ABSTRACT
Title of dissertation: RISK-ANTICIPATED COMMUNITY SUPERVISION
David E. Hu er, Doctor of Philosophy, 2008
Dissertation directed by: Professor Charles F. Wellford
Department of Criminology and Criminal Justice
Many o enders are released conditionally to communities in lieu of jails or prisons be-
cause, for them, the benefits of sustained social ties and community-based treatment are thought
to outweigh any of those brought about by incarceration. There is reason for caution, though, as
their release to some extent jeopardizes public safety. Available research, for instance, convinc-
ingly suggests a sizeable fraction of o enders enters probation yet fails to comply with release
conditions. This steepens the already uphill challenges of o ender management and reintegration
facing supervision agencies.
The underlying goal of this study is the development and validation of an instrument for
informing immediate, risk-anticipated security and treatment assignments among community-
supervised o enders in the District of Columbia. The study examines whether probationers in
the population test positive, provide a bogus specimen, or fail to appear for any drug testing event
as well as whether and, if so, how often they test positive for each of the seven substances (viz., al-
cohol, methadone, amphetamine, cocaine, marijuana, opiates, and phencyclidine) screened by the
Court Services and O ender Supervision Agency for the District of Columbia. It also examines
whether o enders are ultimately convicted given an arrest for a new crime. Analyses also center
on how often supervision- and drug-related violations occur as well as the probabilities and rates
of ultimately terminating unsuccessfully.
These processes are estimated among a random sample of approximately 200 probationers
having terminated their community sentences during the interval beginning on January 1, 2004,
and ending on December 31, 2004. From well over 200 theoretically plausible predictors, this
study identified a very small set that provide the agency with advance notice of the most challeng-
ing groups of o enders. This set of characteristics includes (a) the age at the time of assessment,
(b) the expected length of supervision, (c) the number of substances ever used, (d) whether the
probationer had ever used opiates or phencyclidine, (e) the number of weapons-related convic-
tions, (f) the SFS-98 score, (g) the recommended sentence, (h) the impression of recidivism risk
on the supervising CSO, and (i) local rates of arrests for drug-related and public order crimes.
RISK-ANTICIPATED COMMUNITY SUPERVISION
by
David Eugene Hu er
Dissertation submitted to the Faculty of the Graduate School of the
University of Maryland, College Park in partial fulfillment
of the requirements for the degree of
Doctor of Philosophy
2008
Advisory Committee:
Professor Charles F. Wellford, Chair
Professor Brian D. Johnson
Dr. Calvin C. Johnson
Professor Doris L. MacKenzie
Professor Gary D. Gottfredson
c Copyright by
David Eugene Hu er
2008
Contents
RISK-ANTICIPATED COMMUNITY SUPERVISION 1
LITERATURE REVIEW 5
Individual-level predictors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
Age . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
Education . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
Employment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
Residential stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
Substance use . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
Childhood and family factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
Health . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
Criminal history . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
Contextual predictors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
Sociodemographic and economic characteristics . . . . . . . . . . . . . . . . . . 35
Objective crime measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
Commercialization patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
METHODOLOGY 46
Participants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
Data and Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
Individual-level measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
Contextual measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
RESULTS 77
ii
Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
Substance use . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
Arrest-convictions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
Violations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
Modes of termination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
DISCUSSION 167
Risk-needs Screener 201
SFS-98 207
Crime categories 211
Conditions of supervision 212
Modes of termination 214
Technical details 216
Bootstrapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
Recursive partitioning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
References 220
iii
List of Tables
1 CSOSA Risk-needs Screener. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
2 PSI, sentence and supervision histories . . . . . . . . . . . . . . . . . . . . . . . 63
3 PSI, family characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4 PSI, social characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
5 PSI, health characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
6 U.S. Census, sociodemographic characteristics, 2000 . . . . . . . . . . . . . . . 72
7 CSOSA RNS, social characteristics . . . . . . . . . . . . . . . . . . . . . . . . 78
8 CSOSA RNS, criminal history characteristics . . . . . . . . . . . . . . . . . . . 81
9 CSOSA RNS, substance use and mental health . . . . . . . . . . . . . . . . . . 82
10 PSI, substance use . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
11 PSI, juvenile and adult o ending . . . . . . . . . . . . . . . . . . . . . . . . . . 84
12 PSI, adult convictions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
13 PSI, triggering o ense . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
14 PSI, sentence and supervision performance . . . . . . . . . . . . . . . . . . . . 87
15 PSI, treatment histories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
16 PSI, family characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
17 PSI, social characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
18 PSI, educational and employment characteristics . . . . . . . . . . . . . . . . . 91
19 PSI, health characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
20 U.S. Census, subset of sociodemographic characteristics, 3-Factor solution . . . . 92
21 Densities of retail alcohol licensees, DC, 2004 . . . . . . . . . . . . . . . . . . . 92
22 Arrest rates, DC, 2004 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
23 Description of criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
iv
24 Parameter estimates, MS1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
25 Bias-corrected confidence intervals, MS1 . . . . . . . . . . . . . . . . . . . . . 101
26 Mean rates of positive drug screens . . . . . . . . . . . . . . . . . . . . . . . . . 103
27 Parameter estimates, MS2D . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
28 Bias-corrected confidence intervals, MS2D . . . . . . . . . . . . . . . . . . . . 109
29 Parameter estimates, MS2E . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
30 Bias-corrected confidence intervals, MS2E . . . . . . . . . . . . . . . . . . . . 116
31 Parameter estimates, MS2F . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
32 Bias-corrected confidence intervals, MS2F . . . . . . . . . . . . . . . . . . . . . 121
33 Parameter estimates, MS2G . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
34 Bias-corrected confidence intervals, MS2G . . . . . . . . . . . . . . . . . . . . 126
35 Parameter estimates, MC1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
36 Bias-corrected confidence intervals, MC1 . . . . . . . . . . . . . . . . . . . . . 134
37 Parameter estimates, MC2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
38 Bias-corrected confidence intervals, MC2 . . . . . . . . . . . . . . . . . . . . . 140
39 Frequencies of violations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
40 Parameter estimates, MV1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
41 Bias-corrected confidence intervals, MV1 . . . . . . . . . . . . . . . . . . . . . 147
42 Parameter estimates, MV2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
43 Bias-corrected confidence intervals, MV2 . . . . . . . . . . . . . . . . . . . . . 153
44 Modes of termination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
45 Parameter estimates, MT1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
46 Bias-corrected confidence intervals, MT1 . . . . . . . . . . . . . . . . . . . . . 160
47 Parameter estimates, MT2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
v
48 Bias-corrected confidence intervals, MT2 . . . . . . . . . . . . . . . . . . . . . 166
49 Characteristics associated with increased levels of negative supervision perfor-
mance among Black males supervised in the DC. . . . . . . . . . . . . . . . . . 169
50 Ever used cocaine and marijuana by age . . . . . . . . . . . . . . . . . . . . . . 175
51 Component comparisons of the SFS-98 and the recommended sentence. . . . . . 178
D1 Conditions, general . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
D2 Conditions, special . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
vi
List of Figures
1 Observation periods, minimum and medium supervision levels . . . . . . . . . . 79
2 Observation periods, maximum and intensive supervision levels . . . . . . . . . 80
3 Classification trees, MS1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
4 Calibration, MS1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
5 Predicted probabilities, MS1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
6 Classification trees, MS2D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
7 Calibration, MS2D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
8 Expected rates, MS2D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
9 Classification trees, MS2E . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
10 Calibration, MS2E . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
11 Expected rates, MS2E . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
12 Classification trees, MS2F . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
13 Calibration, MS2F . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
14 Classification trees, MS2G . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
15 Calibration, MS2G . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
16 Classification trees, MC1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
17 Calibration, MC1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
18 Predicted probabilities, MC1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
19 Classification trees, MC2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
20 Calibration, MC2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
21 Predicted probabilities, MC2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
22 Classification trees, MV1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
23 Calibration, MV1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
vii
24 Classification trees, MV2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
25 Calibration, MV2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
26 Classification trees, MT1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
27 Calibration, MT1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
28 Predicted probabilities, MT1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
29 Classification trees, MT2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
30 Calibration, MT2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
viii
RISK-ANTICIPATED COMMUNITY SUPERVISION
Many o enders are released conditionally to communities in lieu of jails or prisons because,
for them, the benefits of sustained social ties and community-based treatment are thought to out-
weigh any of those brought about by incarceration. There is reason for caution, though, as their
release to some extent jeopardizes public safety. Available research, for instance, convincingly
suggests a sizeable fraction of o enders enters probation yet fails to comply with conditions of
release. This often includes absconding, reo ending, failing to pay fines or restitution, or refusing
to attend or complete treatment programs.
Langan and Cunni (1992) for instance, in following a representative sample of 79,000
felony probationers, found within 3 years nearly two-thirds had either been arrested for new felony
charges or charged with violating conditions of supervision. Almost half were either sent to prison
or jail or had absconded. Likewise, the Bureau of Justice Statistics (BJS) estimates roughly 18?
25% of probationers fail to successfully complete supervision (BJS, 2000, 2002, 2003). Recent
estimates indicate the percentage of unsuccessful probationers may be as large as 40% (Glaze
& Bonczar, 2006). In summarizing the BJS Annual Probation Survey for yearend 2004, Glaze
and Bonczar found on average about 16% of probationers are returned to incarceration, a slightly
smaller fraction fails with other outcomes, and roughly 4% abscond. Even successfully completed
terms are punctuated with repetitious violations. For instance, both Clear, Harris, and Baird (1992)
and Bork (1995) found between one-fourth and one-half of probationers that do successfully com-
plete community supervision were not fully compliant with release conditions (see also, Bonczar,
1997; Glaze & Palla, 2004; Gray, Fields, & Maxwell, 2001; Mayzer, Gray, & Maxwell, 2004;
Petersilia, Turner, Kahan, & Peterson, 1985; Petersilia, 1985a, 1985b, 1998).
This steepens the already uphill challenges of o ender management and reintegration facing
supervision agencies. With limited resources, these agencies must identify the most e ective
1
strategies and services for managing and reintegrating an endless stream of o enders. This is a
tremendous task and one replete with uncertainty. Grappling with such issues is the theme of this
research.
Choices facing supervision agencies often necessitate judgments about the future behaviors
of those under their charge, yet research has traditionally provided limited guidance in the early
identification of potentially noncomplying probationers. Recently though several studies have
enriched the understanding of the compliance process by identifying poor performance markers
(see, Gray et al., 2001; Langan & Levin, 2002; MacKenzie & Li, 2002; Mayzer et al., 2004;
Minor, Wells, & Sims, 2003; Silver & Chow-Martin, 2002; F. P. Williams III, McShane, & Dolny,
2000). This study draws heavily on these recent studies.
The underlying goal of this study is the development and validation of an instrument for
informing immediate, risk-anticipated security and treatment assignments among community-
supervised o enders in the District of Columbia. While representing more in some instances,
less in others, here risk represents the propensity for classes of supervised o enders to engage
in negative supervision performance (NSP) which, here, encapsulates key features of community
supervision split across two domains: legal and supervision-specific.
Elements in the legal1 domain include substance use and criminogenic behaviors. This
study examines whether probationers in the population test positive, provide a bogus specimen, or
fail to appear for any drug testing event as well as whether and, if so, how often they test positive
for each of the seven substances (viz., alcohol, methadone, amphetamine, cocaine, marijuana,
1The term ?legal? is used generally for categorizing like behaviors. One element in particular, adult alcohol use,
is legally acceptable; however, grouping this behavior within the legal category maintains continuity within the liter-
ature review and the methodological discussions and it allows for cleaner discrimination between deviant behaviors
and procedural missteps. A fine line separates legal and illegal alcohol intake?it is, of course, illegal under certain
circumstances. It is assumed here, as only a fraction of o enders are screened for alcohol use, the justification for such
screens is inherently deviance-related.
2
opiates, and phencyclidine) screened by the Court Services and O ender Supervision Agency for
the District of Columbia (CSOSA). It also examines whether o enders are ultimately convicted
given an arrest for a new crime during the supervision period and the follow-up period.
Supervision-specific elements include two salient features of supervision performance: con-
dition violations, both supervision- and drug-related, and termination modes. Supervision-related
violations include violations of general and special conditions; drug-related violations include
only the subset of conditions specifically involving alcohol and illegal substances. Analyses cen-
ter on how often supervision- and drug-related violations will occur. Probabilities of terminating
unsuccessfully2 are also estimated as are factors associated with early failures.
Population processes are estimated using both individual- and environmental-level predic-
tors. Individual-level predictors comprise both background and legal characteristics, where back-
ground characteristics include indicators such as age, education and employment patterns, resi-
dential stability, substance use, physical and mental health, and social and family attributes; legal
predictors tap criminal history, aspects of the instant o ense, and characteristics of the instant
sentence. While these predictors capture o ender-specific influences, environment-level forces
capture those potentially a ecting clusters of o enders. These include demographic and socioe-
conomic measures, such as population density, concentrated poverty, and ethnic heterogeneity.
Objective crime patterns, using both overall and crime-disaggregated indices, are also incorpo-
rated as are measures of residential and commercial land use.
This prelude introduces the problem?developing an instrument for assessing risk of NSP
for guiding immediate custodial and treatment decisions?and, in the next few chapters, I begin
linking it with large bodies of theoretical literature and empirical research. I open Chapter 2 by
describing risk assessments in corrections then highlighting conceptual explanations of criminal-
2Potential termination modes include satisfactory expiration or termination, unsatisfactory expiration or termination,
revocation followed by sanctions, revocation followed by reincarceration, absconsion, and death.
3
ity, recidivism, and NSP. I synthesize these patterns in Chapter 3 when describing and justifying
the methods and procedures used in the present study. I describe the results from these procedures
in Chapter 4, and, in the final chapter, summarize these as well as draw out their implications and
caveat these with a discussion of the limitations in this study.
4
LITERATURE REVIEW
Identifying factors contributing to crime rates, understanding the origin, nature, or charac-
teristics of criminal o ending, and ascertaining the causes and correlates of recidivism are core
criminological pursuits. Many studies have isolated contributors to varying regional crime rates;
likewise, criminological research is rich with studies identifying general and o ense-specific cor-
relates of criminal behavior. Correlates are wide-ranging. They include both individual- and
structural-level characteristics drawn variously from biologic, psychologic, sociologic, demo-
graphic, ecologic, and economic sources. Many are consistently associated with, most pertinently,
negative supervision performance (NSP).3
The goal is developing a risk-anticipated instrument for guiding immediate security and
treatment decisions facing the Court Services and O ender Supervision Agency for the District
of Columbia (CSOSA). It draws heavily on literature and research linking both individual- and
structural-level characteristics with subsequent behaviors and relies on a substantial assessment
opportunity provided by information readily available to the CSOSA early in the custodial pro-
cess. A plan to take advantage of these opportunities is developed in the next chapter. Here, I
provide conceptual justifications by highlighting where the weight of the theoretical literature and
empirical research falls with respect to these correlates. Before descending into these details a
background is provided to contextualize this study within its intellectual and methodological her-
itage. It opens with an identification of the purposes of risk assessments then briefly traces their
footing in the field over time.
