Abstract
Increasing evidence suggests that incarceration can improve the social reintegration of inmates in some circumstances. Yet, the mechanisms through which incarceration may favor rehabilitation remain unknown. This study exploits variations in program availability to estimate their effects. We find that programs decrease reincarceration. However, this is mitigated by an increase in future community sentences, usually associated with milder offenses. Programs addressing violence issues, education, and employment exhibit strong effects. Those focusing on addiction and other program types are not found to affect recidivism. These results suggest that specific programs can explain the beneficial effects of incarceration found in the literature.
I. Introduction
Incarceration may impact criminal behavior in several ways. It may prove criminogenic due to greater access to criminal networks and expertise or by lessening an inmate’s capacity to reintegrate the labor market due to social stigma or human capital depletion. Yet, recent evidence suggests that lengthy incarceration in prisons that are rehabilitation-oriented may have beneficial effects on recidivism (Landersø 2015; Bhuller et al. 2020; Lotti 2022; Hjalmarsson and Lindquist 2022; Loeffler and Nagin 2022; Mastrobuoni and Terlizzese 2022). It is often conjectured that programs that focus on education, job skills, or providing psychological assistance might be responsible for such effects. In addition to being potentially beneficial to offenders, such programs could turn out to be cost-efficient from society’s perspective if they reduce future costly incarceration.
Unfortunately, empirical evidence on relative program efficiency is fairly scant. In their meta-analysis on rehabilitation programs for adult offenders, Wilson, Gallagher, and MacKenzie (2000) and Davis et al. (2013) conclude that, while evidence suggests that program participation decreases recidivism and increases employment, the research designs they survey are deemed too poor to provide reliable estimates. The major caveat in most studies is the failure to thoroughly consider selection issues associated with voluntary program participation.
This paper addresses these concerns. We exploit the relative time-varying availability of rehabilitation programs in seven male provincial prisons in Quebec (Canada) to isolate their causal effects on recidivism. Each prison offers an array of programs, ranging from educational and vocational training to addiction and substance use treatments. We find that some programs substantially decrease future incarceration, but that this decrease is partly offset by an increase in future community sentences that are nevertheless associated with less serious offenses.
Our study is based on inmates serving sentences of less than two years.1 This population is particularly relevant for the analysis of rehabilitation programs. Indeed, since prison sentences are relatively short, participants may readily put their newly acquired skills to use. In addition, the offenses for which these individuals are incarcerated are less serious than those of inmates serving lengthier sentences in federal penitentiaries. They are thus perhaps more likely to successfully reintegrate into society. In this respect, Quebec prisons are mandated by law to offer an array of programs to facilitate social rehabilitation. Prison managers collaborate with government bodies, community organizations, and private agencies to develop various programs, such as addiction or violence-related interventions, education and job skill courses, or self-development programs.
In our data, the duration of incarceration is negatively correlated with recidivism, suggesting possible beneficial effects of incarceration. The data also show a positive correlation between the duration of incarceration and the number of programs chosen by individual inmates. While these correlations could arise from the causal effect of programs on recidivism, isolating such an effect is challenging due to voluntary participation. Yet, participation is conditional on program availability, which fluctuates exogenously. Indeed, many programs are administered by local school boards and are seldom available during the summer recess and the December break. In addition, fluctuating prison-specific personnel and funding constraints limit the availability and nature of rehabilitation programs. To exploit these features, we construct an instrumental variable that measures program availability. Our instrumental variable design captures variations in the number of programs a prisoner may possibly enroll in at a specific prison and given the sentence length. We show that the instrument is strongly related to participation but not to other individual characteristics, which suggests it is both strong and as good as random.
Our reduced-form results suggest that a one standard deviation increase in the number of available programs decreases the likelihood of reincarceration in a provincial prison by four percentage points over a five-year window. However, it also slightly increases the probability of a community sentence by two percentage points. Community sentences are used for offenders who have not committed a serious or violent crime or for a crime that does not carry a minimum sentence. As such, our results are consistent with participants committing less serious offenses. We provide further evidence that programs decrease the seriousness of future offenses by showing that greater availability decreases the likelihood of pretrial detention for the next offense, often associated with dangerousness. We also find precise but not statistically significant program effects on reoffenses that lead to federal sentences (that is, sentences of more than two years associated with most serious offenses). Instrumenting participation with program availability, we find that each additional program take-up decreases the probability of reincarceration by as much as seven percentage points over five years. Our results are robust to the inclusion of additional control variables, to alternative recidivism definitions, to variations in the instrument, and to other specifications.
Our heterogeneity analysis shows that programs that focus on violent behavior, education deficiencies, or employment unreadiness all significantly decrease the likelihood of committing a new crime leading to reincarceration. We also find some evidence, though less robust, that self-development programs may decrease incarceration. On the other hand, addiction programs and those related to arts, spirituality, and sports yield precise null effects. While our results suggest that individuals serving longer sentences are most impacted by the number of programs made available, inmates who serve shorter sentences experience modest yet significant benefits.
The shift from incarceration to community-based sentences suggests that programs may help alleviate prison overcrowding and high incarceration costs, as community sentences are less costly. In addition, our results suggest that it may be beneficial to focus resources on programs that exhibit stronger effects. To explore these questions, we conduct a cost–benefit analysis that considers the monetary benefits of reduced incarceration and the costs of providing programs. We find an average benefit-to-cost ratio of three, with programs related to violence issues yielding the highest ratio.
Our work mainly contributes to the growing literature that evaluates different types of programs aimed at offenders at varying stages of their criminal trajectory. An important strand of the literature focuses on programs aimed at young offenders. Heller et al. (2017) provide evidence that behavioral interventions may reduce recidivism among this population. They exploit three randomized controlled trials (RCTs) located in Chicago that focused on at-risk youths and juvenile delinquents and find that participation in tailored programs significantly reduced rearrests and readmissions. Seroczynski et al. (2016) also study similar programs through a RCT design and find it decreases recidivism significantly, whereas Armstrong (2003) find no such effect.
