Recidivism Felony ATI Programs 02 by Administrator

CJA

NEW YORK CITY CRIMINAL JUSTICE AGENCY NEW YORK CITY CRIMINAL USTICE AGENCY

Jerome E. McElroy Executive Director

CRIMINAL RECIDIVISM AMONG FELONY-LEVEL ATI PROGRAM PARTICIPANTS IN NEW YORK CITY

Jukka Savolainen, Ph.D. Project Director / Senior Research Analyst

FINAL REPORT

August 2002

52 Duane Street, New York, NY 10007

(646) 213-2500

CRIMINAL RECIDIVISM AMONG FELONY-LEVEL ATI PROGRAM PARTICIPANTS IN NEW YORK CITY

Jukka Savolainen, Ph.D. Project Director / Senior Research Analyst Wayne Nehwadowich Senior Programmer/Analyst Aïda Tejaratchi Programmer/Analyst Bernice Linen-Reed Administrative Assistant

August 2002

ACKNOWLEDGEMENTS

Mary Eckert, the former Research Director of CJA, developed the foundations for this study. The authors would also like to acknowledge the assistance of other colleagues at CJA who have contributed to the progress of this research project: Raymond Caligiure, Fanny Castro, Mari Curbelo, Joann DeJesus, Barbara Geller Diaz, Marian Gewirtz, Taehyon Kim, Jerome E. McElroy, Mary T. Phillips, Rick Peterson, Elyse Revere Frank Sergi, and Cheryl Welch.

The research has

benefited from contacts with our colleagues at The Vera Institute of Justice.

We are also grateful for the following agencies for providing us with supplemental data: The New York City Department of Corrections, The NYC Department of Probation, The New York State Department of Correctional Services, The NYS Division of Criminal Justice Services, and The NYS Office of Children and Family Services. None of these agencies are responsible for the contents of this report.

TABLE OF CONTENTS List of Tables and Figures............................................................................................... ii INTRODUCTION ............................................................................................................1 1. THE ATI EXIT COHORT ............................................................................................4 Characteristics of the ATI Sample ........................................................................6 2. MATCHED COMPARISON GROUPS ......................................................................12 The Purpose of Matching ..................................................................................12 Comparative Treatments: Incarceration and Probation ......................................14 Matching Criteria ................................................................................................16 Descriptive Statistics ..........................................................................................19 3. TRACKING RECIDIVISM .........................................................................................23 Measures of Criminal Recidivism .......................................................................25 Methods of Analysis ...........................................................................................29 4. FINDINGS ...............................................................................................................36 4.1 Basic Comparisons .........................................................................................36 Does Success in the Program Matter? ...............................................................42 Program-Specific Comparisons..........................................................................45 4.2 Multivariate Models .........................................................................................50 Variables ............................................................................................................52 Findings .............................................................................................................54 5. CONCLUSIONS .......................................................................................................63 Appendix A: The Seven ATI Programs .........................................................................66 Appendix B: How the 7-Digit Matching Variable Was Constructed? .............................67 Appendix C: Program-Specific Comparisons of the Type and Severity of Rearrest .....69 REFERENCES .............................................................................................................69

LIST OF TABLES AND FIGURES TABLES Table 1-1. Table 2-1. Table 3-1. Table 3-2. Table 4-1.

Methods of ATI Targeting and Status at Final Exit ........................................7 Distribution of Matching Characteristics ......................................................21 Length of Tracking in the ATI Sample and Comparisons Groups ................25 Prevalence and Incidence Measures of Recidivism by Gender...................30 Prevalence and Incidence of Recidivism in the ATI Sample and Comparison Groups ....................................................................................37 Table 4-2. Prevalence and Incidence of Recidivism Among Successful ATI Graduates and Matched Comparison Cases ................................................................43 Table 4-3. Cox Proportional Hazard Regression Models of Time to First Rearrest ......55 FIGURES Figure 1-1. Figure 1-2. Figure 3-1. Figure 4-1. Figure 4-2. Figure 4-3. Figure 4-4.

Age and Gender Distribution of ATI Participants..........................................9 Top Indictment Charge Types in the ATI Sample.......................................10 Timing of First Arrest by Gender, Full Sample ...........................................33 Timing of First Arrest in the ATI Sample and Comparison Groups .............41 Program-Specific Comparisons of the Prevalence of Rearrest ..................46 Program-Specific Comparisons of the Incidence of Rearrest .....................48 Effect of Success in the Program and the ATI Regime on Rearrest ...........61

Introduction

Under contract with the Mayorâ&#x20AC;&#x2122;s Office of the Criminal Justice Coordinator (OCJC), the New York City Criminal Justice Agency, Inc. (CJA) has completed a study of criminal recidivism among felony offenders that have participated in an alternative-to-incarceration (ATI) program funded by OCJC. This report describes the findings from this study.

The following set of ATI programs are included in the research: Youth Advocacy Project (YAP), Court Employment Project (CEP), Freedom Program of the Fortune Society, DAMAS Program of the Fortune Society, El Rio Day-Treatment Program, FlameTree Program of the Fortune Society, and Womenâ&#x20AC;&#x2122;s DayTreatment Program of the Project Return Foundation. The basic mission and scope of each program is described briefly in Appendix A. This selection excludes two of the nine felony ATI programs in the Alternative-to-Incarceration Information Services (ATIIS) reporting network: Crossroads Project of the Center for Community Alternatives and STEPS to End Family Violence of the Edwin Gould Services for Children. Crossroads is excluded because, funded through the City Council, it is not subject to the same targeting constraints as OCJCfunded programs. On the other hand, the number of cases that had exited the STEPS program in time to be included in the study sample was too small for meaningful statistical analysis. Given the focus on felony offenders, the

recidivism study does not include cases from the Community Service Sentencing Project (CSSP), an ATI program for misdemeanor cases.

This study also excludes all other correctional interventions available in New York City that may be characterized as “alternatives to incarceration” but do not receive funding from OCJC. Some of these may receive funding from the City Council and/or New York State, while others may operate without any reliance on government support. It is important to understand that the results of this research cannot be generalized beyond the realm of these seven ATI programs. For the sake of parsimony, the remainder of this report reserves the term “ATI program” to refer only to the programs included in this study. Thus, unless specified differently, “ATI participants” henceforth denotes clients of any one of the seven programs listed above. As none of these seven programs have a court presence in Staten Island, this borough of New York City is outside the scope of the study.

This study is focused on the comparative analysis of post-program recidivism. The routine reports by ATIIS publish statistics concerning the rates of recidivism of ATI participants while in the program (e.g., Revere and Curbelo 2001). The current report is concerned about the effects that ATI participation may have on those individuals who complete a program and remain in the community. In order to provide context to these effects, the recidivism of ATI participants is analyzed in comparison with offenders who are similar to the ATI “treatment group” with one exception: they did not participate in any of these seven programs in the

course of their case processing. Specifically, the ATI participants are compared to two traditional sanction categories: those sentenced to straight probation and those sentenced to incarceration (local jail or state prison).

The first part of this report describes the ATI sample selected for this study. This section is followed by a discussion of the matched comparison groups constructed for the ATI sample. The focus of section 3 is on the methodological aspects of the research, including the measurement of recidivism and the methods of analysis. The findings from the research are reported in section 4. The final section of the report provides a summary of the key findings and offers some conclusions concerning the impact of ATI programs on recidivism in New York City.

The ATI Exit Cohort

The ATI population included in this study consists of cases that completed any one of the seven programs between July 1, 1998 and March 31, 2000, and remained in the community. It should be emphasized that the final ATI study sample is not representative of all the cases mandated to participate in these programs. Only those who finish the program, successfully or unsuccessfully, are retained in the sample. Cases of the following description are not included in the post-program recidivism study: •

Cases that never showed up in a program, although mandated

•

Cases ending because of death

•

Cases exited because participant moved to another jurisdiction

•

“Administrative discharges” (CCSS1-screened cases that programs had deemed inappropriate after enrollment)

•

Cases transferred to some other residential or community-based program

•

Cases without valid data on matching characteristics in the UDIIS database

•

Cases exited at sentence resulting in long-term incarceration2

Established in FY98, Centralized Court Screening Service (CCSS) is the predecessor of ATIIS. In addition to the targeting and the data management functions associated with ATIIS, CCSS was in charge of screening and court advocacy for the ATI programs. In July 1999 OCJC decentralized these functions back to the individual programs. 2 For the purposes of this study, the length of incarceration is judged as “long-term” if it prevents an ATI participant from being at risk of recidivating for at least six months. For example, this rule excludes all ATI participants sentenced to state prison upon program exit.

All of these omissions are inevitable because it would not be feasible to study post-program recidivism of ATI cases that did not actually participate in program activities, that are not in the community at risk to recidivate, or that are still participating in some form of program or intervention. While some of these exclusions are rather innocuous, on balance they are likely to underestimate the degree of recidivism in the ATI participant population. Cases that end up being transferred or administratively discharged tend to represent the more difficult end of the rehabilitation continuum. This characterization applies even more strongly to those cases that were terminated because of incarceration, as a number of them ended up incarcerated as a result of in-program recidivism, breach of program rules, and/or violations of other court requirements. A number of these exclusionary outcomes (e.g., mortality, incarceration, and transfer) may be related to factors such as the severity of the drug problem, the presence of a mental health problem, or some other psychosocial characteristic that we will not be able to measure in our research. It is unlikely that matching or any other form of statistical control can address all these sources of bias. This consideration should be kept in mind while interpreting the results of the analysis.

It is also possible that a case may exit from one program only to continue in another. Our study considers only â&#x20AC;&#x153;true exitsâ&#x20AC;?, i.e., cases that are no longer active in any of the seven ATI programs. We do include cases that have been released into an ATI program more than once, as long as they are not active in any of them after March 31, 2000. The sample does also include â&#x20AC;&#x153;monitored

transfersâ&#x20AC;?, i.e., cases that, although transferred to the immediate care of some other service provider, continue to be under the supervision and reporting responsibility of an ATI program. However, only those monitored transfers are included who are no longer active in the ATI program.

Characteristics of the ATI sample Program characteristics The size of this final ATI sample is just slightly over one thousand (1,005). The earliest case (in terms of criminal justice processing) in the sample was first mandated into a program on July 3, 1997; the most recent ATI admission took place on March 14, 2000. A large majority (87 %) of the cases in the sample were released during the operation of CCSS. The remaining 131 cases were released after June 30, 1999. The median length of program participation is six months (for those 40 cases that were released to an ATI more than once, the median length of the total time spent in the system was 7 months).

