should not be used in isolation when evaluating prisons. Recidivism rates vary greatly state-to-state 10 because of different policies around parole, technical violations, arrests, age at which one can be tried as an adult, and the likelihood of being sent to prison after a conviction (Pew 2011). In addition to variation in state policies, Petersila notes, “it is impossible to know whether recidivism indicators reflect more criminality on the part of released prisoners or greater diligence on the part of the police (e.g. more arrests relative to crimes committed)” (Petersila 2003). A low recidivism rate does not always reflect the use of sound release, preparation, and supervision strategies (Pew 2011). For example, it could be a symptom of over incarceration for petty crimes for which people may be less likely to recidivate. To account for these differences, I control for state-to-state variation in my analysis by including state dummy variables and look at both re-arrest and re-confinement for each state in the analysis. Education and Vocational Programs Education programs range from basic literacy classes, to GED attainment courses, to postsecondary education. They are usually general in their aims and focus on increasing knowledge and general skills. Vocational programs are more varied and range from job training to interviewing skills to technical skills in a specific trade. At their core, these programs aim to increase the employability of individuals. In both cases certificates may be awarded for successful completion. Some studies (Harer 1995; Lawrence et al. 2002; Gaes 2008) go into even further specificity when defining types of programs offered, but due to constraints in available data, I will confine my focus to these two overarching program types: education and vocation. 10 See Figure 4 for the re-arrest and re-confinement rates for the states included in this analysis Perry 6 Education and vocation programs have been integrated in the American criminal justice system since 1790 when the Philadelphia Quakers opened the Walnut Street Jail. Their programs focused on reform of offenders by making them penitent and by providing religious texts and guidance (Hall 2006).11 Since then education and vocational training have been a core part of the American penal system. There are many arguments in favor of structured programing in prisons. To start, people argue that employment and crime are negatively correlated (College of Criminology and Criminal Justice, 2006 Annual Report). Thus increasing one’s ability to get a job - through training and education - decreases the chances of committing another crime (Piehl 2000). Another argument in favor of educational and vocational programs is that they normalize the prison experience by rewarding achievement, helping incarcerated individuals reduce their sentences (in some cases), and “nurturing pro-social norms supporting rule/law abiding behavior” (Harer 1995). At the most basic level such programing has value as it keeps incarcerated individuals busy. In 1930 a Senate Judiciary Committee noted, “It is unanimously conceded that idleness in prisons breeds disorder and aggregates criminal tendencies. If there is any hope for reformation and rehabilitation of those convicted it will be founded upon the acquisition be the prisoner of the requisite skill and knowledge to pursue a useful occupation and the development of the habits of industry” (Gaes & Saylor referring to Congressional Record Report No. 529. 71st Cong). Finally, and perhaps the most cited of all, is that incarcerated individuals are often less educated than the general population. “Between 1991 and 1997, the percent of inmates in state prison without a high school diploma or GED remained the same — 40% in 1997 and 41% in 1991” (Harlow 2003). To compare, in 1997 only 18% of the general population age 18 or older had not finished 12th grade (Harlow 2003). This leaves incarcerated individuals even worse off when trying to find employment post release and represents a loss of productivity for the individual and society. 11 For a more extensive history see Hall’s dissertation: Voices Behind Bars (2006) Perry 7 As such, prison education programs provide an outlet to help correct for basic educational deficiencies that might be further holding individuals back. Despite many current studies indicating the success of education and vocation programs 12 (Davis et al. 2013; Pew 2011; Wilson et al. 2003) there is still disagreement about which programs are most effective and how to fund them (Brewster & Sharp 2002; Pew 2011). Additionally, some people still oppose educational and vocational programs, arguing that they do not prepare people for life after prison, they are too little too late, they lack structure and resources to be effective, and they divert resources that could go to non-incarcerated people (Strand 2012; Marc-Taylor). To complicate matters, many studies use different definitions of recidivism or look at very different programs. So while some find benefits in education, others point to vocation for the benefits, though few studies find no impact of either. As program offerings decrease (Table 2) and budgets are cut the question becomes, which programs are the most effective at reducing recidivism? In the following section I go in to greater detail about the existing literature first by looking at several meta-analyses written about educational and vocational programing and recidivism, then by examining select relevant studies. Literature Review Meta-Analysis The literature on the topic of prison recidivism and correctional programs is so extensive that five meta-studies have been written to evaluate the work. Lipton, Martinson, and Wilks wrote the first comprehensive study in 1975. It covered 231 rigorous studies between 1945 and 1967 and, among other topics, examined vocational and educational programs in prisons. It became “the most politically important criminological study of the past half century” (Miller 1989). The authors The Bureau of Justice Statistics includes “in-prison and post-release vocational training and work programs” in the selected list of programs that they know to be effective in reducing recidivism (Bureau of Justice Statistics) 12 Perry 8 came to the ambiguous conclusion that some programs work, while others do not. Still, this was a significant step from Martinson’s 1974 paper, now referred to as the “Nothing Works” paper. Though some say Martinson later revoked this stance that there “is very little reason to hope that we have in fact found a sure way of reducing recidivism” his paper and the subsequent meta-study led to an era of skepticism about rehabilitation (Elliott 2010). As a result “federal- and statesponsored initiatives to address the needs of prisoners were effectively put on the defensive and in some cases curtailed” (Davis et al. 2013). A subsequent study, published in 2000, by Wilson, Gallegher, and MacKenzie, considered “the recidivism outcomes of 33 independent experimental and quasi-experimental evaluations of education, vocation, and work programs and found that program participants recidivate at a lower rate than nonparticipants” (Wilson et al. 2000). Wilson, et al. set high standards for experimental rigor and eliminated poorly run studies from their analysis. On the whole, these studies’ results were heterogeneous indicating that some programs may be highly effective while others may have no effect on future offending behavior (Wilson et al. 2000). The authors assumed a 50% baseline recidivism rate for non-participants, and found that on average participants in some kind of educational or vocational program recidivate at a rate of 39%. It is unclear where the 50% assumption is from, which calls in to question the 39% figure. 13 Further, the authors note that selfselection bias and poor study design still plague many of the evaluated studies. Thus, while their meta-analysis cannot conclusively confirm the benefits of these programs, they claim a positive correlation between program participation and lower recidivism rates (Wilson et al. 2000). As their analysis suggests, looking at selection bias is uncommon in recidivism studies and will thus be included in this analysis. Mackenzie, from the previous meta-analysis, completed a follow-up study in 2006 looking only at programs from 1980 to 2006. She notes “research demonstrates that education programs 13 This is especially odd as research suggests the standard re-confinement rate is closer to 40% and has stayed constant Perry 9 such as basic education, GED, postsecondary and vocational are effective in reducing later recidivism and increasing future employment.” (Mackenzie 2006). Yet she finds huge variation in the quality of vocational programs, which makes it difficult to draw conclusions about their effectiveness (Mackenzie 2006). Unlike strictly educational programs, she finds that vocational training is often paired with other reentry programs so it is difficult to separate the effects. She notes that though the evidence indicates effective programs, more research needs to be done in assessing specifically what makes certain programs more effective (Mackenzie 2006). In 2006 Aos, Miller, and Drake completed another meta-study and found that recidivism was reduced by 7% for participation in educational programs and by 9% for participation in vocational programs. Because many individuals participated in both education and vocation programs, it was difficult to separate the impacts of each program. The authors were, however, able to estimate that for both programs the net benefits per participant were greater than $10,000. The exact derivation of this estimate is complicated but is consistent with the findings in Davis et al. and with the idea that investing in education reduces future costs to society and increases the tax base. A core facet of these studies is the consideration of both education and vocation programs as well as participation in both program types. While Mackenzie finds a greater correlation for education programs, Aos et al. find a greater impact of vocational programs. Both papers suggest that the interaction between types of programs is important in teasing out which program has an effect. Such interactions and varying levels of participation are central to the analysis that follows in this paper. The most rigorous and recent meta-study is the 2013 report by Davis et al. for the RAND Corporation – a think-tank funded, in part, by the U.S. Government. Davis et al. look at 102 effects across 58 studies and find that “between 43.3% and 51.8% of former prisoners were reincarcerated within three years of release, and two-thirds were rearrested within three years of release” (Davis et al. 2013). Applying national recidivism data, this study finds that: Perry 10 Correctional education would be expected to reduce three year re-arrest and reincarceration rates by 13.2 and 13.8 percentage points, respectively. According to these estimates, eight inmates would need to receive correctional education to prevent one additional inmate from being rearrested within three years of release, and seven inmates would need to receive correctional education to prevent one additional inmate from returning to prison within three years. (Davis et al. 2013 ) The authors also provide a brief cost-benefit analysis of prison education programs. They estimate that from a budget standpoint, correctional education programs would “need to reduce the threeyear re-incarceration rate by between 1.9 and 2.6 percentage points” in order to reach the breakeven point (Davis et al. 2013). Their research suggests “that inmates who participate in correctional education programs have a 43 percent lower odds of returning to prison than those who do not” (RAND Press Release) and that people participating in academic or vocational programs had an employment rate 13% higher than those who did not participate. Specifically, “people with vocational training were 28% more likely to be employed after release from prison than [those] who did not receive such training.” In this way, the RAND study provides strong support for both education and vocational programs. These five meta-studies make it clear that there is a significant amount of research being done on the topic of educational and vocational programing and recidivism. 14 These studies also make clear that despite skeptical beginnings, there is some agreement that education and vocational programs reduce recidivism. All of the studies indicate a need for more comprehensive data and studies regarding what specific types of programming are effective. Still several questions remain that motivate this paper: Which programs, educational or vocational, are better at reducing recidivism? Do these impacts still hold when accounting for selection bias – the idea that people who choose to participate might be less likely to recede anyway? Do the impacts of education or vocational training wear off over time? It should be noted that some of this variation could be explained by the program and policy landscape changing over the years. Thus one program type could have been more impactful in the past but since fallen out of favor. 