Studies on programs targeted at adult offenders can be separated into two groups: prison-based and external. Doleac et al. (2020) review three randomized controlled trials of external reentry programs and find, at best, mixed evidence of their effectiveness.2 Blattman, Jamison, and Sheridan (2017) study the effect of behavioral therapy and show that it can reduce violent crimes. Many studies have sought to estimate the effects of prison-based programs, but as mentioned earlier, this literature faces challenges due to selection issues.3 However, recent studies have found encouraging results. Balafoutas et al. (2020) randomly asked inmates to reflect on their incarceration and show that this simple intervention increased the inmates’ social aptitudes. Studying the same setting as the current paper, Arbour (2022) focuses on a behavioral intervention targeted specifically at high-risk offenders and finds large reductions in reoffenses in the short run. Kuziemko (2013) exploits a 1998 policy reform that canceled parole eligibility for convicts in the state of Georgia. She presents evidence that inmates respond by decreasing their rehabilitation effort, including reduced program participation while incarcerated, thereby increasing the likelihood of reoffending. Macdonald (2023) exploits a reform in Arizona that eliminated judges’ discretion over early release. He conjectures that the ensuing increase in recidivism is most likely due to reduced rehabilitation efforts while incarcerated. All in all, scarce empirical evidence suggests that inmates may benefit from rehabilitation programs. Our paper contributes to this literature by focusing explicitly on prison-based interventions and by highlighting the types of programs that are most beneficial.
Our paper also contributes to the broad literature investigating the link between incarceration and post-release outcomes. Incarceration may affect post-release outcomes through various channels, which may explain the mixed results found in the literature.4 Interestingly, recent papers have shown that incarceration in rehabilitation-friendly prisons can be beneficial (Landersø 2015; Bhuller et al. 2020), which could be in part driven by the availability of rehabilitation programs in these types of prison. A related literature compares punitive prisons to rehabilitation-oriented ones, with a particular focus on differences in prison conditions, and provides evidence supporting that rehabilitation-oriented prisons can improve outcomes.5 Our work complements this literature by estimating the effect of several prison-based interventions and by providing evidence that they are indeed a driving force underlying the beneficial effects that prisons may have.
In the following, Section II discusses the setting, the data, and the various definitions of recidivism used in the analysis. In Section III, we present summary statistics and elaborate on the associations between incarceration time, program participation and availability, and recidivism. Section IV presents the main results along with robustness checks and heterogeneity analyses. Section V explores the impact of programs on crime severity and conducts a cost–benefit analysis. We conclude in Section VI.
II. Institutional Details and Data
A. Institutional Details
In Canada, a person charged with a criminal offense may be released while awaiting trial or, depending on the seriousness of the charge or whether it involves violence, be incarcerated until trial (known as “pretrial” detention). If found guilty, the judge determines the type (custodial versus community) and the duration of the sentence. If the sentence is less than two years, it will be served either in a provincial correctional centre (“prison”) or in the community. Sentences of at least two years lead to incarceration in a federal penitentiary. The prison where offenders serve their sentence is not random. The judge may select a specific one based on the nature of the crime, the criminogenic needs of the offender, and the availability of appropriate services and programs.
The most common type of community sentence by far is probation. In Canada, probation corresponds to the release of a convict into the community under a set of stringent conditions and compulsory regular reporting to a probation officer. Examples of conditions include performing community services, not consuming drugs or alcohol, remaining within the court’s jurisdiction, not communicating with a victim of the offense, or not possessing a weapon. Note that probation differs from parole, which is an early release from an incarceration sentence.6 The other form of community sentence is conditional sentencing (also called house arrests), which can be seen as a stricter form of probation. For these sentences, offenders are confined to their residence and receive stricter surveillance. Our data do not allow us to distinguish between probation and conditional sentencing.
Community sentences do not apply to crimes that carry a mandatory minimum sentence. They are typically less costly than incarceration and offer offenders the opportunity to work or attend school.7 In contrast, incarceration often results in the loss of employment and other opportunities that may undermine reintegration upon release. Though provincial incarceration is deemed harsher than community sentences, federal sentences are considered yet harsher.
Our analysis focuses on convicts sentenced to less than two years, with an average sentence of four months. In practice, most will serve only two-thirds of their original sentence.8 We refer to the date corresponding to two-thirds of the sentence as the planned date of release. Prisoners may be released sooner if granted parole, in which case they may be offered other forms of rehabilitation assistance.9
To facilitate their social reintegration, all provincial prisons offer inmates a diversified set of programs, which we discuss below. In addition to delivering programs, rehabilitation-oriented prisons offer healthcare services, counseling, and family support. Furthermore, at the onset of their incarceration, a qualified professional assesses inmates’ physical and mental conditions. Those serving a sentence less than six months will be briefly assessed by a correctional officer. Those serving a longer sentence will be thoroughly assessed and evaluated using the comprehensive standardized risk assessment tool known as the “Level of Service/Case Management Inventory” (LS/CMI).10 Contrary to other jurisdictions, the evaluation in Quebec occurs after the trial and is thus not used for sentencing. The evaluation is meant to provide risk scores and to identify rehabilitative needs.
B. Program Delivery and Availability
Prisons managers collaborate with government bodies, such as the Ministry of Education and the Ministry of Labor, to offer programs tailored to the local labor market. Community organizations and private agencies can also be called upon to develop specific programs, such as addiction interventions and anger management workshops. Participation in such programs is voluntary and may not be mandated by the sentencing judge. However, the judge may request incarceration to the prison deemed best suited to the offender’s criminogenic needs. Once incarcerated, and following their initial assessment discussed above, inmates may be encouraged to enroll in specific programs. Ultimately, prisoners decide whether to participate or not and may enroll in as many programs as available. Program duration varies significantly across and within program types. The majority of the programs are structured to incorporate one to two-hour sessions on a weekly or biweekly basis, typically spanning five to six weeks. Job skills and education programs are often more intensive, and therefore more costly. For instance, to obtain a permit to work on a construction site, an inmate must undergo a minimum of 30 hours of instruction over the course of a week, while other job programs, like mechanics or carpentry courses, are spread over several weeks. Participants enrolled in high school studies dedicate an average of 20–25 hours per week over several weeks to attend classes. Finally, the availability of programs—and their duration—are also prison-specific, as they depend on several factors, such as financial and staffing resources.