Though risk assessments are at the height of fashion today (P. R. Jones, 1996; van Voorhis
& Brown, 1997), forecasting which sets of individuals are most likely to initiate, persist, escalate,
3The term NSP is used here to encapsulate problem behaviors, be they criminogenic or supervision-specific, posed
by supervised o enders.
5
or desist o ending?discovering predictors of future criminality?is far from a recent develop-
ment. These are traditional concerns. Throughout corrections history risk assessments have been
instrumental in their achievement. For instance, they have helped to understand and classify, to
inform decisions bearing on and, invariably, to enhance surveillance and control over corrections
populations (Brennan, 1987; Champion, 1994; Farrington & Tarling, 1985b; S. D. Gottfredson &
Gottfredson, 1985). They have helped isolate the dangerous, the violent, and the insane; they have
helped predict likely delinquents from non-, parole successes from failures, and chronic, habitual
o enders from their counterparts (Burgess, 1928; Cocozza & Steadman, 1976; Glueck & Glueck,
1930, 1950; Greenwood & Abrahamse, 1982; Link, Andrews, & Cullen, 1992; Monahan, 1981).
They have provided considerable value to corrections and, by most accounts, will continue to do
so.
They are specifically valued among community supervision agencies for their ability to both
formalize and guide decisions which mutually tighten agency accountability and augment public
safety and o ender reintegration e orts. Risk assessments lend themselves to the development
of consistent, equitable, e cient, and verifiable decision calculi that can then be evaluated, up-
dated, modified, and refined?perpetually adjusting to, and consequently curtailing, error. Risk
assessments identify and thereafter classify o enders across various measures of risk and need
thus speeding appropriate control and therapeutic responses. They identify special populations,
such as violent, chronic o enders, for whom unique services and control mechanisms are avail-
able. In a similar fashion, they identify minimal risks. This helps ensure only minimally restrictive
mechanisms sustain conformity. They also identify those with special needs, such as educational
or mental health, which then suggests appropriate therapeutic responses. Codified decision rules,
accurate identification and classification, and timely matching of services and control mechanisms
each increase agency accountability, public safety, and o ender reintegration.
6
Isolated examples of criminological risk assessments appear as early as the 1920s (e.g.,
Bruce, Burgess, & Harno, 1928; Burgess, 1928; Glueck & Glueck, 1930; Hart, 1923; Warner,
1923), but it was not until the 1950s, a period Cullen and Gendreau (2001) depict as ripe with cor-
rectional optimism, that they begin appearing all together (see also, Farrington & Tarling, 1985b).
The earliest of these, described variously as anamnestic or idiographic methods (see, Melton,
Petrila, Poythress, & Slobogin, 1997; Morris & Miller, 1985), were largely informed by intu-
ition or personal experience and were thus highly subjective. Based primarily on case studies,
idiographic methods find justification in stability: past behaviors under certain conditions are in-
dicative of future behaviors under similar conditions. Past behaviors were, after all, once future
behaviors themselves. It is no surprise then to find they are among the strongest behavioral pre-
dictors. Still, despite their intuitive charm, idiographic predictions were nonetheless limited first
by the strictly individualized approach?generalizable no further than the case under study?and
second by their exclusive situational approach?restricted only to conditions having already been
observed.
Later methods, though still largely subjective, were increasingly analytic and thus coun-
tered some of the problems with earlier models. Known broadly as clinical assessment, this
tradition, heavily imbued in professional training and experience, was the criminological norm
through the 1950s?a time coinciding with the beginning of the tenebrious end of correctional
optimism (Brennan, 1987; Cullen & Gendreau, 2001). Like idiographic methods, clinical assess-
ments were also based on case-by-case observations and they, too, were highly subjective. Most
often they were conducted by a single clinician who, by piecing together information derived from
unstructured interviews and detailed case reports, estimated likelihoods a particular case might ex-
perience a given outcome. Because of this single-case method, risk factors for one case could, and
in many instances did, vary widely across cases and even within cases assessed by multiple ob-
7
servers. Distinctive is their embodiment of professional?particularly psychological?guidance.
Clinical assessments were based on measures deemed relevant by the observing clinician. Though
at times they incorporated batteries of psychometric tests they were, for the most part, neither
standardized nor quantified (Brennan, 1987; Grove & Meehl, 1996; Grove, Zald, Lebow, Snitz, &
Nelson, 2000; Morris & Miller, 1985). And while they represent a more formal approach than id-
iographic methods, they too were mired in problems. Indeed, success among both idiographic and
clinical assessments pales in comparison to success found among the statistical assessments that
soon came to dominate the field (for review see, Cocozza & Steadman, 1976; P. R. Jones, 1996;
Kahneman & Tversky, 1973; Lidz, Mulvey, & Gardner, 1993; Meehl, 1954; Menzies, Webster,
McMain, Staley, & Scaglione, 1994; Monahan & Steadman, 1994; Monahan, 1981; Mossman,
1994; Quinsey & Maguire, 1986).
The clinical assessment heyday coincides with a period Cullen and Gendreau (2001) say
viewed the notion of criminal rehabilitation most favorably. Many accepted the notion that dan-
gerous, chronic, habitual o enders could be pre-identified and early selected for special control
and therapeutic strategies. The considerable support for the rehabilitation of o enders on one
hand coupled with a general willingness to identify and classify o ender subgroups on the other
created conditions risk assessments naturally satisfy: they bridge the stated goals of corrections
with a means to do so.
While optimistic through the 1950s, the atmosphere later turned increasingly critical. It was
during the decades that followed when evidence challenging extant correctional practice began
prompting many to reconsider correctional mainstays. Rehabilitation was among the most notable
of these (Cullen & Gendreau, 2001). The theme resonating was that correctional treatment was in-
e ective. For some, the interventions were at fault (Bailey, 1966; D. Lipton, Martinson, & Wilks,
1975; Martinson, 1974; Robinson & Smith, 1971). Others blamed the setting (Foucault, 1977;
8
Irwin & Cressey, 1962; Sykes & Messinger, 1960). Regardless, a shared disdain for correctional
treatment was taking hold. Prisons became conspicuous targets. From within, violent inmate
uprisings began attracting attention to inequities and injustices. Outside their walls, academics
began pointing to pervasive racial disparities and wide sentencing variabilities (R. L. Austin &
Allen, 2000; D?Alessio & Stolzenberg, 2003; D. A. Smith, Visher, & Davidson, 1984; Spohn,
2000; Zatz & Hagan, 1985; Zatz, 1984).
Along with attacks on the rehabilitative front, fear was rising as a result of both real and
perceived crime increases. This generated a growing sense of intolerance toward crime and crim-
inals begetting later demands for coherent and consistent identification, processing, sentencing,
and control of serious, persistent o enders (Brennan, 1987; Fogel, 1975; von Hirsch, 1976;
P. R. Jones, 1996). By the 1970s, these sentiments, well-founded or not, were firmly established.
The passing of this decade witnessed the development of a new correctional mindset and the near
abandonment of one its core ideals. Rehabilitation lost its foothold, its public favor; as it fell,
inextricable praxes, like risk assessments, followed?not only because embracing them tacitly
embraced the philosophy in which they were grounded, but also because evidence of their ostensi-
ble inadequacies was mounting. In place of rehabilitation corrections began instead emphasizing
retribution and crime control. This shift was in a large way reflective of increasingly conservative
politics.
The 1980s confluence of rising crime and economic constraint fostered a conservative po-
litical shift. Crime was booming, the economy was busting, and this together began pushing many
into advocating tougher and less expensive crime control strategies. The long-term consequences
have been enormous. It essentially triggered a pervasive burden still overwhelming corrections.
Of particular interest is the ironic sparking of continued research into the e ectiveness of cor-
9
rectional treatment, non-standard methods of o ender management, and statistical prediction of
which I describe next.
To rehabilitation, the American industry had collectively turned a blind, skeptical eye. De-
spite the benumbing nihilism rehabilitation research continued (see, Cullen & Gendreau, 2001).
This was continued mainly at the hands of a minority of rehabilitation adherents trained in psy-
chology and measurement (e.g., Lipsey, Cordray, & Berger, 1981; T. Palmer, 1975) and was, for
the most part, of Canadian origin. It was, nonetheless, largely overshadowed by the unfortunate
American emanation.
Skepticism persisted at least until the late-1980s when the findings aired by persevering
rehabilitation adherents were becoming increasingly di cult to ignore. Proponents consistently
a rmed the contrary of much of the dialogue that had provided rehabilitation?s deathblow only
decades earlier (D. A. Andrews et al., 1990; Cullen & Gilbert, 1982; Gendreau, Goggin, & Pa-
parozzi, 1996; Gendreau & Ross, 1979, 1987). Rather than accepting the nihilistic view on pure
faith, many began gradually adopting?once again?an optimistic stance. Theoretical advances
in rehabilitation began slowly reemerging across the country (Cullen & Gendreau, 2001).
Springing from these advances was an increasingly clear understanding of risk factors and
the distinctions among them, a greater understanding of o ending and recidivism, and growth
in rigorous assessment methods. Traditionally, distinctions have been made between static and
dynamic risk factors (D. Andrews & Bonta, 1998). On the one hand, static risk factors are ei-
ther aspects of o enders or their past that are predictive of criminality but not subject to change
(e.g., age, gender, early family factors, and, often one of the strongest factors, criminal history;
D. Andrews and Bonta; S. D. Gottfredson and Gottfredson). These have been and will doubtlessly
continue to be longtime staples of corrections research. On the other, dynamic risk factors re-
flect circumstances, attitudes, and behaviors that are mutable and, consequently, likely targets for
10
intervention (D. Andrews & Bonta, 1998). Dynamic factors are subdivided respective to their
association with criminal behavior: criminogenic dynamic risk factors (or simply criminogenic
needs; e.g., antisocial personality, substance abuse, or low self-control) have been empirically
linked to criminal behaviors; non-criminogenic needs (e.g., low self-esteem) have not.
Both static risk factors and criminogenic needs are linked to criminality. Further, studies
show criminogenic needs predict criminal behaviors as well as static factors do (Champion, 1994;
Gendreau, Goggin, & Paparozzi, 1996; Glover, Nicholson, Hemmati, Bernfeld, & Quinsey, 2002).
They di er, however, in their expected responsiveness to treatment. It is this di erence that makes
static factors less desirable than criminogenic needs for guiding treatment plans. Because they
cannot, for the most part, change readily in response to treatment, static risk factors are unlikely
intervention targets. There is simply little expected utility in targeting them. Criminogenic needs,
because they can change readily, are indeed likely intervention targets (D. Andrews & Bonta,
1998). They are related to criminality and reasonably responsive to treatment; targeting them is
expected to provide at least some rehabilitative benefit.
It is precisely this mutability, however, that makes them less attractive for immediate cus-
todial decisions than static risk factors (cf., Quinsey, Coleman, Jones, & Altrows, 1997; Zamble
& Quinsey, 1997). Assessing criminogenic needs implies both longitudinal data and substantial
observation. This is likely unavailable when making immediate decisions. For these decisions, in-
formation is typically limited to the readily accessible such as that available from police records,
Presentence Investigation report (PSI) reports, court documents, jail or prison profiles, and previ-
ous supervision information.
As these conceptual distinctions were becoming well-understood, criminologists continued
investigating factors implicated in the o ending and recidivism process. This prompted continued
experimentation with innovative o ender management strategies. Early studies of persistent of-
11
fending (e.g., Wolfgang, Figlio, & Sellin, 1972) had established the notion that a small number of
o enders were disproportionately responsible for crime. Given large di erences in individual of-
fending rates, strategies targeting only high-risk o enders for incarceration appeared viable. The
problem thus became identifying and controlling serious, persistent o enders.
Selective incapacitation strategies became widely accepted justifications for incarceration
and continued to be seen as such as more studies replicated the findings of Wolfgang et al. (1972).
In the very least, though intuitively appealing, identifying chronic, high-risk o enders proved
elusive. Worse, it had the potential to and, by most accounts, did create more problems than it
solved (see, Bernard & Ritti, 1991). Prisons and ultimately all corrections populations distended
as the stream of inmates sent ?up the river? became progressively larger and less manageable.
Community supervision agencies were expectably burdened?not only by sheer numbers but also
by economic constraint. The number of o enders agencies managed was rising; their budgets
were not (Petersilia, 1985b). This prodded many agencies to search for alternatives for meeting
increasing service demands (P. R. Jones, 1996).
Unlike the prison industry community corrections agencies were unable to build them-
selves out of the crisis. Boxed in, they could acquiesce by, for instance, pragmatically shifting
interests toward o ender management rather than treatment (Feeley & Simon, 1992). Or they
could innovate. They could, for instance, increase the development of and subsequent reliance on
non-traditional correctional practices.
Boot camps or intensive supervision programs are among these innovations (for compre-
hensive reviews, see MacKenzie, 2000; Petersilia, 1998). These and the many strategies not men-
tioned share a common thread with risk assessments. Namely, the compromise between public
safety and public cost. Balancing the two is something for which risk assessments are vital; they
allow agencies to allocate resources e ciently by reserving extensive surveillance and control
12
mechanisms for those presenting the highest risk and more costly treatment programs for those
presenting the greatest need. As it turned out, risk assessments reemerged in the 1990s as vogue
correctional tools (see, P. R. Jones, 1996; van Voorhis & Brown, 1997). Much of this attraction
was epiphenomenal.
While conceptual developments were broadening and more corrections agencies were toy-
ing with non-standard means of o ender management, methodological advances were leading
to increasingly precise assessment and classification tools (Champion, 1994). The seeds were
planted in the 1950s when pitfalls associated with existing instruments were urging for more rigor-
ous methods (Glaser, 1987; S. D. Gottfredson & Gottfredson, 1979; Meehl, 1954; Sawyer, 1966).
As increasingly precise methodologies began appearing, clinical assessments became nearly uni-
versally rejected by criminologists. An alternative, more formal approach, known interchangeably
as statistical, mechanical, or actuarial assessment, began its ascent. These methods, characterized
mainly by their objectivity and formality, markedly improved practice consequently contributing
to a greater acceptance of risk assessments (Brennan, 1987) at just the time a newfound attraction
to correctional rehabilitation was emerging. As criminology began embracing?once again?the
rehabilitation philosophy, the scope and precision of risk assessments were both increasing. Where
early e orts were largely driven by security concerns, contemporary e orts encompass security
as well as programmatic goals. Along those lines, contemporary risk assessments?unlike earlier
models?explicitly incorporated characteristics that are related to criminal behavior, changeable,
and thus amenable to treatment. By the 1990s, this broader variant had made its way into prevail-
ing fashion (Champion, 1994; P. R. Jones, 1996).