14 Perry 11 Selected Studies There are countless studies 15 that look exclusively at education programs. Some look at one particular program and compare the “treated” group to the prison average or group of nonparticipants. For example Thorpe et al. (1984) study individuals earning secondary degrees or certificates in prison in the state of New York. The authors find that 14% of the individuals with a degree or certificate were returned to custody during the survey period, while 20% of the overall population returned to confinement. The authors note that the significantly lower return rate of this sample may “be jointly attributed to the offenders’ motivation and capabilities as well as the impact of the program.” Thus, while the authors acknowledge the omitted variables bias in the selection process, they do not account for it in their study. While the Thorpe study lacks a genuine control group, Tyler et al. (2007) look at a pool of high school dropouts in Florida who were admitted to prison around the same time; some choose to earn their General Education Development (GED) during their time in prison and others do not. The authors use panels of quarterly earnings and find that the effect of GED certification declines over time. While, this paper focuses on employment, not recidivism, it indicates the potential for decreasing returns to prison education and is one of the few studies that explicitly compares individuals with similar educational backgrounds. A 1994 study by Adams et al. sampled 14,411 incarcerated individuals released from prison between 1991 and 1992 in Texas. The authors find that the more hours of participation in a program, the lower the likelihood of recidivism, but only for those “most educationally disadvantaged” inmates. Outside of this cohort, effects were not significant. Their findings suggest that initial level of education can be an important factor in understanding the impacts of educational programs. This is supported by Cho and Tyler (2013) who find that participation in For greater detail on the variety of studies that have been used to test education and vocation programs in prisons see https://www.ncjrs.gov/works/chapter6.htm and http://www2.ed.gov/offices/OVAE/AdultEd/OCE/19abstracts.html 15 Perry 12 adult basic education (ABE) is associated with higher post release earnings and employment rates, especially for minorities. They do not find, however, any positive effects of participation on reducing recidivism. Unfortunately, ICPSR3355 does not have information about prior education so I use instrumental variable analysis in attempts to tease out these omitted variables. Cho and Tyler (2013) also suggest that level of program completion is important, despite not being included in many studies. ICPSR3355 does include information about the level of participation and thus this will be taken into account in the analysis. These types of state and program specific studies are common in the literature but do not lend well to policy or cross-state comparison. Note that the three above studies all use different periods to define recidivism, rather than the BJS standard of 1,096 days. A study that provides more program comparison by Jenkins et al. (1993) compares various types of education programs in prisons. The study looked at ABE, GED, vocational education and post-secondary students in Maryland. The study finds a positive and significant benefit of education for students in ABE, GED, and vocational programs, but has inclusive finding for post-secondary students due to small sample size. The benefits include reduced recidivism, increased employability, and higher wages when compared to similar individuals who did not participate in any programing. There are few vocation-only studies in the literature. One particularly notable one in 1997 by Gaes and Saylor covered over 7,000 individuals in Federal Prisons over four years. It specifically looked at UNICOR 16 workers and vocational students as compared to a matched group of individuals with no such experience. They find UNICOR and vocational students have better institutional and half-way house adjustment, are less likely to return to federal prison, are more likely to be employed, and earn more money than the comparison group. This suggests that vocational programs can have significant impacts on post release outcomes, including recidivism. UNICOR workers are given vocational training and work on producing goods and services for the government. Individuals are paid and usually make less than $2 an hour. 16 Perry 13 Even fewer studies compare the differences between educational and vocational outcomes. Anderson et al. (1988) focus on both education and vocational programs. 760 individuals in Illinois were studied for one year following their release via reports from their parole officers. The individuals were divided into four groups while in prison: (1) vocational training, (2) vocational and academic training, (3) no vocational training, and (4) only academic work. Vocational (1) and vocational and academic (2) groups had higher employment and fewer arrests than the other groups. Those who received no vocational training had the highest crime rate; the academic only group had the lowest employment rate. Due to parole officer’s lack of knowledge regarding employment type of the individuals, the study admits to possible data bias. Still, it provides valuable insight and is one of few studies to look at education and vocation as well as employment and rearrest. Specifically it provides evidence that re-arrest rates can be significantly reduced for individuals participating in vocational programs in Illinois. A 1998 study by Mitchell Jancic finds lower recidivism rates for people earning their GED, participating in correctional educational and vocational programs, completing high school requirements, and completing or participating in post-secondary programs. His study looks at seven published studies conducted in a variety of states but it is difficult to draw conclusions across states or programs because each study measures education and vocation differently. My paper seeks to address such cross-state discrepancies by using a centrally compiled dataset. An Oklahoma statewide study by Brewster and Sharp (2002) looks at survival times (time until recidivism) for individuals participating in GED programs or vocational training. They find that completion of GED programs was strongly associated with longer survival times outside prison and that completion of vocational-technical training was linked to shorter survival times when compared to people participating in no programs. 17 While this study is strong and raises Other studies of Oklahoma find a similar result for vocation training, likely because of how individuals are enrolled in vocational programs and because often completion of such a program is a prerequisite for receiving parole. 17 Perry 14 interesting questions about vocational programs and the power of choosing to enroll, it is unable to exclude people who die, look at people who recidivate out of state, or account for level of program completion – all of which are accounted for in the analysis that follows. The paper most similar to the current analysis by Zarona Ismailova (2007) uses ICPSR3355 and looks at participation in education and vocational classes as they relate to recidivism. Using logistic regression analysis she finds that vocational training alone does not significantly impact recidivism, while education, or some combination of the vocational and educational training, is correlated with lower recidivism (Ismailova 2007). She controls for criminal history and several demographic variables, but does not account for state-to-state differences or selection bias in the analysis. Additionally, she does not consider the level of program completion. My paper adds to the abundant existing literature by looking at a nationally representative dataset of state prisons and specifically comparing the difference between educational and vocational programs. Unlike other papers that do this (Ismailova 2007; Brewster and Sharp 2002; Anderson et al. 1988) I control for state-to-state variation, look at different levels of program completion, and consider two measures of recidivism. I attempt to control for selection bias with instrumental variable two staged least squares analysis. Additionally, I look at the impact of education and vocational programs on time until recidivism using a Cox Proportional Hazard Model. Perry 15 Data This paper uses data from the Bureau of Justice Statistics Special Report: Recidivism of Prisoners Released in 1994. The data is managed by the Interuniversity Consortium for Political and Social Research (ICPSR) and is arguably the most comprehensive study of recidivism at the individual level available in the United States. 18 In 1998 the Bureau of Justice Statistics asked 15 State Departments of Corrections 19 to release basic records and criminal information on prisoners released in 1994. Of those, only five states recorded participation in education and vocational programs.20 Across all 15 states 38,624 incarcerated individuals were sampled; because of the limited data on education and vocation programs21, only 11,824 will be included in this study. Files were collected to be representative for each state across 13 offense categories using a stratified random sample by state in 12 of the 13 offense categories; in the case of sex-offenders (the remaining category), all files were pulled. All individuals were then matched, based on State and FBI identification numbers, to all appropriate Record of Arrest and Prosecution (RAP) sheets and state criminal history repositories. Files on each incarcerated individual released in 1994 were thus created, along with their arrest and adjudication history from their first ever arrest through When weighted, the sample of 33,796 ex-inmates represents 272,111 people: two-thirds of all State prisoners released with sentences greater than a year in 1994 (Langen & Levin). Still, for the purpose of this analysis weights will not be needed as they do not produce statistically different results according to the data codebook: Analysis done to assess the effect of non-adjustment found the effect to be negligible, both at the national level and at the individual State level. At the national level, adjusted weights produced a re-arrest rate of 67.5% (187,280 / 277,360 = 67.52%), virtually identical to the published rate (183,554 / 272,111 = 67.45%). Even among those individual States that required the greatest weight adjustment – Florida and Illinois - similarly negligible differences were found between rates produced by unadjusted weights and those from adjusted weights. (Codebook) 19 The fifteen states included in the study are AZ, CA, FL, IL, MD, MI, MN, NJ, NY, NC, OH, OR, TX, and VI. 20 The five states that reported educational and vocational programs in 1994 are FL, IL, NY, NC, and TX. 18 As is done in the BJS analysis only individuals where analysis=1 are included. This ensures that all individuals meet the following requirements listed in the Codebook: - had a RAP sheet - did not die in the 3 year follow up period - had a sentence was greater than 1 year - was not released via escaping, on appeal, to a warrant, for a transfer, or for administrative reasons 21 Perry 16 1,096 days (or three years) after their release in 1994. 22 Descriptive statistics can be seen in Appendix Table 8. A note on time period: At the time of data collection, the U.S. Congresses was passing many “tough on crime” policies. For example, in 1993 the Violent Crime Control and Law Enforcement Act was passed followed by the Higher Education Reauthorization Act of 1994 these acts essentially terminated Pell grant funding for post-secondary correctional education projects. Despite the fact that less than 5% of Pell money was directed towards individuals in prison, Congress felt it was inappropriate to subsidize the education of “criminals” (Ubah & Robinson 2003). With Pell grants cut, there remain four other ways individuals can get funding for higher education while in prison: Federal Perkins Funds, private foundation grants, private funds (their own or those of family members), and state-based education grants (Ubah & Robinson 2003). Because these acts were passed in 1994 it is unlikely that they had an impact on the cohort of individuals in ICPSR3355. Still, it speaks to the growing skepticism surrounding correctional education and vocation programs in the early 1990s. Additionally, it should be noted that in 1994 many agencies were transitioning toward electronic documentation and more extensive use of the internet. This may explain why only five of 15 states in the study recorded information about educational and vocational program participation to the BJS. 22 For example, the earliest arrest in the five sates of interest is in 1918. Perry 17 Methods PART ONE: Predicting Recidivism Probit analysis is used to estimate the probability with which an individual recidivates based on his/her participation in an educational or vocational program and other control variables. Interaction terms are generated to capture the variation of the 16 different combinations of program participation, with no participation in either as the base state. These states are mutually exclusive – an individual can fall into only one of the 16 categories below. TABLE 3: INTERACTION DUMMY VARIABLES FOR PARTICIPATION IN EDUCATION AND VOCATION PROGRAMS Level of Education Participation Level of Vocation Participation Completed Education Program Completed Vocation Program Partially Completed Vocation Program Unknown Completion of Vocation Program Did not enroll in Vocation Program Completed Vocation Program Partially Completed Vocation Program Unknown Completion of Vocation Program Did not enroll in Vocation Program Completed Vocation Program Partially Completed Vocation Program Unknown Completion of Vocation Program Did not enroll in Vocation Program Completed Vocation Program Partially Completed Vocation Program Unknown Completion of Vocation Program Did not enroll in Vocation Program Partially Completed Education Program Unknown Completion of Education Program Did not enroll in Education Program Percent 4.88 6.83 0.06 4.93 4.38 18.44 0.08 7.29 1.15 0.07 1.57 6.42 0.91 1.38 3.25 37.72 As noted earlier, this paper looks at two very different measurements of recidivism: rearrest and re-confinement. In the five-state sample, 61.77% of the individuals are re-arrested and 43.06% return to prison or jail within 1,096 days of their release. To account for variation, certain observable characteristics are used as controls: term served, race, sex, prior arrest history, age, and Perry 18 criminal offense category.23 Also, a dummy variable for each state is included to help account for the state-to-state variations in funding, policies, and crime rates – something not done in the Ismailova paper (2007). The models are based on the following equation: EQUATION 1: PROBIT MODEL FOR RE-ARREST AND RE-CONFINEMENT 15 4 𝑅𝑒𝑐𝑖𝑑𝑖𝑣𝑖𝑠𝑚𝑖 = 𝛽0 + ∑ 𝛽1𝑗 𝑒𝑑𝑢𝑣𝑜𝑐𝑗 + 𝛽2 𝑡𝑚𝑠𝑟𝑣 + 𝛽3 𝑛𝑜𝑛𝑤ℎ𝑖𝑡𝑒 + 𝛽4 𝑠𝑒𝑥 + 𝛽5 𝑝𝑝𝑟𝑖𝑑 + ∑ 𝛽6𝑘 𝑠𝑡𝑎𝑡𝑒𝑘 𝑗=1 12 𝑘=1 6 + ∑ 𝛽7𝑚 𝑠𝑚𝑝𝑜𝑓𝑓𝑚 + ∑ 𝛽8𝑛 𝑎𝑔𝑒𝑟𝑎𝑛𝑔𝑒𝑛 𝑚=1 𝑛=1 Where:24 𝑅𝑒𝑐𝑖𝑑𝑖𝑣𝑖𝑠𝑚𝑖 is a dummy variable representing two measures of recidivism: re-arrest and re-confinement within 1,096 days after release. 