Prison managers have a legal obligation to deliver such programs.11 However, the law does not provide any guidelines as to the number of programs that must be offered nor as to their content and duration. In practice, these are determined at the prison level and can vary over time for various reasons.
To understand the underlying causes of fluctuating rehabilitation program availability in prisons, we conducted ten semi-guided interviews with prison counselors overseeing these programs. We asked the two following questions via email or over the phone: (i) “Based on your experience, what affects the availability of programs on a daily basis?” and (ii) “To your knowledge, are there fewer programs during specific periods of the year?” According to all counselors, the most significant factor affecting program availability was the operational downtime during summer months, a period marked by school board recesses and staff vacations, or other holidays. All counselors agreed that such downtime led to notable reductions in program availability. One counselor reported numerous instances recorded in prisoners’ files as “[The prisoner] wanted to participate, but no programs were available” because of such downtime. The variability in prison staff and budgetary limitations were also mentioned by most counselors as causes of limited availability. Additionally, some counselors mentioned constraints imposed by the physical infrastructure of prisons, with renovations often making suitable facilities unavailable. These insights are apparent in the data. In Section III, we show that program availability fluctuates within prisons and decreases during the summer and the holiday season.
C. Data
The data used in this paper are drawn from the Administrative Correctional Files (DACOR) of the Province of Quebec Ministry of Public Security.12 DACOR contains detailed information on all daily convicts incarcerated in all provincial prisons. In addition to judicial information, DACOR documents the type of crime and individual characteristics, such as age, number of dependents, and Indigenous status. At first offense, convicts are assigned a unique identifier that is used to generate a historical record of their interactions with the judicial system.
For each sentence, DACOR records the start and end dates of the sentence and a code for the most serious charge associated with a given crime.13 Pretrial detention is also reported and matched to individual records. Importantly, even though our paper focuses on releases from provincial prisons, all future reoffenses are recorded in the correctional files irrespective of whether they are served in the community, in a provincial prison or in a federal penitentiary.
DACOR also includes all individual LS/CMI assessments for offenders serving for at least six months. The data contain the evaluation of all 43 items of the assessment tool, which are aggregated into eight risk scores. The tool documents risk factors related to: (i) criminal history, (ii) consumption of alcohol and drugs, (iii) educational attainment, (iv) lack of adequate leisure, (v) family issues, (vi) social relationship issues, (vii) anti-social patterns, and (viii) pro-criminal attitudes. These risk scores are highly informative proxies for the propensity to recidivate.14
As mentioned above, rehabilitation programs are prison-specific and managed locally. We obtained detailed program enrollment files from seven prisons over different time frames. These were merged to the DACOR files using individual identifiers.15 The files comprise as many as 145 different programs. These were aggregated into six categories following advice from several program managers: self-development, violence, addiction, education, job skills, and “others,” which includes leisure, spirituality, and sports activities. The stratification was based on the criminogenic risks and needs targeted by each. Thus, we observe the list of all the programs a prisoner enrolled in, but not their duration.
D. Measuring Recidivism
The definition of recidivism differs in the literature and across jurisdictions.16 In this paper, we consider various definitions of recidivism, all based on sentencing for a new crime occurring after the planned date of release. These are:
• A new provincial incarceration;
• A new federal incarceration;
• A community sentence;
• Any of the above.
As we show in our robustness analysis, our results are virtually unchanged if the follow-up period starts at the actual rather than the planned release date.17 Our results are also robust to using the beginning of the sentence as a starting date for computing recidivism. We vary the follow-up period from one to five years, discarding censored observations in each case.
III. Descriptive Statistics
Table 1 reports the average and standard deviation of all explanatory variables for program participants and nonparticipants separately. The types of crime are aggregated into four categories: assault, burglary and theft, drug-related, and “others.”18 Reported crimes are those deemed most serious for a given sentence and are thus mutually exclusive. Participants are more likely to be charged for assault or burglary and theft, but less likely for drug-related or other. Indigenous convicts represent 6.6 percent of the sample but as much as 9.4 percent of all participants. This is perhaps due to the fact some prisons have designed specific programs for this population. Interestingly, participation does not vary much across age groups.
Summary Statistics for Programs Participants and Nonparticipants
The average number of incarceration days is greater for participants both during pretrial and posttrial detention. The relationship between participation and posttrial incarceration is intuitive: prisoners with longer sentences have more opportunities to enroll (see below). The relationship between pretrial detention and participation in less intuitive, given that convicts do not typically enroll during pretrial detention.19 However, according to Bourgon and Grech (2011), this correlation can be explained by the fact that more serious crimes tend to entail lengthy and complex legal procedures and, as such, may postpone the trial date.
The “LS/CMI: Evaluation” variable indicates whether an inmate was evaluated with the LS/CMI assessment tool. Participants are more likely to be evaluated than nonparticipants (69.5 percent versus 37.7 percent). This is not surprising since participants have longer posttrial sentences and given that the evaluation is conducted only if the sentence is greater than six months. The next set of rows report the average of the total LS/CMI score, as well as those of its eight components.20 Most LS/CMI scores are smaller among participants. While some differences are statistically significant, they are nevertheless small relative to their standard deviations.
The intensity of participation varies considerably among participants. Figure 1 depicts the distribution of program enrollment per participant. Approximately 44 percent enroll in a single program, while around 19 percent (11 percent) enroll in two (three) programs. It is not uncommon to observe yet more program enrollment per prison spell. Table 2 reports participation statistics broken down by program category. Column a focuses on the unconditional participation rates per program while Column b reports the conditional distribution of program enrollment. Roughly 19 percent of convicts participate in one program, the most common being education, self-development, and addiction programs. Over the course of their stay in prison, participants enroll in three different programs on average.
Frequency of Program Enrollment—Participants Only
Notes: Around 44 percent of participants participated in only one program. The figure is limited to 20 program participation to preserve scaling. Less than 0.05 percent of participants participated in more than 20 programs.