Characterized by objectivity, formality, and empirical rigor, statistical risk assessments are
undoubted advancements over earlier tools (D. M. Gottfredson, 1987; Meehl, 1954; Monahan et
al., 2001; Mossman, 1994; Quinsey, Harris, Rice, & Cormier, 1998; Rice & Harris, 1995; Sawyer,
13
1966; van Voorhis & Brown, 1997). Assessments are largely derived from objective evaluations of
predictors following predefined rules. Predictors typically comprise a set of fixed risk factors (e.g.,
age, gender, age of onset) and a range of criminal history variables (e.g., versatility, frequency,
prior parole failure, security classification, o ense severity, sentence length, o ense type; see,
Brown, 2002). Decision rules define how predictors are selected and mathematically weighted
with the ultimate goal of maximizing the statistical association with the criterion. Compared to
clinical assessments, when validated and implemented properly, statistical assessments are more
accurate and the instruments on which they are based demonstrate higher reliabilities (Brennan,
1987; Cocozza & Steadman, 1976; Grove & Meehl, 1996; Grove et al., 2000; P. R. Jones, 1996;
Lidz et al., 1993; Meehl, 1954; Menzies et al., 1994; Monahan & Steadman, 1994; Monahan,
1981; Morris & Miller, 1985; Mossman, 1994; Quinsey & Maguire, 1986).
Up to this point, I have contextualized risk assessments within corrections research, de-
scribed their historical trends, and sketched out current methods. The next section turns to an
examination of individual-level factors consistently correlated with criminality, recidivism, and
NSP. After reviewing the individual factors, contextual factors are discussed.
Individual-level predictors
This section is divided into two groups, the first describes personal and social attributes of
o enders that are associated with crime, recidivism, and negative supervision performance (NSP),
such as age, education, employment, residential mobility, substance use, health, and family fac-
tors, and the second describes criminal and supervision histories and highlights how these char-
acteristics are linked to future NSP. Net of other characteristics, findings from empirical studies
imply NSP will be more prevalent among younger, poorly educated o enders having unstable
employment and residential histories, alcohol and substance abuse problems, strained early and
14
current family relationships, those with health problems?particularly with injuries resulting from
assaults, those with the earliest and most extensive involvement with the juvenile justice sys-
tem, those with lengthy histories of criminal justice involvement, and those whose supervision is
the longest and most intensive. NSP will also be more common among probationers residing in
communities characterized by economic disadvantage, ethnic heterogeneity, immigration concen-
tration, residential instability, high crime, mixed-use land patterns, and high alcohol availability.
Discussion opens with evidence bearing on the link between age and o ending, recidivism, and
NSP.
Age
Age is linked inextricably with both o ending and recidivism in such a way as to expect
higher levels of NSP among the most youthful o enders. Two issues are entwined. First, an
early onset of problem behaviors, conduct disorders, and delinquency predict persistent o ending.
Second, younger o enders are disproportionately represented in age distributions of o enders,
recidivists, and failing supervisees. The focus for now is on the latter issue; discussion of the
former is deferred to a later section describing how juvenile and criminal histories anticipate NSP.
As a nearly undisputed criminological mainstay, criminal behavior is concentrated among
adolescence and young adults ages 12 to 30 (Dembo et al., 1995; Farrington, 1986; Gendreau,
Little, & Goggin, 1996; M. R. Gottfredson & Hirschi, 1986; Ho man & Beck, 1984; Matza, 1964;
Osgood, Johnston, O?Malley, & Bachman, 1989; Sampson & Laub, 1993; Wolfgang, Thornberry,
& Figlio, 1987). It increases gradually with age then tapers through young adulthood. Along
these lines, the likelihood of NSP is highest among younger o enders and studies consistently
find unsuccessfully completing community sentences is more likely among younger o enders
(Clarke, Lin, & Wallace, 1988; Cloninger & Guze, 1973; Dembo, Williams, Schmeidler, et al.,
15
1991; Harer, 1994; Harrison & Gfroerer, 1992; Irish, 1989; Morgan, 1993; Rhodes, 1986; Sims
& Jones, 1997; Whitehead, 1991; cf., Benedict & Hu -Corzine, 1997).
Some argue this reflects a natural inclination toward and an ultimate ?burn-out? or ?matura-
tional reform? from criminal participation (Ho man & Beck, 1984; Matza, 1964). Others point to
the age-graded changes in social influences and institutions (e.g., Sampson & Laub, 1993). It may
also indicate, as some suggest, less e ective adjustment to the requirements and demands of super-
vision, such as sustaining contact with supervision o cers, among younger o enders (MacKenzie,
Shaw, & Souryal, 1992; F. P. Williams III et al., 2000; cf., McReynolds, 1987; Schwaner, 1997).
Characteristics associated with age ultimately place younger o enders at greater risk of
NSP. The predictor discussed in the next section is partly dependent on youthfulness and, like
age, is also a prominent and well-established correlate of o ending and NSP.
Education
Educational performance, commitment to educational goals, and educational attainment
are inversely associated with criminal justice involvement. Schools are secondary socializing
institutions acting to reinforce social values and a large body of research suggests those perform-
ing poorly in early academic settings are involved more so in o ending than their counterparts
(Hindelang, 1973; Hirschi, 1969; Kruttschnitt, Heath, & Ward, 1986; Sampson & Laub, 1993;
Ward & Tittle, 1994; and see, Horney, Osgood, & Marshall, 1995).
Poor performance weakens commitment to conventional educational goals which in turn
limits the strength of otherwise inhibiting control mechanisms (Hirschi, 1969). As A. K. Cohen
(1955) argued, to the extent these goals are positively valued, the inability to achieve status and
acceptance in the educational sphere might generate negative a ect. This distances youth from
conventionality. Those exhibiting the least commitment to educational goals are the most crime
16
prone (Agnew & White, 1992; Kruttschnitt et al., 1986; Sampson & Laub, 1993; L. Zhang &
Messner, 1996). Brezina concurs. He argues those experiencing such a ect cope using an array
of cognitive, behavioral, and emotional strategies, increased o ending being one such mechanism
(see, Agnew, 1985, 1989, 1992; Brezina, 1996; Farnworth & Leiber, 1989; Thornberry, Moore, &
Christenson, 1985a).
The least committed also face higher risks of dropping out of high school, truncating, of-
ten immutably, their educational attainment (see, Farrington, 1997; Harrison & Gfroerer, 1992;
Jarjoura, 1996; Thornberry, Moore, & Christenson, 1985b). This is relevant as research indicates
the likelihood of criminal involvement decreases as educational attainment increases (Beck et al.,
1993; S. D. Gottfredson & Gottfredson, 1979; Quinsey et al., 1998; Thornberry et al., 1985b) It is
a decisive predictor of NSP (Gray et al., 2001; Harer, 1994; Irish, 1989; Landis, Merger, & Wol ,
1969; Mayzer et al., 2004; Morgan, 1993; Rhodes, 1986; Roundtree, Edwards, & Parker, 1984;
Silver & Chow-Martin, 2002; Sims & Jones, 1997). Silver and Chow-Martin (2002), for example,
found the likelihood of rearrest within five years was well-predicted by whether o enders finished
high school.
Educational characteristics are thus important factors in assessing risk of NSP. Stemming
partly from a tenuous hold on conventional values, something instilled early in life and reinforced
by the educational system, poor performance early in the educational process, a lack of commit-
ment to its goals, and lowered levels of attainment are expected to increase risk. A similar e ect is
expected throughout adulthood in the legitimate employment sector, which is the topic of the next
section.
17
Employment
Certain employment characteristics are consistently associated with NSP. Unemployment
and job instability in particular are both associated with elevated o ending (Farrington, 1986;
Thornberry & Christenson, 1984; Thornberry & Farnworth, 1982) and recidivism (Uggen, 2000).
This is widely replicated and, by most accounts, o enders sharing these attributes present a higher
risk of NSP.
The legitimate labor force is a conventional setting that Sampson and Laub (1993) note
likely encourages conformity (also see, Warr, 1998). Thornberry and Christenson (1984) continue
that a commitment to these conventional goals reduces criminal involvement by simultaneously
increasing its costs and decreasing both available time for, and any rewards likely generated from,
nonconformity. As involvement in illegal activities increases, legitimate opportunities shrink. This
interchange suggests unemployment will have a positive e ect on NSP and that this e ect should
become increasingly stronger as unemployment spells become more frequent.
Being unemployed is associated with rearrests, technical violations, and absconding (J. Austin
& Litsky, 1982; Gray et al., 2001; Harer, 1994; Irish, 1989; M. Jones, 1995; Landis et al., 1969;
MacKenzie & Li, 2002; Morgan, 1993; Silver & Chow-Martin, 2002; Sims & Jones, 1997;
F. P. Williams III et al., 2000; cf., Roundtree et al., 1984). Irish (1989) for instance examined
a randomly sampled cohort of probationers discharged in 1982 and found unemployment was as-
sociated with higher rearrest rates. Among a subsample of n = 562 probationers (n = 349 [62%]
of whom were employed; n = 213 [38%] were unemployed) he distilled information bearing
on o ender characteristics, program adjustment, and supervision outcomes from a wide array of
sources, such as PSIs, supervision case records, and police arrest reports. Criteria included rear-
rest, supervision violation, and supervision adjustment. Out of all the predictors, unemployment
was one of the most influential. Specifically, o enders unemployed at the time of arrest or sen-
18
tencing had higher probabilities of arrest while supervised or shortly after supervision had ended
than their counterparts. In addition, probationers unemployed during their sentence were more
likely to violate supervision conditions.
Employment instability increases the chances of NSP. For instance, numerous job changes
and frequent unemployment spells have been linked to both o ending (Farrington, 1997) and re-
cidivism (S. D. Gottfredson & Gottfredson, 1979; D. M. Gottfredson, Wilkins, & Ho man, 1978).
As unstable work histories are characteristically shared among the criminally active (Laub & Lau-
ritsen, 1994), this will likely a ect supervision performance. Landis et al. (1969) examined adult
felons in California (n = 791) and found among those failing probation, employment instability
was a decisive predictor. Similarly, F. P. Williams III et al. (2000) found employment instability
was among the most important predictors of absconsion among a relatively large, random sample
of parolees in California (n = 4047). In fact, stability measures in general were the most influ-
ential predictors of absconsion within the first year of parole. These included occupational and
residential stabilities (and see, J. Austin & Litsky, 1982; MacKenzie & Li, 2002; Mayzer et al.,
2004).
Sampson and Laub (1993) argue that beyond mere stability, the quality of employment
moderates its impact on o ending. Specifically, higher wage, higher quality, and more satisfying
occupations are those in which conforming values are most likely instilled (see, Allan & Stef-
fensmeier, 1989; Huiras, Uggen, & McMorris, 2000). Naturally, selection artifacts are potential
explanations. As Uggen and Sta warn, ?. . . the best recidivism risks may be most likely to self-
select into higher quality jobs, but they would be less likely than other people to recidivate even in
the absence of employment? (2001, p. 3). Nevertheless, the evidence bearing on the interrelation-
ship among employment status, stability, and o ending suggests it is a clear factor when assessing
risk of NSP. In terms of continued substance use, convictions, violations, and modes of termina-
19
tion, employment instability will likely exert the strongest e ect on convictions, violations, and
modes of termination. A measure with a similar focus on continuity also shown to correlate with
o ending and recidivism is residential mobility.
Residential stability
The frequency with which o enders move is related to o ending and supervision perfor-
mance. This likely involves underdeveloped informal social controls (see, Kasarda & Janowitz,
1974; Sampson, 1988), as those changing residences often are unlikely to develop strong interper-
sonal ties with their neighbors and other members of their community. They are also less likely to
have a wide opportunity?and, perhaps, willingness?to participate in community activities and
organizations (Sampson, 1988). This social isolation is thought to leave unchecked pressures to
deviate otherwise dampened by informal social controls.
Using data derived from the British Crime Survey, Sampson (1988) examined (a) the re-
lationship of community residential stability on local friendship ties, community attachment, and
social activity patterns and (b) the influence of community characteristics on individual behavior.
In ways Hirschi (1969) anticipated, Sampson found among n = 10905 residents and n = 238
localities, those with the longest tenure in the community were more likely to have developed
dense local friendships, to be strongly attached to their community, and to participate more often
in community organizations.
Residential stability is also associated with NSP. For instance, F. P. Williams III et al.
(2000) found parolees with unstable living arrangements were more likely to abscond. Among the
sample (n = 4047) they examined, roughly one-fourth ultimately absconded. Modeling abscon-
sion within the first year of supervision, the most influential characteristics were (a) unstable living
conditions, (b) frequent unemployment, (c) previous parole violations, (d) low stakes in confor-
20
mity, (e) frequent prior arrests, (f) being single, and (g) having previous felonies. J. Austin and
Litsky (1982) corroborate these findings, at least among probationers. They examined n = 12526
o enders supervised in Nevada. Of these, n = 338 eventually absconded. For probationers, initial
assessment scores were predictive of later absconsion the main contributors to which included fre-
quent address changes, low motivation for change, youthfulness, unemployment, prior probation
sentences, and prior revocations (and see, Mayzer et al., 2004).
It is expected then that residential instability will be key in discriminating probationers by
risk of NSP. Another set of factors consistently able to discriminate among high and low risk
o enders are measures of substance use and abuse.
Substance use
Use and abuse of both alcohol and illegal substances are consistently associated with of-
fending and NSP. I discuss both in this section beginning with the former.
Alcohol is empirically bound with o ending, including aggression and violence, as well as
with general criminality, recidivism, and NSP. Linking mechanisms include both individual- and
structural-level characteristics, where processes at the individual level bear primarily on indirect
pharmacological consequences of consumption; at the structural-level, on ecological inducements
implicated in criminality.
Intoxication has known pharmacological antecedents. Contemporary research suggests
these e ects are complicated within a host of mediating and moderating factors (see, Fagan, 1990;
Miczek et al., 1994; Reiss & Roth, 1993). Behavioral consequences are contingent on qualities of
use, such as habituation and intensity, tolerances, concentrations in the brain, and whether these
concentrations are rising or falling. They also depend on individual attributes, such as personali-
21
ties, behavioral histories, and expectancies and on endocrinological, genetical, and neurobiologi-
cal characteristics.
It is implicated in a large fraction of crimes and, as expected, a large fraction of o enders
are alcohol consumers (Greenfeld, 1998). The most common crimes committed by intoxicated
o enders are public-order o enses and assaults, both of which are prevalent among probationers
(see, Bonczar, 1997; Glaze & Palla, 2005; Greenfeld, 1998). Based on analyses of administrative
records and personal interviews, Mumola and Bonczar (1998) report up to 40% of probationers
were under the influence of alcohol at the time of their instant o ense.
Cognitive distortions may intensify already heightened aggressive predispositions thus am-
plifying combativeness (Miczek et al., 1994). This might explain research finding aggression is a
more frequent recourse among non-abstainers than abstainers (Boyum & Kleiman, 1995): there is
a strong, statistical association between alcohol and both aggression and violence. Indeed, intox-
ication is a central antecedent to murder, assault, and rape (see, Reiss & Roth, 1993; Q. Zhang,
Loeber, & Stouthamer-Loeber, 1997).
Alcohol intoxication is thus expected to predict aggressive and violent aspects of NSP.
This is, understandably, where the bulk of alcohol-crime research is centered. Just as clearly,
studies link alcohol with other crime types including, obviously, driving while intoxicated, public
intoxication, and liquor law violations as well as o enses void inherent alcohol characteristics,
such as public-order and property crimes and certain o ending patterns.