𝑒𝑑𝑢𝑣𝑜𝑐𝑗 represents the sixteen aforementioned interaction terms (Table 3) in which no education and no vocation is the base state and thus excluded from the above regression. 𝑡𝑚𝑠𝑟𝑣 is the length in months of the term served for which the individual was released in 1994. 𝑛𝑜𝑛𝑤ℎ𝑖𝑡𝑒 is a dummy variable where a value of one includes black, American Indian/Aleutian, Asian/Pacific Islander, and “other”. 𝑠𝑒𝑥 is a dummy variable for the sex of the released individual (1 = male; 0 = female). 𝑝𝑝𝑟𝑖𝑑 is a dummy variable indicating whether or not the individual had any prior sentences before the sentence that got him/her into the study. 23 24 Full descriptions of the controls can be found in Appendix Table 7 Details on the variable interpretations can be found in Appendix Table 7 Perry 19 𝑠𝑡𝑎𝑡𝑒 represents the state in which the individual was released. These are dummy variables for each of four states: Florida, Illinois, New York, and North Carolina. Texas is the base state. 𝑠𝑚𝑝𝑜𝑓𝑓 are 12 dummy variables for which offense category the individual is convicted and serving the 1994 sentence; when there are multiple, the most serious is used. The base state crime level is homicide. A detailed description of the offense levels can be found in Appendix 7. 𝑎𝑔𝑒𝑟𝑎𝑛𝑔𝑒 indicates one of six age groups based on the individuals’ age at release. The base state is age for individuals aged 14-17 at release. Coefficients ∑15 𝑗=1 𝛽1𝑗 are of primary interest. Negative and significant coefficients indicate that some combinations of vocational and educational programs are negatively correlated with the two dependent variables (𝑅𝑒𝑐𝑖𝑑𝑖𝑣𝑖𝑠𝑚𝑖 ) and thus that these programs may help reduce the likelihood that someone recidivates.25 All models use robust standard errors, to correct for any problems caused by potential heteroskedasticity. At the State Level Table 4 and 5 give the frequencies for each level of participation for both education and vocation across all five states. They suggest that an aggregate level analysis might not fully represent the underlying differences in education and vocation offerings across states. For example, in Florida all individuals participate in either an education or vocational program, but not both. In Texas, information is available for all individuals – there are no people recorded as “unknown” or “questionably completed.” As such this subsection dives further into each state to see which, if any, have programs that seem to be driving the aggregate results. The same probit models for re-arrest 25 Because correlation does not causation, this should be taken lightly Perry 20 and re-confinement are used but are adjusted to fit each state according to how education and vocation program participation has been recorded. TABLE 4: PARTICIPATION IN EDUCATION PROGRAMS ACROSS STATES Completed Partially Questionably Did Not Unknown Completed Completed Participate Participation Total FL 749 1,815 0 0 0 2,564 IL 168 119 323 1,707 0 2,317 NY 464 582 454 412 554 2,466 NC 22 15 516 1,347 147 2,047 TX 473 782 0 1,175 0 2,430 11,824 TABLE 5: PARTICIPATION IN VOCATIONAL PROGRAMS ACROSS STATES Completed Partially Completed Questionably Did Not Unknown Completed Participate Participation FL 475 2,089 0 0 0 2,564 IL 13 15 108 2,181 0 2,317 NY 463 573 4 651 775 2,466 NC 26 20 473 1,400 128 2,047 TX 319 297 0 1,814 0 2,430 Total 11,824 Perry 21 PART TWO: Selection Bias Selection bias is a problem for many studies looking at educational and vocational programs in prisons. Because these programs are often voluntary; only those individuals who are motived to participate will do so.26 This motivation factor could make these individual less likely to recidivate regardless of participation. Prior education level could also influence motivation to participate as individuals might seek education in accordance with their educational or vocational background.27 Basically, it is difficult to tell if it is the program or the person that leads to better release success and a lower chance of recidivism. In this way the selection bias of choosing to participate in a program could also be considered an omitted variable bias as there are omitted variables (motivation and prior education level) that impact the choice of whether or not to participate. One way to adjust for this is to proxy for the variables that cannot be measured with measurable instrumental variables in a two stage least squares analysis. In the case of individual participation in education or vocation programs, there are two unobserved omitted variables that could impact participation in the programs: motivation and prior education level. Due to restrictions in the data sntln (sentence length), sntln2 (sentence length squared), and nfrctns (number of infractions) will be used to proxy for motivation and educational history. Below are the justifications for each instrument: If sentence length is too short 28 individuals might not see the benefit of educational or vocational programs as program completion seems unlikely. When sentences are longer (to an unknown point) people may feel such participation could provide necessary stimulation and be a way to earn “good time.” Plus, individuals are more likely to be able to fully participate in a program. One could also argue, however, that long sentences remove any It appears that the programs in FL were not, in fact, voluntary as all individuals participated in either educational or vocational programs. See Table 4 and Table 5. 27 Unfortunately this is not recorded in ICPSR3355 28 Recall all individuals are incarcerated for at least one year as a prerequisite for inclusion in this analysis 26 Perry 22 incentive to be prepared at release as that time is far in the future. People with very long sentences might not see the benefit. To account for the non-linear incentives sntln2 is used. Nfrctns is a binary variable indicating whether or not the individual was formally disciplined for a prison-rule infraction. This can help proxy for the individual’s motivation as individuals who are more motivated to get out of prison and do well are also more likely to abide by rules in prison and try to earn “good time.” Because this model is more complicated than the probit model in Part 1, four groups, rather than 16, are used for participation in education and vocational training. The four groups are described below in Table 6 and the proportion of each group type can be seen in Figure 1. TABLE 6: PARTICIPATION IN EDUCATION AND VOCATION PROGRAMS, VARIABLE DESCRIPTIONS Variable Name Education Only Vocation Only Both None Description People who have only participated (com, quest, or part) in education programs People who have only participated (com, quest, or part) in vocation programs People who have participated (com, quest, or part) both in education and vocation programs People who have participated in neither an education or vocation program; this is the base group FIGURE 1: PERCENT PARTICIPATION IN EDUCATION AND VOCATION PROGRAMS, ALL STATES 16.63% Education Only Vocation Only 44.46% Both None 33.96% 4.95% Perry 23 For a two stage least squares estimator, first the endogenous variables of interest (eduonly, voconly, edu_voc) are regressed on the exogenous variables in the system plus the instruments. The resulting coefficients are then used to predict the variables of interest which are then used in a regression explaining recidivism. According to Stock and Watson (2011), all instruments must be both relevant and exogenous. For the single regressor, X, with instrument, Z, an instrument is considered relevant if 𝑐𝑜𝑟𝑟(𝑍𝑖 , 𝑋𝑖 ) ≠ 0. For the same case, the exogenous requirement is met if 𝑐𝑜𝑟𝑟(𝑍𝑖 , 𝑢𝑖 ) = 0 (Stock & Watson 2011). If there were only one regressor of interest simply looking at the F-statistic on the first stage regression would be sufficient to test the weakness (or relevance) of these variables. However, there are three variables of interest (eduonly, voconly, edu_voc ) so instead the CraggDonald Wald F-statistic is used. Baum, Schaffer, and Stillman suggest that a minimum F-value of 10 indicates strength of the instruments. Also, because this model is exactly identified (there is the same number of regressors as instruments) the over identifying restrictions test for exogeneity is not possible (Stock & Watson 2011), so there is no test for exogeneity. The Instrumental Variable Two Stage Lease Squares Regression (IV 2SLS) is then compared to a standard Ordinary Least Squares (OLS) regression using the three categories of participation (as described in Table 6) to see the impacts of accounting for motivation and to see if these omitted variables are driving the results. This is done separately only for those states that seem to have successful programs as demonstrated in the previous section, as opposed to at the aggregate five state level. Perry 24 PART THREE: Time until Recidivism While the previous sections examine the binary outcome of recidivating or not recidivating, this section looks specifically at when within the 1,096 day period people recidivate. This is of interest as it indicates the effect overtime for such programs. A common way to predict time until recidivism is with hazard models which estimate the likelihood of a given event (recidivism) at time 𝑡. Such models have been used since the early 1970s in this field and much earlier in medical treatment analysis. As early as 1974 Stollmack and Harris used hazard analysis to predict recidivism (Schmidt & Witte 1988). In the beginning, the majority of recidivism hazard model studies used parametric models and did not consider explanatory variables. 29 Only in 1977 did Schmidt and Witte consider different distribution models and explanatory variables when predicting time until recidivism. Since then both parametric and proportional hazard models 30 have been applied to recidivism. Schmitt and Witte provide an indepth analysis of different models used to predict recidivism by doing an extensive literature review and applying a variety of hazard models to a 1978 and 1980 cohort of individuals released from state prisons in North Carolina. Their analysis serves as a guide to understand which proportional model is the best fit for a given dataset. Schmidt and Witte (1988) find that including explanatory variables is helpful in predicting recidivism, especially across cohorts. For this reason, and because this paper is concerned with the differing effects of education and vocation programs on recidivism only models that allow for explanatory variables will be considered; models that try to fit data based on certain distributions will not be used. 31 Per recommendation by Schmidt and Witte in their book, Predicting Recidivism Using Survival Models, the Cox Proportional Hazard Model will be used in this paper. Though they Schmit & Witte cite the following; Stollmack & Harris (1974), Maltz & McCleary (1977), Maltz (1984), Maltz, McCleary, & Pollock (1979) 30 Rhodes and Matsuba (1984) 31 Such functional forms common in this field are exponential, log-normal, log-logistic, Weibull, and LaGuerre 29 Perry 25 find there are more complicated models that more accurately predict the North Carolina cohorts. They find value in the Cox model because it “allows one to estimate the effects of individual characteristics on survival times without having to assume a particular form for the distribution function.” (Schmidt & Witte 1988). Other studies also employ the Cox Proportional Hazard Model.32 The hazard function, conceptually is: 33 ℎ(𝑡) = 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑖𝑛𝑑𝑖𝑣𝑖𝑑𝑢𝑎𝑙𝑠 𝑟𝑒𝑐𝑖𝑑𝑖𝑣𝑎𝑡𝑖𝑛𝑔 𝑎𝑡 𝑡𝑖𝑚𝑒 𝑡 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑖𝑛𝑑𝑖𝑣𝑖𝑑𝑢𝑎𝑙𝑠 𝑠𝑢𝑟𝑣𝑖𝑣𝑖𝑛𝑔 𝑎𝑡 𝑡𝑖𝑚𝑒 𝑡 If an individual has a row