Participation Statistics by Program Category
Figure 2 plots the rates of recidivism by participation status defined over one- to five-year intervals after the planned date of release. The figure distinguishes between reoffenses that lead to a community, a prison (provincial), or a penitentiary (federal) sentence. Overall, program participants have lower rates of recidivism save for federal sentences, which rarely occur in our data (bottom-left panel). Interestingly, the differences between the two groups are stable across intervals with respect to provincial incarceration but taper off with respect to community sentences. The bottom-right panel shows the rates of recidivism that lead to any sentence.21
Rates of Recidivism
Notes: This figure shows the rates of recidivism within one to five years after the planned release from a provincial prison. The four panels vary the type of sentence considered as recidivism. The p-values test the differences in proportions between recidivism rates for participants and non-participants.
Lengthy sentences will naturally tend to provide greater opportunities to enroll in programs, which could generate a negative relationship between sentence duration and recidivism if programs work. In our data, we indeed observe that longer sentences correlate with higher participation and lower recidivism.22 Several other factors may affect program participation. Financial resources and labor constraints vary across prisons and time, so not all convicts may face the same opportunities. To see this, we computed the number of programs a prisoner could potentially enroll in conditional on the duration of incarceration during their sentence. Figure 3 draws a scatter plot of the number of such programs with respect to duration separately for each prison in our data. As shown, program availability varies significantly across prisons. For example, a prisoner serving a 100-day sentence could in some cases have access to fewer than five different programs. In other cases, as many as 10–20 programs could have been offered for a similar sentence duration.
Incarceration Length and Available Programs, by Prison
Notes: The figure presents a scatter plot of the number of programs available that a prisoner may enroll in over the course of their incarceration for each prison in our data. The lines are smoothed local polynomial regressions with a bandwidth of 30 days.
Importantly, the prison in which a prisoner is incarcerated is not random, nor is the sentence duration. Thus, the across-prison variations in program availability and those that arise from variations in sentence duration cannot be used to isolate the effects of programs. Yet, a series of other factors may exogenously affect program availability. One source of such variation is the timing of the incarceration. Indeed, fewer programs are usually available during the summer recess and the holiday season, as shown in Figure 4. Thus, a convict who must serve a sentence that overlaps the months of June–July or November–December will very likely be offered fewer opportunities. In what follows, we construct an instrument that captures such exogenous variations in program availability.
Average Number of Available Programs by Sentencing Month
Notes: This figure presents the average number of programs available for a prisoner to enroll in during theirs sentence conditional on the month the sentence starts.
IV. Estimation and Results
A. Research Design
Program participation is endogenous. To estimate the causal impact of participation on recidivism, we use the availability of programs as an instrumental variable. More precisely, our instrument corresponds to the number of available programs over the entire course of one’s incarceration. The instrument is constructed from the program enrollment files, which contain the registration date, the specific program and the personal identifier of all enrollees. Figure 5 illustrates this. Assume a timeline such that we observe three enrollees in Program A, three in Program B, four in Program C, and four in Program D. Given the timing and the duration of their incarceration, Prisoner 1 can only enroll in Programs B and C, and Prisoner 2, in Programs C and D. In this example, the instrumental variable is equal to two for both of them. Following this, we compute our instrument zip as the number of programs in prison p for which we observed any enrollment, from the beginning of the sentence of individual i to their release date. We test the robustness of our results to this definition in Section IV.D.
Computation of the Instrumental Variable
Notes: Each dot represents an individual enrolling in a specific program. Prisoner 1 can enroll in two programs (B and C) based on their sentence spell and observed enrollments. Prisoner 2 can also enroll in two programs (C and D). Thus, the number of available programs is 2 for both prisoners.
We begin with a series of reduced-form regressions to measure the intent-to-treat effects of program availability on recidivism. Consider inmate i who is incarcerated in prison p and sentenced during year t. We use the following specification:
1where yipt is a measure of recidivism, zip is the instrument, and si is the duration of incarceration during the sentence. In all specifications, we include prison fixed effects (αp) and year of sentence fixed effects (αt). Additionally, we include prison-specific length of stay effects (γp). We refer to these controls (αt, αp, and γpsi) as the set of randomization controls. We include these for the following reasons. First, some prisons tend to offer more programs. However, as mentioned previously, prison allocation is not random and will often be based on the offender’s characteristics. This explains our choice to use prison fixed effects and exploit within-prison variations in the instrument. Second, as shown in Figure 3, prisoners serving lengthier sentences are mechanically exposed to more programs at a rate that is prison-specific. As sentence duration is not random, our estimation must not exploit variations in zip arising from si. Conditioning on prison-specific length of stay effects (γpsi) is thus crucial in ruling out that our instrument captures such variations. Finally, the year of sentence fixed effects αt ensure that the remaining variations in our instrument arise from timely fluctuations rather than longer-term time trends. By conditioning on this set, the instrument leverages the within-prison vertical (that is, conditional on sentence duration) variations in program availability that are visible in Figure 3.
All regressions include a vector of individual characteristics, Xi′, that includes up to all the variables listed in Table 1. We cluster the standard errors at the prisoner level since some experience multiple spells—we verify the robustness of our findings when clustering at the prison–year level as well.23 We also test the robustness of our results to a specification that includes only the first incarceration spell observable for each individual. The distribution of zip is standardized to have a standard deviation of one. Therefore, β must be interpreted as the marginal effect of increasing program availability by one standard deviation on recidivism (σz = 11 in our sample).
We next estimate a series of two-stage least squares (2SLS) regressions. In these specifications, the number of programs in which inmate i participates, nitp, is instrumented using zip. The first- and second-stage regressions are given by:
2
3B. Validity of the Instrument
Table 3 reports estimation results of the first-stage regression, Equation 2. Column 1 only includes the randomization controls. Column 2 adds the additional control variables, Xi. Column 3 is the same as Column 2 except that it focuses on individuals that are observed up to five years upon release. In all three specifications, we find that the instrument is strongly correlated with program enrollment, as demonstrated by the high and significant F-statistics.
First-Stage Regression: Number of Programs
Table 4 next investigates the link between individual characteristics and program enrollment (nitp), on the one hand, and individual characteristics and program availability (zip) on the other hand. As reported, some explanatory variables have an impact on the number of programs enrolled in during a prison spell. Drug-related crimes and assault are found to have a positive impact while Indigenous convicts enroll in fewer programs than the non-Indigenous. In addition, inmates spending more time in pretrial detention tend to participate in more programs. However, Column 2 shows that program availability (the instrument) is unrelated to individual characteristics save for drug-related crimes and to Indigenous status. Nevertheless, the resulting F-statistic is as little as 1.59 and not statistically significant, which indicates that the variations in program availability that we exploit for identification are unrelated to individual characteristics. Column 3 uses all control variables to predict the number of programs an inmate enrolls in.24 The predicted values are not correlated to the instrument, suggesting that the characteristics are balanced across participation status.