Public-order crimes comprise almost one-third of all crimes reported to the police (Greenfeld,
1998; Stitt & Giacopassi, 1992), and, generally, higher levels of use are associated with increased
o ending (see, Seltzer & Langford, 1984; Shupe, 1954; Stitt & Giacopassi, 1992; Q. Zhang et
al., 1997). Alcohol is similarly implicated in property crimes. For instance, in analyzing detailed
case histories of convicted property o enders, Cordilia found consumption was often implicated
22
in the ?unplanned, low-profit, high-risk crime? characteristic of casual property o enders (1985,
p. 170). Not only was alcohol more accepted among casual o enders, it was verily encouraged,
as, for them, it was functional. It bunched the loose-knit groups, helped sustain criminal activity,
and facilitated repetitious o ending.
Consumption and intoxication, while empirically linked with crime and NSP are, unfortu-
nately, unlikely candidates for predicting NSP as this process depends so strongly on situational
factors. There is, however, evidence suggestive of the less situational linkage between alcohol and
NSP.
Suggestive of the high prevalence of alcoholism among convicted property o enders, Cordilia
(1985) notes that the lifestyles of many leads them into a downward spiral of alcoholism, eventual
exclusion from organized o ending groups and, ultimately, homelessness or prison. Also, among
known violent o enders, both alcohol use and abuse are disproportionately high and, likewise,
violent crimes are unexpectedly prevalent among those with alcohol dependencies (Miczek et al.,
1994).
There is also evidence suggesting specific crime patterns, such as persistent o ending and
repeated probation violations, are linked with alcohol. Higher consumption is generally associ-
ated with greater risk (J. Austin & Litsky, 1982; Cordilia, 1985; Farrington & Hawkins, 1991;
S. D. Gottfredson & Gottfredson, 1979; Harer, 1994; MacKenzie, Browning, Priu, Skroban, &
Smith, 1998; MacKenzie & Li, 2002; Rice & Harris, 1995; Schmidt & Witte, 1988, 1989).
Among an entry cohort of felony probationers in Virginia (n = 125), MacKenzie and Li
(2002) found those drinking excessively were more likely to persist o ending. Similarly, in an
analyses of both self- and o cially-reported behavior among n = 126 probationers, MacKenzie
et al. (1998) found o ending increased during months characterized by heavy alcohol use. These
patterns are widespread among abusers. Among federal parolees (n = 1205), for instance, Harer
23
(1994) found 3-year recidivism rates were higher among those reporting alcohol dependency
compared to their counterparts. And among parolees and, especially, probationers in Nevada
(n = 12526), J. Austin and Litsky (1982) found absconders had substantially higher levels of al-
cohol abuse compared to non-absconders. Similarly, in their analysis of North Carolina releasees,
Schmidt and Witte (1988) found alcoholism was an important characteristic common among rein-
carcerated parolees.
Thus, among supervised populations, those with problem consumption patterns may pose
greater risks. This may be seen in increased alcohol-related as well as public-order and property
crimes. As well, problem drinkers are more likely to be persistent o enders, repeat probation
violators, and absconders.
Closely paralleling these findings are those describing the interplay among illegal sub-
stances, o ending, and NSP. Illegal substance use and abuse is a pervasive American problem
(Substance Abuse and Mental Health Services Administration [SAMHSA], 2002). Beyond its
inherent unlawfulness, research strongly links illegal substances with criminality (M. R. Chaiken
& Chaiken, 1987; Nurco, 1979). Drug use in general is associated with higher o ending levels
(Anglin & Speckart, 1988; Clayton & Tuchfeld, 1982; D. C. McBride & McCoy, 1981; Mc-
Glothin, 1979; Stacy & Newcomb, 1995). Recreational users, for example, are more crime prone
than abstainers (Dembo et al., 1995; Elliott, Huizinga, & Menard, 1989; L. Gardner & Shoemaker,
1989; Newcomb & Bentler, 1986; Stice, Myers, & Brown, 1998). Addiction and dependency as
well are linked to both o ending (Bland, Newman, Thompson, & Dyck, 1998; Inciardi, 1980; Mc-
Glothin, 1979; Speckart & Anglin, 1985) and recidivism (S. D. Gottfredson & Gottfredson, 1979;
J. Palmer & Carlson, 1976; Wish & Johnson, 1986) levels (and see, Dembo, Williams, Getreu, et
al., 1991; Dobinson & Ward, 1986; Greenfield & Weisner, 1995; Guze, Wolfgram, McKinney, &
Cantwell, 1968; Swanson, Holzer, Ganju, & Jono, 1990).
24
This implies prevalent drug use among incarcerated and supervised o enders, which is
supported by research (Mumola, 1999; Mumola & Bonczar, 1998). Drug use and abuse have
also been linked to recidivism and NSP. For example, findings from analyses of pretrial releasees
suggests those testing positive for drugs at arrest posed heightened rearrest risks (D. A. Smith &
Polsenberg, 1992). Among probationers, Benedict and Hu -Corzine (1997) found a history of
drug use?especially higher levels?was related to increased risk of rearrest (see also, MacKenzie
et al., 1998; MacKenzie & Li, 2002; Silver & Chow-Martin, 2002). Likewise, MacKenzie et al.
(1998); MacKenzie and Li (2002); Silver and Chow-Martin (2002) found probation absconders
had higher drug use levels (and see, Baird, Storrs, & Connelly, 1984; Gray et al., 2001; Harer,
1994; Schmidt & Witte, 1988).
In summarizing, alcohol and drug use and abuse have both been linked with criminality and
NSP. Use and abuse histories will thus inform assessments of supervision performance, where,
namely, those probationers either currently or with a history of using and abusing alcohol or illegal
substances are expected to present greater risks of NSP.
Childhood and family factors
This section describes literature and research regarding the linkages among childhood and
family factors, criminality, and NSP beginning with the earliest factors. Certain childhood expe-
riences, such as early economic conditions and family practices, influence later experiences, like
family relationships; aspects of both are associated with supervision performance.
Ample research connects certain early life influences with problem experiences in child-
hood, adolescence, and into adulthood. To this end, there is considerable stability in behaviors.
As, among others, M. R. Gottfredson and Hirschi (1990) contend, childhood experiences have life
long consequences (see also, Wilson & Hernnstein, 1985). This is not to say, however, that these
25
influences are irreversible (see, Sampson & Laub, 1993). Nevertheless, I draw attention here to
lowered economic conditions and early family characteristics and discuss how these are expected
to inform assessments of NSP.
For Merton (1957), success is a value shared across American culture. The means for
achieving success, however, are only narrowly accessible to those in the lower strata. Anomic
adaptations are likely when these aspirations go unmet. Resorting to unconventional means for
achieving success, such as criminality, are among such adaptations.
Yet, motivation and aspiration alone do not fully account for adaptive behaviors. The impact
of this is expressed in Cloward and Ohlin?s (1960) classic blending of anomie and di erential
association traditions. They point out that, just as it is for legitimate opportunities, illegitimate
opportunities are themselves structured unevenly and, further, the values and skills needed to take
advantage of these opportunities must be learned within the environment.
Minority boys especially those in the lower classes bear the brunt of this, as they are the
ones most likely deprived of educational and occupational opportunities, exposed to high levels
of deviance, and to have little at stake inhibiting deviance. This makes them most susceptible to
early delinquent onset. If illegitimate opportunities do arise and they acquire the skills to take
advantage of them, deviance likely represents the least costly and most rational alternative to
conventionality. Moreover, such criminal precociousness is one of the strongest predictors of
persistence as I demonstrate in later section.
The family is a learning context and, easily, exposure to deviance in the family contributes
to the learning of necessary skills and motivations for deviance. Parent criminality disrupts family
functioning in more way than one. It obstructs the development of strong parent-child ties; more-
over, it might limit?or, in the case of incarceration, even remove?the capacity of one or both
parents to monitor and supervise their children (Farrington, 2000; Hirschi, 1983). Even more,
26
M. R. Gottfredson and Hirschi (1990) expect criminally involved parents will limit their involve-
ment in parenting; Sampson and Laub (1993) expect, if they are involved, their child rearing skills
are likely severely limited. So, in addition to providing an environment conducive to learning
criminality, deviant parents may also fail to inculcate safeguards promoting conventionality. This
likely explains evidence that those having criminally involved parents are more likely themselves
to become criminal (see, McCord, 1979; Robins, 1978).
Other early family aspects associated with later criminality are large family size (Farrington,
1997; Hirschi, 1994; and see, Gove, Hughes, & Galle, 1979; Sampson & Laub, 1993; Stark,
1987), family conflict and dissolution (Farrington, 2000; Kolvin, Miller, Fleeting, & Kolvin, 1988;
Thornberry, Smith, Craig Rivera and, & Stouthamer-Loeber, 1999; Wells & Rankin, 1991; cf.,
Kruttschnitt et al., 1986), and maltreatment (Dembo et al., 1995; Dembo, Williams, Schmeidler,
et al., 1991; Kazdin, 1995; Kruttschnitt et al., 1986; C. Smith & Thornberry, 1995; Widom, 1989).
Immediate family characteristics, such as marital quality and family involvement, are also
associated with o ending NSP. In general, such investments in conventional society insulate
against o ending (Farrington, 1989; Sampson & Laub, 1990; Sampson, Raudenbush, & Earls,
1997). Lower o ending levels are found among married versus non-married o enders (Horney et
al., 1995; Laub, Nagin, & Sampson, 1998; Sampson & Laub, 1990, 1993; West, 1982); yet, the
linking mechanism is somewhat cloudy.
For one, evidence suggests it is not merely being married but also the quality of the relation-
ship that matters (Laub et al., 1998; Sampson & Laub, 1993). Sampson and Laub (1993) contend
marital cohesiveness and strong bonds of a ection are more influential than is the status (and see,
Glueck & Glueck, 1950). More generally, marriage may only signify a more basic attribute, such
as self-control (M. R. Gottfredson & Hirschi, 1990; and see, Wilson & Hernnstein, 1985). Those
most prone to criminality may chose not to marry, be unable to find a willing mate, or be able to
27
sustain a marriage. Those most heavily involved in criminality may be unwilling to enter into or
sustain marriages. For instance, M. R. Gottfredson and Hirschi (1990) assert, to the extent such
relationships impinge on criminality, o enders likely eschew or abandon conventional ties. This
includes those with spouses and children. The institution itself may influence o ending in other
ways. For instance, rather than functioning through informal social controls, it could function
by insulating one from unconventional opportunities and learning environments (see, Warr, 1993,
1998).
Conceptual ambiguities notwithstanding, risk of NSP is lower among married versus non-
married o enders (Clarke et al., 1988; Landis et al., 1969; Morgan, 1993; Sims & Jones, 1997;
cf. Gray et al., 2001; Mayzer et al., 2004; Roundtree et al., 1984). Successful supervision per-
formance is thought to depend as much or more on having an agreeable marital climate, one
characterized by mutual trust and obligation, than simply on being married (Laub et al., 1998).
Sharing a residence with a spouse suggests an agreeable climate. Being married and cohabiting is
associated with reduced criminality. For instance, MacKenzie and Li (2002) found probationers
living with their spouses were less likely to commit non-drug related crimes than their counter-
parts. Among federal inmates, Harer (1994) similarly found married releasees cohabiting with
their spouses posed less recidivism risk than those living under di erent arrangements (see also,
Petersilia, 1985a).
Of the instant family factors, marriage and family relations are implicated in the production
of NSP. Given conceptual ambiguities, how precisely key processes are captured will surely dic-
tate predictive merit. Still, strained family relations and a diminished social support system likely
influence violations and termination modes. And even though relatively little research has exam-
ined early childhood factors among supervision populations, these too are likely salient factors in
anticipating NSP.
28
Health
Certain health-related characteristics anticipate performance while supervised. The two
dimensions considered here are physical and mental health. Aspects of both associate positively
with o ending and recidivism, but the strength of the evidence bearing on these latter is more
prevalent and thus more compelling.
Empirical findings suggest the less healthy and more injury prone are involved more so
criminally than their counterparts. To begin, the prevalence of certain communicable diseases is
higher among correctional populations (National Commission on Correctional Health Care, 2002).
In addition, there is evidence that injuries and accidents are more common among the crime prone
population (Shepherd, Farrington, & Potts, 2002). The injuries having the strongest relationship
with subsequent criminality are those resulting from assaults (Farrington, 1995).
Evidence?notwithstanding want?indicates less healthy people may pose higher supervi-
sion risks. Unfortunately, physical health and supervision performance have not been adequately
investigated. Implicitly measures of physical health and especially those related to assaultive in-
juries will vary with supervision performance. Although the evidence bearing on physical health
and o ending is limited, research on mental health and crime abounds.
An often disputed assertion is that those with mental disorders are more likely to engage
in criminal behaviors than those without such disorders (see, Link, Andrews, & Cullen, 1992;
Monahan, 1992). This is an extremely di cult relationship to confirm, yet one indeed meriting
attention. Research since the mid-1960s suggests those diagnosed with any of the major forms
of mental illness are more likely than their non-diagnosed counterparts to become crime involved
(Bland et al., 1998; Estro , Dackis, Gold, & Pottash, 1985; Johnston & O?Malley, 1986; Lin et al.,
1996; Link, Andrews, & Cullen, 1992; Link, Cullen, & Wozniak, 1992; Monahan, 1992; Teplin,
Abram, & McClelland, 1996; Teplin, 1990).
29
The Bureau of Justice Statistics (BJS) reports mental disorders tend to be fairly common
among corrections populations (Ditton, 1999). Roughly 16% each of state prisoners, jail inmates,
and probationers are mentally ill. Considering the research connecting it with criminality, there
will likely be a link between mental health and supervision performance although, as a risk factor,
its influence may be moderate at best compared to the e ect associated with other factors.
The health factors discussed here included physical and mental health. There is little reason
to ignore either in making risk assessments, though substantially more empirical work has been
devoted to the link between mental health and crime. Even so, there is reason to expect its predic-
tive ability will be overshadowed. Some of this, as alluded to earlier, stems from strong empirical
relationships with other personal characteristics. Indeed, the rest will likely stem from criminal
history and instant sentence characteristics that will likely bubble up as the most influential pre-
dictors. These are discussed next.
Criminal history
Discussion thus far has centered on o ender attributes. It shifts slightly here and centers
on their past behaviors and, particularly, any criminal justice responses these might have elicited.
I describe characteristics empirically associated with o ending and NSP beginning with earliest
of these, such as early and extensive involvement with the juvenile justice system, then moving
toward more recent characteristics, such as previous criminal justice involvement and behavior
while under custody, and qualities of the instant supervision sentence.
Past behavior is axiomatically the best predictor of future behavior. Behavioral patterns
from early childhood, through adolescence, and well into adulthood are often quite similar (see,
M. R. Gottfredson & Hirschi, 1990; Sampson & Laub, 1993), and researchers invariably find
previous delinquency and criminality are among the strongest predictors of future o ending (Irish,
30
1989; Roundtree et al., 1984). Nevertheless, in a well described paradox, most delinquents do not
become adult criminals (Robins, 1978). There are, however, early o ending characteristics that
well-discriminate likely persisting youths from their counterparts. In particular, those o ending
the earliest are most likely to persist and escalate; moreover, extensive delinquency involvement
foreshadows serious adult criminality (Baird et al., 1984; Dean, Brame, & Piquero, 1996; Laub &
Lauritsen, 1994; Piquero & Buka, 2002; Piquero & Chung, 2001).