vector of information X then the proportional hazard model assumes a hazard rate of: ℎ(𝑡|𝑋) = ℎ0 (𝑡)𝑒 𝑋𝛽 Where ℎ0 (𝑡) is the baseline hazard function that sets all parameters equal to 0, a purely theoretical value. Suppose individual 𝑖 recidivates at time 𝑡𝑖 then (if there are 𝑁 total individuals, 𝑛 of whom recidivate within the 1,096 day time period) for any time 𝑡, 𝑅(𝑡) is the set of all 𝑖 = 1,2 … 𝑁 such that the individual 𝑖 was still at risk just prior to time 𝑡. For any observed failure time, 𝑡𝑖 , the probability of individual 𝑖 failing is given by: ℎ(𝑡𝑖 |𝑋𝑖 ) ∑𝑗∈𝑅(𝑡𝑖) ℎ(𝑡𝑖 |𝑋𝑗 ) = 𝑒 𝑋𝑖𝛽 ∑𝑗∈𝑅(𝑡𝑖) 𝑒 𝑋𝑗𝛽 To find the partial likelihood function, the above equation is multiplied over all 𝑛 failure times. Below, 𝐿 , is the probability of any individual failing (in this case recidivating): 𝑛 𝐿 = ∏{ 𝑖=1 32 33 𝑒 𝑋𝑖 𝛽 ∑𝑗∈𝑅(𝑡𝑖) 𝑒 𝑋𝑗𝛽 } Used by Rhodes and Matsuba (1984) and Sherman and Berk (1984) The explanation following here forward draws heavily from Schmidt & Witte, 1988, pg 84 Perry 26 Looking at the hazard ratios given by the above will enable an understanding of the isolated impact of participation in education only, vocation only, or both in reducing the likelihood of re-arrest and re-confinement at any given time, 𝑡. Hazard rates below one indicate decreases in the rate of recidivism for the 1,096 day period, while values greater than one indicate increases in the rate of recidivism for the period. Perry 27 Results PART ONE: Predicting Recidivism Re-arrest: At first glance there are significant and negative impacts of educational and vocational programs on re-arrest. Figure 2, on the following page, presents the coefficients on the variables of interest for re-arrest at the aggregate level. Longer bars indicate greater impact and stars indicate significance level. As seen in Figure 2, eight of the fifteen participation combinations have significant negative relationships with re-arrest at the 10% level or higher. This seems to suggest that educational and vocational programs are helping reduce re-arrest. Still, the story is muddled; completing both programs (educom_voccom) for example decreases the probability of rearrest by .052 and is significant at the 10% level, completing only a vocational program (eduno_voccom) has no effect, and completing only an education program (educom_vocno) reduces the probability of re-arrest by .044. Even more confusing is the significant and positive impact of enrolling in and questionably completing both an education and a vocation program (eduquest_vocquest), which suggests that such individuals are more likely to be re-arrested than those whom have had no education or vocational programming (eduno_vocno). The largest significant impact is for individuals who questionably complete an education program and never participate in vocational programs (eduquest_vocno); these individuals have a decreased probability of re-arrest by .108. Given that the probability of re-arrest in the sample is 62.6%, these reductions are significant. Still, it is unclear from this analysis which program type has the greatest impact on re-arrest. Perry 28 FIGURE 2: SIGNIFICANCE OF RE-ARREST COEFFICIENTS, ALL STATES *p<.10; **p<.05; ***p<.01 Values are reported from dprobit, thus they indicate changes in probabilities For a full model with controls reported, see Appendix Table 1 n=10,157 Perry 29 Re-Confinement: The results for re-confinement at the aggregate level are similar to those for re-arrest. Figure 3 presents the coefficients on the variables of interest for re-confinement at the aggregate level. As with Figure 2, longer bars indicate greater impact and stars indicate significance. Of the fifteen participation combinations, ten have a negative and significant impact on reconfinement. The greatest impact can be seen for individuals completing part of an education program and questionably completing a vocational program (edupart_vocquest); for such people the probability of re-confinement decreases by .2507, significant at the 10% level. 34 The probability of re-confinement in the five state sample is 43.06% (lower than the 61.77% for re-arrest). Just as for re-arrest, completing both education and vocation (educom_voccom) has significant reductions, as does completing only educational programs (educom_vocno), but completing vocation only does not (eduno_voccom). These data seem to suggest that there are generally significant impacts of education and vocation programs on reducing re-confinement. While the results for both re-arrest and re-confinement seem to indicate that education and vocation programs reduce both re-arrest and re-confinement, it is difficult to draw specific conclusion about which programs in particular are more effective. As noted in the Methods section this could be due to underlying differences between the states; this is explored in the following section. One possible explanation for the significance of questionable and partial completion is that individuals who enroll in programs and otherwise earn “good time” may be released early, before they can finish the programs. In this way, failing to complete a program could be a positive signal. 34 Perry 30 FIGURE 3: SIGNIFICANCE OF RE-CONFINEMENT COEFFICIENTS, ALL STATES *p<.10; **p<.05; ***p<.01 Values are reported from dprobit, thus they indicate changes in probabilities For a full model with controls reported, see Appendix Table 1 n=10,157 Perry 31 At the State Level It is possible that certain states are driving the aggregated results. This could be due to differences in program effectiveness, recording standards (see Tables 4 and 5), or re-arrest and reconfinement rates. Figure 4 indicates the re-arrest rate and re-confinement rate by state, with the total sample below. While the re-arrest rate is higher than the re-confinement rate in all states, the difference ranges from less than half a percentage point (NY) to nearly 40 percentage points (FL). This variation suggests significant differences in conditions across states and begs for deeper state level analysis. FIGURE 4: RECIDIVISM BY STATE Values indicate the percent of individuals re-arrested or re-confined within 1,096 days of their release n is the total number of individuals in the data from each state Perry 32 The results from the probit model in Equation 1 run independently for each of the five states can be seen for re-arrest in Appendix 2 and for re-confinement in Appendix 3. 35 Of the five states, only Illinois and New York demonstrate programs that are significantly related to recidivism; as such, much of the following analysis focuses heavily on these two states. In Florida, program participation (either education or vocation) is mandatory which means there is no viable comparison group within the state. North Carolina, as seen in Tables 4 and 5, has very low levels of participation in both education and vocation programs. Texas, while more evenly distributed across participation levels, has few significant results in the state level probit analysis. For these reasons, and because Illinois and New York indicate significant relationships between education and vocation programs and recidivism, they are the focus of the state level analysis that follows. 35 Note that because states record different levels of program completion the base condition is different for each state Perry 33 Re-Arrest: Educational and vocational programs have a greater impact on reducing rearrest in Illinois than in New York. These impacts are demonstrated visually in Figure 5 which depicts the coefficients on each level of participation; the longer the line the greater the impact. In Illinois where the baseline probability of re-arrest is .7085, there is a .456 decrease in the probability of re-arrest for individuals who complete vocation programs without participating in education programs (eduno_voccom), and a .141 decrease in probability of re-arrest for people who complete education programs without participating in vocation programs (educom_vocno). In Illinois people participate only in one type of program (there are no people in both education and vocation) though programs are not mandatory. This could reflect a specialization of the prisons (each prison only offers education or vocation programs) or a rule that requires individuals to participate in one program only. Whichever reason, all program combinations are highly significant at least at the 5% level. Between the two, vocational programs seem to have a greater effect across the board as the three vocational coefficients are larger than the three educational coefficients. In New York, where the baseline probability of re-arrest is .6119, the impacts are smaller but still significant. In New York, nine of the eleven participation combinations are significant, but the story is less clear as to which program is the most effective. The largest impact can be found for people who participate in and complete vocational programs without participating in educational programs (eduno_voccom); they see a .304 reduction in the probability of re-arrest. This is nearly the same as the impact for people completing educational programs and participating in but not completing vocational programs (educom_vocpart) that see a decrease in the likelihood of re-arrest of .296. In New York the results are smaller than for Illinois, but spread across more types of participation. Perry 34 FIGURE 5: SIGNIFICANCE OF RE-ARREST COEFFICIENTS, NEW YORK & ILLINOIS *p<.10; **p<.05; ***p<.01 Values are reported from dprobit, thus they indicate changes in probabilities For a full model with controls reported, see Appendix Table 2 Illinois: n=2,124 New York: n=1,193 Perry 35 Re-Confinement: Just as New York and Illinois drove the results for re-arrest, they also drive the results for re-confinement, seen in Appendix 3. The coefficients on the variables of interest are visually displayed in Figure 6; the longer the line the greater the impact. In Illinois the impact of education and vocational programs is again larger than in New York. Recall the baseline re-confinement rate in Illinois is 39.12%. For all types of participation in Illinois, there is at least a .2636 decrease in the probability of re-confinement significant at the 5% level. The largest impact can be seen for individuals who questionably complete an education program and do not participate in vocation programs (eduquest_vocno); they have a reduction of .386 in the likelihood of re-confinement. Because the baseline rate is 39.12% this reduction brings the probability of re-confinement close to zero. It is less clear which program has greater impacts, education or vocation, but it seems education is slightly greater (the lower three blue lines in Figure 6 are slightly longer than the top three blue lines). Still, the differences are small and both programs have significant and large reductions in the likelihood of re-confinement. In New York, where the baseline probability of re-confinement is 58.9%, the greatest impact can be seen for individuals who complete a vocational program and part of an educational program (edupart_voccom). For such individuals the probability of re-confinement is decreased by .2558. Overall, both education and vocation programs seem to have negative and significant impacts on reconfinement – seven of the eleven combinations are significant at the 5% level or higher. Again, it is not obvious which program has the overall greater impact, but it seems that vocational programs are slightly more impactful. Perry 36 FIGURE 6: SIGNIFICANCE OF RE-CONFINEMENT COEFFICIENTS, NEW YORK & ILLINOIS *p<.10; **p<.05; ***p<.01 Values are reported from dprobit, thus they indicate changes in probabilities For a full model with controls reported, see Appendix Table 3 Illinois: n=2,124 New York: n=1,193 Perry 37 PART TWO: Selection Bias As the results from Part 1 demonstrate, the aggregate results obfuscate the deeper programmatic variation across states. For this reason, selection bias is only considered for those states with programs that have programs which significantly reduce recidivism: Illinois and New York. First, it is important to check the validity of the instruments: sentence length, sentence length squared, and number of infractions. Instrumental variables must be both relevant and exogenous (Stock & Watson). Recall that because there are three regressors (eduonly, voconly, edu_voc ) a Cragg-Donald Wald F-statistic is used to test for whether or not the instruments are “relevant.” For Illinois, eduonly and voconly are the only two regressor and sntln and sntln2 are used as instruments. This is because the number of infractions (nfrctns) is not recorded in Illinois, and there are no observations for edu_voc. For Illinois, the Cragg-Donald F-statistic is 2.303. For New York, all three regressors are included and the resulting F-Statistic is .007. 