Balance Tests
We further test the validity of the instrument by focusing on individuals who are incarcerated more than once during the time period covered by the data. We investigate the relationship between the instrument in a given sentence and its value in the previous sentence. If the two are correlated, this would imply that the availability of programs depends on unobservable time-invariant individual characteristics. Columns 4 and 5 use only observations for which we observe a past incarceration for the individual. Column 4 shows that program availability is unrelated to past program availability after conditioning on our randomization controls. Column 5 adds the other controls and finds similar results, providing additional support for our instrument’s validity.
As a final balance test, we apply the procedure proposed in Pei, Pischke, and Schwandt (2019). This involves separately regressing each characteristic against the instrument. Online Appendix Table B.1 details the findings. Two out of 11 characteristics statistically correlate with the instrument. Yet, when adjusting the p-values for multiple hypothesis testing for different outcomes (Romano and Wolf 2005; Clarke, Romano, and Wolf 2020), none of the coefficients remain statistically significant.
C. Main Results
We investigate the link between program take-up and four different types of recidivism: (i) incarceration in a provincial facility, (ii) community sentences, (iii) federal sentences, and (iv) all sentences. Moreover, we examine the occurrence of recidivism within different time frames, ranging from one year to five years after planned release. In the robustness analysis below, we further consider alternative start dates for measuring recidivism, including the onset of the sentence and the date of actual release.
Table 5 presents the results. Panel A focuses on reincarceration in provincial prisons. The ordinary least squares (OLS) estimates show a negative association between recidivism and the number of programs taken up during a spell, with each program being associated with a decrease in reincarceration of 0.7 percentage points within three years. The results are also suggestive of a negative causal relationship: the reduced-form estimates indicate that a one standard deviation increase in program availability reduces the likelihood of reincarceration by about four percentage points within the first three years, which translates into a decrease of 12 percent relative to the mean. The 2SLS estimates show a statistically significant negative effect but of greater magnitude than the OLS estimates. Specifically, for each additional program participation, the probability of recidivism decreases by about six percentage points on average, a decrease of 19 percent per program.
Effect of the Number of Programs on Recidivism
The previous beneficial effects of program participation do not carry over to reoffenses that lead to community sentences. According to Panel B, while the OLS estimates are all negative, the reduced-form and 2SLS estimators yield opposite results. Thus, a one standard deviation increase in program availability is found to increase the likelihood of future community sentences by one to three percentage points, or about 8 percent relative to the mean. The 2SLS yields estimates of similar magnitude. It might be conjectured that the rehabilitative programs induce participants to commit less serious offenses once released from prison—we test this channel more formally in Section V.
Recall from Figure 2 that federal reincarcerations are very few and that they are usually associated with serious offenses. Panel C focuses on these. Although all point estimates suggest that program availability and that participation decrease the likelihood of reoffending, all are close to zero and not statistically significant. The reduced-form estimator yields precise null effects. For example, a 95 percent confidence interval of the marginal effect of one standard deviation increase in program availability over a three-year period ranges from −0.8 to 0.3 percentage points. Finally, we consider all sentences in Panel D. Note that, as mentioned previously, the rates for “all sentences” do not correspond to the sum of the three others because one can recidivate multiple times within a given time period (see Footnote 21). Program availability and participation are found to have no effect on overall recidivism.25
Taken together, the estimates of Table 5 unearth interesting program effects. Globally, reintegration programs are found to impact reoffenses very little. Yet, they do impact prison reincarceration negatively and community sentences positively. This is likely beneficial from society’s perspective because community sentences are much less costly and are typically given for less serious offenses compared to those that resulted in the initial imprisonment. It is also worth noting that the positive impact of reintegration programs on reducing prison reincarceration can lead to significant reductions in prison overcrowding.
D. Robustness
The Online Appendix reports a series of robustness tests. These include omitting the control variables from the regression (Online Appendix Table B.3) or alternatively including the eight additional LS/CMI risk scores and focusing on the subset of evaluated offenders (Online Appendix Table B.4). We also compute prison–year clustered standard errors (Online Appendix Table B.5) and limit the sample to the first occurrence of each individual (Online Appendix Table B.6). The results are robust to all these alternative specifications.
Our instrument exploits prison-specific time variations in program availability. Although we control for year fixed effects, there remains the possibility that inmates who are incarcerated at different times during the same year might differ in their unobserved characteristics. To control for such potential seasonal effects, we also included month fixed effects corresponding to the first month of incarceration (Online Appendix Table B.7). Further, our original instrument measures the number of available programs during the entire incarceration spell. One might be concerned about the release date being endogenous. We thus construct an alternative instrument that counts the number of available programs up to the planned release date (Online Appendix Table B.8). Once again, the estimates are robust to these specifications.
In our primary regressions, recidivism outcomes were measured starting at two-thirds of the individual sentences, that is, the planned date of release. We investigate the consequences of measuring recidivism starting at the beginning of the sentence or at the actual release date in Online Appendix Tables B.9 and B.10. The results remain unchanged regardless of the start time we use.
Finally, Online Appendix Table B.11 replicates our main regression results, focusing only on observations for which we observe a previous incarceration for the individual and adding the value of the instrument associated with the previous incarceration as a control variable. The results remain largely unchanged, suggesting that our main conclusions hold for the subsample of offenders who are incarcerated more than once and that past program availability is unrelated to the outcomes.
E. Heterogeneity
Our baseline estimates use the number of available programs of any type as an instrumental variable. To estimate the effects of each specific program category (self-development, education, violence, job skills, addiction, and other), we construct an instrument for each. We then estimate the reduced-form model (Equation 1) separately for each category using the appropriate normalized zip. We control for prisoner characteristics and fixed effects in the same manner as in our main specification.