As indicated by Table 1-1 on Page 7, with 360 admissions and 350 completions, CEP is the largest source of participants in the sample, followed by Freedom (229/231) and YAP (104/104). Fewer than 50% of the cases were released to an ATI program by way of statistical targeting process. The most typical source of outside referral was either a judge or a defense counsel. Program-specific variation in these patterns is very minor. A majority of the cases (63%) in the recidivism sample completed the program successfully at

Total

Reason for UT Non-attendance Breach of program rule Non-complience with treatment Failure at transfer placement Physical illness Incarceration Rearrest Warrant issued Other

Completion status at final exit Successful completetion (SC) Unsuccessful termination (UT)

3 25 18 1 1 0 2 0

% 1 43 50

100 53 6 12 1 0 0 21 5 4

350 100

81 43 5 10 1 0 0 17 2 3

20 6 2 1 360 100 CEP N % 204 58 146 42

11 90 66 3 3 0 7 0

Source of referral DA Defense counsel Judge Family/Self ATI Program Treatment court Other Source not known

Court screening Mechanism unknown Total

CEP N 5 153 180

Method of targeting Manual targeting Computer targeting Referral 6 20 23 0 3 0 2 0

100 27 0 9 18 0 0 0 18 27

67 100

11 3 0 1 2 0 0 0 2 3

4 6 2 3 66 100 DAMAS N % 46 69 21 31

4 13 15 0 2 0 1 0

DAMAS N % 1 2 25 38 35 53 13 16 3 0 0 13 5 21

100 59 0 7 3 3 14 3 0 10

77 100

29 17 0 2 1 1 4 1 0 3

0 0 0 0 80 100 El Rio N % 33 43 44 57

10 13 2 0 0 10 4 17

El Rio N % 0 0 24 30 56 70

5 22 17 2 1 0 5 0

15 52 48 1 2 0 2 1

7 23 21 0 1 0 1 0

Freedom N % 3 1 101 44 121 53

136

31 16 2 3 2 0 1 6 1 0

18 13 1 2 0 0 2 0 0 0

100 231

100 52 6 10 6 0 3 19 3 0

100

100 72 6 11 0 0 11 0 0 0

3 2 3 1 0 0 1 0 126 100 229 100 FlameTree Freedom N % N % 87 64 180 78 49 36 51 22

6 28 21 3 1 0 6 0

FlameTree N % 2 2 56 44 65 52

Table 1-1. Methods of ATI Targeting and Status at Final Exit. Program-Specific Statistics.

23 13 5 0 5 8 8 5

% 8 15 65 2 8 36 0 0 0 1 0

YAP N 19 38 47 2 8 35 0 0 0 1 0

% 18 37 45

57 209 190 7 10 13 24 20

5.7 20.8 18.9 0.7 1.0 1.3 2.4 2.0

Total N % 33 3.3 403 40.1 530 52.7

20 10 0 4 0 0 3 0 3 1

19 7 1 0 0 0 2 2 6 1

100 104

100 50 0 20 0 0 15 0 15 5

209 109 9 22 6 1 12 26 14 10

100 1,005

100 37 5 0 0 0 11 11 32 5

100

100 52.2 4.3 10.5 2.9 0.5 5.7 12.4 6.7 4.8

4 10 0 0 34 3.4 1 3 0 0 5 0.5 40 100 104 100 1,005 100 Return YAP Total N % N % N % 16 40 65 63 631 62.8 24 60 39 38 374 37.2

9 5 2 0 2 3 3 2

Return N 3 6 26

final exit.3 Freedom is associated with the highest success rate at 78 %, while El Rio (43 %) and Project Return (40 %) are the programs with the lowest rates of successful completion. Reasons for unsuccessful terminations (UT) are not well documented in the data because this information was not required by OCJC until February 1999. However, among those UT cases for which this information is available (N=209, 56%), the leading cause of termination is non-attendance (52%) followed by re-arrest (12%).

Participant characteristics With 311 (31%) cases, the Supreme Court of Manhattan is the largest source of ATI participants in the sample, followed by the Bronx (30%), and Brooklyn (27%); 13 percent of the cases were processed in Queens. At the time of arrest, fewer than 5 percent of the people in the sample had a permanent address outside the five boroughs. The ethnic composition of the ATI participant sample is dominated by Blacks and Hispanics: 53 % and 42 %, respectively; there are only 34 (nonHispanic) White individuals and not a single person of Asian ethnicity in this sample. On the basis of the ROR-interview, 49% of the ATI participants in the recidivism sample are either employed or enrolled in school or a training program. A chart describing the age and gender composition of the sample is provided on Page 9 (Figure 1-1). Not surprisingly, an overwhelming majority (84 gender are between 17 and 20.

These statistics are not valid estimates of the likelihood of successfully completing these ATI programs. Given the way in which the recidivism sample is selected, these success rates do not describe any incoming ATI cohort. The research sample is based on an exit cohort of participants that are discharged into the community upon program completion. By design, participants who fail in the program are systemically underrepresented in the recidivism sample.

Figure 1-1: The Age Distribution of ATI Participants by Gender.

Number of ATI Participants

180 160 140 120 100

Men Women

80 60 40 20

0 Age in Single-Years

In order to be eligible for ATI treatment, a felony offender must fit a certain criminal justice profile. One goal of the ATI system in New York City is to displace jail or prison time. As a result, these programs tend to target defendants who are facing a sentence of six months or more in incarceration. On the other hand, because incarceration is mandatory for certain categories of offender status, the ATI programs do not tend to get defendants charged with A-level felonies, felonious sex offenders, or those with multiple felony convictions, among others. These constraints in the ATI selection process generate a degree of homogeneity in the sample. It is not representative of the general client population of the NYC criminal justice system, but reflects a segment of the â&#x20AC;&#x153;middle rangeâ&#x20AC;? felony offenders.

60% of the ATI cases were charged with a B-felony in the Supreme Court arraignment; over 95% of them fall between B- to D-level felonies; only five cases were charged with an A-felony, the most severe class of criminal offending. A more detailed analysis of the offense types reinforces the impression of homogeneity with respect to the criminal justice profile in this sample: Only two articles in the New York Penal Law, robbery (§ 160) and criminal possession/sale of controlled substance (§ 220), suffice to cover about 3/4 of the entire range of arrest charges in this sample of ATI participants. The distribution of the top charges by the type of offense is displayed in Figure 1-2 below.

Figure 1-2. The Distribution of the Top Indictment Charge by CJA Type Classification in the ATI Participant Sample (%). Percent of All Cases (N=1,005) 0

10 7.1

Harm to persons … persons and property

5.3 3.8

Drug Other

40.6

Weapon Property crime

40.8 2.4

Instead of listing all the sections in the Penal Law, it is based on an aggregated typology. This chart replicates the bipolar distribution of the offense distribution revealed by the examination of the PL sections. The CJA type category entitled

“harm to persons and property” is dominated by robberies, whereas the “drug” category entails the charges under PL 220. The third most frequent offense type is “harm to persons” (7.1%), which involves assault charges. A majority (62%) of the individuals in the sample had at least one prior adult arrest. However, a minority had been previously convicted of either a felony (7%) or a misdemeanor (18%).

Matched Comparison Groups

The purpose of matching The basic research question of this study is quite simple: do offenders who have completed an ATI program end up recidivating more frequently or less frequently than they would have, had they not participated? In terms of research methodology, randomized experiment would constitute the most efficient way to study this question. The starting point would be a pool of offenders that qualify as program participants. By way of random selection, half of these people would be assigned to the appropriate ATI program, while the other half would be given more standard treatments, such as incarceration or probation. Since, by virtue of design, the only systematic difference between these groups would be the nature of the sanction, any differences in recidivism could be attributed to this single variable. Although experiments of this description are not uncommon in policyoriented criminal justice research (e.g., Harrell, Cavanagh, and Roman 2000), our study has not been designed along these lines.

In the context of non-experimental research, the use of statistical controls is the standard way of achieving sound comparisons. This approach calls for some form of multivariate analysis, such as cross-tabulation or multiple regression, that divides the data into smaller units of comparison standardized with respect to any number of relevant characteristics. In random experiments, the comparability between groups is ensured prior to the treatment, whereas in multivariate

analysis statistical controls are applied as a post-treatment remedy. The latter approach is intrinsically inferior for at least two methodologically compelling reasons. First, the data needed to control for all the relevant differences between the treatment group and the comparison group may not be available. Second, the size of the samples may be too limited to allow for a simultaneous use of several control variables.

Although our research does rely heavily on techniques of multivariate analysis, the design of this study can be best characterized as quasi-experimental, a mix between a random experiment and a purely non-experimental situation (Posavac and Carey 1997, 160-180). The basic advantage that we have over a strictly nonexperimental design is the fact that we have considerable control over the composition of the comparison group. We have ample information about the pretreatment characteristics of the ATI participants: their age, sex, prior record, the nature of the offense, etc. With the help of this information, we can develop a set of criteria for choosing our comparison group to match the profile of the treatment group. To be sure, we can match only on those characteristics for which we have data available. Fortunately, CJAâ&#x20AC;&#x2122;s database features a very comprehensive pool of suitable data elements. As a matter of principle, this approach can never guarantee a perfect match. However, all the variables that we have deemed salient for our purposes are at our disposal.

A study of this description is â&#x20AC;&#x153;quasi-experimentalâ&#x20AC;? in that the composition of the comparison group is constrained to be similar to the treatment group in all but one respect, participation in an ATI program. The presence of matching comparison groups reduces the number of control variables required in the analysis. This property is particularly desirable from the perspective of programspecific comparisons. For example, in the current version of the research sample, there are only 66 cases in DAMAS and 40 cases in Project Return. Samples of that size do not lend themselves to the kinds of detailed breakdowns that exclusive reliance on multivariate analysis would require.

Comparative treatments: incarceration and probation Since ATI stands for alternative-to-incarceration, the semantically correct population base for drawing the comparison group consists of those offenders who were sentenced to incarceration. However, there is reason to believe that a number of clients released to an ATI program would not have ended up serving time in the absence of the ATI option. For example, some judges do not seem to distinguish ATI programs from other community-based sanction alternatives that are not set out to displace jail time. Thus, in lieu of an ATI program placement, some participants could have been diverted into some other type of drug treatment or employment program. Also, a number of individuals who meet the statistical ATI targeting criteria, but are screened out in the process, are known to receive a sentence of probation. (Since participation in a residential treatment program may be imposed as a condition of probation, in some cases an ATI may

represent a more lenient sanction than probation, at least in terms of confinement.) Finally, the process of targeting defendants for ATI advocacy, prior to disposition and sentence, is subject to error both in targeting cases that would not have been incarcerated and in failing to target some that would have been. In light of these considerations, we deem it prudent to draw comparison groups from two basic populations: those sentenced to incarceration, jail or prison, and those sentenced to probation. The effect of â&#x20AC;&#x153;ATI treatmentâ&#x20AC;? on recidivism is compared to both of these traditional sanctions.