This is true for NSP as well. For instance, Roundtree et al. (1984) found significant di er-
ences in probation outcome with respect to age at first arrest. They compared probationers having
been arrested at ages 18 and younger to those without such early arrests and found a larger propor-
tion of revokees among those with younger ages at first arrest. J. Austin and Litsky (1982) found a
similar pattern regarding absconsion among parolees and probationers. Likewise, MacKenzie et al.
(1992) found those younger at first exposure to the criminal justice system performed poorer while
supervised than did their counterparts. In addition, when examining a randomly sampled exit co-
hort of probationers discharged from probation in 1982 in Nassau County, New York, Irish (1989)
found probationers having the most extensive juvenile records were among those performing the
poorest both during and after community sentences. Silver and Chow-Martin (2002) similarly
found extensive juvenile records predict both rearrest and reincarceration. Taken together, these
findings suggest early and extensive involvement with the juvenile justice system will foreshadow
NSP.
Past arrests are often used predictors of later criminal involvement. Measured typically by
an item dichotomously capturing whether an o ender was arrested or not either ever in the past or
during some finite temporal window or by a measure summing the number of prior arrests. Given
the lowered standards of proof, arrests are only suggestive. Convictions, because of the more
stringent burden of proof, are arguably better measures (see, Maltz, 1984). They demonstrate that
31
not only was crime was committed but also that the given o ender was implicated. A high number
of previous convictions strongly indicates patterned criminality (Petersilia, 1985a; Quinsey et al.,
1998; Silver & Chow-Martin, 2002; Whitehead, 1991). In some studies, convictions dominate
among explanations. For example, M. Jones (1995) found the number of previous misdemeanor
convictions was one of the strongest predictors of probation failure (see also, Mayzer et al., 2004).
Similarly, Schwaner (1997) found having more adult and juvenile convictions was associated with
increased likelihoods of absconding (see also, J. Austin & Litsky, 1982; J. Austin, Quigley, &
Cuvelier, 1989).
Dispositions have unique predictive abilities of their own. For the most part, risk of NSP
increases as histories of incarcerations or community sentences lengthen. For example, Silver and
Chow-Martin (2002) found as the number of prior incarceration sentences increased, so too did
risk of recidivism. For instance, Schwaner (1997) found that parolees with lengthier incarceration
histories had greater chances of absconding. And chances of a absconding also increase as the
number of previous supervision sentences increase (J. Austin & Litsky, 1982).
Other criminal history aspects are also telling, such as specializing in particular crime types
and engaging in more serious types of o enses. Chronic property o enders, for instance, are
more likely to specialize in continued, property-related o ending than other crime types (Bartell
& Thomas, 1977; Petersilia et al., 1985). In fact, Bartell and Thomas (1977) found the number of
past arrests for burglary is one of the strongest predictors of probation failures. Similarly, Petersilia
et al. (1985) found property o enders returned to crime faster and more often than those whose
sentencing o enses included robbery or drug-related o enses (see also, Cunni , 1986; Irish, 1989;
McGaha, Fichter, & Hirschburg, 1987; Sims & Jones, 1997; Vito, 1987). The seriousness of pre-
vious criminal involvement is also important. This is usually measured by non-violent to violent
or misdemeanor to felony comparisons among arrests and convictions. In general, o enders with
32
histories of violence present lower risks of reo ending than do, say, chronic property o enders.
Morgan (1993), for instance, found histories of felonious o ending were linked to lowered proba-
tion success. Both Mayzer et al. (2004) and F. P. Williams III et al. (2000) found similar results.
Regarding performance while in custody, institutional and post-release behaviors are often
quite similar. Misconduct while incarcerated, for instance, is a strong predictor of subsequent
recidivism (Brown, 2002; Harer, 1994). Among released federal inmates, higher levels of insti-
tutional misconduct were associated with higher levels of post-release recidivism (Harer, 1994).
Also, Mayzer et al. (2004) found as the number of condition violations increased, so did the likeli-
hood of revocation and absconsion among supervised o enders. Further, the timing of violations
discriminated between these groups, where absconders were more likely to experience violations
early in supervision (Mayzer et al., 2004). Having previously failed while supervised in the com-
munity has been linked with later supervision performance. J. Austin et al. (1989), for example,
found those with previous parole or probation revocations were more likely to be rearrested and
to abscond during the instant sentence. One?s criminal justice status at the time of arrest is also
linked with NSP. For instance, federal inmates under community supervision when committing
the crimes for which they were sentenced were more likely to be rearrested upon release than their
counterparts not under community supervision (Harer, 1994).
With respect to the triggering o ense, both technical violations and continued o ending are
more likely among o enders sentenced to violent crimes. Bork (1995) for example found o end-
ers serving community sentences for robbery were more likely to violate probation conditions.
Gray et al. (2001) similarly found probationers serving sentences for assaultive crimes were more
likely to commit technical o enses and had a higher probability of committing new crimes. And
Harer (1994) found among federal prison releasees, those sentenced for crimes against persons
33
(i.e., robbery, homicide, manslaughter, sex o enses) had the highest rates of returning to crime
upon release; those sentenced for fraud or drug tra cking, the lowest.
Qualities of community sentences, such as intensity and treatment exposure, are also as-
sociated with supervision performance (see, Benedict & Hu -Corzine, 1997; MacKenzie, 1991;
Mayzer et al., 2004; Rhodes, 1986). Benedict and Hu -Corzine (1997) found, at least among
whites, the more intensive and rigorous the supervision, the higher was the risk of failure. Longer
and more stringent probation requirements were linked similarly by Mayzer et al. (2004) with
increased risk of later absconsion or revocation. And, MacKenzie and Souryal (1994) note in-
creased levels of both self- and o cially-reported crimes during months where the supervision
o cers contacted the families and employers of the o enders. Much recidivism research indi-
cates risk increases as sentences lengthen (Kronick, Lambert, & Lambert, 1998; MacKenzie et al.,
1992; Morgan, 1993; Roundtree et al., 1984; Sims & Jones, 1997; cf., Benedict & Hu -Corzine,
1997). Kronick et al. found years of sentence to be associated with revocations among parolees,
where longer sentences increased the likelihood of violating. Also, examining the months elapsed
until community sentence ended, Sims and Jones found those having longer terms had higher
probabilities of failing. Similarly, MacKenzie et al. found those with longer community sentences
adjusted more poorly than their counterparts. There is also evidence suggesting longer community
sentences may enhance the likelihood of absconding (McReynolds, 1987).
Treatment exposure has been linked with NSP as well. Minor et al. (2003) for instance,
found o enders ordered to mental health treatment were more likely to subsequently violate con-
ditions. This may reflect similarities among o enders sharing such orders, something that places
them as a group at higher risk. It might also be that the conditions were so demanding that it nearly
assured unattainability. On the other hand, MacKenzie et al. (1998) found among those ordered to
drug treatment, self-reported crime increased in months where treatment was missed.
34
Throughout this section I highlighted the considerable stability in behaviors over time and
described how the relationships between past behaviors and their consequent responses will influ-
ence NSP. In particular, those with the highest risks of NSP are likely the ones with the earliest
exposure to and most extensive involvement with the juvenile justice system. The salience of
adult arrests and, more importantly, convictions and dispositions was introduced, with the bulk
of research suggesting those with lengthier conviction and custodial histories will likely pose the
highest risk of NSP. Patterns within these histories, such as chronic property o ending, engaging
in more serious o ense types, and custodial misconduct, also anticipate NSP. Finally, evidence
bearing on aspects of the instant sentence, such as triggering o ense, intensity, and treatment ex-
posure, were discussed with the weight of evidence suggesting these factors are salient for risk
assessments.
Contextual predictors
A compositional explanation of local crime sees increases developing in areas dispropor-
tionately comprising residents manifesting characteristics known to correlate with criminality and
NSP, such as those introduced in the foregoing section. Another side sees the behaviors of res-
idents not as a cause but rather a consequence of environmental characteristics. I introduce a
few of these in this section. These are, namely, local sociodemographic, economic, crime, and
commercialization characteristics.
Sociodemographic and economic characteristics
Sociodemographic and economic characteristics associated with crime and NSP include
(a) dense population, (b) high racial and ethnic heterogeneity, (c) disproportionate age structure,
(d) high family disruption, (e) residential instability, (f) unemployment, and (g) income inequality.
35
I begin this section with a discussion of population density and its expected influence on social
interaction, the prevalence of deviant values, and the development of moral cynicism.
Generally, o ending increases commensurately with relative population (Galle, Gove, &
McPherson, 1972; Freedman, 1975). Such densities vary positively with specific crime forms,
such as violent (Kposowa, Breault, & Harrison, 1995; Sampson & Lauritsen, 1994; D. A. Smith
& Jarjoura, 1988; Stack, 1983) and property (Jackson, 1984; Schuerman & Kobrin, 1984; Stack,
1983). As area populations become inordinately dense residents may be more able to cloak them-
selves in anonymity making for less personal social interactions undermining the development of
the strong informal social ties thought to exist in cohesive neighborhoods (Sampson et al., 1997).
Residents are less willing to participate in community organizations, monitor their neighbors, su-
pervise local children, intervene when crimes are committed, or request assistance from the police
which, in turn, diminishes collective e orts to maintain crime-free communities. At the same
time, increases in population densities strain public resources, which, itself, undermines the crime
inhibiting force otherwise felt through formal mechanisms of control (Sampson et al., 1997).
The environment structures the content of social interactions and the presence of crimino-
genic attitudes and definitions are unavoidably high when relative populations are dense (see,
Akers, 1998). It is not just that denser areas tend to have more deviant residents and thus a higher
presence of prodeviant definitions, as Stark (1987) contends they do, but also that the high con-
centration of people in places leads to a heightened awareness among the residents about one
another (see also, Sampson & Raudenbush, 2004). It is simply harder, Stark contends, to con-
ceal deviance when others so easily witness or learn of morally discreditable behaviors. Sims and
Jones?s (1997) and Harer?s (1994) research, for instance, demonstrate that as the relative size of
the population in which o enders live or are released into increases, success while supervised be-
comes less likely. In some studies, relative population emerges as one of the strongest predictors
36
of recidivism among supervised o enders (Silver & Chow-Martin, 2002). Inextricably connected
to dense populations is an expected increase in racial and ethnic diversity. As I discuss next, this
too is a likely influence on NSP.
Heterogeneity?typically operationalized as the relative number of Black or non-white res-
idents within the population (Bursik, 1986; Sampson, 1986)?is regularly linked with regional
crime and o ending patterns (Harries, 1974; Messner, 1983). Certain forms of crime appear more
susceptible to such shifts. Areas with higher concentrations of Black residents, for example, tend
to have higher rates of violent crime (Sampson, 1986; Sampson & Raudenbush, 2004). Some
have explained these shifts as culturally-induced phenomena (Wolfgang & Ferracuti, 1967; and
see Curtis, 1975; Fischer, 1978; Gastil, 1971; Hackney, 1969; Loftin & Hill, 1974; Messner,
1983; Silberman, 1978). Others contend heterogeneity impedes the establishment of common
values and that cultural conflicts or inconsistencies underlie di erential o ending patterns (Sellin,
1938; D. A. Smith & Jarjoura, 1988; Wirth, 1956/1964).
Given evidence linking age with o ending, it is not surprising that crime rates emulate size
changes in the proportion of teens and young adults composing the population (Loftin & Hill,
1974; Pogue, 1975; Wilson & Hernnstein, 1985). Regardless of their ages, though, o enders
living in areas disproportionately comprising teens and young adults might be at higher risk of
NSP. This is thought to be influenced less so by the direct e ect of the disproportionate presence
of a high crime group and more so by the indirect influence it has on other community processes
and, ultimately, on the strength of crime inhibiting social controls. It might influence, for instance,
family structure, mobility patterns, and employment and income characteristics.
Family structure is typically conceptualized in terms of either generalized intactness, often
measured by the divorce rate, or in terms of control-contributiveness, often measured by the pro-
portion of single- and female-headed households. A high prevalence of family disorder is thought
37
to undermine community e orts to realize common goals, like staving o encroaching crime and
disorder. Sampson (1986) has shown that these areas have weakened non-coercive means of self-
regulation because participation levels in networks such as friendship and community organiza-
tions, thought to mediate formal controls, are less prevalent; those existing, less forceful (and see,
Sampson & Raudenbush, 2001). Supervised o enders residing in areas characterized by high
family disruption, because they are bound less by social control mechanisms are freer to engage
in NSP than those living in areas characterized by greater family cohesion. Not too distant from
this control-theoretic framework is the notion that family structure influences routine activities. In
particular, it typifies an aspect of guardianship (see, L. Cohen & Felson, 1979). Single-headed
compared to partnered households contribute at most half as much guardianship potential. With
only one family member there will be less contact with other residents, less supervision exercised
over peer groups, and less intervening for the sake of community goals. Consistent with this,
researchers have shown that as the proportion of divorced residents increases, o ending does as
well (Choldin & Roncek, 1976; Land, McCall, & Cohen, 1990; Sampson, 1986). At times it exerts
considerable unique impact (e.g., Messner & Tardi , 1986).
Neighborhood transience is a related concept that is also likely implicated in the production
of NSP. Its expected e ects are similar to those stemming from disrupted families in that both
are thought to undermine non-coercive means of informal social control. It has typically been
measured as the proportion of residents living in the same dwelling for the last five years or as
a function of the ratios of renters to owners (e.g., Heitgerd & Bursik, 1987; Sampson, 1986;
Sampson et al., 1997). It is an often included structural covariate on the grounds that high levels
are thought to disrupt primary relationships and inhibit the development of strong institutional
ties. It weakens community integration by limiting friendship and organizational participation
and, as a consequence, by reducing collective regulatory e orts (Bursik & Grasmick, 1993). High
38
transience areas will be less able to secure and mobilize resources and will thus be less able to
favorably a ect policy, such as local business and policing practices. Likewise, residents are
less likely to develop and sustain strong attachments and they are less likely to intervene for the
common good. Residents of areas having high levels of mobility tend to have the highest o ending
levels; likewise, areas characterized by high levels of mobility tend also to have the highest rates
of o ending (Sampson & Lauritsen, 1994; D. A. Smith & Jarjoura, 1988). NSP is also expected
to vary with local economic factors as they, too, have shown to vary with regional crime and
o ending rates.
Regional unemployment rates vary positively with o ending (see, Cantor & Land, 1985;
Carroll & Jackson, 1983). The evidence bearing on this issue, however, is of uncertain significance
as many studies have been unable to isolate unemployment-o ending e ects (e.g., Danziger &
Wheeler, 1975) and few studies have examined the impact of local unemployment on NSP. An
exception is Harer?s (1994). He found as the size of unemployed residents in a region grew, the
likelihood of recidivism among parolees released into those areas increased (see also, Kubrin &
Stewart, 2006). More compelling research links crime, at least certain forms, with aspects of
relative wealth, such as income inequality. I describe this next.