For the sake of brevity, we focus on recidivism leading to a provincial incarceration, as our main results showed the strongest and most statistically significant effects of programs for this outcome.26 Figure 6 displays the estimated effects of the different program types on the probability of being reincarcerated. As shown, most programs decrease reincarceration irrespective of the time frame. Furthermore, programs that address violence issues yield marginally larger effects compared to other program types. On the other hand, programs that focus on addiction and those in the other category have precise null effects on recidivism. As a robustness test, we estimate the effects of program availability for each program type in a single joint regression, rather than in separate regressions.27 The results (Online Appendix Figure A.3) depict a similar picture, save for self-development programs, for which the effects are not significant. Overall, our robust findings are that programs that address violence issues, education programs, and, to a lesser extent, job skill programs, seem to decrease recidivism.
Heterogeneity by Program Type
Notes: Each coefficient is obtained with a reduced-form regression of the reincarceration outcome on the number of available programs of each category, standardized by its standard deviation. All regressions include the set of randomization controls (prison- and year fixed effects, and the sentence duration × prison fixed effects) and full controls (type of crime, Indigenous status, age (categorical), indicator for at least a dependent, and number of days of pretrial detention).
The program effects potentially differ across inmates. Knowing who benefits most is essential for optimal resource allocation. For instance, programs may affect those serving short and long sentences differently. This information can then be used to tailor programs to individual needs and to maximize their impact.
We implement our baseline approach by estimating reduced-form regressions based on the total number of available programs and divide the sample into three equal-sized groups according to sentence length. In order to account for the fact that those with longer sentences have more opportunities to train, we standardize the instrument according to the number of months spent in prison.28 The instrument thus measures the marginal effect of increasing the number of available programs per month on reincarceration. Figure 7 depicts the parameter estimates. Program availability benefits all inmates irrespective of sentence length, though those serving longer sentences benefit most. Specifically, increasing monthly program availability by one reduces the likelihood of reincarceration by 0.25 percentage point for those serving short sentences (7–33 days) and by as much as one percentage point for those serving long sentences (112 days +).
Heterogeneity by Sentence Duration
Notes: Each coefficient is obtained with a reduced-form regression of the reincarceration outcome on the number of available programs by month of incarceration. All regressions include the set of randomization controls (prison- and year fixed effects, and the sentence duration × prison fixed effects) and full controls (type of crime, Indigenous status, age (categorical), indicator for at least a dependent, and number of days of pretrial detention).
Online Appendix Figure A.4 reports reduced-form estimates of the baseline specification stratified according to several characteristics. Interestingly, there does not appear to be any particular heterogeneous response to program availability among the groups we consider: the dynamic pattern is the same for each age group (below or above the median age of 35), each crime category, and each LS/CMI risk group (below or above the median risk score of 27).
Finally, we combine the heterogeneity assessments by characteristic and program type, conducting a reduced-form regression for each intersecting category. The outcome is reincarceration within a three-year span. This analysis, shown in Online Appendix Figure A.5, largely corroborates our previous findings. Most program types exhibit a significant impact, except for addiction programs and those in the other category. While these regressions are based on smaller sample sizes, it is interesting to observe that education programs seem less effective for low-risk inmates but highly beneficial for offenders convicted of assault. Nevertheless, the results for each program type are largely consistent across characteristics, and all point estimates for programs focusing on self-development, education, violence, and job skills indicate a negative impact on reincarceration rates.
V. Additional Analyses
In this section, we provide further evidence that the shift of future sentences from incarceration to community arises from programs reducing crime severity. We also evaluate the potential monetary benefits from this shift.
A. Effects on Pretrial Detention and Crime Severity
Our results suggest that programs may reduce crime severity and shift future sentences from incarceration to community correction. However, an alternative explanation could be that judges consider past program participation when determining sentence type, causing a higher propensity to grant a community sentence if the offender participated in the past.29 Therefore, more evidence is needed to conclude that programs affect crime severity.
Measuring the severity of a crime in our context is challenging, as no severity measure exists for the categorization of crimes we observe. An intuitive proxy for severity could be the sentence duration. However, community sentences are typically lengthier even though considered less severe. Therefore, such a proxy would be ill-suited to compare crime severity between community and incarceration sentences. In our context, a relevant indicator of crime severity is the occurrence of pretrial detention. Pretrial detention often signals an individual considered too dangerous to be released before trial.30 Importantly, as the decision to resort to pretrial detention happens before the trial, the effect of programs on this outcome cannot capture the sentencing judge’s consideration of past participation.
To estimate the effect of program availability on the likelihood of pretrial detention, we regress a dummy variable for pretrial detention on our instruments, controlling for our randomization controls and individual characteristics. As shown on Online Appendix Figure A.6, we find that, with the exception of addiction and other program types, the availability of most types of programs significantly reduces the probability of future pretrial detention. This pattern holds true across the entire sample, as well as if we focus on individuals who recidivate within one or five years. These results suggest that program successfully diminish the probability of offenders being deemed too dangerous to be released before a trial for future offenses.
B. Cost–Benefit Analysis
The net benefits of the shift from incarceration to community sentences are ambiguous. Indeed, community sentences, while less costly on a daily basis, are usually lengthier. Furthermore, as discussed in Section II.B, there is substantial heterogeneity in duration across program types. We thus conduct a simple, conservative, cost–benefit analysis of programs’ effects on sentencing costs taking into account these differences. The conservative assumption is that the programs’ benefits only arise from diverting individuals from prison to community sentences.31 More precisely, we use our reduced-form estimates of program availability on recidivism leading to a provincial incarceration within five years and assume this effect represents a shift from incarceration to community sentences. We compute program costs based on the salaries of professional counselors and on the typical heterogeneous duration of the programs. We use the average number of individuals in each program type to convert these total costs into cost per participant. Online Appendix C provides details and complete results.