36 Baum, Schaffer, and Stillman (2007) suggest that a minimum F-statistic of 10 indicates strong instruments. Clearly then, the suggested instruments are weak. No other instruments included in the data, however, provide a better estimate so these will be used but with great caution. Further, since in both regressions the model is exactly identified, the over identifying restrictions test for exogeneity is not possible (Stock & Watson 2011). Because the instruments are weak the following results should be taken very lightly. Note that these values are the same for re-arrest and re-confinement because values are calculated based on the relationship between regressors and instruments without consideration to the dependent variable 36 Perry 38 Re-Arrest: Table 7 shows the coefficients of interest for re-arrest in Illinois and New York. Columns 2 and 4 are for the two staged least squares instrumental variable regression. These columns represent the impact of such programs when controlling for motivation. Columns 3 and 5 contain the OLS regression output not controlling for motivation. For both Illinois and New York the instrumental variable regressions indicate that the effect of these programs is largely stemming from motivation. The education and vocational programs are significant when included in the OLS regression (column 3 and 5), but not significant in the instrumental variable two stage least squares regression. This suggests that the effect from these programs is largely because those individuals who elect to participate are more motivated then those who do not and are thus more likely to avoid re-arrest after release. However, these large coefficients and standard errors imply that the weak instruments have biased the magnitudes of the coefficients and their standard errors upwards. Note that this problem is especially prominent for the New York sample where the Fstatistic is particularly low. For the OLS results, in New York, the coefficients for all three levels of participation are similar and an F-test (F=.16, p=.8553) cannot reject the hypothesis that the estimated effects are equal. 37 In Illinois, however, the null hypothesis that programs have equal impact can be rejected at the 5% significance level. 38 This suggests that in Illinois vocational programs have greater impact on re-arrest than educational programs, while in New York the programs are more similar in their impacts on re-arrest. The following null hypotheses are also tested: eduonly=voconly, voconly=edu_voc, and eduonly=edu_voc . None can be rejected. 38 Because there are no observations for edu_voc in Illinois, only eduonly=voconly is tested. 37 Perry 39 Re-Confinement: Table 8 shows the coefficients of interest for re-confinement in Illinois and New York. Columns 2 and 4 are for the two staged least squares instrumental variable regression. These columns represent the impact of such programs when controlling for motivation. Columns 3 and 5 contain the OLS regression output not controlling for motivation. The results for re-confinement are near identical to those for re-arrest. The most notable difference is the negative and significant coefficient for voconly in Illinois. This suggests that even when controlling for motivation of people selecting to participate, there are program impacts on re-confinement. Still, this is likely biased upward as the instruments are incredibly weak. The instrumental variable regression for New York suggests that though education and vocational programs have an impact, as demonstrated by column 5 of Table 8, and this is likely because participants are more motivated than individuals who do not participate in the programs. As with re-arrest, the large coefficients and standard errors suggest that the weak instruments have biased the magnitudes of the coefficients and their standard errors upwards. For the OLS regressions testing the equality of coefficients on eduonly, voconly, and edu_voc 39 give a p-values of .229 and .269 for Illinois and New York respectively. Thus the null hypotheses that the estimated effects of the levels of participation are equal cannot be rejected. 40 Though all programs have significant impacts, the F-tests suggest that in Illinois and New York there are no significant differences in the effects of different program types on re-confinement. Basically, for both re-arrest and re-confinement there is reason to believe that motivation is driving the results (the exception being voconly in Illinois). These results must be taken lightly as the instruments are weak. This analysis also suggests that the effects of different programs may not be significant. Recall this variable only exists for New York The following null hypotheses are tested: eduonly=voconly=edu_voc, eduonly=voconly, voconly=edu_voc, and eduonly=edu_voc. None can be rejected. 39 40 Perry 40 TABLE 7: INSTRUMENTAL VARIABLE ANALYSIS: RE-ARREST, ILLINOIS & NEW YORK Illinois Variable eduonly voconly edu_voc n Instrumental Variable Estimation New York OLS Regression Coef. -.856 (.634) -.379 (.393) -.184 (.023) -.278 (.042) -- -- 2,127 Instrumental Variable Estimation *** *** 2,124 -2.55 (11.95) -3.63 (20.67) -1.89 (3.63) OLS Regression Coef. -.157 (.044) -.126 (.060) -.156 (.037) 1,039 *** ** *** 1,039 *p<.10; **p<.05; ***p<.01 Numbers reported in parenthesis are robust standard errors For a full model with controls reported, see Appendix Table 4 TABLE 8: INSTRUMENTAL VARIABLE ANALYSIS: RE-CONFINEMENT, ILLINOIS & NEW YORK Illinois Variable eduonly voconly edu_voc n Instrumental Variable Estimation -.245 (.454) -.521 (.257) -2,127 ** New York OLS Regression Coef. -.421 (.019) -.389 (.029) -2,124 *p<.10; **p<.05; ***p<.01 Numbers reported in parenthesis are robust standard errors For a full model with controls reported, see Appendix Table 3 Instrumental Variable Estimation *** *** -3.39 (14.23) -.886 (24.64) -1.71 (4.30) 1,039 OLS Regression Coef. -.100 (.044) -.103 (.059) -.155 (.039) 1,039 ** * *** Perry 41 PART THREE: Hazard Model The Cox Proportional Hazard Model teases out the relationship of the control variables over time when predicting the likelihood that an individual recidivates. Of individuals who are rearrested within three years, over 15% are re-arrested within their first month out of prison; within the first year that increases to 38.95% of individuals. For re-confinement, around 5% return to prison or jail within one month of their release; that number jumps to 42.66% within one year. This can be seen in Figure 7A and 7B, which show the distribution of re-arrest and re-confinement, respectively, in the 1,096 day follow-up period. Each bar represents the percent of individuals who experience their first post-release arrest (Figure 7A) or re-confinement (Figure 7B) in a given month after release. These figures suggest that time until recidivism is another important dimension to study. It is possible, for example, that the impact of education and vocation programs wear off or that they effectively delay the time until recidivism. FIGURE 7: DAYS UNTIL FIRST RECIDIVISM, ALL STATES Figure A: Re-Arrest Figure B: Re-Confinement These histograms represent the percent of individuals who experience their first post-release arrest (Figure A) or reconfinement (Figure B) within 1,096 days of their 1994 release. Because the period is three years, each of the 36 bars represent one month. The horizontal axis indicates days until recidivism. n=10,157 Perry 42 Table 9 includes the hazard estimates on the variables of interest at the aggregate level for both re-arrest and re-confinement. As seen in Table 9, below, all three levels of participation have estimated hazard rates of less than one, which indicate a decrease in the likelihood of re-arrest and re-confinement in every period. TABLE 9: HAZARD RATIO, RE-ARREST AND RE-CONFINEMENT, ALL STATES Variable eduonly voconly edu_voc n Re-Arrest Hazard Ratio .790 (.030) .781 (.049) .858 (.042) 10,157 *p<.10; **p<.05; ***p<.01 For a full model with controls reported, see Appendix Table 5 Numbers reported in parenthesis are robust standard errors *** *** *** Re-Confinement Hazard Ratio .656 (.039) .726 (.065) .745 (.051) 10,157 *** *** *** Perry 43 Re-Arrest: The results of the Cox Proportional Hazard model, in Table 9, indicate that all three levels of program participation significantly reduce the rate at which people return to prison. For people who participate in only education programs, holding all else constant, the rate of rearrest is reduced by 21%.41 For people only participating in vocational programs the reduction is 21.9% and for people participating in both the reduction is 14.2% - all significant at the 1% level. This suggests that education and vocational programs decreased the rate at which individuals are re-arrested, effectively delaying their return to the criminal justice system. After conducting a Chi-squared test to see if there are significant differences between the estimated effects of the three program types in the hazard analysis, the null hypothesis of equal impact cannot be rejected at the five-state level. This indicates that all programs have similar results. It should be noted that the difference between the coefficients on eduonly and edu_voc is significant at the 12% level, but that no other program effects are significantly different. One way to visualize the reductions in frequency of re-arrest is with Kaplan-Meier curves. These curves, seen in Figure 8, demonstrate the survival functions for each type of participation. A given point on the line indicates that 𝑥 days after release 𝑦% of individuals have not yet recidivated. The slope of the line indicates how quickly people return to prison. The below figure suggests that individuals with neither education nor vocation participation (𝑛𝑜𝑛𝑒) have the steepest survival curve, meaning their likelihood of re-arrest is highest. People in both programs have an initial drop, but level out after the first few days. Also, the below curves demonstrate that less than half of the sample stays out of re-arrest, as seen by the ending point on all four graphs. 41 (1-hazard ratio)*100 Perry 44 FIGURE 8: RE-ARREST KAPLAN-MEIER SURVIVAL ESTIMATES BY PROGRAM TYPE, ALL STATES Y axis is the percent “surviving” – not being re-arrested Analysis Time, the X axis, is the number of days since release Perry 45 Re-Confinement: The relationship between re-confinement and participation in education or vocation program is similar to that of re-arrest, as seen in the last column of Table 13. For individuals participating only in an educational program, holding all else constant, there is a 34.4% reduction in the rate of re-confinement and a 27.4% reduction for people participating only in a vocational program; both results are significant at the 1% level. Individuals participating, to some degree, in both programs have a reduction of 25.5% in the rate of re-confinement also significant at the 1% level. This suggests that educational and vocational programs significantly reduce the rate at which individuals are back behind bars. This is in line with the findings in Brewster and Sharp which suggest that education programs increase survival times, but contrary to their finding that vocational programs reduce survival time. This analysis suggests that both programs lead to longer survival times (a lower rate of re-confinement). After conducting a Chi-squared test to see if there are significant differences between the estimated effects of the three program types in the hazard models, the null hypothesis of equal impact cannot be rejected at the five-state level. This indicates that all programs have similar results. While the difference between eduonly and edu_voc is significant at the 10% level, no other programs are significantly different. Figure 9 has the four Kaplan-Meir survival estimates by program type for re-confinement. The steepest survival curve is for people participating in neither program; it is also the one with the lowest end point. Note that these slopes are generally less steep than the slopes for re-arrest (Figure 8) indicating that the rate of re-confinement is slower than the rate of re-arrest. Also note that the end points are higher for re-confinement, representing that fact that fewer people are reconfined than re-arrested. The similar slopes on the Kaplan-Meir survival estimates for eduonly, voconly, and both confirm the findings of the Chi-squared test that suggests there is no difference between the estimated effects of the program types on re-confinement. Perry 46 FIGURE 9: RE-CONFINEMENT KAPLAN-MEIER SURVIVAL ESTIMATES BY PROGRAM TYPE, ALL STATES Y axis is the percent “surviving” – not being re-confined Analysis Time, the X axis, is the number of days since release Perry 47 At the State Level It is prudent to look specifically at the states that drove the analysis in Part 1: Illinois and New York, as the results in these states may provide better hazard estimates than the aggregate level analysis. Re-Arrest: Table 10 has the results of the Cox Proportional Hazard model for the variables of interest in Illinois and New York.42 The results indicate that all relevant levels of program participation significantly reduce the rate at which people are re-arrested in both Illinois and New York. In Illinois, people who participate in education programs, holding all else constant, have a reduced rate of re-arrest by 50.1%.43 For people only participating in vocational programs the reduction is 61.7%. In New York the reductions are smaller but still significant. For individuals exclusively in education programs the reduction in the rate of re-arrest is 28.1%; nearly identical to the 28.6% reduction for people only in vocational programs. People in both programs (edu_voc) have a rate of re-arrest reduced by 32. 