The probation comparison group is drawn from the population of felony cases sentenced to straight probation between July 1, 1998 and March 31, 2000. The probationers are, thus, discharged into the community to serve their sentence at the same time as the ATI participants exited their programs. In other words, the ATI group and the probationers are at risk of recidivating around the same time. Standardized period at risk is one way in which the comparison groups are constrained to be similar to the ATI sample. This ensures that any differences in the rearrest rate between the two groups cannot be explained in terms of some period-specific influences, such as sudden crackdowns by the police. The treatment group and the control groups are subject to the same set of trends. The probation data are provided by the NYC Department of Probation (DOP).

The comparison group of incarcerated felons consists of two distinct penal categories: jail inmates and state prisoners. The data on jail inmates are provided

by the NYC Department of Corrections (DOC) while the prison data come from the NYS Department of Correctional Services (DOCS) and the Office of Children and Family Services (OCFS; this source covers juvenile felony offenders serving time in a state facility). The two basic categories of incarceration, prison and jail, are treated as separate groups in this study; a separate matched comparison group will be developed for each sanction. In accordance with the principle of standardized period at risk, these comparison groups are drawn from a population of felony offenders discharged from a state facility or a local jail between July 1, 1998 and March 31, 2000.4

The matching criteria To reiterate, the purpose of matching is to make the comparison group resemble the treatment group of ATI participants in all relevant respects: if 20 percent of the ATI participants are women, and 70 percent are between ages 14 and 25, the comparison group should reflect the same mix of gender and age. Methodologically speaking, matching is but an alternative way of imposing statistical controls. In the case of multivariate analysis, the samples are given, and the purpose of statistical controls is to deal with heterogeneity between the treatment group and the comparison group. In matching, we use information concerning the characteristics of the ATI treatment group to select comparison 4

The data from the DOCS do not include cases that were sentenced as Youthful Offenders (YOs) under the principle that such records are sealed from the public. However, we were able to reach data sharing agreements with the other source agencies (DOP, DOC, and OCFS), which provided us with access to all types of cases, including YOs. As the ATI sample does include cases eligible for YO treatment, the absence of such cases in one comparison group is unfortunate. On the other hand, all the variables that determine the YO status (age, prior record, and the nature of the instant offense) are a part of our matching criteria. This should minimize any bias caused by the exclusion of YO cases in the DOCS file.

groups in a way that minimizes heterogeneity between the two sets of samples. As stated above, an important advantage with this strategy is that it saves â&#x20AC;&#x153;degrees of freedomâ&#x20AC;? in the course of analysis, i.e., it enables meaningful comparisons without the need for complex multivariate modeling. If two samples have an identical gender composition and age structure, any differences between their rates of recidivism cannot be caused by gender or age, and there would be no need to attend to those variables in the analysis.

In the case of a random experiment, the similarity of the treatment and control group is a built-in property of the research design. In the process of generating matched comparison groups, the similarity must be achieved by choosing cases that share characteristics with the ATI cases. Obviously, no two cases are exactly alike. In choosing the criteria for matching, it is necessary to distinguish between relevant and irrelevant (or less relevant) characteristics. To simulate an experimental design, the cases in the comparison group should be composed of individuals that could have been admitted to one of these felony ATI programs. In other words, the nature and the severity of their offense, their criminal history, etc. should be similar to those of the cases in the ATI sample. In addition to this criterion, our choice of matching criteria is informed by the goal of minimizing variation in demographic characteristics that are salient predictors of recidivism, such as age and gender.

The characteristics we used in matching are as follows: Instant offense: -

Top charge type at Supreme Court arraignment

Top charge severity at Supreme Court arraignment

Borough of prosecution

Detention status leaving Criminal Court arraignment

Criminal history: -

Prior arrest

Prior convictions

Demographic: -

Age at the time of arrest

Gender

On the basis of these characteristics, we created a 7-digit variable, each value of which indicates a unique combination of the above characteristics5. The ATI participant sample was first described in terms of this variable. For example, one CEP participant scored 1042314 on the matching variable. This code indicates that he is a male at ages 18-19 with one prior misdemeanor conviction; he was prosecuted in the Bronx, charged with a B-felony drug offense, and he was detained leaving Criminal Court arraignment.

The next step was to select an identical case from each comparison group. In some cases, an identical match was not available. To find the closest possible match we would first look for an identical case from a different borough of prosecution: if the ATI case was from Manhattan, we would look for an identical case in the Bronx (and vice versa); if the ATI case was from Brooklyn, we would 5

See Appendix B for details.

look for a match in Queens (and vice versa). If this did not produce a match, we would try relaxing our age criteria, limiting the search to cases at only one age grade above or below the target case. If none of these modifications worked, we would not look any further.

Descriptive statistics We were not able to find acceptable matches for each ATI case in the sample. However, a large majority (79%) of the cases in the ATI sample have a match in each of the three comparison groups: probation, jail, and state prison (including the OCFS). 16 percent did not have a match in one comparison group and 4 percent missed a match in two groups. Only nine ATI cases (less than one percent) had no match in any pool of comparison. The proportion of unmatched cases is lowest for the state prison group: only 3.5 percent of the ATI cases did not have an acceptable match among cases discharged from either a state prison or an OCFS facility. With the shortage of matches for 17.7 percent of the ATI cases, the DOC had the lowest number of hits. 6.2 percent of the ATI cases are without a matched comparison in the probation data.

The purpose of generating comparison groups for the ATI cases is to provide a context for evaluating the impact of these programs. Any result regarding the recidivism of ATI participants is quite meaningless without knowing the equivalent statistics pertaining to a comparable group of felony offenders outside the ATI system. Hence, in the analyses that follow, all the comparisons between

the ATI sample and its comparison group involve only those ATI cases for which there is an acceptable match. As a result, the size of the ATI comparison samples varies depending on the comparison group: the ATI sample is largest (N = 970) in comparisons with the state prison sample and smallest in comparisons with the jail sample (N = 827). Retaining ATI cases without a matched equivalent in the comparison group would defeat the very purpose of matched comparison. As only nine ATI cases have no match in any of the three comparison groups, more than 99 % of the cases in the ATI sample are included in these comparisons. Table 2-1 on Page 21 describes the composition of the resulting ATI samples and their comparison groups with respect to the matching characteristics. The purpose of this table is to describe the degree to which the ATI sample is similar to its comparison group with respect to the matching characteristics. An identical match is one where there are no differences between the ATI sample and its comparison group. The first row of percentages in Table 2-1 is an example of a perfect match with respect to the proportion of cases charged with an A level felony: the percentage of such cases in each comparison group is exactly the same as in the respective ATI sample. Although the same outcome is not true for all the categories of charge severity, the distributions are remarkably similar across the board.

Indeed, each of the three comparison groups matches very closely across all of the seven characteristics. There are only two instances where the ATI sample differs from its comparison group by a statistically significant (p â&#x2030;¤ .05) margin.

Table 2-1: The Distribution of the Matching Characteristics in the ATI Sample and Its Comparison Groups (%). Matching Characteristic Charge severity A Felony B Felony C Felony D Felony E Felony

1st Comparison ATI Probation % % 0.2 0.2 61.4 60.9 19.9 20.4 14.8 12.9 3.6 5.6

2nd Comparison ATI Jail % % 0.4 0.4 60.5 59.4 19.6 20.1 15.6 14.0 4.0 6.2

3rd Comparison ATI Prison % % 0.5 0.5 62.2 62.7 19.5 19.6 14.3 14.2 3.5 3.1

Charge type Harm to persons â&#x20AC;Ś and property Weapon Property crime Drug Theft intangible Obstructing justice

% 5.6 39.8 5.5 3.6 44.8 0.2 0.5

% 5.5 39.9 5.5 3.6 44.8 0.2 0.5

% 5.1 36.9 6.2 4.0 47.3 0.1 0.5

% 4.7 36.9 6.2 4.1 47.2 0.5 0.5

% 7.1 40.1 5.2 3.5 43.5 0.2 0.4

% 7.6 42.3 4.2 3.0 42.8 0.2 0.0

Borough Brooklyn Manhattan Queens Bronx

% 25.9 30.2 13.5 30.4

% 25.7 30.2 13.7 30.4

% 26.6 28.5 13.2 31.7

% 27.4 29.0 13.0 30.5

% 26.4 30.6 13.0 30.0

% 26.2 30.9 13.0 29.9

% Detained at arraignment

85.6

86.2

85.7

86.2

86.9

Prior criminal history First adult arrest Prior adult arrest(s) One prior misd.conviction 2+ prior misd.convictions One prior felony conviction 2+ prior felony convictions

% 39.0 43.3 4.9 6.9 3.7 2.2

% 39.2 43.1 4.9 6.9 3.7 2.2

% 32.5 47.0 5.8 8.0 3.9 2.8

% 37.5 43.3 4.8 7.4 4.3 2.6

% 37.0 45.4 3.9 6.8 4.3 2.6

Age 14-15 16-17 18-19 20-24 25-39 40+

% 10.3 34.6 19.8 11.9 14.6 8.8

% 7.9 36.9 20.0 11.5 15.0 8.7

% 6.0 34.8 21.3 14.0 15.0 8.8

% 1.5 35.7 24.5 13.3 17.9 7.1

% 10.8 33.2 19.9 12.7 14.6 8.8

% 7.0 38.8 19.3 12.6 14.6 7.7

% Women Total number of cases

14.0 943

7.5 827

14.6 970

14.5 970

Both of these situations involve age at arrest as the matching characteristic: Compared to cases discharged from jail or prison, the ATI participants are more likely to have cases arrested at the very youngest age category (14-15). This difference is particularly pronounced in the jail comparison.

With sample sizes of this magnitude, relatively trivial differences between groups survive the standard criteria of statistical significance (i.e., are deemed as an accurate reflection of the differences in the population from which the sample has been drawn). In light of this fact, absence of statistically significant differences can be viewed as an indication of a â&#x20AC;&#x153;practicallyâ&#x20AC;? exact match between the two groups. According to this argument, these comparison groups constitute an exact match with their respective ATI samples, with the exception of age in the jail and prison samples. The lack of precise age match in these two comparison groups will be taken into account in the analysis.

Tracking recidivism

Time at risk and the logic of post-program tracking The focus of the study is on post-program recidivism: what is the rate of recidivism among those who have exited an ATI program compared to a similar group of felony offenders who have experienced a different treatment? To address this question, we start tracking the participant sample upon program exit, and compare their recidivism record to matched samples of cases released into the community during the same period. In the case of the prison and jail comparison groups, the tracking starts from the date of discharge from incarceration. In the probation sample, the tracking begins on the sentence date.