Our findings reveal that, on average, programs deliver benefits that surpass their costs by a factor of three. The results are consistent across various characteristics, resonating with the heterogeneity results. Offenders with shorter sentences have a benefit-to-cost ratio ranging between 6 and 7, while those with longer sentences have a ratio of 1.51. Evaluating different program types, we note that programs targeting violence-related issues yield the highest benefits. Programs in the addiction and other categories exhibit negative ratios, though these are not precisely estimated. Though these results could seem high, they are not inconsistent with previous estimations in other contexts.32
VI. Conclusion
Recent evidence suggests that lengthy incarceration may favor rehabilitation (Landersø 2015; Bhuller et al. 2020; Hjalmarsson and Lindquist 2022) and that prison conditions play an important role (Lotti 2022; Mastrobuoni and Terlizzese 2022; Tobón 2022). Many have argued that rehabilitation programs may drive these effects. Our paper supports their claim. Using rich data from provincial prisons in Quebec, Canada, we investigate different forms of recidivism (community sentence, provincial prison, federal penitentiary), with each defined over several time windows ranging from one to five years. We address selection issues using an instrumental variable defined as prison-wide program availability while incarcerated. We find that the availability of rehabilitation programs addressing violent behavior or education and employment deficiencies can substantially decrease reincarceration. These findings are robust to several tests. Interestingly, we find no evidence of heterogeneous response to program availability by age group, type of crime, or (LS/CMI) risk profile.
Our results unearth an interesting feature of recidivism that is seldom investigated in the literature. While rehabilitation programs reduce the risk of reincarceration, they also increase the likelihood of future community sentences. We conjecture that rehabilitative programs may induce participants to commit less serious offenses once released from prison. This idea is supported by the significant decrease in the probability of pretrial detention when programs are more widely available. On the whole, prison-based programs are likely socially beneficial since community sentences are much less costly, are resorted to for less serious offenses, and can help alleviate the burden on the criminal justice system. We estimate that providing more opportunities to participate in various programs would result in significant net benefits, even if considering only the monetary benefits of reduced incarceration. For instance, increasing the number of programs by two during summer—such that program availability would match that of other seasons—would yield net benefits of $237 per participant. Furthermore, our heterogeneity cost–benefit analysis suggests that making these two additional programs focus on violence and job skills would yield maximal net benefits.
In this paper, the efficiency of rehabilitation programs is strictly gauged against future reoffenses. Investigating their long-term impacts on employment, housing, welfare dependency, or health is a worthwhile research avenue. Inducing convicts to serve community service rather than incarceration sentences may yield additional benefits by helping them maintain connections with their families, support networks, and employers.
Acknowledgments
The authors thank Quebec’s Ministry of Public Security, in particular Bernard Chéné, Isabelle Paquet, and Camille Blouin, for providing key institutional details and for crucial assistance with the data. They thank Julie Côté for organizing our visit at the Quebec Detention Center. They thank seminar participants at the Public Economics and Policy Seminar at HEC (Université de Lausanne), the Law, Institutions and Economics in Nanterre (Université Paris Nanterre), the Southern Economic Association, and the 2021 Conference of the Society of Labor Economists (SOLE), as well as three anonymous referees for comments that greatly improved the paper. The views expressed in this paper are those of the authors and do not necessarily represent those of the Ministry. A previous version of this paper has circulated under the title “Prison Rehabilitation Programs: Efficiency and Targeting.” The authors acknowledge the financial support of the Research Chair on the Evaluation of Public Policies. This research was also partially supported by the Australian Research Council’s Centre of Excellence for Children and Families over the Life Course (Project ID CE200100025). The authors declare that they have no duality of interest to disclose. This paper uses confidential data from administrative correctional files of the Ministry of Public Security of Quebec. Researchers wishing to access the data must contact the Ministry and submit a proposal at the following address: direction-recherche{at}msp.gouv.qc.ca.
Footnotes
↵1. In Canada, offenders sentenced to less than two years serve their sentence in a provincial prison, while those with sentences of two years or more are sent to a federal penitentiary.
↵2. Other studies find mixed results. For instance, Cook et al. (2015) conduct a RCT on high-risk offenders in Wisconsin. They find a decrease in the likelihood of a rearrest but no difference in the likelihood of reincarceration.
↵3. In her literature review, Doleac (2023) notes that “future work exploiting natural experiments or field experiments that avoid selection bias would be valuable for determining the power of educational programs–alone or in combination with other [prison-based] interventions—to encourage desistance from crime.” Davis et al. (2013) and Visher, Winterfield, and Coggeshall (2005) recognize the same gap in the literature.
↵4. Some papers find that lengthy prison sentences may deplete human capital and alter an inmate’s capacity to reintegrate into the labor market (Lochner 2004; Aizer and Doyle 2015; Mueller-Smith 2015). Others argue that prison spells might act as a school of crime in which networks of criminals share knowledge and influence one another (Bayer, Hjalmarsson, and Pozen 2009; Stevenson 2017). Conversely, incarceration may decrease crime through incapacitation (Owens 2009; Buonanno and Raphael 2013; Barbarino and Mastrobuoni 2014) or deterrence (Chen and Shapiro 2007; Mastrobuoni and Rivers 2016). Others who have studied the effect of incarceration (or incarceration duration) include Rose and Shem-Tov (2021) and Norris, Pecenco, and Weaver (2021), who find that incarceration decreases reoffending, and Kling (2006) and Green and Winik (2010), who find no effect on economic outcomes or recidivism, respectively.
↵5. Such conditions may include the composition of peers, occupancy rate, employee-to-inmate ratios, and prison activities, which have been suggested to affect recidivism (Chen and Shapiro 2007; Drago, Galbiati and Vertova 2011; van Ginneken and Palmen 2023; Yu et al. 2022). Lotti (2022) finds that a shift from strict warehousing to rehabilitating of young offenders decreased recidivism substantially. Mastrobuoni and Terlizzese (2022) find that rehabilitation-oriented (open) prison regimes yield lower recidivism relative to harsh (closed) prisons. Tobón (2022) shows that moving inmates from older to newer prisons—which are less crowded and offer better living conditions, services, and rehabilitation programs—substantially reduces recidivism. Finally, Hjalmarsson and Lindquist (2022) show that time spent in prisons that offer various programs allows inmates with mental health issues to engage in therapy, resulting in long-lasting health benefits and reduced recidivism.
↵6. Unlike probation, parole is not part of the sentence or determined during the trial. Rather, all prisoners with incarceration spells of at least six months may apply for parole while incarcerated. As such, parole is part of an incarceration sentence and is not captured by our measures of community sentences in this paper.