7%. TABLE 10: HAZARD RATIO FOR RE-ARREST, ILLINOIS & NEW YORK Variable eduonly voconly Illinois Hazard Ratio .499 (.034) .383 (.056) *** *** edu_voc -- n 2,124 New York Hazard Ratio .719 (.082) .714 (.113) .673 (.068) *** ** *** 1,193 *p<.10; **p<.05; ***p<.01 For a full model with controls reported, see Appendix Table 6 Numbers reported in parenthesis are robust standard errors 42 43 Recall that individuals in Illinois participate in only one type of programming so edu_voc has zero observations (1-hazard ratio)*100 Perry 48 Figure 10A and 10B, on the following page, have the Kaplan-Meier curves for Illinois and New York respectively. The figures suggest that individuals with no education or vocation participation (𝑛𝑜𝑛𝑒) have the fastest rate of re-arrest and, on the whole, and have the highest rate of re-arrest in the 1,096 day period (evidenced by the low ending value of the line) for both states. In Illinois the rate of re-arrest is lowest for individuals participating only in vocational programs, whereas in New York the hazard estimates for the education only and vocation only cohorts are nearly identical – evidenced by the nearly identical curves in Figure 10B. The Chi-squared test of the null hypothesis that estimated effects of eduonly=voconly in Illinois can reject the null hypothesis with a p-value of .084. In New York, however, none of the four possible null hypotheses 44 can be rejected indicating that there are no notable differences between only education, only vocation, and some combination of the two. This suggests that vocational programs in Illinois have a significantly greater impact on reducing re-arrest than do education programs, but that the type of programming matters less in New York. 44 Null hypotheses include: eduonly=voconly=edu_voc, eduonly=voconly, eduonly=edu_voc, and voconly=edu_voc Perry 49 FIGURE 10A: SURVIVAL MODELS FOR RE-ARREST, ILLINOIS FIGURE 10B: SURVIVAL MODELS FOR RE-ARREST, NEW YORK Y axis is the percent “surviving” – not being re-arrested Analysis Time, the X axis, is the number of days since release Perry 50 Re-Confinement: Table 11 has the results of the Cox Proportional Hazard model for the variables of interest in Illinois and New York.45 The results indicate that four of the five relevant levels of program participation have statistically significant reductions on the rate of reconfinement. In Illinois, people who participate in education programs, holding all else constant, have a reduced rate of re-arrest by 91.9%.46 For people only participating in vocational programs the reduction is 83.5%. These are by far the largest reductions observed in the analysis. In New York the reductions are smaller and the effect for people participating in educational programs is less significant.47 People in vocation programs only, however see a 31.8% reduction in the rate of re-confinement, similar to people participating in both programs that see a decrease of 30%; both significant at the 5% level or more. TABLE 11: HAZARD RATIO FOR RE-CONFINEMENT, ILLINOIS & NEW YORK Variable eduonly voconly Illinois Hazard Ratio .081 (.021) .165 (.068) *** *** edu_voc -- n 2,124 New York Hazard Ratio .855 (.118) .682 (.132) .700 (.087) *p<.10; **p<.05; ***p<.01 For a full model with controls reported, see Appendix Table 6 Numbers reported in parenthesis are robust standard errors Recall that individuals in Illinois participate in only one type of programming (1-hazard ratio)*100 47 It is significant at the 25% level 45 46 1,193 ** *** Perry 51 In Illinois, the hazard model suggests that the greatest effect is from education program participation, but the Chi-squared test indicates that the null hypothesis that the effect of eduonly is equal to the effect of voconly can only be rejected at the 14% significance level. For New York the Chi-Squared tests reveal that eduonly and edu_voc are different at the 10% level, with the addition of a vocational program to an educational program the rate of re-confinement is reduced. Further, the magnitude of the coefficients suggest that vocational programs have the greater effect at reducing the rate of re-confinement for incarcerated individuals in New York when compared to education programs but with small sample size of participants in voconly, the Chi-squared tests do not reveal a statistically significant difference. Because of the Chi-squared results, this should be taken lightly. Figure 11A and 11B have the Kaplan-Meier curves for Illinois and New York respectively. Notably, the re-confinement rates for individuals in Illinois participating in education or vocation programs are very low – nearly horizontal lines. These curves are especially shallow when compared to the curve for individuals with no such programming. In New York the differences between survival curves are not as noticeable, though the rate of re-confinement for people with no programming is steeper than for people with some kind of programming. Perry 52 FIGURE 11A: SURVIVAL MODELS FOR RE-CONFINEMENT, ILLINOIS FIGURE 11B: SURVIVAL MODELS FOR RE-CONFINEMENT, NEW YORK Y axis is the percent “surviving” – not being re-confined Analysis Time, the X axis, is the number of days since release Perry 53 Conclusion This study adds to the existing literature and suggests educational and vocational programs are correlated with significant reductions in recidivism. I find greater influence from vocational programs in the probit analysis (Part 1) and the hazard analysis (Part 3) than from educational programs. That is not to say that educational programs do not have an impact on re-arrest and reconfinement, they clearly do, just that the impact of vocational programs is greater. This is contrary to many previous studies, specifically the Ismailova paper, which uses the same dataset and finds a greater impact from education programs. Unlike previous studies, my analysis addresses both re-arrest and re-confinement of the same individuals and tracks them for the BJS standard of 1,096 days after release. I find reductions for both re-arrest and re-confinement, with greater decreases for re-confinement. This analysis also suggests that it is important to do separate state analyses as New York and Illinois largely drive the results for the five-state group. While state-by-state case studies are important, there is still a need for standardized data. There is reason to believe that motivation might be influencing the results found in Part 1, but the weakness of the instrumental variables means these results cannot be taken with any great certainty. Part 2 may suggest that the programs themselves are not inherently decreasing re-arrest and re-confinement (outside of vocational programming’s impact on re-confinement). Still, this begs the question: is selection bias bad? As Part 1 demonstrates the people who participate in education and vocation programs have significantly reduced rates of recidivism. Perhaps these programs allow individuals to signal their motivation, gain “good time,” and stay mentally engaged in prison. Saylor and Gaes (1997) note that allowing incarcerated individuals the outlet for their motivation is essential; eliminating programs because the people who elect to take them are possibly more motivated and thus are less likely to be re-arrested or re-confined denies people the Perry 54 chance to succeed. In addition to providing an outlet for motivation and helping to reduce recidivism, there are also more intrinsic advantages to providing individuals with access to education and vocational training; such as literacy, mental stimulation, decreased boredom, and confidence building (Ubah & Robinson 2003). Because selection bias into programs is often cited as a big issue in many studies thinking about which instruments could help proxy for motivation would be helpful for any future follow-up studies. It should be remembered, however, that selection bias in this case could still be a justification in favor of education and vocation programs. In addition to adding to the literature by attempting to control for selection bias, this study has a unique approach to hazard analysis. While many studies look at hazard analysis and recidivism, few also look at the role of education and vocation programs in influencing survival times. I find that both education and vocation programs increase survival times for re-arrest and reconfinement, with the larger effects for re-confinement. Between the two program types, the greater impact tends to be for vocational programs which again points to the success of these programs found in Part 1. This analysis suggests that such programs help to reduce the rate of rearrest and re-confinement over time and effectively delay the incidence of recidivism. What makes the Illinois and New York programs successful? One notable finding from this study is the variation between states in having effective programs. Illinois and New York in particular had particularly effective programs in 1994 for reducing recidivism of individuals through educational and vocational programs. What exactly made their programs more effective is an area for future study and outside of the scope of this analysis. Ubah and Robinson (2003) snuggest that Illinois provides more state funding for people in secondary education programs, enabling more low-income incarcerated individuals to earn a secondary degree in prison. New York was also quick to find alternative funding when Federal Pell Grants were no longer available to incarcerated individuals (Ubah & Robinson 2003). Still, more research on these two states and the ways in which the other states in the BJS study have updated their programs and changed Perry 55 monitoring practices would provide a more grounded understanding of what programs work and for whom. Limits and Constraints: This study highlights the lack of comparable data on education and vocation programs across states. Given that ICPSR3355 is the official BJS dataset on recidivism it is startling that only five of 15 states had information on the education and vocational program participation of individuals in their system. The data is nearly 20 years old, since the individuals release in 1994 there have likely been significant policy, technology, and programmatic changes across the states. Such poor data tracking makes it difficult to project costs, compare programs, or make policy improvements. While there are numerous state level studies arguing for or against vocation and education programs there is no common language or data collection. Some states track enrollment but not level of completion, others are very detailed but only on one or two facilities. It is also possible that individual institutions are recording this information but not recording it to the BJS in a standard way, or at least that this was the case in 1994 and no such data is presently available to the public. As correctional spending increases and recidivism rates stay the same, it is time to re-evaluate what can be done to better the system. To move the already robust literature forward, there needs to be better centrally compiled data that can be compared year over year and across states. As noted previously, the limited data also prevents including information regarding prior incarceration or post release information such as education, employment, and income. Having such information at entry would provide for better controls, while having the information for the 1,096 day follow-up period would strengthen the probit and hazard analysis. The Cox Proportional Hazard model in particular can account for changes in the follow-up period to help better predict the hazard rates. The inclusion of such information in future data-sets, though cumbersome to collect, would prove incredibly helpful in creating more accurate models. Perry 56 Should funding to correctional education and vocation programs be increased? Despite the limits of this study, my results offer support for both educational and vocational programs as they help to reduce re-arrest and re-confinement. While my results tend to favor vocational programs, the benefits of both are worth noting. Given the increasing prison population in America and the corresponding spending increases, greater efforts toward reducing recidivism should be taken. Educational and vocational programs provide such reductions and should thus be improved and made more available. Perry 57 References 2006 Annual Report to the Florida Department of Education: Juvenile Justice Educational Enhancement Program. Publication. College of Criminology and Criminal Justice, 2006. Adams, K., Bennett, K. J., Flanagan, T. J., Marquart, J.W., Cuvelier, S. J., Fritsch, E., Gerber, J., Longmire, D., and Burton, V. (1994) "A Large-Scale Multidimensional Test of the Effect of Prison Education Programs on Offenders' Behavior." Prison Journal 74:433-99. Anderson, Dennis B., Sara L. Anderson, and Randall E. Schumacker. "Correctional Education A Way to Stay Out: Recommendations for Illinois and a Report of the Anderson Study." Illinois Council on Vocational Education. 