The selection of the specific follow-up period is a product of two conflicting aims: length of tracking and the recentness of the data. On the one hand, it would be desirable to track recidivism over a long period of time, at least a full year. The longer the span, the more confident we can be about any pattern of differences. On the other hand, if the minimum follow-up period is set at 365 days, the most recent data point in the sample would have to be a case that exited one year earlier; and because one must allow for additional time for these cases to be processed through the NYC criminal justice system, the most recent data point in the study would be dated by roughly one and a half years. As one important consequence, the ATI sample of such description would have very few cases from the post-CCSS period of the ATI operations.

As a compromise between these two goals, the ATI sample consists of cases that have exited their program(s) between July 1, 1998 and March 31, 2000, and the post-program recidivism of this sample is tracked until September 30, 2000. Thus, in theory, the possible length of the follow-up period in this study ranges from 6 to 27 months. However, to reduce heterogeneity in the sample, no case will be tracked for longer than one year.

The post-program tracking of the ATI cases is complicated by the fact that some of them are incarcerated upon program exit. For example, they may have been rearrested while in the program and then terminated from the program while remaining in detention. To the extent that such cases are eventually discharged into the community in time to be at risk of post-program recidivism, they will be included in the ATI sample. However, we will not start tracking their recidivism until they have been released from DOC (or equivalent juvenile) custody. On the other hand, a large number of the ATI cases have not been sentenced by the time of program exit. Some of these cases may be sentenced to incarceration during the follow-up period, in which case we stop tracking their recidivism as soon as they are incarcerated (if the total length of their incarceration during the tracking period exceeds 30 days).

The cases in the probation sample are tracked starting from their sentence date and the cases in the jail and prison samples are tracked from the date of discharge. In both situations, the cases are tracked for up to one year but no

further than September 30, 2000. As reported in Table 3-1, the average length of the time at risk varies from 344 to 332 days across the four categories of penal intervention. The average length in the ATI sample is 340 days, which is only one day above the mean for the entire sample. The median length of the tracking period is 365 days in each group; over 70 % of the cases in the study are tracked for a full year.

Table 3-1: Length of Tracking in the ATI Sample and Its Matched Comparison Groups. Mean 340

St. Dev. 50

Median 365

N 1005

Probation

332

365

943

State Prison

340

365

970

City Jail

344

365

827

Total

339

365

3745

ATI

A standardized period at risk is an important aspect of this study: each group is tracked for an equal length of time during the same period in the same geographic area. This design strengthens the quasi-experimental nature of the research methodology.

Measures of criminal recidivism Recidivism refers to criminal offending following a penal intervention. In this study, criminal offending is measured in terms of prevalence, incidence, and timing. Prevalence refers to the static aspect of an outcome. For example, the percentage of people with AIDS in the population indicates the prevalence of this

disease. On the other hand, the number of new cases diagnosed with AIDS within a given year indicates the incidence of this disease in the population. In the case of recidivism, prevalence refers to the number of cases that recidivate (at least once); the incidence of recidivism refers to the number of such incidents. A population has a low prevalence of recidivism if only one person recidivates; if the person recidivates 100 times, this population may have a high incidence of recidivism. In the context of longitudinal research, recidivism can also be studied from a dynamic perspective: how long does it take before a subject recidivates? The specific methods involved in the analysis of prevalence, incidence, and the timing of recidivism will be addressed later.

The analysis of criminal recidivism will take into account the type and severity of the offense. The categories of offense type include violent, drug, and property crime. The degrees of severity vary between felony, misdemeanor, and infraction. Due to data constraints, the research is limited to those crimes that come to the attention of the New York City criminal justice system. Failing to capture offenses that are not recorded or take place outside the jurisdiction of NYC, these statistics are bound to underestimate the true level of recidivism. However, it is unlikely that this source of error varies between the ATI participant sample and its comparison groups. Moreover, given that these programs are funded by the City of New York, it is clear that the focus of the study should be on their impact on criminal activity in New York City. The recidivism data will be collected at four levels of case processing: arrest, Criminal Court arraignment,

Supreme Court indictment (if applicable), and conviction. Only the top charge at each level will be included in the dependent variable.

Our ability to track the cases in the sample for rearrests in New York City is based on the NYSIID code associated with the defendant’s instant case.6 This method is problematic for tracking the recidivism among Juvenile Offenders (JO’s) for two reasons. First, to the extent that the rearrest will be prosecuted in Family Court as a Juvenile Delinquency case, it will not be recorded in the UDIIS, which is our only source of rearrest data. Second, the NYSIID code associated with a JO case does not carry over to arrests occurring at age 16 or older. Thus, to the extent that an individual charged as a JO on the instant case turns 16 during the tracking period, we will not be able to link the JO NYSIID code to any rearrest, as this will be associated with a new, adult NYSIID code. As a result, our measures of recidivism underestimate the number of rearrests among Juvenile Offenders. The Youth Advocacy Project (YAP), the ATI program for JO cases, is the program most affected by this problem. We should expect the rate of recidivism captured by our measures to be exceptionally low among YAP participants as well as among cases discharged from an OCFS facility (JO equivalent of state prison). We do have the adult NYSIID code for those YAP participants who were assigned one due to an in-program rearrest. However, there are only 15 such cases in the sample.

New York State Identification and Intelligence Division (NYSIID) maintains a file, also known as a “rap sheet”, of the criminal history of people who have been arrested in New York State.

Finally, as a rule, recidivism will be treated as a terminal event that removes the subject from the risk of future recidivism. Thus the focus of the analysis will be on the first arrest. To be sure, a number of the subjects who recidivate early in the tracking period are likely to do it more than once. Thus, treating recidivism as a terminal event underestimates the true volume of the phenomenon. However, because detention is one possible consequence of recidivism â&#x20AC;&#x201C; especially the kind of recidivism that results in an arrest, arraignment, and conviction â&#x20AC;&#x201C; it would be analytically burdensome to treat it as a repeatable event. This would require calculating the time at risk for each subject in a way that adjusts for time spent in detention. Although there is nothing wrong with this approach in principle, it is less than optimal in practice. First, this strategy would have required additional data requests and other potentially time-consuming procedures that would not have been realistic within the project schedule. Second, since recidivism is measured uniformly across each subject category, the focus on first instances will not bias the pattern findings; it will not under- or overestimate the effect of ATI participation on recidivism. Third, to be on the safe side, we do report data on the total number of arrests and convictions in the contexts of basic comparisons.

The analysis of the timing of recidivism will focus exclusively on the incident of the first arrest, specifically, the number of days between the onset of the time at risk and the incidence of the first arrest. On the basis of this information, we

calculate two measures, survival rate (for descriptive analyses) and hazard ratio (for multivariate analyses) to capture group-level differences in the timing of recidivism. Each of these two measures is described in detail in the following section, which also provides empirical illustrations of each type of recidivism measure discussed in this chapter.

Methods of analysis Basic comparisons The presence of individually matched comparison groups secures relatively firm results by means of fairly simple descriptive statistics. The analyses of prevalence will be conducted with cross-tables and the robustness of the findings will be evaluated in terms of the Ď&#x2021;2 -test of statistical significance. The incidence data will be analyzed by comparing the mean scores using F-tests of statistical significance. In both situations, the basic comparisons report the values of the recidivism measures in the ATI treatment group and in its matched comparison groups. This basic mode of analysis will then be elaborated by conducting matched comparisons within selected sub-categories of ATI participants. Table 3-2 below provides gender-specific descriptive statistics on the prevalence and incidence measures of recidivism for the entire study sample, pooling ATI cases with the cases in each comparison group.

The first row of statistics indicates that men are almost twice as likely as women to have been rearrested during the tracking period. This difference persists through each level of case processing, arraignment and conviction. Slightly over

Table 3-2: Prevalence and Incidence Measures of Recidivism by Gender. ATI and Comparison Samples Combined. 1. Prevalence measures % Arrested % Arraigned on 1st arrest % Convicted on 1st arrest Severity of first arrest % A-B Felony % C-E Felony % A Misdemeanor % B Misdemeanor % Other/unknown Type of first arrest % Harm to persons % … and property % Drug % Property crime % Weapon % Other/unknown 2. Incidence measures Average number of … arrests … arraignments (CC) … convictions in CC … indictments (SC) … convictions in SC N

Men 46.1 44.0 29.3

Women 23.9 22.4 15.9

12.0 10.1 13.8 9.2 0.7

3.9 1.1 13.8 4.3 0.9

5.5 4.6 20.9 7.9 2.0 5.0

1.9 0.2 10.8 4.5 0.4 6.0

0.81 0.77 0.41 0.16 0.13 3281

0.35 0.33 0.22 0.00 0.00 464

65 percent of the first arrests resulted in a conviction in either Criminal or Supreme Court. (As some of these cases may still have been pending at the time

these data were collected, this not an accurate estimate of the true conviction rate.) The data on charge severity indicates that most of the gender difference in recidivism stems from the higher prevalence of felony arrests among men. Women are almost as likely as men to have been arrested with a lesser charge. Drug charge is the most prevalent type of charge associated with the first arrest of both men and women. The incidence measures reported in Table 3-2 describe the average number of incidents at different stages of case processing. Attending to the total number of incidents amplifies the magnitude of the gender difference in recidivism to some degree. The average frequency of arrests among men is slightly more than twice as high as among women. This implies that a higher proportion of men than women in the sample were arrested at least twice.

Descriptive (bivariate) analyses of the timing of recidivism will be conducted by graphing the estimates of the Kaplan-Meier survivor rates, which indicate the probability of â&#x20AC;&#x153;surviving the risk of recidivismâ&#x20AC;? beyond a given point in time (Hosmer and Lemeshow 1999, 27-72). Let us assume that we are tracking recidivism over 5 time points and that the rates of recidivism and survival during this period are as follows: At time ... 1 2 3 4 5

... this many recidivated 10 4 6 5 10

... out of this many at risk 100 90 86 80 75

What is the probability of surviving recidivism beyond the first time point? 10 out 100 did not survive, which means that 90 percent did. Correspondingly, as 4 out of 90 failed to survive beyond the second time point, the probability of survival is 86/90 = .96. This latter probability is called conditional, because it assumes that one has survived beyond the first time point. The unconditional probability of surviving beyond the second time point is the combined probability of surviving beyond the first and the second points in time: .90 x .96 = 86%. The KaplanMeier estimates of the survivor function (S) for these data are displayed below:

Time Recidivated 1 10 2 4 3 6 4 5 5 10

Risk Set 100 90 86 80 75

P .90 .96 .93 .94 .87

S .90 .86 .80 .75 .65

The statistical significance of the difference between the survival curves associated with the ATI treatment groups and their matched comparison groups will be tested by the Breslow (generalized Wilcoxon) test of the equality of the estimated survival functions. This test is based on a Ď&#x2021;2 -statistic.