Wealth is unevenly distributed across society and, largely, as it becomes more disparate
regional crime rates increase. Some find, for instance, residential areas having the highest aggre-
gated incomes tend to have relatively lower crime rates (LaFree & Drass, 1996; Sampson, 1986;
M. D. Smith & Parker, 1980) or, similarly, that areas having higher concentrations of lower income
residents tend to have higher crime rates (Crutchfield, Geerkin, & Grove, 1982; D. A. Smith &
Jarjoura, 1988). Some studies, however, have either been unable to identify a relationship between
wealth and o ending (Messner & Tardi , 1986; Patterson, 1991) or have found contradicting evi-
dence (Messner, 1983; Rosenfeld, 1986).
39
Insight anent this uncertainty is gleaned from studies disaggregating the e ects of income
di erentials by crime types. Income inequality is associated with higher levels of violent o ending
(Loftin & Parker, 1985; Messner & Tardi , 1986; Patterson, 1991; Taylor & Covington, 1988)
and, to a lesser extent, property o ending (Crutchfield et al., 1982).
It is thought then that NSP will be more common among those probationers residing in
densely populated areas, as these areas increase anonymity and impersonalized social interaction,
exposure to deviant values, and moral cynicism. Higher levels of NSP are also expected among
probationers whose residential areas are characterized by high racial and ethnic heterogeneity, a
disproportionate presence of teens and young adults, a disrupted family structure, and residential
instability. Finally, NSP is expected more so among those residing in areas having high unem-
ployment levels or income inequality than their counterparts. Another likely contextual influence
on NSP is the presence of local crime. I discuss relevant literature and research in the next section.
Objective crime measures
Ample research indicates criminal activities cluster in place and time (e.g., Sherman, Gartin,
& Buerger, 1989). Compositional explanations suggest this reflects the spatial distribution of pre-
disposed individuals (e.g., Wilson & Hernnstein, 1985), but environmental characteristics are also
likely implicated. Mentioned thus far are certain sociodemographic and economic characteris-
tics. The argument continues by pointing out likely influences on NSP stemming from local crime
levels. I begin by introducing a simple and, in most cases, rightfully avoided explanation, that
residents of high-crime communities are also those that are responsible for a large fraction of the
local crime, then provide reasons for not wholly avoiding ecological inference. I then describe di-
rect influences on NSP expected from exposure to criminal activities as well as indirect influences
40
expected through stigmatization, increased crime tolerances, and weakened formal and informal
controls.
A simple explanation posits those residents of high-crime communities are also those that
are responsible for a large fraction of the local crime. This is presumptuous. Moreover, inferences
such as these invariably knuckle under criticisms: they assume each resident within a community
shares those characteristics represented by the community as a whole. This fallacy is widely
known and, for inferential work, should be approached cautiously. On the other hand, these are
much less prohibitive concerns in risk assessments. Not all residents of an area are criminally
active. Rather, a small group of active o enders likely accounts for a large fraction of crime.
There is persuasive evidence suggesting supervised o enders, as they have each been implicated
criminally in one way or another, are members of this small but active group.
Although a non-ignorable amount of crime in an area is committed by outsiders (Rand,
1986), Bursik and Grasmick (1993) show o enders commit a large fraction of their crimes within
their own communities. It could reflect, as Reiss (1986) suggests, o enders seeking to minimize
costs associated with o ending including those associated with target selection and rarely venture
into unfamiliar areas to o end (and see, Boggs, 1965; Carter & Hill, 1978; Gould, 1969; Reppeto,
1974). There is reason then to expect that, first, at least some neighborhood crime is generated by
its own residents and, second, that supervised o enders are likely among the criminally active.
More complex explanations turn on direct influences from exposure to criminal activities
and indirect influences through stigmatization, increased crime tolerances, and weakened formal
and informal controls. Community crime greatly a ects the likelihood of learning criminal behav-
iors. Akers (1998) argues the clustering of crime reflects the extent sociocultural traditions and
control systems provide learning environments conducive to deviance. Structural characteristics
underlie o ending variabilities, but, importantly, they do so by a ecting the process by which indi-
41
viduals learn to commit or refrain from criminal acts. Residents of crime-ridden areas more likely
witness or hear of law violations than those living in less crime-ridden areas and the perceptions
residents attach to these behaviors influences their own. How rewarding or justifiable one sees
criminality, how certain one believes the consequences, and how strongly one identifies with those
committing or espousing definitions favorable to crime each influence the probability of engaging
in criminal behavior (Akers, 1998). Thus, the level of crime in an area might influence subsequent
NSP directly by providing a learning context conducive to criminal behaviors. Indirect e ects are
also likely.
High crime rates stigmatize communities and degrade the moral standing of its residents and
potentially undermines strong social ties (Stark, 1987).They also influence tolerances for deviance,
which vary regionally. There may be areas where crime flourishes, areas where it dies o quickly,
and areas where it somehow never germinates. Deviant patterns somehow come to dominate
social interactions. It may be, like the incivilities hypothesis suggests (see, Kelling & Coles,
1996; Taylor, 1999; Wilson & Kelling, 1982), disorder creeps into areas, most likely those with
insu cient social capital to resist its growth, and eventually overwhelms them. Left unchecked,
disorder begets continued disorder and, ultimately, makes conditions ripe for rising crime.
This stigma may also influence local law enforcement practices. As areas become more
crime entrenched, the law becomes di erentially enforced. Police, knowing elimination is un-
likely, might come to tolerate higher levels of crime in some areas hoping at the very least for
containment (D. A. Smith, 1986). This sets o an enduring trend of localization of both o end-
ing and o enders: it signals criminality?s general tolerance, entices motivated non-residents, and
compels conventional residents to disinvest (Stark, 1987).
Local crime patterns may thus prove critical in assessing supervised o enders for risk of
NSP. Being the likely participants of community crime and being surrounded by negative stigma,
42
criminal attitudes and definitions, and high tolerances for crime, puts a mechanistic explanation
behind why those from high crime areas might be higher risks of NSP than those residing in com-
munities with lower crime areas. Another regional aspect deserving attention is commercialization
patterns, which I discuss next.
Commercialization patterns
This section describes literature and research bearing on commercialization patterns and
o ending then draws inferences regarding how these characteristics might contribute to NSP. The
focus is on the relative densities of businesses within geographic space and, because they are
thought to uniquely influence NSP, I make the distinction between businesses of a general nature
and those primarily involved in the sale or distribution of alcohol.
There are intrinsic, crime generating characteristics in areas wherein residential units are
coexistent with or adjacent to commercial areas. Such mixed-use areas manifest higher rates of
deviance, disorder, and crime, and this is mainly attributed to its influence on social control mech-
anisms, criminal attitudes and opportunities, and perceptions of neighborhoods and residents.(see,
Kelling & Coles, 1996; Reiss, 1986; Sampson & Raudenbush, 2001; Skogan, 1992; Stark, 1987;
Taylor, 1999; Wilson & Kelling, 1982).
As residential areas begin taking on a more commercial constitution a higher proportion of
non-residents intermingle with residents who, as anonymity increases, become less able or willing
to exert control over and thus contribute to order maintenance (Sampson, 1986, 1988). Mixed-use
areas invite a greater blend of people than otherwise expected were it not for the commercial draw.
This overwhelms formal means of control (Sampson & Raudenbush, 1999, 2001).
They also tend to have denser, less stable, and less economically advantaged populations
and, by most accounts, are the least desirable and most a ordable residential areas (Stark, 1987).
43
Tending toward dilapidation, this feature is adopted by the residents and ascribed by those with
whom residents interact (see, Sampson & Raudenbush, 2004; Stark, 1987). Those residing in
mixed-use areas have fewer reasons to conform; those ascribing, fewer reasons to interact. Mixed-
use areas likely elevate exposure to deviant people and patterns. Exposure to deviance greatly
a ects the likelihood of learning such behaviors. Its residents are more likely witness or hear
of deviance; this can influence the probability of engaging in criminal behavior (Akers, 1998).
Be they patrons, the businesses themselves, or even the architectural design of the areas, mixed-
use areas also present more criminal opportunities. They provide motivated o enders with easily
accessible and attractive targets and living close to these areas provides them with intimate knowl-
edge of areas having the highest criminal opportunities and the lowest risk of detection (Stark,
1987).
There is reason to believe an alcohol-related component of some businesses will further
influence the production NSP. Crimes, many of which are alcohol-related, cluster in time and
space and much of this coincides with densely available alcohol (see, Cochran, Rowan, Blount,
Heide, & Sellers, 1998; Costanza, Bankston, & Shihadeh, 2001; D. M. Gorman, Speer, & Grue-
newald, 2001; Gyimah-Brempong, 2001; R. Lipton & Gruenewald, 2002; Scribner, Cohen, Ka-
plan, & Allen, 1999; Scribner, MacKinnon, & Dwyer, 1995; Sherman et al., 1989; Speer, Gorman,
Labouvie, & Ontkush, 1998; Stitt & Giacopassi, 1992).
A greater presence of alcohol retailers indulges non-abstainers. Unfettered proclivity and
near-ubiquitousness increases their chances of consumption which in turn raises the potential for
uncharacteristic behaviors (see, Gruenewald, Ponicki, & Holder, 1993; cf. Fitzgerald & Mulford,
1993). Nonetheless, alcohol-related problems are simply not explained fully via pharmacological
e ects `a la disinhibition (see, Room & Collins, 1983). Rather, they hinge on the combination of
44
these e ects with situational and sociocultural characteristics (see, Bushman, 1997; Fagan, 1990;
Gustafson, 1994; Parker & Rebhun, 1995; Reiss & Roth, 1993).
Alcohol-related businesses increase criminal opportunities. Some o enders might savvily
prey on intoxicated victims. Roncek and Maier (1991), for instance, highlight the expectedness
that patrons may have cash or other desirable items and, further, that intoxicating e ects may
diminish the abilities of these patrons to protect themselves and others. Social norms delineate ac-
ceptable behavior with respect to alcohol consumption and intoxication (Linsky, Colby, & Straus,
1986; Parker & Auerhahn, 1998; Skog, 1985; Wiseman, 1991). These norms may encourage or
discourage either (Parker, 1993). As MacAndrew and Edgerton eloquently state, ?The way people
comport themselves when they are drunk is determined not by alcohol?s toxic assault upon the
seat of moral judgment, conscience, or the like, but by what their society makes of and imparts
to them concerning the state of drunkenness? (1969, p. 165). An unmixed presence of alcohol-
related businesses may signal to residents and outsiders alike that consumption and intoxication
are acceptable if not encouraged and the greater visibility of taverns, bars, and liquor stores?and,
naturally, of consumption and intoxication?minimizes alcohol-related stigma.
To iterate, the goal here is developing a risk-based instrument for guiding immediate secu-
rity and treatment decisions facing the CSOSA, and, up to this point, I have contextualized such
instruments within corrections research, described their historical trends, and sketched out the cur-
rent methods. I then spent the bulk of the chapter on conceptual justifications for including certain
individual- and structural-level characteristics as part of this instrument. These include, in particu-
lar, age, education, employment, residential stability, substance use, childhood and family factors,
health, and criminal history as well as contextual characteristics such as sociodemographic and
economic characteristics, objective crime measures, and commercialization patterns. In the next
chapter I describe the methods and procedures used in the present study.
45
METHODOLOGY
The present task is describing the linkage between the methods and procedures used in
developing an instrument for informing immediate, risk-anticipated security and treatment as-
signments among community-supervised o enders in the District of Columbia and the literatures
and researches discussed in the previous chapter. It begins with the research design. To preface,
measures will not and, for that matter, rightly cannot be manipulated. Nor are any treatments or
interventions implemented. Of interest instead are the population relationships subordinated in
the production of negative supervision performance (NSP). The correlational design used here is
appropriate as it lends itself to explorations of such complex systems and aids in disentangling the
relative importance of involved predictors.4 What follows is a description of the participants, the
data sources and measures, and the procedures.
Participants
Appropriate samples in assessment studies are representative of the population for whom
inferences are made. This di ers, naturally, from samples representative of the general population
(see, S. D. Gottfredson & Gottfredson, 1986), and it suggests those for whom estimates apply are
unambiguous (see, J. M. Chaiken, Chaiken, & Rhodes, 1994). Considering these two points the
population for this study was confined to only the most typical probationers yet to be supervised
4These designs are nevertheless weaker than, say, experimental or prospective, longitudinal designs. Internal validity
can be questionable and the magnitude that observed rather than omitted predictors influence variation in the criterion
can be di cult, if not impossible, to unravel. Likely, the most momentous consequence is the inability to infer causation.
Correlational designs merely attest to associations; neither can antecedent or consequential relationships be identified
nor can assurances of nonspuriousness be drawn. Despite these weaknesses correlational designs are attractive (see,
Holmes & Taggart, 1990). One reason for this centers on their high level of external validity as, assuming appropriate
sampling and modeling techniques, derived conclusions tend to generalize well. Their potential contributions to both
theoretical literature and predictive research are also attractive?particularly when limiting predictors to those that are
theoretically causal. The extent to which predictors and criteria covary is estimable and findings, in turn, inform both
theory and research.
46
by the Court Services and O ender Supervision Agency for the District of Columbia (CSOSA).5
Because it was a theoretical population an enumeration was impossible. Estimates were instead
derived from observed behaviors among a sample of probationers having already served their
sentences. In the next few passages I describe the procedures that ensured, as much as possible,
the sample adequately represented the population.
The underlying goal was guiding immediate, risk-anticipated custodial and treatment de-
cisions facing the CSOSA by developing an instrument applicable to the majority of incoming
o enders. The majority of o enders supervised by the agency are Black males having been sen-
tenced to regular probation;6 the sampling frame was thusly restricted to this group and, because
of this, the instrument developed here provides a general assessment tool.
Such precautions, while limiting, especially cross-jurisdictional, generalizeability, reduced
variability in and thus size requirements for the sample. This stems from the constraining e ect on
the number of covariate patterns.7 At the same time they also militated against both the presence
of and the techniques for accommodating missing data: o enders sentenced to the most common
forms of supervision tend to have fewer data inconsistencies than those sentenced to special and,
especially, less intensive forms of supervision.
5The CSOSA is an independent federal executive branch agency that was established underx11232 of the National
Capital Revitalization and Self-Government Improvement Act of 1997. Its mission is to increase public safety, prevent
crime, reduce recidivism, and support the fair administration of justice in close collaboration with the community. It
comprises the Pretrial Services Agency and the Community Supervision Programs which, together, provide pretrial
and community supervision services to over 15,000 DC residents serving probation, parole, and supervised release
sentences.
6The term ?regular? defines a specific, indeed, the most common, sentence among those supervised by the CSOSA.
Compared to sentences typical among ?special? sentences, such as those imposed on sex o enders, those with mental or
physical disabilities, or those supervised as part of the interstate compact?none of which are included in this study?
regular sentences stand out in their generality. Demographic and supervision level data describing the 2004 exit cohort
are available upon request.
7These and other technical points are discussed further in a later section.