↵7. In Quebec, in 2018, the average inmate daily cost is $254 (Statistics Canada 2022). The daily cost of probation is $49 (Parliamentary Budget Officer 2018).
↵8. In order to incentivize good behavior, time may be credited towards release upon serving two-thirds of a sentence.
↵9. Arbour and Marchand (2023) study the effect of parole among this population and find it can yield rehabilitative effects. Parolees represents less than 9 percent of our sample. They only serve one-third of their sentence unless reincarcerated due to a breach of conditions. Fewer than 40 percent of convicts are eligible for a hearing, of whom only 22 percent are granted parole.
↵10. The LS/CMI is a widely implemented proprietary assessment tool in North America (Andrews, Bonta, and Wormith 2004) and is the culmination of ulterior versions of tools within the “Level of Supervision Inventory” family. Officers must complete the evaluation seven days at the latest before the sixth of the sentence or within 45 days after sentencing, whichever comes first. If an inmate is reincarcerated, a new evaluation is not deemed necessary if a prior evaluation is still considered appropriate and was completed less than two years prior. Only trained officers can conduct the assessment. See Vose, Cullen, and Smith (2008) for a review of the LS/CMI.
↵11. Article 21 of the Act Respecting the Québec Correctional System (2023) states: “The Minister shall develop and offer programs and services to encourage offenders to develop an awareness of the consequences of their behavior and initiate a personal process focusing on developing their sense of responsibility.” Article 22 reads: “The Minister shall see to it that the offenders’ access to specialized programs and services offered by community-based resources to foster their reintegration into the community and support their rehabilitation is facilitated. Such programs and services are designed to initiate the process of solving the problems associated with the delinquency of the offenders, in particular problems of domestic violence, sexual deviance, pedophilia, alcoholism, and substance abuse.”
↵12. In French: Dossiers Administratifs CORrectionnels.
↵13. In practice, offenders can be sentenced for each charge. Our data only contain the sentence for the most serious charge for which the offender was found guilty.
↵14. All risk scores are highly predictive of recidivism in our sample. See our previous version of this paper for a detailed analysis (Arbour, Lacroix, and Marchand 2021).
↵15. Twenty-two prisons were active over our sample window. Eighteen are still active as of this writing; some were closed, and others were merged into new and more modern establishments. Among the seven prisons for which we have program participation files, six are among the eight largest in terms of caseloads. The other two are relatively small establishments.
↵16. For instance, Kuziemko (2013) defines recidivism as any reincarceration occurring within three years. In contrast, Bhuller et al. (2020) define recidivism as the event of being charged with at least one crime during a given period. The Quebec Ministry of Public Security as well as Public Safety Canada both consider as recidivism any new sentence for a different offense occurring within two years following to end of the previous sentence. Alternatively, the United States National Institute of Justice defines recidivism as any subsequent involvement with the criminal justice system within three years following release, whether or not a new sentence was issued.
↵17. This is not surprising given that the majority of inmates are released near the planned date. Recall that fewer than 9 percent of individuals in our sample are granted parole. See Footnote 9 for further details.
↵18. The “other” category comprises crimes such as fraud, traffic-related offenses, prostitution-related crimes, and possession of illegal weapons.
↵19. From extensive discussions with prison personnel, one reason inmates seldom enroll in any program during pretrial is the fear it may be interpreted as an admission of guilt.
↵20. Note that the comparisons are conditional on being evaluated. Hence the smaller number of observations.
↵21. The rates associated with “All sentences” do not correspond to the sum of the other three. This is because one can recidivate several times over a given period, each time with a different sentence type. For example, consider someone who recidivates twice within a given time window. He may first receive a community sentence, then be incarcerated. The dependent variables we measure are: yinc = 1, ycom = 1, yfed = 0, yall = 1. Thus, yinc + ycom + yfed ≠ yall.
↵22. Online Appendix Figure A.1 plots the relationship between incarceration length and recidivism one and five years after planned release and shows a negative correlation between the two. Online Appendix Figure A.2 plots the average number of program enrollments and shows that it positively correlates with incarceration length.
↵23. More precisely, 75 percent of individuals appear only once, 15 percent appear twice, and 10 percent appear three times or more.
↵24. We thank a referee for this suggestion. The predictions are obtained with an OLS regression of the number of program participation on all observed characteristics and the randomization controls.
↵25. We further investigate the effects of program availability on recidivism at the intensive margin. That is, we estimate the effect of one additional program conditional on participating in at least one program. Our results are reported in Online Appendix Table B.2 and are consistent with the main results. Each additional program reduces the likelihood of reincarceration and increases that of community sentencing slightly, although the latter is not statistically robust to post-release intervals.
↵26. We find no heterogeneous effects when considering the other outcomes. The entire set of results is available upon request.
↵27. The idea behind this test is that, if prisoners are limited in the number of programs they can participate in, then participating in one program could imply not participating in other programs. In that case, in the separate regressions, the effect of a wider availability of one program type could capture the effect of a decrease of participation in other programs. We thank an anonymous referee for raising this possibility.
↵28. More precisely, we divide the instrument by the number of days of incarceration multiplied by 30.
↵29. We thank an anonymous referee for raising the possibility.
↵30. When the police arrest an individual, they may either release the person with a court summons or incarcerate the person immediately. If incarcerated, the person will appear before a judge within 24 hours. This judge will decide if the person is to be released with certain conditions until the trial or be detained until the trial. This decision is supposed to be based on the crime’s seriousness, the individual’s potential danger to the public, their past criminal record, or the possibility that the individual does not show up to the trial. Thus, past program participation is unlikely to be assessed at this stage.
↵31. We view this assumption as conservative because programs may yield benefits beyond the monetary costs associated with the type of future sentences. If programs reduce crime severity, one could expect additional benefits, such as reduced social costs associated with crimes or positive employment effects.
↵32. For instance, Heller et al. (2017) found a significantly higher benefit-to-cost ratio of 30 for a program aimed at juveniles. However, it is important to note that their analysis incorporated the social costs of crime, which are not included in our analysis. See also Zarkin et al. (2012).
- Received October 2021.
- Accepted December 2023.
This open access article is distributed under the terms of the CC-BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0) and is freely available online at: https://jhr.uwpress.org.
References
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.