As an illustration of the survival curve, Figure 3-1 (below) plots the Kaplan-Meier estimates of the first arrest for men and women in the pooled sample. Indicating the probability of surviving the risk of re-arrest during the full tracking period, the end points of these curves correspond to the prevalence measure reported in Table 3-2. By definition, both men and women start with an untarnished record: the probability of â&#x20AC;&#x153;surviving the risk of recidivismâ&#x20AC;? is 1.00 at the beginning of

tracking. After the first 22 days, 5 percent of both men and women have failed, i.e., the probability of survival has dropped to 95 %. Following this short period, the two curves begin to diverge significantly. At 200 days, the probability of survival is about .85 among women and only .65 among men. Figure 3-1: Kaplan-Meier Survival Estimates for the Timing of First Arrest by Gender. ATI and Comparison Groups Combined.

K-M Estimate

0.9

Men Women 0.7

85 10 6 12 7 14 8 16 9 19 0 21 1 23 2 25 3 27 4 29 5 31 6 33 7 35 8

1 22 43

0.5

Days

Multivariate analysis As argued above, given the level of detail exercised in the matching process, we should be able to associate a great deal of confidence in the findings generated by descriptive analyses. However, a number of research interests will be served best with the more powerful tools of multivariate modeling.

First, in order to arrive at a limited number of matching categories, variables used in matching were not measured with maximum specificity. For example, continuous variables, such as age, were collapsed into larger units. Although it is not very likely that the level of measurement will have significant impact on the pattern of findings, the ability to handle more precise measures is nevertheless one advantage with multivariate estimation. Second, and by far a more important aspect of multivariate modeling is that, unlike descriptive comparisons, it yields information about the impact on recidivism of all the variables included in the analysis. While the focus of the research is on the effects of ATI participation, it would be wasteful to ignore any results we may generate about other correlates of recidivism. Third, to the extent that the matching process has failed to include all the salient factors that need to be held constant in the estimation of the true ATI effect on post-program recidivism, it will be necessary to resort to multivariate analysis.

To avoid redundancy in the analysis, the multivariate modeling will be limited to the longitudinal measures of recidivism. In other words, the multivariate analysis of recidivism will be conducted from the perspective of timing but not of incidence or prevalence. (Subsuming prevalence, the longitudinal perspective captures the phenomenon of recidivism in a more complete manner.) Cox proportional hazards regression is the method of choice for estimating multivariate models of this description (Hosmer and Lemeshow 1999, 87-112). Designed for the

analysis of discrete event data over time, it can be characterized as a longitudinal version of logistic regression.

The concept of hazard rate is key to interpreting findings generated by this technique. Basically a mirror image of the survival rate, it indicates the probability of experiencing an event, such as recidivism, during a given time interval. Cox proportional hazards regression yields coefficients expressed in terms of hazard ratios. For example, if the hazard rate of recidivism is .23 among men and .11 among women, the hazard ratio for men is .23 /.11 = 2.09, which means that men are over two times more likely to recidivate than women. Unlike standard regression, this procedure does not require any assumptions about the shape of the hazard curve over time. However, the basic Cox model does assume that it is the same for each subject in the analysis. If necessary, this assumption can be modified by conducting stratified analyses.

Findings

4.1

Basic comparisons

The first part of this chapter reports findings from basic (non-multivariate) comparisons between the ATI sample and its matched comparison groups. This part attends to the full spectrum of the recidivism measures described in the previous chapter: charge type, charge severity, dimension (prevalence, incidence, and timing), and the level of case processing. This section will also feature some disaggregated analyses, including program-specific comparisons.

Table 4-1 below reports comparative statistics concerning the prevalence and incidence of rearrests. Note that the ATI sample is compared separately to each â&#x20AC;&#x153;control groupâ&#x20AC;?. As explained earlier, this stems from the fact that none of the three comparison groups feature an acceptable match to each ATI case in the study sample. For example, there are 970 cases in the ATI sample used in the prison comparison because 35 ATI cases did not have an acceptable match in the prison sample.

The first set of recidivism measures in Table 4-1 describes the prevalence of rearrests: the percentage ever arrested during the tracking period and selected characteristics of the first rearrest. As indicated by the first panel of comparisons, 40.6 % among the ATI participants and 41.6 % of the matched sample of probationers were rearrested at least once. This one percentage point difference

* p ≤ .05; ** p ≤ .01; ns = statistically not significant

violent crime includes robbery and assault

2. Incidence (mean scores) Average number of … Rearrests Arraigned rearrests Indicted rearrests Convictions on a rearrest Total number of cases

ATI 0.65 0.61 0.13 0.42 943

Recidivism measure 1. Prevalence (%) Rearrested Arraigned on first rearrest Convicted on first rearrest First rearrest felony First rearrest violent crime a First rearrest drug charge Probation 0.69 0.65 0.13 0.44 943

1st Comparison ATI Probation % % 40.6 41.6 38.5 39.6 25.0 25.3 20.6 18.6 13.9 12.0 16.4 17.4 p ns ns ns ns

p ns ns ns ns ns ns

ATI 0.73 0.68 0.15 0.47 827

Jail 1.08 1.05 0.21 0.79 827

2nd Comparison ATI Jail % % 44.4 52.8 42.0 52.0 27.6 37.1 23.2 22.2 15.6 12.1 18.0 25.5 p ** ** ** **

p ** ** ** ns * ** ATI 0.65 0.61 0.13 0.42 970

Prison 0.64 0.60 0.12 0.42 970

3rd Comparison ATI Prison % % 40.3 40.1 37.6 37.6 24.3 24.8 21.0 18.5 14.1 9.6 16.3 20.6 p ns ns ns ns

p ns ns ns ns ** *

Table 4-1: The Prevalence and Incidence of Recidivism in the ATI Sample and its Matched Comparison Groups.

is not even close to being statistically significant. In other words, there are no differences in the prevalence of rearrests between the ATI participants and the probation group. This fact does not change with the level of case processing: In both groups, 2 percentage points of the initial rearrest rate is shaved off at arraignment; 25 percent of the cases in each group are convicted as a result of the first arrest. A somewhat higher share of the ATI cases are charged with a felony but this is not statistically significant either. Finally, no statistically significant differences emerge in comparisons concerning the type of the arrest charge.

The second set of recidivism measures in Table 4-1 has to do with the frequency, or incidence, of rearrests. Attending to this dimension of recidivism does not add to the pattern generated by the analysis of prevalence. There are no differences in recidivism between the ATI participants and their matched equivalents in the probation sample. With one minor qualification, the same holds true for the prison comparisons, the last panel of Table 4-1. In the ATI sample, the first arrest is more likely to be a violent crime (robbery or assault), whereas the cases discharged from prison are more likely to be charged with a drug offense. However, these differences are not particularly large, and, more critically, there are no differences whatsoever in the overall prevalence or incidence of rearrests between these two groups.

By contrast, the ATI graduates fare much better in comparison to their counterparts in the jail sample. Cases sentenced to jail are 10 percentage points more likely to have been rearrested, arraigned, and convicted. Although there is no difference in the prevalence of a felony charge in the first arrest, the incidence of arrests leading to an indictment in Supreme Court is significantly higher in the jail comparison group. Overall, felons discharged from the City jail are rearrested 1 Â˝ times as frequently as the ATI cases.

In light of the findings reported in Table 4-1, the recidivism behavior of the ATI participants does not differ from that of cases sentenced to prison or probation. On the other hand, both the prevalence and incidence of recidivism is significantly higher among those who have served a jail sentence. A careful reader may recall that the jail comparison group features a particularly small number of individuals in the very youngest age category, 14 to 15; the proportion of cases in these ages is four times higher in the corresponding ATI sample. Could this explain the major differences that emerged in this comparison?

The impact of age composition on recidivism will be addressed more fully later in the context of multivariate analysis. However, as a preliminary step to that direction, we compared the basic prevalence and incidence measures between the ATI and jail groups without the cases in the youngest age category. This modification to the data had no effect on the pattern of findings. We observed statistically significant differences of the same magnitude in favor of the ATI

sample. It appears that the shortage of cases from the very youngest age group in the jail sample does not explain their higher rates of recidivism.

Results concerning the timing of recidivism are reported in Figure 4-1 (next page). The Kaplan-Meier estimates portrayed in these charts describe the cumulative probabilities of surviving the risk of recidivism during the tracking period. By definition, each category of cases starts with a 100 percent survival rate. The deeper the descent of the slope, the quicker the rate of recidivism, and the lower it goes, the higher the total rate (prevalence) of recidivism. As indicated by Charts a and b in Figure 4-1, the survival curves of the probation and prison comparison groups are practically indistinguishable from those of the respective ATI treatment groups. This visual impression is confirmed by the fact that, in both comparisons, the difference between the two curves is not statistically significant. In the case of probation-comparison, the two lines are virtually identical; the prison group is more different from the ATI participants.

Consistent with the results based on the two static measures of recidivism (incidence and prevalence), the jail-comparison is the only one yielding statistically significant findings (Chart c). Interestingly, the pace of recidivism is very similar between the two groups during the first 100 days of tracking. The major difference between the ATI and the jail cases emerge between 120 and 280 days. The jail group continues to recidivate at the same high rate up to day 200 (approximately). In the ATI sample, the slope starts to level off much earlier.

0.25

0.35

0.45

0.55

0.25

0.35

0.45

0.55

0.65

0.75

0.85

0.95

c) Jail comparison (p = .01)

Days

0.65

0 24

K-M Estimate

Days

Jail ATI

Probation ATI

0.25

0.35

0.45

0.55

0.65

0.75

0.85

0.95

K-M Estimate

b) Prison comparison (p = .15)

Days

0 12

0.75

0 18

0.85

0 24

0.95

0 30

24 0

12 0

0 30

a) Probation comparison (p = .83)

36 0

Prison ATI

Figure 4-1: Timing of First Arrest in the ATI sample and its Matched Comparison Groups. Kaplan-Meier Survival Estimates.

0 36 30 0

18 0

0 36

Towards the end of the tracking period (280-300 days), the two groups return to equal rates of recidivism. However, due to a longer period of sharp decline, the jail sample ends up with a significantly lower cumulative survival rate.

Each of the three dimensions of recidivism â&#x20AC;&#x201C; prevalence, incidence, and timing â&#x20AC;&#x201C; point to the same direction: Cases discharged from jail do significantly worse than cases that have participated in an ATI program. On the other hand, ATI participants do not recidivate any more or any less than probationers and cases released from a state prison.