47
The sampling frame was also restricted to only those having already terminated sentences.
An exit cohort enabled observation at least until sentences terminated, and, as emphasized previ-
ously, the processes underlying how probationers ultimately concluded their sentences was central
to this assessment.
Another restriction relates to exposure. Focal criteria occurred largely within the observa-
tion period which comprises the supervision period, which is, formally, the interval beginning on
the first day of an o ender?s sentence and ending on the last day of the sentence, and the post-
supervision period, the interval beginning on the day after the last day of the sentence and ending
on the last day of observation period.
To maximize exposure, the lower limit of the observation period reached as far as possible
into agency history and, as it turned out, this reach was bounded by the second issue: data quality.
As it coordinated new with existing functions the CSOSA felt growing pains in several areas. Data
integrity was one.
A migration from an early to their current case management system was attempted during
which the agency su ered isolated yet substantial data loss; information predating this migration is
of dubious quality. As such, their analyses would be dicey. Direct population of data in the current
case management system, Supervision and Management Automated Record Tracking (SMART),
began regularly by the first few months of 2002, and, given an average sentence for Black, male
probationers hovering near 2 years, the exit cohort selection parameter had to be at least 2 years
after this to adequately capture data from the beginning of the supervision period.
To minimize risks associated with problematic data and at the same time allow for the
longest possible follow-up, probationers were selected from among those terminating sentences
during the interval beginning January 1, 2004, and ending December 31, 2004. They were fol-
48
lowed until December 31, 2006. This allowed between 23 and 35 months of post-supervision
observation.
Two related qualifications follow. The first bears on sample size requirements; the second,
case selection. Risk assessments require large samples for optimal performance and, while obvi-
ously subjective, approximate guidelines defining large exist nonetheless. Some suggest at least
500 cases are required (e.g., P. R. Jones, 1996). This assumes, however, a general population.
As I discussed earlier, because the studied population was relatively homogenized, a smaller, ran-
domized sample was adequately representative. Even more, the absolute sample size?at least
for the modeling strategies used here?is of lesser concern than, say, the number of events either
per covariate pattern or per parameter (Hosmer & Lemeshow, 2000; Peduzzi, Concato, Kemper,
Holford, & Feinstein, 1996). As a general rule, at least 10 events are needed per covariate pattern
(Hosmer & Lemeshow, 2000; Peduzzi et al., 1996). Here, the expected sample size was N  200
Black male probationers8 and I ensured the ratios of events to covariate patterns were within these
rules of thumb.
Models developed here were validated by bootstrapping, a validation technique outclassing
both of the more common validation approaches: data splitting and cross-validation techniques
(Efron & Tibshirani, 1993). Like cross-validation, bootstrapping is a resampling technique. It
involves estimating expected variability from numerous, random samples drawn with replacement
from the same, original sample (Efron, 1983; Efron & Tibshirani, 1993; S. D. Gottfredson &
Gottfredson, 1986; Linnet, 1989; Monahan et al., 2001). When bootstrapping regression models,
population parameters are estimated by first repetitively sampling observed data, with replace-
ment, estimating the parameter among each, then calculating intervals around statistics by pooling
8As this is comparatively modest given P. R. Jones?s (1996) suggestions, an explanation is necessary. Much of the
analytic work here relies on information derived from Presentence Investigation reports (PSIs), which, as I show in the
next section, do not readily lend themselves to analysis; incorporating more than a handful is unapproachable.
49
and averaging subsample estimates (Efron, 1979, 1982, 1983; Efron & Tibshirani, 1991, 1993;
Fox, 1997). Done this way, bootstrapping provides nearly unbiased estimates of predictive ac-
curacy, is more e cient than cross-validation, makes full use of available data, and allows for
post-estimation optimization adjustments (see, Harrell & Lee, 1985; Harrell, Lee, & Mark, 1996).
It is thus preferable among alternative measures of internal validity.9
The restricted population, randomized sample, and validation techniques o set, at least par-
tially, the concern regarding the relatively small sample. To iterate, the population is limited to the
the most typical o enders yet to be supervised in the District of Columbia by the CSOSA. These
are, namely, Black males sentenced to regular probation. Population estimates are bootstrapped
from characteristics observed among a random sample of roughly 200 probationers having termi-
nated their sentences during the interval beginning on January 1, 2004, and ending on December
31, 2004. With the description of the participants complete attention turns now to the remaining
methodological elements including the data, measures, and procedures, and I pick this up with a
description of the data sources and the measures derived thereof.
Data and Measures
Both individual- and environmental-level data were gathered.10 Individual-level data de-
scribe aspects unique to each probationer; environmental-level data describe contextual aspects
that are, potentially, shared among probationers.11
9Monahan et al. (2001) recently used this approach when examining n = 939 patients from the MacArthur Risk As-
sessment Study. Their criterion was serious violence in the community within 20 weeks of discharge. So as not to limit
the data available for analyses, they bootstrapped parameter estimates. This entailed constructing 1,000 subsamples
from their original data, fitting their model to each subsample, then summarizing across estimates.
10To link these two levels, the primary residence of each probationer was recovered from the CSOSA housing data
and then geocoded to a point within the x-y space defining the DC. Each of these points were in turn aggregated to the
U.S. Census Bureau (Census) block-group (BG) level defined by the Census, and it is precisely this level at which the
environmental data were summarized.
11At least, that is, among those living within the same BG. And, indeed, there were 75 o enders living in the same
BG as at least one other o ender.
50
Measures were obtained from three broad sources including (a) the CSOSA, (b) the Census,
and (c) local regulatory and criminal justice agencies. These sources along with the measures
derived from each are described next beginning with individual-level measures.
Individual-level measures
Individual-level measures were obtained from the CSOSA and included (a) the Risk-needs
Screener, (b) Presentence Investigation reports, and (c) Supervision and Management Automated
Record Tracking database.
Risk-needs Screener. The Risk-needs Screener (RNS) is an instrument originally designed
by the Community Supervision Services (CSS) and Community Justice Programs (CJP) o ces of
the CSOSA. It is described fully in Appendix A; the specific measures used here are shown in
Table 1. As readily seen, the RNS encapsulates many of the individual-level features described in
Chapter 2 to well-predict NSP such as age, educational level, and employment stability. It gropes
in the dark, however, when it comes to operationalizing these features. Because of this, I relied
heavily on data recovered from the PSIs.
Missingness was a relatively minor issue with respect to the RNS yet one still needing pre-
modeling attention. Three o enders were completely missing the RNS data; a 4th o ender was
missing a single RNS value?the item capturing original o ense rnsOOi?which was proxied
from the PSI. One of the 3 with completely missing screener data was simultaneously missing the
PSI completely. This o ender was dropped from analyses thus reducing the sample to n = 199
Black male probationers. Most items for the other two o enders were proxied from data in either
the PSI or the SMART database. I elaborate on this next.
51
Table 1
Description and representation of items comprising the CSOSA Risk-needs Screener.
Symbol Description Coding
rnsAgei Age at the time of assessment [0;100)
rnsEdui Educational level a
rnsS S Ni Significant relationships 0 = None, 1 = One, 2 = Two or more
rnsRLi Recent loss 1 = Yes, 0 Otherwise
rnsEmpi Employment changes/year b
rnsResi Residential moves/year b
rnsPVi Prior violent o ense 0 = None, 1 = One, 2 = Two or more
rnsNPAi Prior adult arrests c
rnsPS i Prior supervision failures 0 = None, 1 = One to two, 2 = Three or More
rnsFAi Frequency of arrests/year 0 = None, 1 = One, 2 = Two or more
rnsAFi Age at first arrest d
rnsPCi Prior convictions 0 = None, 1 = One to five, 2 = Six or more
rnsCS Ai Current substance abuse 1 = Yes, 0 Otherwise
rnsHS Ai Prior substance abuse 1 = Yes, 0 Otherwise
rnsCMDi Current mental disorder 1 = Yes, 0 Otherwise
rnsHMDi History of mental disorder 1 = Yes, 0 Otherwise
rnsImpi CSO Impression 1 = Low, 2 = Medium, 3= High
rnsLOCi Level of cooperation 1 = Fully, 2 = Non-, 3 = Restrained
rnsOOi Originating o ense 1 = Drug-related, 2 = Non-violent, 3 = Violent
rnsPDi Physical disabilities 1 = Yes, 0 Otherwise
a 1 = 10th or Below, 2 = 11th, 3 = High School or GED, 4 = Some college.
b 0 = Currently or recently incarcerated or in a shelter; 1 = Two or fewer; 2 = Three or more.
c 1 = Two or less, 2 = Three to four, 3 = Five, 4 = Six or more. d 1 = Ages 15 and younger, 2 = Ages 16
to 17, 3 = Ages 18 to 25, 4 = Ages 26 and older.
52
Specifically, for i = 1;2 o enders with missing educational data, education level rnsEdui
was replaced with
rnsEdui =
8>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>><
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
:
psiEduGrdCmi  10; 10th or below;
psiEduGrdCmi = 11; 11th;
psiEduGrdCmi = 12
or psiEduGEDi = true; HS/GED;
psiEduGrdCmi > 12; Some college;
where psiEduGrdCmi captures the highest grade completed and psiEduGEDi captures whether,
if not a high school graduate, a GED was earned.12
Significant relationships rnsS S Ni, having original levels of no relationships, relationship
with 1 person, and relationships with 2 or more people, was replaced with
rnsS S Ni =
8>>
>>>>
>>>>
>>>>
><
>>>>
>>>>
>>>>
>>>:
xi < 1; no relationships;
xi = 1; relationship with 1 person;
xi > 1; relationship with 2 or more people;
where xi represents the sum of the PSI variables capturing whether the PSI writer finds a sup-
portive social network psiFamS S Ni, whether the o ender has sustained contact with his mother
psiFamCntmi or father psiFamCnt fi until the instant arrest or, if either is deceased, until their
time of death.13
12Unless otherwise noted, values of true and false are coerced to integers as true
false7!10 throughout.
13Unless otherwise noted, values of yes and no are coerced to integers as yesno7!1
0 throughout.
53
Residential changes within the previous year rnsResi was replaced with
rnsResi =
8>>
>>>>
>>>>
>>>>
><
>>>>
>>>>
>>>>
>>>:
hi + ii  1; Currently/recently incarcerated;
xi > 2; 3 or more moves;
xi  2; 2 or fewer moves;
where hi and ii are calculated from SMART housing tables and indicate, respectfully, the instant
supervision period began within 30 days of either a discharge from a halfway house or a custodial
sentence; xi reflects the number of unique addresses for each o ender within a backwards 2-year
window available in the SMART housing tables.
Employment changes/year rnsEmpi was replaced with
rnsEmpi =
8>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>><
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
:
hi + ii  1; Currently/recently incarcerated;
psiEmpCurmi  12 or0
BBBB
BBBB
BBBB
BBBB
BBBB
BBBB
BBBB
B@
psiEmpCurmi < 12
and psiEmpLessi = FALSE
and psiEmpS tai
3fUnemployed, Erratic/Odd jobsg
9>>
>>>>
>>>>
>>>>
>>>>
>=
>>>>
>>>>
>>>>
>>>>
>>>;
; 2 or fewer jobs ;
psiEmpS tai 2fUnemployed, Erratic/Odd jobsg; 3 or more jobs/unemployed;
where hi and ii are calculated from the SMART housing tables and indicate, respectfully, the in-
stant supervision period began within 30 days of either a discharge from a halfway house or a cus-
todial sentence; psiEmpCurmi14 captures the number of months at the current job; psiEmpLessi
captures whether there are any jobs within the previous year with a duration of less than 30 days;
psiEmpS tai best characterizes employment status at time of instant o ense.
14The item psiEmpCurmi was dropped from the analyses as it was missing values for 111=199 = 0:56 of the o enders.
Luckily, it was non-missing for the 2 o enders missing rnsEmpi.
54
Prior violent o ense rnsPVi was replaced with
rnsPVi =
8>>
>>>>
>>>>
><
>>>>
>>>>
>>>:
vci < 1; None;
vci = 1; 1;
vci > 1; 2 or More;
where vci represents the sum of the number of adult convictions psiCrmAdlCnvVioi and juvenile
o enses psiCrmJuvVioi involving violence.
Prior adult arrests rnsNPAi, having original levels of 2 or less, 3 to 4, 5, and 6 or more, was
replaced with a categorical transformation of the PSI item capturing the number of adult cases
psiCrmAdlCasni, where
rnsNPAi =
8>>
>>>>
>>>>
>>>>
>>><
>>>>
>>>>
>>>>
>>>>
>:
0 psiCrmAdlCasni  4; 2 or less;
4 < psiCrmAdlCasni  8; 3 to 4;
8 < psiCrmAdlCasni  12; 5;
12 < psiCrmAdlCasni; 6 or more:
Prior convictions rnsPCi, having original levels of None, 1 to 5, and 6 or more, was replaced
with the PSI measure capturing the number of adult convictions psiCrmAdlCnvi,
rnsPCi =
8>>
>>>>
>>>>
><
>>>>
>>>>
>>>:
psiCrmAdlCnvi < 1; None;
1 psiCrmAdlCnvi  7; 1 to 5;
7 < psiCrmAdlCnvi; 6 or more:
Prior supervision failures rnsPS i, having original levels of None, 1 to 2, and 3 or more, was
replaced with
rnsPS i =
8>>
>>>>
>>>>
><
>>>>
>>>>
>>>:
psiS upRevi < 1; None;
1 psiS upRevi  2; 1 to 2;
2 < psiS upRevi; 3 or more;
where psiS upRevi captures the number of previous supervision failures.
55
Current substance abuse rnsCS Ai, recorded as either Yes or No, was replaced with the
minimum of either dtmsPosi or 1  rnsCS Ai = min(dtmsPosi;1) , where dtmsPosi represents the
sum of the number of positive drug screens taken 30 days before and after the screener interview.
Values greater than zero were mapped to Yes.
Prior substance abuse rnsHS Ai, also coded as either Yes or No in the screener, was replaced
with rnsHS Ai = min(psiEUi;1) where psiEUi represents the sum of the PSI variables capturing
whether the o ender admits to ever using alcohol psiS ubAlEUi, amphetamines psiS ubAmEUi,
cocaine psiS ubCoEUi, opiates psiS ubHeEUi, marijuana psiS ubMaEUi, opiates psiS ubOpEUi,
or PCP psiS ubPcEUi.
History of mental disorder rnsHMDi was replaced with the PSI item capturing whether the
o ender has been diagnosed with a mental illness psiMedMdDxi. If so, rnsHMDi = Yes.
Physical disabilities rnsPDi was replaced with
rnsPDi =
8>>
>>>>
<>
>>>>
>:
psiPDi  1; Yes;
psiPDi  1; No;
where psiPDi is the sum of the number of disabilities psiMedDisi and injuries psiMedIn ji.
Current o ense rnsOOi, having original levels of drug-related, non-violent, or violent, was
replaced with
rnsOOi =
8>>
>>>>
>>>>
><
>>>>
>>>>
>>>:
Oi = any violent; Violent;
Oi = Otherwise, any drug-related; Drug-Related;
Otherwise; Non-violent;
where Oi represents an item taken from SMART categorizing > 2000 o ense codes into 1 of
roughly 30 broad o ense groups.