1988. Angrist, Joshua D., and Jorn-Steffen Pischke. "Ivreg2 Update." Mostly Harmless Econometrics. , 20 Feb. 2010. Aos, Steve, Marna Miller, and Elizabeth Drake. (2006). Evidence-Based Public Policy Options to Reduce Future Prison Construction, Criminal Justice Costs, and Crime Rates. Olympia: Washington State Institute for Public Policy. Batchelder, J. S., and J. M. Pippert. "Hard Time or Idle Time: Factors Affecting Inmate Choices between Participation in Prison Work and Education Programs." The Prison Journal 82.2 (2002): 269-80. Baum, Christopher F., Mark E. Schaffer, and Steven Stillman. "Enhanced Routines for IV/GMM Estimation and Testing." The Stata Journal 7.4 (2007): 465-506. Brewster, D. R., and S. F. Sharp. "Educational Programs and Recidivism in Oklahoma: Another Look." The Prison Journal 82.3 (2002): 314-34. Gregg, Cherri. "Once Called Ex-Cons, Philadelphia ‘Returning Citizens’ Begin Six Weeks Of Reintegration - CBS Philly." CBS Philly., 13 Nov. 2013. Cho, R. M., and J. H. Tyler. "Does Prison-Based Adult Basic Education Improve Post Release Outcomes for Male Prisoners in Florida?" Crime & Delinquency 59.7 (2013): 975-1005. Codebook: Recidivism of Prisoners Released in 1994. ICPSR 3355. U.S. Department of Justice. Bureau of Justice Statistics. Inter-University Consortium for Political and Social Research, 25 Sept. 2002. Coley, Richard J., and Paul E. Barton. Locked Up and Locked Out: An Educational Perspective on the U.S. Prison Population. Rep. Policy Evaluation and Research Center: Educational Testing Service, Feb. 2006. Conan, Neal, Alan Johnson, Bill Seitz, and Michael Thompson. "Programs Keep Inmates From Returning To Prison." NPR. NPR, 10 Oct. 2012. Crayton, Anna, and Suzanne Rebecca Neusteter. "The Current State of Correctional Education." The Urban Institute (2008): 1-29. The Urban Institute: Reentry Roundtable on Education. The Urban Institute, May 2008. Perry 58 Davis, Lois M., Robert Bozick, Jennifer L. Steele, Jessica Saunders and Jeremy N. V. Miles. Evaluating the Effectiveness of Correctional Education: A Meta-Analysis of Programs That Provide Education to Incarcerated Adults. Santa Monica, CA: RAND Corporation, 2013. Elliott, Delbert, and Steve Aos. "A Life of Unintended Consequences." Prevention Action. N.p., 08 July 2010. Gaes, Gerald G. "The Impact of Prison Education Programs on Post Release Outcomes”. The Urban Institute (2008): 1-30. The Urban Institute: Reentry Roundtable on Education. The Urban Institute, Feb. 2008. Gaes, Gerald G., and William G. Saylor. "Training Inmates through Industrial Work Participation and Vocational and Apprenticeship Instruction." Corrections Management Quarterly (1997): 3243. Aspen Publishers. Gaes, Gerald G., and William G. Saylor. "Post Release Employment Project: Summary of Preliminary Findings." Federal Bureau of Prisons, Office of Research and Evaluation. Washington D. C. June 27, 1991. Hall, Renee, "Voices Behind Bars: Correctional Education from the Perspective of the Prisoner Student" (2006). University of New Orleans Theses and Dissertations. Paper 384 Harlow, Caroline Wolf, Ph.D. Education and Correctional Populations. Rep. no. NCJ 195670. Bureau of Justice Statistics: U.S. Department of Justice, Jan. 2003. Harer, Miles D. Prison Education Program Participation and Recidivism: A Test of the Normalization Hypothesis. Rep. Federal Bureau of Prisons Office of Research and Evaluation, May 1995. Ho, Taiping, Katie Knutson, Susan Lockwood, and John M. Nally. "The Effect of Correctional Education on Postrelease Employment and Recidivism: A 5-Year Follow-Up Study in the State of Indiana." Crime & Delinquency 58.3 (2012): 380-96. Hughes, Timothy, and Doris James Wilson. Reentry Trends in the U.S. Rep. Bureau of Justice Statistics: U.S. Department of Justice, 25 Oct. 2002. Ismailova, Zarona. Prison Education Program Participation and Recidivism. Thesis. Duquesne University: McAnulty College and Graduate School of Liberal Arts, Mar. 2007. Gumberg Library Digital Collections. Jancic, Mitchell. "Does Correctional Education Have an Effect on Recidivism?"Journal of Correctional Education 49.4 (1998): 152-61. Jenkins, H. David., Jennifer Pendry, and Stephen J. Steurer. "A Post Release Follow-up of Correctional Education Program Completers Released in 1990-1991." Maryland State Department of Education. 1993. Langan, Patrick A., and David J. Levin. Recidivism of Prisoners Released in 1994. Rep. no. NCJ 193427. Washington, DC: U.S. Department of Justice, Bureau of Justice Statistics, 2002. Lawerence, Sarah, Daniel P. Mears, Gelenn Dubin, and Jeremy Travis. The Practice and Promise of Prison Programming. Rep. The Urban Institute, May 2002. Perry 59 Lipton, Douglas S., Robert Martinson, and Judith Wilks. The Effectiveness of Correctional Treatment: A Survey of Treatment Evaluation Studies. New York: Praeger, 1975. Print. MacKenzie, Doris L. What Works in Corrections: Reducing the Criminal Activities of Offenders and Delinquents. New York: Cambridge University Press, 2006. Print. Maltz, Michael. D., and Richard McCleary. "The Mathematics of Behavioral Change: Recidivism and Construct Validity." Evaluation Review 1.3 (1977): 421-38. Maltz, Michael D., Richard McCleary, and Stephen P. Pollock. "Recidivism and Likelihood Functions: A Reply to Stollmack." Evaluation Quarterly 3.1 (1979): 124-31. Maltz, Michael D. "Recidivism." Academic Press, Inc. (1984) Manglesdorf, Paul, and Dave Owens. Can They Survive? Thesis. Haverford College. Marc-Taylor, Jon. "Pell Grants for Prisoners." PELL GRANTS FOR PRISONERS. Laird Carlson. Miller, Jerome G. "The Debate on Rehabilitating Criminals: Is It True That Nothing Works?" The Debate on Rehabilitating Criminals: Is It True That Nothing Works? Washington Post, Mar. 1989. Mitchell, Michael N. Data Management Using Stata: A Practical Handbook. College Station, TX: Stata, 2010. Print. New York State Department of Correctional Services (1989) Analysis of return rates of the inmate college program participants. Albany, Department of Correctional Services. Petersilia, Joan. When Prisoners Come Home: Parole and Prisoner Reentry. Oxford: Oxford UP, 2003. Print. Pew Center on the States. State of Recidivism: The Revolving Door of America’s Prisons. Rep. Washington, DC: The Pew Charitable Trusts, April 2011. Piehl, Anne Morrison. "Economic Conditions, Work, and Crime." Handbook of Crime and Punishment. Cary, NC: Oxford University Press, 2000. 302-19. EBrary. "Press Release: Education and Vocational Training in Prisons Reduces Recidivism, Improves Job Outlook." RAND Corporation., 22 Aug. 2013. Rhodes, W., and S. Matsuba. "Pretrial Release in Federal Courts: A Structural Model with Selectivity and Qualitative Dependent Variables." Evaluation Review 8.5 (1984): 692-704. Schmidt, Peter, and Ann D. Witte. Predicting Recidivism Using Survival Models. New York: SpringerVerlag, 1988. Print. Schaffer, Mark E. "Stata: Data Analysis and Statistical Software." IVREG2 with Two Endogenous Variables: CD/KP F-statvs.APF. Stata, 2 Feb. 2013. Sherman, Lawrence W., and Richard A. Berk. "The Specific Deterrent Effects of Arrest for Domestic Assault." American Sociological Review 49.2 (1984): 261 Sill, Mike. "Chapter 5: Cox Proportional Hazards Model." Perry 60 Stata. Stata Survival Analysis and Epidemiological Table: Reference Manual: Release 10. College Station, TX: StataCorp LP, 2007. Print. Stock, James H., and Mark W. Watson. "Chapter 12: Instrumental Variables Regression." Introduction to Econometrics. 3rd ed. Boston: Pearson/Addison Wesley, 2011. 419-68. Print. Stollmack, Stephen, and Carl M. Harris. "Failure-Rate Analysis Applied to Recidivism Data." Operations Research 22.6 (1974): 1192-205. Strand, Paul. "Punishing Taxpayers: US Prison System Run Amok?" Christian Broadcasting Network. 1 Mar. 2012. Thorpe, Thomas, Donald MacDonald, and Gerald Bala. "Follow-Up Study of Offenders Who Earn College Degrees While Incarcerated in New York State." Journal of Correctional Education 35.3 (1984): 86-89. Tyler, John H. and Jeffrey R. Kling (2007) Prison-based education and reentry into the mainstream labor market. In Shawn Bushway, Michael A. Stall, and David F. Weiman (Eds.), Barriers to Reentry: The Labor Market for Released prisoners in Post-Industrial America, New York: Russel Sage Foundation, 227-256. Ubah, Charles B. A., and Robert L. Robinson Jr. "A Grounded Look at the Debate Over Prison-Based Education: Optimistic Theory Versus Pessimistic Worldview." The Prison Journal 83.2 (2003): 115-29. Visher, Christy A., Ph.D., and Pamela K. Lattimore, Ph.D. "Major Study Examines Prisoners and Their Reentry Needs." National Institute of Justice Journal 258 (n.d.): 30-33. National Institute of Justice. National Institute of Justice, Oct. 2007. Warren, Jennifer, Susan Urahn, Richard Jerome, Jake Horowitz, and Joe Gavrilovich. One in 31: The Long Reach of American Corrections. Rep. Pew Center on the States: The Pew Charitable Trusts, Mar. 2009. Warren, Jennifer, Susan Grange, Lori Grange, and Tim Lynch. One in 100: Behind Bars in America 2008. Rep. Pew Center on the States: The Pew Charitable Trusts, Feb. 2008. "What Have We Learned From Evaluations of Reentry Programs?" BJA Center for Program Evaluation and Performance Measurement. Bureau of Justice Assistance, Office of Justice Programs, U.S. Department of Justice,. Wilson, David B., Catherine A. Gallagher, and Doris L. MacKenzie. "A Meta-Analysis of CorrectionsBased Education, Vocation, and Work Programs for Adult Offenders." Journal of Research in Crime and Delinquency 37.4 (2000): 347-68. SAGE Social Science Collections. Perry 61 Appendices Appendix Table 1: Re-Arrest & Re-Confinement, All States Variable educom_voccom edupart_voccom eduquest_voccom eduno_voccom educom_vocpart edupart_vocpart eduquest_vocpart eduno_vocpart educom_vocquest edupart_vocquest eduquest_vocquest eduno_vocquest educom_vocno edupart_vocno eduquest_vocno tmsrv nonwhite sex pprid FL IL NY NC Re-Arrest dF/dx -.052 * (029) -.076 *** (.030) -.011 (.048) -.074 (.059) -.068 ** (.029) -.080 (.025) -.022 *** (.064) -.041 (.045) .191 (.172) -.211 (.169) .079 ** (.038) -.086 *** (.031) -.044 * (.026) -.050 ** (.022) -.108 *** (.023) -.0002 (.0001) .120 *** (.011) .099 *** (.021) .120 *** (.011) .273 *** (.017) .172 *** (.014) .095 *** (.017) -.054 *** (.019) Re-Confined dF/dx -.079 *** (.027) -.141 *** (.026) .028 (.053) -.054 (.053) -.117 *** (.026) -.118 *** (.023) -.074 (.058) -.096 ** (.042) .078 (.224) -.251 * (.110) .009 (.041) -.105 *** (.026) -.131 *** (.022) -.106 *** (.020) -.174 *** (.017) .0003 (.0001) .097 *** (.011) .070 *** (.020) .120 *** (.011) .032 (.023) -.030 * (.017) .218 *** (.020) .019 (.019) Perry 62 SMPOFF2 SMPOFF3 SMPOFF4 SMPOFF5 SMPOFF6 SMPOFF7 SMPOFF8 SMPOFF9 SMPOFF10 SMPOFF11 SMPOFF12 SMPOFF13 age rage 2 age rage 3 age rage 4 age rage 5 age rage 6 age rage 7 .100 (.029) .184 (.026) .184 (.026) .232 (.023) .251 (.012) .219 (.028) .204 (.026) .172 (.028) .173 (.037) .201 (.028) .196 (.026) .216 (.025) -.169 (.085) -.240 (.085) -.286 (.083) -.347 (.080) -.414 (.073) -.485 (.065) *p<.10; **p<.05; ***p<.01 Values are reported from dprobit, thus they indicate changes in probabilities Numbers reported in parenthesis are robust standard errors n=10,157 *** *** *** *** *** *** *** *** *** *** *** *** ** *** *** *** *** *** .117 (.034) .209 (.037) .174 (.038) .242 (.036) .269 (.036) .209 (.046) .149 (.041) .148 (.036) .192 (.052) .149 (.036) .238 (.039) .215 (.040) -.033 (.071) -.089 (.068) -.111 (.068) -.149 (.065) -.200 (.058) -.263 (.051) *** *** *** *** ** *** *** *** *** *** *** *** ** *** *** Perry 63 Appendix Table 2: Re-Arrest by State Variable Florida Illinois New York North Carolina Texas Probability of re-arrest 72.8% 70.9% 61.2% 56.1% 50.5% educom_voccom -.049 (.045) edupart_voccom eduquest_voccom eduno_voccom educom_vocpart edupart_vocpart -.456 (.145) -.033 (.035) -.057 (.029) *** ** eduquest_vocpart eduno_vocpart -.327 (.159) ** -.149 (.063) -.222 (.089) -.132 (.067) -.304 (.119) -.296 (.069) -.205 (.058) -.115 (.085) -.086 (.079) ** ** ** ** eduquest_vocquest educom_vocno edupart_vocno eduquest_vocno tmsrv nonwhite -.0003 (.0002) .081 (.019) *** *** *** *** *** *** -.164 (.079) -.168 (.063) -.198 (.079) .0003 (.0004) .125 (.032) .034 (.058) .036 (.078) .026 (.078) *** edupart_vocquest -.296 (.054) -.141 (.045) -.194 (.052) -.259 (.035) .0001 (.0003) .236 (.023) .182 (.125) .233 (.125) *** educom_vocquest eduno_vocquest .010 (.048) -.106 (.044) .039 (.080) ** *** *** *** .122 (.173) -.025 (.175) .165 (.248) -.142 (.175) .118 (.047) .015 (.038) -.103 (.144) .047 (.184) .041 (.035) .0009 (.001) .113 (.027) ** ** ** .031 (.037) .018 (.028) *** -.0007 (.0003) .057 (.022) * *** Perry 64 sex pprid SMPOFF2 SMPOFF3 SMPOFF4 SMPOFF5 SMPOFF6 SMPOFF7 SMPOFF8 SMPOFF9 SMPOFF10 SMPOFF11 SMPOFF12 SMPOFF13 age rage 2 age rage 3 age rage 4 age rage 5 age rage 6 age rage 7 .098 (.042) .085 (.019) .141 (.045) .172 (.033) .185 (.031) .169 (.034) .229 (.023) .171 (.038) .210 (.027) .187 (.033) .158 (.049) .195 (.037) .209 (.026) .206 (.027) -.004 (.180) -.080 (.194) -.129 (.202) -.173 (.208) -.283 (.219) -.359 (.215) *** *** *** *** *** *** *** *** *** *** ** *** *** *** * .067 (.049) .142 (.022) .118 (.053) .119 (.051) .158 (.043) .163 (.043) .210 (.034) .191 (.039) .181 (.043) .109 (.053) .130 (.067) .148 (.050) .168 (.042) .118 (.053) -.975 (.003) -.956 (.004) -.948 (.004) -.916 (.006) -.862 (.007) -.879 (.007) *** ** ** *** *** *** *** *** * ** *** * *** *** *** *** *** *** .093 (.066) .142 (.031) .037 (.090) .289 (.057) .085 (.097) .294 (.057) .279 (.061) .134 (.116) .134 (.116) .226 (.074) -.063 (.165) .106 (.103) .134 (.099) .265 (.067) -.933 (.0070 -.956 (.005) -.974 (.003) -.915 (.008) -.838 (.011) -.826 (.010) *** *** *** *** * *** *** *** *** *** *** *** .111 (.042) .159 (.025) .140 (.080) .290 (.063) .252 (.070) .365 (.051) .327 (.058) .338 (.054) .279 (.066) .243 (.073) .304 (.070) .291 (.065) .291 (.063) .341 (.052) -.207 (.105) -.258 (.103) -.275 (.102) -.333 (.096) -.382 (.086) -.447 (.073) *** *** * *** *** *** *** *** *** *** *** *** *** *** ** ** *** *** *** *** .095 (.046) .043 (.022) .012 (.062) .058 (.071) .170 (.066) .184 (.064) .197 (.064) .211 (.079) .177 (.072) .110 (.065) .146 (.093) .175 (.072) .069 (.075) .148 (.072) -.325 (.184) -.351 (.179) -.385 (.174) -.419 (.156) -.458 (.121) -.496 (.110) ** ** ** *** *** ** ** ** ** ** * * ** *** *** Perry 65 Appendix Table 3: Re-Confinement by State Variable Probability of re-confinement educom_voccom Florida 34.55% .080 (.048) Illinois 39.1% New York 58.9% * edupart_voccom eduquest_voccom eduno_voccom educom_vocpart edupart_vocpart -.264 (.067) ** -.006 (.038) .003 (.034) eduquest_vocpart eduno_vocpart -.317 (.032) *** -.179 (.061) -.256 (.085) -.059 (.070) -.109 (.127) -.214 (.069) -.213 (.055) -.100 (.080) -.147 (.075) *** .218 (.133) .332 (.126) *** edupart_vocquest eduquest_vocquest educom_vocno edupart_vocno eduquest_vocno tmsrv nonwhite -.0007 (.0003) .112 (.021) ** ** -.324 (.217) -.327 (.020) -.304 (.022) -.386 (.016) .0003 (.0004) .109 (.024) *** *** *** *** *** -.163 (.072) -.133 (.060) -.047 (.080) .004 (.0005) .067 (.032) ** *** ** ** ** *** ** Texas 37.2% -.074 (.043) -.051 (.042) *** educom_vocquest eduno_vocquest North Carolina 47.9% -.046 (.179) .073 (.178) .023 (.262) -.180 (.163) .140 (.049) .059 (.038) .096 (.141) .012 (.198) .041 (.036) -.001 (.001) .159 (.026) -.023 (.077) .017 (.055) .022 (.041) *** .062 (.079) *** .011 (.035) .013 (.027) *** -.0004 (.0003) .028 (.021) Perry 66 sex pprid SMPOFF2 SMPOFF3 SMPOFF4 SMPOFF5 SMPOFF6 SMPOFF7 SMPOFF8 SMPOFF9 SMPOFF10 SMPOFF11 SMPOFF12 SMPOFF13 age rage 2 age rage 3 age rage 4 age rage 5 age rage 6 age rage 7 .075 (.036) .134 (.022) .096 (.066) .171 (.080) .150 (.080) .175 (.078) .244 (.077) .171 (.098) .100 (.085) .122 (.074) .206 (.113) .157 (.123) .252 (.078) .280 (.082) -.053 (.152) -.114 (.143) -.126 (.141) -.174 (.129) -.175 (.122) -.273 (.091) ** ** ** ** ** *** * *** *** ** -.048 (.055) .064 (.028) .162 (.080) .224 (.092) .203 (.095) .277 (.090) .318 (.088) .241 (.108) .115 (.107) .149 (.088) .215 (.127) .061 (.113) .260 (.095) .071 (.100) .989 (.002) .955 (.004) .944 (.005) .891 (.007) .804 (.010) .818 (.009) ** ** ** ** *** *** ** * * *** *** *** *** *** *** *** -.010 (.063) .113 (.037) .251 (.098) .382 (.043) .296 (.072) .386 (.046) .404 (.040) .316 (.068) .332 (.059) .353 (.065) .179 (.132) .270 (.079) .308 (.066) .390 (.031) -.932 (.007) -.955 (.005) -.955 (.005) -.907 (.008) -.824 (.010) -.810 (.012) *** ** *** *** *** *** *** *** *** ** *** *** *** *** *** *** *** *** .151 (.039) .162 (.025) .014 (.090) .235 (.087) .201 (.090) .350 (.073) .314 (.077) .301 (.090) .200 (.092) .176 (.087) .282 (.010) .212 (.097) .212 (.093) .333 (.077) -.067 (.100) -.117 (.099) -.113 (.099) -.138 (.098) -.233 (.090) -.319 (.077) *** *** *** ** *** *** *** ** ** ** ** ** *** ** *** .068 (.042) .077 (.021) .108 (.061) .081 (.062) .115 (.073) .053 (.070) .107 (.071) .066 (.091) .055 (.077) .030 (.065) .072 (.099) .050 (.079) .122 (.077) .017 (.076) -.012 (.161) -.073 (.154) -.080 (.154) -.107 (.149) -.177 (.131) -.183 (.131) *** * Perry 67 Appendix Table 4: Selection Bias, All States Re-Arrest Variable eduonly voconly edu_voc tmsrv nonwhite sex pprid FL IL NY NC SMPOFF2 SMPOFF3 SMPOFF4 SMPOFF5 SMPOFF6 SMPOFF7 SMPOFF8 SMPOFF9 Instrumental Variable Estimation -3.867 (2.31) -1.551 (4.931) .830 (2.322) -.001 (.001) .092 (.034) -.099 (.135) .113 (.048) -1.620 (2.036) .212 (.662) -.619 (.404) -.392 (.800) .020 (.166) .047 (.160) .166 (.118) .239 (.147) .062 (.216) .003 (.214) .072 (.218) .093 (.117) * *** ** Re-Confinement OLS Regression Coef. -.045 (.013) -.053 (.021) -.015 (.015) -.0001 (.0002) .110 (.009) .094 (.018) .119 (.009) .235 (.017) .170 (.014) .062 (.014) -.045 (.015) .091 (.027) .186 (.030) .194 (.031) .264 (.030) .263 (.030) .267 (.038) .211 (.033) .182 (.043) *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** Instrumental Variable Estimation -5.798 (3.37) -2.313 (7.206) 1.288 (3.391) -.001 (.001) .070 (.050) -.220 (.195) .111 (.071) -2.839 (2.97) .064 (.965) -.835 (.588) -.490 (1.163) -.012 (.242) -.047 (.235) .113 (.175) .189 (.217) -.060 (.319) -.201 (.315) -.078 (.321) -.005 (.173) OLS Regression Coef. * -.118 (.013) -.074 (.020) -.061 (.015) .0005 (.0001) .096 (.009) .069 (.018) .119 (.010) -.010 (.018) -.020 (.015) .183 (.014) . 013 (.015) .097 (.023) .162 (.028) .153 (.029) .227 (.028) .238 (.028) .190 (.038) .130 (.032) .128 (.026) *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** Perry 68 SMPOFF10 SMPOFF11 SMPOFF12 SMPOFF13 group 2 group 3 group 4 group 5 group 6 group 7 _cons .229 (.214) .083 (.166) -.005 (.194) .238 (.156) .165 (.270) .108 (.257) .100 (.266) .080 (.275) -.086 (.246) -.062 (.307) 1.424 (.778) * .222 (.036) .222 (.036) .211 (.032) .228 (.033) -.135 (.049) -.204 (.050) -.241 (.050) -.288 (.050) -.361 (.051) -.447 (.051) .432 (.059) *p<.10; **p<.05; ***p<.01 Numbers reported in parenthesis are robust standard errors *** *** *** *** *** *** *** *** *** *** *** .240 (.318) -.078 (.245) -.100 (.286) .201 (.230) .379 (.374) .335 (.374) .361 (.387) .371 (.401) .174 (.358) .266 (.447) 1.733 (1.130) .167 (.044) .125 (.034) .220 (.031) .182 (.031) -.054 (.068) -.116 (.068) -.132 (.068) -.167 .069 -.224 (.069) -.299 (.069) .263 (.074) *** *** *** *** * ** *** *** *** *** Perry 69 Appendix Table 5: Cox Proportional Model Hazard, All States Variable eduonly voconly edu_voc tmsrv nonwhite sex pprid FL IL NY NC SMPOFF2 SMPOFF3 SMPOFF4 SMPOFF5 SMPOFF6 SMPOFF7 SMPOFF8 SMPOFF9 SMPOFF10 SMPOFF11 Re-Arrest Hazard Ratio .790 (.030) .781 (.049) .858 (.042) .999 (.0005) 1.344 (.036) 1.335 (.071) 1.396 (.038) 2.335 (.121) 1.679 (.069) 1.391 (.069) .876 (.039) 1.416 (.138) 1.950 (.204) 1.857 (.197) 2.193 (.227) 2.452 (.255) 2.205 (.273) 2.039 (2.039) 1.746 (.177) 1.746 (.237) 1.912 *** *** *** ** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** ReConfinement Hazard Ratio .656 (.039) .726 (.065) .745 (.051) .999 (.0008) 1.535 (.068) 1.134 (.089) 1.607 (.068) 1.129 (.094) 1.156 (.077) 2.802 (.195) 1.389 (.089) 1.529 (.304) 2.780 (.570) 2.389 (.495) 3.473 (.705) 3.896 (.791) 2.991 (.684) 2.440 (.520) 2.500 (.502) 2.076 (.517) 2.463 *** *** *** *** *** ** *** *** ** *** *** *** *** *** *** *** *** *** Perry 70 SMPOFF12 SMPOFF13 age group 2 age group 3 age group 4 age group 5 age group 6 age group 7 n (.237) 1.989 (.218) 2.042 (.226) .594 (.093) .496 (.078) .437 (.069) .386 (.061) .313 (.051) .233 (.038) 10,157 *** *** *** *** *** *** *** *** (.573) 2.903 (.641) 3.012 (.641) .876 (.210) .775 (.187) .637 (.154) .627 (.153) .499 (.125) .325 (.084) 10,157 *** *** * * *** *** Perry 71 Appendix Table 6: Cox Proportional Hazard Model, Illinois & New York Illinois Variable eduonly voconly edu_voc tmsrv nonwhite sex pprid SMPOFF2 SMPOFF3 SMPOFF4 SMPOFF5 SMPOFF6 SMPOFF7 SMPOFF8 SMPOFF9 SMPOFF10 SMPOFF11 SMPOFF12 SMPOFF13 age group 2 Re-Arrest Hazard Ratio .499 (.034) .383 (.056) -.999 (.001) 1.946 (.120) 1.218 (.153) 1.542 (.096) 1.392 (.290) 1.762 (.391) 1.924 (.435) 2.138 (.475) 2.529 (.565) 2.558 (.669) 2.094 (.523) 1.554 (.337) 1.496 (.423) 1.719 (.492) 1.939 (.459) 1.622 (.386) .133 *** *** *** *** *** *** *** *** *** *** ** * *** ** ** New York ReConfinement Hazard Ratio .081 (.021) .165 (.068) -- *** *** .999 (.002) 2.389 *** (.293) 1.121 (.254) 1.373 *** (.141) 1.166 (.493) 1832 (.796) 1.384 (.622) 1.943 (.849) 2.534 (1.101) 1.489 (.764) 1.645 (.779) 1.889 (.809) 1.717 (.878) .588 (.424) 1.715 (.795) .968 (.474) see note Re-Arrest Hazard Ratio .719 (.082) .714 (.112) .673 (.068) 1.000 (.001) 1.35 (.111) 1.299 (.215) 1.528 (.122) .941 (.284) 1.185 (.706) 1.165 (.385) 2.616 (.840) 2.485 (.812) 1.508 (.627) 1.401 (.476) 1.685 (.541) .781 (.363) 1.261 (477) 1.466 (.525) 2.089 (.726) .256 ReConfinement Hazard Ratio *** ** *** *** *** ** *** *** * ** ** .855 (.118) .682 (.132) .700 (.087) 1.001 (.002) 1.358 (.139) .961 (.177) 1.676 (.162) 1.693 (.872) 4.810 (2.549) 2.013 (1.095) 5.338 (2.829) 5.937 (3.169) 3.491 (2.119) 3.105 (1.694) 3.398 (1.804) .846 (.654) 2.076 (1.25) 3.016 (1.701) 4.915 (2.708) .119 ** *** *** *** *** *** *** ** ** ** ** *** *** Perry 72 age group 3 age group 4 age group 5 age group 6 age group 7 n (.134) .098 (.099) .089 (.090) .073 (.074) .065 (.0669) .045 (.045) 2,124 ** ** *** *** *** .901 (.105) .744 (.099) .732 (.119) .616 (.123) .459 (.112) 2,127 ** ** *** (.152) .204 (.121) .161 (.096) .154 (.093) .128 (.078) .088 (.055) 1,193 *** *** *** *** *** (.072) .110 (.067) .075 (.045) .085 (.052) .064 (.040) .060 (.038) *** *** *** *** *** 1,193 *p<.10; **p<.05; ***p<.01 Numbers reported in parenthesis are robust standard errors For IL re-confinement age group 1 and 2 were combined to make the base state as there were too few observations in group 1, I believe this is what causes the increase in 3 observations. Perry 73 Appendix Table 7: Description of Variables The following descriptions are adapted from the ICPSR3355 Codebook. Variable Long Name of Variable Description Value & Meaning in Analysis Dependent Variables REARRD Number of Rearrests Dummy variable indicating whether the individual was rearrested at any time within 3 years following release 1= re-arrested 0= not re-arrested RCNFITV Re-confined – Including Technical Violations Dummy variable indicating whether the individual was resentenced to prison or jail for any reason (new sentence or technical violation) within the 3 years following release 1= re-confined 0= not re-confined . = not applicable Indicates whether or not the individual took part in an education program while serving the sentence that ended in 1994 Indicates whether or not the individual took part in a vocation program while serving the sentence that ended in 1994 1= participated & completed 2= participated, not completed 3= participated, unknown completion 4= did not participate . = unknown Is the total maximum sentence length (in moths) for all offenses. If sentences are concurrent this records the longest sentence. This is used to create SNTLN2, the square of sentence length. Indicates whether or not the .=not applicable Variables of Interest 48 EDUCAT Education Courses VOCAT Vocational Courses 1= participated & completed 2= participated, not completed 3= participated, unknown completion 4= did not participate . = unknown Instrumental Variables SNTLN Sentence Length NFRCTNS Number of 1= has a record The above variables were used to generate dummies for each possible outcome. The dummies were then interacted to create 16 new variables. The dummies are listed in Table 3 and are mutually exclusive. The base state in the interactions was eduno_vocno, meaning no participation in education or vocational programs. 48 Perry 74 Infractions individual was formally disciplines for a prison-rule infraction. 0= does not have a record .=unknown Time served (in months) before release Sex of released individual integers >12 Control Variables TMSRV SEX Time Served for 1994 Imprisonment Sex PPRID Prior Sentences to Prison SMPOFF13 Sample Offense 13 Categories RLAGE Age Range STATE State Releasing Prisoner in 1994 RACE Race 1= male 0= female . = unknown Dummy variable indicating 1= prior sentences to prison whether, prior to the sentence 0= no prior sentence to prison . = not applicable that got him/her into the study, he/she had previously received a prison sentence for another offense The offense (13 levels) for 1= homicide which the sampled individual 2= rape/sexual assault was in prison 3= robbery 4= agg assult 5= burglary 6= larceny –mvt 7= ffe (fraud, forgery, or embezzlement) 8= drug possession 9= drug trafficking 10= weapons 11= dui 12= other public order 13= other Age of prisoner in months at 1= 14 to 17 years old 2= 18 to 24 years old the date of release in 1994 3= 25 to 29 years old broken into seven distinct 4= 30 to 34 years old groups. 5= 35 to 39 years old 6= 40 to 44 years old 7= 45 years and older 9= unknown State from which the 1= alabama 2= alaska individual was released in … 1994. There are 15 states 54= puerto rico included in the study but variables for all 55 states and 55= virgin islands territories Race of released individual. 1= White This variable is used to create 2= Black 3= American Indian/Aleutian a dummy for NONWHITE where people who identify as 4= Asian/Pacific Islander 5= Other Black, American . = unknown Perry 75 Indian/Aleutian, Asian/Pacific Islander, or Other are coded as “nonwhite” Date Variables A0##MO Month of Arrest Month arrested in cycle ## A0##DA A0##YR J0##MO Day of Arrest Year of Arrest Month of Adjudication Day of arrest in cycle ## Year of arrest in cycle ## Month of Adjudication in cycle ## J0##DA Day of Adjudication J0##YR Year of Adjudication J0##PJP1 Type of Sentence Day of Adjudication in cycle ## Year of Adjudication in cycle ## Prison, jail or probation for first adjudicated offense cycle 1= January 2= February … 12=December integers 1-31 integers 1918 - 1998 1= January 2= February … 12=December integers 1-31 integers 1936 - 1999 1= prison 2= jail 3= probation-fine-other 4= confined, type unknown 5= convicted, sentence blank 6= not convicted 8= not applicable 9= unknown Perry 76 Appendix Table 8: Descriptive Statistics The below provides key descriptive statistics for the five-state sample. Number of Observations Mean Standard Deviation Min Max Sex (where male=1) 11,824 .9361 .3445 0 1 Non-white 11,352 .5347 .4988 0 1 Prior Arrest History 11,824 .4075 .4914 0 1 Term Served (in months) 11,815 27.946 33.6417 .03 436.43 Sentence Length (in months) 11,824 85.954 85.3665 0 1,800 Number of Infractions 9,153 .4156 .4929 0 1