Does success in the program matter? The sample of ATI participants included in the above comparisons includes all cases that exited from these programs and remained in the community. This selection includes cases that did not meet the criteria of successful completion, e.g., cases that had a poor attendance record and cases that did not comply with the treatment plan. To the extent that program participation is expected to have an impact on recidivism, it makes sense to distinguish between those who are deemed as successful graduates by the programs and those who failed. As reported in Table 1-1, 37 % of the ATI cases in the sample were labeled as unsuccessful. To examine the impact of the quality of program participation, we compared the recidivism of the successful ATI participants to the matched samples of cases from each of the three comparison groups. The results from these analyses are reported below (Table 4-2). Note that, in these comparisons,

* p ≤ .05; ** p ≤ .01; ns = statistically not significant

violent crime includes robbery and assault

2. Incidence (mean scores) Average number of … Rearrests Indicted rearrests Total number of cases

ATI 0.64 0.13 596

Recidivism measure 1. Prevalence (%) Rearrested First rearrest felony First rearrest violent crime First rearrest drug charge p ns ns ns ns

Probation p 0.68 ns 0.13 ns 595

1st Comparison ATI Probation % % 38.9 41.0 17.8 18.3 12.6 11.1 16.4 18.3

ATI 0.73 0.15 512

Jail 1.03 0.20 511

2nd Comparison ATI Jail % % 43.4 50.9 20.5 21.3 14.1 10.4 19.1 25.2

p ** ns

p * ns ns *

ATI 0.64 0.14 610

Prison 0.61 0.11 610

3rd Comparison ATI Prison % % 38.7 38.4 18.5 16.6 12.8 8.9 16.2 19.7

p ns ns

p ns ns * ns

Table 4-2: The Prevalence and Incidence of Recidivism among Successful ATI Graduates and Their Matched Comparison Cases.

the composition of the comparison groups is also different; they include only those cases that match the â&#x20AC;&#x153;successfulâ&#x20AC;? ATI cases.

The findings from these successful-only comparisons are remarkably similar to the findings based on the full samples. The jail comparison remains the only context with statistically significant differences in key measures of recidivism. Due to the loss in sample size, some differences regarding the type and severity of rearrest charges are no longer statistically significant. However, in terms of substance, Table 4-2 replicates the findings from Table 4-1. These findings imply that success in the program does not matter to the impact of ATI participation on criminal recidivism. This statement does not mean that those ATI participants who complete the program successfully are as likely to recidivate as those who are terminated unsuccessfully. However to study the effect of program success, one must compare successful ATI graduates to similar cases in the comparison groups, not to dissimilar cases in the ATI sample. As it happens, successful graduates are somewhat less likely to have been arrested during the tracking period: 39 % of them were arrested compared to 43 % among the unsuccessful exits. However, this difference is not statistically significant. This potentially surprising finding may have to do with the fact that the least successful ATI participants are not included in the recidivism study because they were incarcerated, either because of in-program rearrest or some other major breach of a program requirement. Thus, the group of unsuccessful ATI graduates

retained in the recidivism sample may represent a relatively â&#x20AC;&#x153;successfulâ&#x20AC;? segment of the unsuccessful category.

Program-specific comparisons In the above analyses, the seven ATI programs were treated as a single group. Although the resulting findings provide a very good idea about the aggregate effects of these programs, they may also hide important differences between the programs. In what follows, we report results from program-specific comparisons. As before, these comparisons are conducted separately for each comparison group (probation, prison, and jail). In addition, each comparison-group-specific setting is divided into program-specific sub-samples. For example, the first set of program-specific comparisons features cases from the ATI and the probation samples; the comparisons concerning CEP include all CEP cases from the ATI sample and all those cases from the probation sample that match CEP cases. The program-specific comparisons are limited to two dimensions of recidivism, prevalence and incidence. (The multivariate analyses will feature programspecific analyses of the timing of recidivism.)

The program-specific results concerning the prevalence of recidivism are displayed in Figure 4-2 (below). The first chart (a) in Figure 4-2 involves comparisons with the probation sample. CEP, DAMAS, and FlameTree are associated with a higher prevalence of rearrest than the matched group of

Figure 4-2: Program-Specific Comparisons of the Prevalence of Rearrest. a) Probation Percent rearrested

Context of comparison

CEP DAMAS El Rio

Probation ATI

Flametree Freedom Project Return* YAP

b) Prison Percent rearrested 0

Context of comparison

CEP* DAMAS El Rio

Prison

Flametree

ATI

Freedom Project Return* YAP*

c) Jail Percent rearrested 0

Context of comparison

CEP DAMAS El Rio Flametree*

Jail ATI

Freedom Project Return ^ YAP*

* p â&#x2030;¤ .10 ; ^ N < 50

probationers. However, none of these differences is statistically significant.7 Project Return, a women’s drug treatment program, is the only program that is statistically different from its counterparts in the probation sample. Project Return participants are 2.4 times less likely to have been rearrested. Moving to the next chart (b – prison comparison), we observe the same effect in association with Project Return. In addition, the four percentage points’ difference to the advantage of YAP and the rather trivial (1.7 points) difference against CEP are also statistically significant. YAP is a 12-month program that offers services to felony offenders at ages 14-15; CEP is a 6-month program providing educational and vocational services to clients at ages 16-19. In the jail comparisons, only two programs are different from their comparison group with statistical significance. The participants in FlameTree, a drug program for the general adult population, recidivate more than their counterparts in the jail sample, whereas YAP cases recidivate less. Note that the relative magnitude of the difference in the comparison involving Project Return is the same as before. However, since this sub-sample involves only 48 cases altogether, this difference does not meet the 10 percent criterion of statistical significance. There simply are not enough cases in the jail sample to match the client base of Project Return; a few more cases would have resulted in a statistically significant effect.

Figure 4-3 (below) presents a parallel set of findings concerning the average frequency (incidence) of rearrests. In the probation comparisons (Chart a), there

Given the reduced size of the resulting samples, we apply a lower standard of statistical significance (p LQ WKH program-specific analyses. The number of cases featured in each comparison is reported in Appendix C. 7

Figure 4-3: Program-Specific Comparisons of the Incidence of Rearrest. a) Probation Mean number of rearrests

Context of comparison

0.2

0.4

0.6

0.8

1.2

CEP DAMAS El Rio

Probation

Flametree

ATI

Freedom Project Return YAP

b) Prison Mean number of rearrests 0

0.2

0.4

0.6

0.8

1.2

Context of comparison

CEP* DAMAS El Rio

Prison

Flametree

ATI

Freedom Project Return YAP*

c) Jail Mean number of rearrests

Context of comparison

0.2

0.4

0.6

0.8

1.2

1.4

CEP* DAMAS El Rio* Flametree* Freedom*

Jail ATI

Project Return ^ YAP*

* p â&#x2030;¤ .10 ; ^ N < 50

are no statistically significant program-specific differences in the incidence of rearrest. Compared to their matched prison samples (Chart b), CEP cases are rearrested more frequently and YAP graduates less frequently. These are the only statistically significant results from this chart. Finally, cases discharged from jail are associated with higher incidence of rearrest than cases from any of the seven ATI programs. With the exception of DAMAS and Project Return, all of these differences are statistically significant. The main implication of this last set of comparisons must be that the jail comparison group features an exceptionally large number of cases that were rearrested more than once.

In general, the program-specific comparisons sustain the basic pattern of findings based on the aggregate analysis. However, these data do reveal two important results that were previously undetected. First, Project Return appears to be very effective in reducing the rate of recidivism among its participants. This effect prevails regardless of the comparison group. Second, participation in YAP results in lower levels of recidivism than a sentence of incarceration, prison or jail. This finding holds across both dimensions of recidivism, prevalence and incidence. Program-specific comparisons concerning the type and severity of arrest charge are presented in Appendix C (Table C-1).

YAP is an ATI program that serves Juvenile Offenders (ages 14-15). As reported in Table 2-1, both jail and prison samples have significantly fewer cases in this age group than the corresponding ATI samples. In the course of matching, cases

that did not have a match in the same age category were assigned a match from the next age category, either below or above. In the case of the youngest age group, the next category is always the one above. Although this may not seem like a major distortion in age composition, it may have caused serious bias in the measurement of recidivism. As explained in Section 3, based on NYSIID link in the UDIIS database, our rearrest data fail to capture a significant share of rearrests by cases with a JO status on the instant case. As there are fewer such cases in the comparison groups, it is likely that the superior performance by YAP may reflect a systematic error in our measures of rearrest. To study this and other more subtle aspects of these data, we turn to multivariate analysis.

4.1

Multivariate models

As explained in the section on methodology, multivariate modeling serves several purposes in this study. First, It will improve the level of statistical control in two ways: (1) modeling all the matching characteristics as independent variables will correct any imperfections in the outcome of the matching procedure, such as the lack of precise age match in the comparison groups; (2) we are now able to introduce additional control variables ignored in the matching grid. Second, by virtue of estimating the effects of a large number of independent variables, multivariate models yield information about other predictors of recidivism than ATI participation and other penal interventions. Third, in the presence of sufficient controls, it is now possible to include those ATI cases that did not have an acceptable match in a comparison group. Fourth, the multivariate approach allows us to compare recidivism rates across each penal intervention

type in a single model. In other words, we can conduct global comparisons instead of the group-by-group comparisons characterizing the previous chapter.

As discussed in Section 2, one downside with multivariate analysis has to do with inefficiency in the face of small samples. This is bound to have an impact on program-specific findings: it is highly unlikely that programs with a small number of participants in the sample (DAMAS, El Rio, and Project Return) will be able to generate statistically significant effects in the presence of a large number of control variables. This fact should be taken into account when interpreting the results from the multivariate models.

As explained in Section 3, the multivariate models estimated in this study focus on the timing of first arrest. Specifically, the dependent variable is the hazard of rearrest, which indicates a group-level probability for an event to occur during a given time interval. The regression coefficients are expressed in terms of hazard ratios: the size of the hazard (probability) relative to the size of the hazard in some reference group with a fixed hazard of 1.00. For example, if women are treated as a reference category, their hazard ratio of recidivism is 1.00, by definition. Given that men are more likely than women to recidivate, the hazard ratio for men is likely to be higher than 1.00. If it is 2.00, men are twice as likely to recidivate. Any value below 1.00 implies that the risk of recidivism in that group is lower than in the reference category. The lowest possible hazard ratio is zero.

Variables We begin the multivariate analysis by modeling all the variables used in matching as control variables. The main independent variable is the nature of penal intervention, which consists of four basic categories: prison, jail, probation, and ATI. In program-specific analyses, this variable features eleven categories: seven ATI programs, probation, jail, and two prison groups, adult and juvenile (OCFS). (The purpose of separating OCFS into a distinct category is to single out a more valid comparison group for YAP. Serving cases charged as Juvenile Offenders, both of these groups are subject to the same source of error in the tracking of rearrests.8) Models of this description essentially replicate the prior analyses with more sophisticated methods.

As the second step in the modeling procedure, we introduce a new set of control variables in the mix. The purpose of matching was to control for characteristics that were presumed to be of critical importance in order to achieve comparable samples. A particular focus in this effort was to make sure that the cases in the comparison groups have a similar offender profile as the cases in the ATI sample, i.e., that we are dealing with comparable criminal justice client populations. Attending to just gender and age, the matching grid is less comprehensive when it comes to socio-demographic background variables. Four of the six control variables introduced in this section were chosen to supplement this dimension of measures. To tap the economic status of the subjects, we rely

See page 27 for details.

on two indicators. Indicating full-time participation in the labor market or schooling, first of these variables measures ties to the conventional economy. Under the assumption that being represented by a private attorney (other than a court-appointed one) correlates positively with access to economic resources, we use this information as the second measure of economic status.

As reported in Section 1, 95 % of the cases in the ATI sample are either AfricanAmerican or Hispanic. To attend to the possible effects of ethnic composition, we will include it as another new control variable. Dictated by both policy needs and data limitations, this study is concerned only with rearrests that occur within the jurisdiction of New York City. As the likelihood that a subject will recidivate in NYC is likely to depend on his or her region of residence, we will control for this variable as well. Based on data collected at the time of the arrest on the instant case, this variable is less than perfect measure of the residential situation during the tracking period, especially among those released from prison.

In addition to the socio-economic variables described above, we include one new indicator of criminal justice system status. A careful reader may recall a footnote in Section 2 explaining the absence of Youthful Offenders in the (adult) prison sample. Given that all of the variables influencing the YO-status were included as matching characteristics, this was not deemed as a fatal flaw. However, to be quite accurate, one relevant variable was not taken into account in matching: in order to be eligible for YO treatment in sentencing, one must not have been

given such treatment previously. To correct this omission in the research, we model previous YO-status as a control variable. Finally, it is possible for a person to be arrested for an incident that took place a long time ago. Specifically, it is possible that some of the first rearrests captured in our data reflect incidents that occurred before the beginning of tracking. Because the CJA database does not record the date of the incident for arrests that occurred prior to November 1998, we were not able use this information for each rearrest. However, for most first rearrests in the data, we are able to identify cases where the incident took place prior to tracking.

Findings The results from the first set of multivariate models are displayed in Table 4-3 below. These two models describe findings from program-specific analyses; the first variable in Table 4-3 is penal intervention, which features eleven categories. Probation is the first category and the reference category for interpreting the hazard ratios. Model 1 in Table 4-3 controls for all the variables used in matching. The new control variables introduced above are added to the analysis in Model 2. Note also that Table 4-3 provides frequencies of each variable included in the analysis. Due to missing data on some of the new control variables, the number of cases drops from 3,720 in Model 1 to 3,619 in Model 2.

From the perspective of this study, the most important findings in Table 4-3 are coefficients associated with the different categories of penal intervention. With

Table 4-3: Cox Proportional Hazard Regression Models of Time to First Rearrest (Hazard Ratios). Independent variables 1. Penal intervention Probation (reference) Prison Jail OCFS CEP DAMAS El Rio FlameTree Freedom Project Return YAP 2. Control variables Sex Women (reference) Men Age 14-15 16-17 18-19 20-24 25-39 40+ (ref.) a Detention status Not detained (ref.) Detained Charge severity a A-B felony C-E felony (ref.) Charge type a Harm to persons (ref.) â&#x20AC;Ś and property Weapon Property crime Drug Other a Prior criminal history No prior arrests (ref.) Prior arrest(s) 1 misd. conviction 2+ misd. convictions 1 felony conviction 2+ felony convictions Borough of prosecution a Brooklyn (reference) Manhattan Queens Bronx

MODEL 1 p = .000 1.00 1.12 1.35 * 0.03 * 1.15 0.99 0.77 1.18 1.01 0.52 0.33 * p = .000 1.00 1.83 p = .000 2.87 2.97 2.41 2.11 1.63 1.00 p = .617 1.00 1.04 p = .536 0.96 1.00 p = .001 1.00 1.86 1.87 1.98 2.04 2.33 p = .000 1.00 1.66 1.75 2.3 2.66 1.82 p = .011 1.00 0.95 1.16 1.14

N 943 883 823 66 350 67 77 136 231 40 104

464 * 3256 * * * * *

313 908 775 799 601 324 497 3223 2298 1422

214 * 1499 * 197 * 129 * 1655 * 26

* * * * *

1374 1651 183 269 150 93 991 1127 486 1116

MODEL 2 p = .000 N 1.00 931 1.11 864 1.34 * 793 0.03 * 60 1.17 341 0.87 66 0.83 73 1.19 134 1.02 225 0.58 36 0.30 * 96 p = .000 1.00 1.78 p = .000 3.01 3.02 2.50 2.19 1.73 1.00 p = .710 1.00 1.03 p = .581 0.96 1.00 p = .002 1.00 1.83 1.90 2.00 2.00 2.31 p = .000 1.00 1.60 1.62 2.15 2.53 1.55 p = .019 1.00 1.01 1.20 1.18

447 * 3172 * * * * *

302 885 767 776 578 311 489 3130 2249 1370

212 * 1445 * 192 * 125 * 1619 * 26

* * * * *

1332 1624 181 252 142 88

961 1098 * 469 * 1091

Table 4-3. (Continued) 2. Control variables (cont.) Incident prior to tracking No (reference) Yes Previous YO status a Yes (reference) No Not recorded a Private attorney No Yes (reference) Ethnicity Black (reference) White Hispanic, White Hispanic, Black Hispanic, other Other At work or school? a Yes, unverified (ref.) Yes, verified No, unverfied No, verified Unresolved Area of residence a NYC (reference) Metro area Outside Metro area N

MODEL 1

3720

MODEL 2 p = .000 1.00 6.92 p = .600 1.00 0.93 0.92 p = .500 0.94 1.00 p = .002 1.00 0.59 0.98 0.75 0.88 0.64 p = .100 1.00 0.96 1.16 1.12 1.07 p = .001 1.00 0.53 0.63

N 3593 * 26 375 2090 1154 3283 336 2001 124 423 * 216 780 * 75 *

877 789 * 1261 552 140

* *

3398 41 180 3619

These variables are characteristics surrounding the situation of the instant offense; they do not refer to a case stemming from a rearrest. * p < .05

probation as the reference category, the substance of these findings must be interpreted accordingly. These models provide two types of information concerning the statistical significance of the estimated effects. First, the p-value of a variable is displayed above the coefficients pertaining to its categories. For example, the p-value associated with the penal treatment variable is .000, which means that it is highly significant (the probability that this pattern of effects does not characterize the population is less than one per 1,000). On the other hand, an asterisk next to a coefficient indicates that this hazard ratio is different from that of the reference category at .05-level of statistical significance. Both statistics are useful for evaluating the statistical significance of the findings. The p-value associated with the variable indicates the extent to which the distinctions implied by this concept matters to recidivism. The statistical significance of the categories of a variable points to the specific values that are most responsible for its effect. For example, the fact that jail, OCFS, and YAP are associated with statistically significant coefficients in Model 1 implies that much of the explanatory power associated with this variable can be reduced to these three categories.

Five of the eleven categories of penal intervention in Model 1 have a hazard ratio higher than 1: prison, jail, CEP, FlameTree, and Freedom. This finding means that the probability of recidivism in these groups is higher than among probationers. However, the jail category is the only one with a statistically significant coefficient.

The only ATI program associated with a statistically significant effect is YAP: the hazard of rearrest among the participants in YAP is almost three times lower than among probationers. However, note that the effect associated with OCFS is ten times stronger than that of YAP. Given that cases discharged from an OCFS facility represent the only comparison group that is subject to the same limitations in the measurement of rearrest data as YAP cases, it cannot be concluded that the level of recidivism among YAP participants is particularly low. In fact, it is quite high in comparison to OCFS. On the other hand, the number of cases in the OCFS category is too low (66) to warrant any firm conclusions. A larger sample of cases and more accurate measures are required in order to adequately address JO-specific questions.

As anticipated, the effect associated with Project Return in Model 1 does not meet a conventional criterion of statistical significance.9 However, given the magnitude of the effect (hazard ratio = .52), the fact that the dummy variable as a whole is statistically significant, the small number of cases in this category, and the pattern of findings based on matched comparisons, we are prone to assign a great deal of confidence in the conclusion that Project Return participants are indeed significantly less likely to recidivate than comparable cases in other categories. Adding a new set of control variables in Model 2 has no impact on the findings concerning the effects associated with different categories of penal intervention. (Although the hazard ratio for Project Return increases somewhat in 9

The actual p-value of this coefficient is .16, which, given the degrees of freedom, might be considered acceptable.

Model 2, due to a further loss in degrees of freedom, this effect becomes statistically even less significant.)

As an auxiliary outcome of multivariate analysis, Table 4-3 contains a wealth of information concerning other determinants of recidivism in this sample. Eight of the thirteen control variables included in the analysis are statistically significant (p

predictors of the risk of rearrest. To start with the more trivial findings, men are close to twice as likely as women to recidivate; subjects in the two youngest age group are almost three times more likely to have been rearrested than those in the oldest category; prior record of arrests and convictions is positively related to the risk of recidivism; and those residing outside New York City are less likely to have been rearrested there.

The variable indicating whether or not the incident leading to a rearrest occurred prior to tracking is statistically the strongest “predictor” of recidivism by far. Those rearrested for an incident that took place prior to the beginning of tracking are nearly 7 times more likely to have been rearrested. The mere reading of the previous sentence should reveal the artificial nature of this “finding”. By definition, the category associated with this tremendously large hazard ratio consists solely of cases that were rearrested. It is therefore not surprising that they are “more likely” to have been rearrested. (The purpose of this variable was to control for possible differences in this aspect of first rearrest across different categories of penal intervention.)

One interesting finding from these analyses has to do with charge type. Cases that were charged with an assault (“harm to persons”) are significantly less likely to recidivate than cases charged with any other offense type. A satisfactory explanation of this finding is beyond the scope of this study. However, it can be speculated that a larger portion of assaults may represent “expressive” or noninstrumental crimes, whereas crimes like robbery and drug offenses may be more instrumental in nature, i.e., means to an end such as supporting a drug habit or making a living. It is possible that the sources of instrumental crime are more fundamentally linked to the social and psychological conditions of the offender, and that these conditions are less likely to change as a result of any form of intervention. A more parsimonious explanation could argue that a disproportionate share of the cases in the violent crime category are JO’s. According to this perspective, the lower hazard of rearrest in this category simply reflects systematic bias in the measurement of recidivism. As an argument against this interpretation, it should be pointed out that these models control for a number of variables tapping JO status (YAP, OCFS, and age).

As the last step in the multivariate modeling, we return to aggregate ATI-wide analysis. First, to verify the somewhat surprising finding concerning the impact of success in the program, we estimated a model that features separate coefficients for those ATI cases that were unsuccessfully terminated (UT) and those that successfully completed (SC) a program. In addition, we estimated another model, which distinguishes between those ATI cases that were mandated into a

program during the centralized (CCSS) regime of the ATI operations in contrast to those who entered the system under current regime (ATIIS). The findings from these two models are summarized in Figure 4-4 below. Figure 4-4: The Effect of Successful vs. Unsuccessful Program Participation (SC/UT) and the Effect of CCSS vs. ATIIS regimes on Rearrest (Hazard Ratios).

0.94

ATI, SC ATI, UT

1.06

1.01

ATI, CCSS 0.81

ATI, ATIIS 0.70

0.90

1.10

1.30

Hazard Ratio Note: both of these comparisons are based on multivariate models featuring the full set of control variables.

As previously, the hazard ratios in Figure 4-4 are calculated with probation as the reference category. First comparison concerns the effect of success in the program. The multivariate analysis confirms the finding reported on the basis of matched comparisons: unsuccessful participants are somewhat more likely to recidivate than successful ones. However, this effect is not statistically significant. Although the magnitude of the difference between CCSS and ATIIS

regimes is larger, neither one of these coefficients is sufficiently different from that of probation to meet an acceptable standard of statistical significance.

Conclusions

The purpose of this study is to evaluate the impact of ATI participation on the level and nature of recidivism following program exit. In order to provide context for the findings, the recidivism outcomes of ATI participants were compared to those of a matched sample of cases from three different categories of penal intervention: probation, jail, and prison. The analyses attended to three dimensions of recidivism: prevalence, incidence, and timing of rearrest. The research featured data at four levels of case processing: arrest, arraignment, indictment, and conviction. The measures of recidivism were also disaggregated by the type and severity of rearrest. We performed both ATI-wide and programspecific comparisons.

The results concerning the recidivism of ATI participants vary sharply depending on the comparison group. There are no significant differences in recidivism between ATI participants and cases sentenced to probation or state prison. On the other hand, ATI participants recidivate significantly less than cases discharged from incarceration served in a New York City jail. Despite some minor variation in program-specific results, none of the seven ATI programs included in the analysis deviates from this basic pattern.10

The results concerning YAP are biased by inadequate measures of rearrest, and the promising findings regarding Project Return need to be replicated with a larger sample of cases.

Thus, the main conclusion of this study is that cases sentenced to jail are more likely to recidivate than ATI participants or cases sentenced to probation or prison. An adequate explanation of this finding is beyond the scope this study, the sole focus of which is on the ATI participants. However, one can speculate some possible explanations that could be addressed in future research. One characteristic that seems to distinguish between jail cases from the three other categories is the absence of post-release supervision. Cases discharged from prison have to report to their parole officer and probationers to the probation officer (probation is also a typical sentence for an ATI participant who remains in the community). This explanation must be qualified by the fact that a number of cases in the jail sample received a split sentence of jail and probation, which means that, following the discharge from jail, they are subjected to similar conditions as probationers. Unfortunately, our data do not allow us to identify the â&#x20AC;&#x153;splitsâ&#x20AC;? in the jail sample. However, the fact remains that cases released from jail are less likely to be under community-based supervision.

Another perspective could argue that the results concerning the inferior performance of the jail intervention is a function of a certain sub-population that either does not exist in or is handled with more success by other interventions. For example, a history of mental health problems is one important characteristic that precludes a defendant from participating in any of the seven ATI programs included in this study. Thus, a jail-bound case that would otherwise qualify for an

ATI treatment will end up in jail. On the other hand, it is possible that state facilities have better resources to treat clients with mental health problems.

Finally, according to another robust finding from this study, cases deemed as successful in the program did not recidivate any less than cases that were unsuccessfully terminated. To be sure, simply by virtue of remaining in the community upon program exit, the sample of unsuccessful ATI cases included in this study represents the â&#x20AC;&#x153;successfulâ&#x20AC;? segment of the unsuccessful population. However, this finding clearly suggests that the qualitatively major difference between success and failure in an ATI program does not translate into a significant difference in recidivism.

APPENDIX A The Seven Alternative-to-Incarceration Felony Programs: •

Court Employment Project (CEP) of CASES was contracted to provide six months of community-based supervision with in-house educational and vocational services for the sixteen to nineteen year-old population likely to receive an incarcerative sentence of more than six months on a felony conviction.

•

Freedom Program of the Fortune Society was contracted to provide six months to a year of community-based supervision with in-house educational and vocational services for the nineteen and older population likely to receive an incarcerative sentence of six months or more on a felony conviction.

•

El Rio Day-Treatment Program (El Rio) of the Osborne Association was contracted to provide six months to a year of day-treatment substance abuse services to the eighteen and older population residing in upper Manhattan and the Bronx and who were likely to receive an incarcerative sentence of six months or more on a felony conviction.

• FlameTree Program of the Fortune Society was contracted to provide six months to a year of day-treatment substance abuse services to the eighteen and older population residents of lower Manhattan, Brooklyn and Queens, and who were likely to receive an incarcerative sentence of six months or more on a felony conviction. •

Youth Advocacy Project (YAP) of the Center for Community Alternatives (CCA) was contracted to provide a year of community-based supervision and other rehabilitative services to the thirteen to fifteen year old population likely to receive an incarcerative sentence of six months or more on a felony conviction.

•

DAMAS Program of the Fortune Society was contracted to provide six months to a year of community-based supervision and other gender-specific services to women likely to receive an incarcerative sentence of six months or more on a felony conviction.

•

Women’s Day Treatment Program of the Project Return Foundation (Project Return) was contracted to provide outpatient substance abuse treatment services for women likely to receive an incarcerative sentence of six months or more on a felony conviction.

Appendix B How the 7-Digit String Matching Variable Was Constructed The information used to match cases is based on the first nine variables described in Table B-1 below. In order to reduce this information into a single matching variable, several of these original variables were first combined and/or recoded. For example, age was grouped into six categories. The resulting new variables are described in the latter part of Table B-1 (variables 10-12). The 7digit variable used in the matching process is based on the following set of variables: sex (1), type and severity (10), prior criminal history (12), age group (11), detention status (3), and the borough of prosecution (9). The resulting string variable combines these six variables; it features 7-digits because â&#x20AC;&#x2DC;type and severityâ&#x20AC;&#x2122; is a two-digit variable and all the others are measured in single digits. Table B-1: Variables used to create the matching string.

Original Variables 1. Sex 2. Age 3. Release status following Criminal Court arraignment 4. Severity of top charge in Supreme Court arraignment 5. CJA Type of top charge in SC Arraignment

Values 1=Male, 2=Female 14 to 66 1=Detained, 0=Not detained

Recoded Variables 10. Type and severity of top charge (4+5)

New Values 1 to 13, representing a specific combination of charge type And severity 1= 14-15, 2=16-17, 3=18-19, 4=20-24, 5=25-39, 6=40+ 1= First arrest, 2=Prior arrest(s), 3=Prior misdemeanor conviction, 4=2+ prior misdemeanor convictions, 5=Prior felony conviction, 6=2+ prior felony convictions

A to E Felony

1= Harm to persons, 2=Harm to persons and property, crime, 3=Weapon, 4=Property, 5=Drug, 6=Theft intangible, 9=Obstructing justice 6. First adult arrest? 1=Yes, 0=No 7. Number of prior misdemeanor convictions 0 to 34 8. Number of prior felony convictions 0 to 4 a 9. Borough of prosecution Bronx, Brooklyn, Manhattan, Queens

11. Age group (2) 12. Prior criminal history (6+7+8)

None of the seven ATI programs have a court presence in Staten Island, which excludes cases from this borough from the study.

28.9 17.3

0.20 329

20.4 18.3

0.19 329

Recidivism measure 1. Prevalence % with 1st arrest … felony charge drug charge 2. Incidence Average number of indicted rearrests N

a) Jail comparisons

Recidivism measure 1. Prevalence % with 1st arrest … felony charge drug charge 2. Incidence Average number of indicted rearrests N 0.21 379

0.12 379

0.22 294

0.30 294

0.10 30

0.01 76

13.2 15.8

0.01 75

13.0 15.6

0.15 73

12.3 23.3

0.01 73

13.7 16.4

EL RIO Prison ATI

0.11 75

18.7 21.3

EL RIO Prison ATI

0.13 76

19.7 23.7

0.15 129

24.8 24.8

0.15 133

25.0 25.0

0.20 131

24.4 24.0

0.15 131

25.2 24.4

Flametree Prison ATI

0.17 133

22.6 24.8

Flametree Prison ATI

0.16 129

23.3 20.9

0.12 218

20.2 19.3

0.13 212

20.8 19.0

0.17 221

20.8 21.7

0.12 221

19.9 19.0

Freedom Prison ATI

0.11 212

18.4 20.3

Freedom Prison ATI

0.11 218

22.0 21.6

0.01 39

2.6 7.7

0.01 38

2.5 7.5

0.00 24

0.0 4.2

0.13 24

4.2 12.5

Project Return Prison ATI

0.00 38

5.3 18.4

Project Return Prison ATI

0.10 39

10.3 23.1

0.00 95

8.4 1.1

0.0 83

9.6 1.0

0.19 54

18.5 20.4

0.01 54

11.3 1.9

YAP Prison ATI

0.14 83

26.5 15.7

YAP Prison ATI

0.00 95

9.4 2.1

Freedom Project Return YAP Probation ATI Probation ATI Probation ATI

Note : comparisons typed in bold face are statistically significant at the 10 % level.

0.10 30

6.7 23.3

31.9 18.0

28.9 30.3

13.3 20.0

DAMAS Prison ATI

0.00 50

CEP Prison ATI

0.01 50

6.0 14.0

30.8 17.5

18.2 21.4

6.0 13.4

DAMAS Prison ATI

0.01 57

7.0 14.0

CEP Prison ATI

0.00 57

3.5 1.8

EL RIO CEP DAMAS Flametree Probation ATI Probation ATI Probation ATI Probation ATI

a) Prison comparisons

Recidivism measure 1. Prevalence % with 1st arrest … felony charge drug charge 2. Incidence Average number of indicted rearrests N

a) Probation comparisons

Table C-1: Program-Specific Comparisons of the Type and Severity of Rearrests.

Appendix C

References Harrell, Adele; Shannon Cavanagh; and John Roman 2000. Evaluation of the D.C. Superior Court Drug Intervention Programs. Washington, D.C.: U.S. Department of Justice, Office of Justice programs, National Institute of Justice (Research in Brief). Hosmer, David W. Jr. and Stanley Lemeshow 1999. Applied Survival Analysis: Regression Modeling of Time to Event Data. New York, NY: John Wiley & Sons, Inc. Posavac, Emil J. and Raymond G. Carey 1997. Program Evaluation: Methods and Case Studies. Upper Saddle River, NJ: Prentice-Hall, Inc. Revere, Elyse and Mary Curbelo 2001, Alternative-to-Incarceration Information Services First Half Fiscal Year 2001: Six-Month Report. New York, NY: New York City Criminal Justice Agency.