Proxies were unavailable for the RNS items capturing recent loss rnsRLi, level of cooper-
ation during the interview rnsLOCi, current mental disorder rnsCMDi, age at first arrest rnsAFi,
56
frequency of arrests/year rnsFAi, and the impression of risk on the o cer administering the inter-
view rnsIMPi. All were imputed using methods described next.
Missing values for rnsRLi, rnsLOCi, rnsCMDi, rnsAFi, rnsFAi, and rnsIMPi were es-
timated and imputed using random draws from the conditional distributions of the nonmissing
values on each given the values across the other screener variables. Specifically, 5 imputes were
derived from random draws from the conditional distributions of the nonmissing values of each
target measure given the values across the other variables. This resulted in a 5-length vector of
imputes for each probationer-value. This vector represents, for each o ender, the ?best guess? esti-
mate of the true value with an added stochastic component. This random residual is added in such
a way that conditional variances for the target variable are comparable to those of the nonmissing
values. For descriptive purposes the average of these imputes are reported; when modeling, each
impute is used and the resulting coe cients and standard errors are adjusted for imputation.
There were further adjustments made to the RNS variables for all i = 1;2;:::;N probation-
ers. For example, The items rnsOOi and rnsPDi were excluded in favor of measures collected
from the PSIs. The items rnsResi and rnsEmpi were coded near-identically?each having three
levels, the first for both capturing whether the o ender was currently or recently released from
incarceration or was residing in a shelter at the time of the screening. As including both would
likely introduce redundancy and needlessly absorb degrees of freedom, these two items were col-
lapsed into one summary measure rnsS tai. The item rnsS tai reflects whether either of rnsResi
or rnsEmpi indicated the o ender was currently or recently released from incarceration or was
residing in a shelter at the time of the screening, in which case rnsS tai 7! 0; both rnsResi and
rnsEmpi indicated the o ender had experienced 2 or fewer changes in either condition, in which
case rnsS tai 7! 1; and, otherwise, if either of rnsResi or rnsEmpi indicated the o ender had
experienced 3 or more such changes, then rnsS tai 7!2.
57
The two items capturing current rnsCS Ai and past rnsHS Ai substance abuse were signif-
icantly related.15 A large proportion (167=199 = 0:84) of o enders admit prior substance abuse
as compared their counterparts (32=199 = 0:16). This imbalance was less pronounced as it con-
cerned current substance abuse where the proportion of o enders admitting abuse (78=199 = 0:39)
was considerably smaller than that among those not admitting (121=199 = 0:61). These two items
were collapsed into a single measure capturing whether the ith o ender had a history of substance
abuse. If not, rnsDrgi 7! 0. Otherwise, if the o ender did have a history of substance abuse but
no indication of current substance abuse then rnsDrgi 7! 1; if the o ender had both a history of
and indications of current substance abuse then rnsDrgi 7!2.
A similar reduction was used for the items capturing a history of rnsHMDi and a current
rnsCMDi mental disorder. Here, though, the o enders having either condition was extremely rare:
10=199 = 0:05 reported a history of and 6=199 = 0:03 a current mental disorder. A single indicator
was created capturing whether the ith o ender had either a history of or a current mental disorder
and, if so, rnsMHi 7!1; Otherwise, 0.
The item rnsLOCi was collapsed into a dichotomous indicator rnsFullCoopi. Most of-
fenders (170=199 = 0:85) were classified as fully cooperative; very few as either noncooperative
(6=199 = 0:03) or restrained (23=199 = 0:12). Given this, the last two levels of rnsLOCi were
combined resulting in an indicator of whether o enders were fully cooperative or not.
The last two levels of rnsNPA were collapsed thus dividing the sample into probationers
with 2 or less 68=199 = 0:34 3 to 4 45=199 = 0:23 or 5 or more 86=199 = 0:43 prior adult arrests.
Similarly, the last two levels of rnsPS were collapsed thus dividing the sample into probationers
having none 120=199 = 0:60 versus 1 or more 79=199 = 0:40 prior supervision failures.
15 2 = 8:8893; d f = 1; p = 0:00.
58
Presentence Investigation report. The PSIs provide an unmatched picture of personal and
social aspects and the most comprehensive description available of both the triggering event and
the criminal and supervision histories.16 Typically they describe (a) general and demographic at-
tributes; (b) information about the instant o ense; (c) the defendant?s statement about the instant
o ense; (d) criminal and supervision histories; (e) educational, vocational, and employment char-
acteristics; (f) family and social backgrounds; (g) health profile; (h) substance use, abuse, and
treatment histories; and (i) both sentencing recommendations and any notable features that might
inform such decisions. Fortunately, all PSIs17 are securely stored on the agency network. Their
unseemly format, however, makes incorporating more than a handful unapproachable: they are
literally disjoined from remaining agency functions, authored largely without content or structure
prescription, and are thus less conformable to warehousing than those data derived from, for exam-
ple, surveys or realtime data. They are seen primarily as output and, once complete, are essentially
buried in the agency network.18 This overlooks an opportunity for informing agency decisions.19
Their richness warrants recovering as much information as feasible. To this end, an instru-
ment was developed for extracting the most common PSI features that have also been shown to
vary with NSP. In the next few passages I describe the specific items obtained from the PSIs.20
16For most sentenced o enders the CSOSA provides the sentencing authority with a PSI. Because of their intensive-
ness, these reports are not typically ordered for minor o enders.
17At least those ordered by the sentencing authority on or after the first few months of 2002
18PSI authors save reports in a subfolder within the agency network share identified, typically, by the author?s last
name. The reports themselves are often either Microsoft Word documents or Adobe Portable Document Formats with
filenames comprising case numbers or other identifiers.
19There are, of course, sophisticated text-mining applications capable of squeezing out much of the information
contained in these reports. There are no plans for using these procedures here. The reasoning stems from, first, a
project that focuses precisely on extracting PSI information is concurrent with the present study. Any e orts done here
would be redundant?and, likely, inferior. Second, there are plans to integrate the PSI authoring and recording within,
or closer to, SMART. Once implemented, the reports will be more easily incorporated into quantitative studies.
20There were originally close to 200 items extracted from the PSIs, but roughly half were either completely missing
or constant across all o enders and were thus excluded.
59
Two sets of 6 items were obtained from the PSIs to capture whether the ith o ender ever
used ( j = 1) and, if so, admitted to a problem with ( j = 2) each K substance
k =
8>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
<>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>:
1; alcohol;
2; amphetamine;
3; cocaine;
4; marijuana;
5; opiates; and
6; phencyclidine
as psiS ubik j. For each, 1 = Yes, 0 Otherwise.
Summary measures were derived to capture the number of drugs out of the 7 the ith o ender
ever used psiS ubEU1i, which ranged [0;+1), and whether or not (1 = Yes, 0 Otherwise) the
o ender ever used any of the 7 psiS ubEU2i, and, likewise, the number of drugs out of the 7 that
the o ender admits to having a problem with psiS ubAP1i, which ranged [0;+1), and whether or
not (1 = Yes, 0 Otherwise) the o ender admits to a problem with any of the 7 psiS ubAP2i.
Juvenile adjudication and confinement characteristics were also recovered from the PSIs.
These were, specifically, the number of juvenile cases psiCrmJuvCasni, the number of juvenile
adjudications psiCrmJuvAd ji, and the number of juvenile confinements of length greater than 30
days psiCrmJuvConi. Each of these items took on values within the interval [0;+1).
Criminal history measures include the total number of adult cases psiCrmAdlCasni, the
number of adult convictions psiCrmAdlCnvni, and k = 1;2;:::;5 crime-disaggregated measures
60
across
k =
8>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>><
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
:
1; violence;
2; property;
3; drugs;
4; weapons; and
5; NSP
as psiCrmAdlCnvik. These last items potentially range within the interval [0;+1).
Additional items captured the nature of the originating o ense.21 These individual items
were reduced before model estimation to a hierarchical measure, triggering o ense psiO f f ensei,
capturing whether the instant o ense involved (a) violence, or, if not, (b) weapons, or, if not,
(c) property, or, if not, (d) drugs, or, if not, (e) public order, or, if not, (f) sex crimes, or, otherwise,
(g) unclassified crimes.
PSI authors provide accounts of the originating o ense from the perspectives of both the
arresting authority as well as the o ender and, typically, note any discrepancies. Several items
were constructed to capture this including whether the o ender agrees with the arresting authority
account psiS taAgri. Also, if the o ender denies responsibility, whether intoxication psiS taBladi,
21These included whether the instant o ense involved absconsion psinspai, bail-reform charges psiNspBrai, bail
jumping psiNspJumpi, or other NSP charges psiNspOthi; using drugs psiO f f Drgusi, buying drugs psiO f f Drgbui,
possessing drugs psiO f f Drgpoi, selling drugs psiO f f Drgsei, drug paraphernalia psiO f f Drgpai, or some other, un-
categorized, drug o ense psiO f f Drgoti; theft psiO f f Prothi, autotheft psiO f f Proati, burglary psiO f f Probui, stolen
property psiO f f Prospi, destruction of property psiO f f Prodpi, forgery psiO f f Pro f oi, or some other property o ense
psiO f f Prooti; disorderly conduct psiO f f Pubdii, gambling psiO f f Pubgai, vagrancy psiO f f Pubvai, public drunk-
enness psiO f f Pubpdi, or some other public order crime psiO f f Puboti; prostitution psiO f f S expri or some other
sex crimes psiO f f S exoti; murder psiO f f Viomui, rape psiO f f Viorai, robbery psiO f f Vioroi, arson psiO f f Proari,
assault psiO f f Vioasi, or some other violent o ense psiO f f Viooti; or discharging a gun psiO f f Weadii or some other
gun-related o ense psiO f f Weaoti. For each, 1 = Yes, 0 Otherwise.
61
injury psiS taBlaii, mental disorder psiS taBlami, the police psiS taBlapi, self psiS taBlasi, or the
victim psiS taBlavi was blamed for the o ense instead. For each, 1 = Yes, 0 Otherwise.
Additional items captured whether the police suspected the ith o ender was under the in-
fluence of drugs psiO f f DrgS usi or alcohol psiO f f IntS usi at the time of the arrest and whether
drug psiO f f DrgVeri or alcohol psiO f f IntVeri use was field-verified; whether the o ender tested
positive for psiS ubLuPosAi or drugs psiS ubLuPosDi at lock-up; whether someone other than the
o ender was physically injured as a result of the o ense psiO f f In ji; and whether the police re-
covered money from the o ender or the area wherein the crime occurred psiO f f Moni. For each,
1 = Yes, 0 Otherwise.
Two items capture?given a previous incarceration sentence?whether any screens were
positive for alcohol psiS ubInPosAi or other substances psiS ubInPosD, 1 = Yes, 0 Otherwise.
A notable feature recovered from the PSIs is the Salient Factor Score (SFS) which is the
recidivism prediction instrument used by the United States Parole Commission (USPC) (see, Ho -
man & Beck, 1974). The SFS has been extensively validated and is known to be quite accurate
(Blumstein, Cohen, Roth, & Visher, 1986; Janus, 1985). SFS 98?the most recent revision?is
calculated here as psiS FS i and used as a comparative tool (see, United States Parole Commission
[USPC], 2003). Its calculation is described in Appendix B.
The recommended sentence psiS Ri was also recovered from the PSIs. Such recommenda-
tions took on values of either incarceration, probation, or split-sentence.
Data bearing on sentence and supervision histories were collected from the PSIs. This
included information about criminal justice status at the time of arrest, previous community su-
pervision and incarceration sentences, and previous acts of NSP. These items are shown in Ta-
ble 2. Also obtained were times capturing the number of times (values ranged within [0;+1)) the
ith o ender was referred to drug psiS upT xdri, detoxification psiS upT xdxi, and mental health
62
Table 2
Presentence Investigation report, sentence and supervision histories.
Symbol Description Coding
psiS upS tai Status, time of instant arrest a
psiCrmAdlS upN pri Number, previous probation sentences [0;+1)
psiCrmAdlS upN pai Number, previous parole supervision sentences [0;+1)
psiCrmAdlInci Number, ICsb [0;+1)
psiS upRevi Number, previous supervision failures [0;+1)
psiS upAbsi Number, previous supervision absconsions [0;+1)
psiS upWari Number, previous supervision warrants [0;+1)
psiAdlS upRevi Number, previous supervision revocations [0;+1)
a Either fugitive, probation, parole, or free from control.
b Incarceration sentence for length greater than 30 days.
psiS upT xmhi treatment and whether the ith o ender had previously been exposed to either drug
or alcohol treatment psiS ubT XAnyi, 1 = Yes, 0 Otherwise.
Family characteristics obtained from PSIs capture early and current family structure and
support. Parent marital status at birth, their involvement through childhood, and whether there
is sustained contact, for example, are among those shown in Table 3. The item capturing parent
marital status at birth psiFamBiri had sparse levels. At birth, parents of o enders were largely
either married and living together (92=199 = 0:46), unmarried and living apart (78=199 = 0:39), or
unmarried but cohabiting (27=199 = 0:14). The parents of the remaining 199 (92+78+27)=2=199 = 0:01
o enders were married and living apart and divorced and living apart. This item was reduced to
psiFamMarBiri =
8>>
>>>>
>>>>
>>>>
><
>>>>
>>>>
>>>>
>>>:
0; not cohabiting, married or unmarried;
1; cohabiting, unmarried,;
2; cohabiting, married,:
63
where cohabiting collapses the levels Married-living together and Unmarried-cohabiting; married
collapses levels Married-living together, Married-living apart, and Divorced-living apart.
Several indicators were created to reduce the dimensionality in the data. For example, the
indicator parentInvolve was created to represent whether either mother, father, or both were un-
involved in parenting or that both were involved 1 = Yes, 0 Otherwise. The indicator parentAlive
was created to represent whether both parents were alive, 1 = Yes, 0 Otherwise. The indica-
tor parentContact was created to represent whether there was sustained contact with both par-
ents, 1 = Yes, 0 Otherwise. raisedBy was created to indicate that the o ender was raised
by either a single mother, single father, or an extended family, 1 = Yes, 0 Otherwise. The
items psiFamAbuNG, psiFamAbuPH, and psiFamAbuS X were summarized with the indicator
anyAbuse, 1 = Yes, 0 Otherwise. Finally, psiS ocMarS ta and psiS ocMarDiv were represented
by the indicator psiS ocMar to represent whether the o ender was cohabiting, either married or
unmarried, or not.
Certain social characteristics were also obtained. These items, shown in Table 4, capture
marital status, dependents, and whether the ith o ender lives with dependent children.
There were also several educational characteristics obtained from the PSIs. These included
for i = 1;2;:::;N, highest grade attempted psiEduGrdAti and completed psiEduGrdCmi educa-
tional years and, if not a high school graduate, whether a GED was earned psiEduGEDi (1 = Yes,
0 Otherwise).
A pair of items capture whether in the previous year there were any employment stints of
less psiEmpLessi or more psiEmpMorei than 30 days (1 = Yes, 0 Otherwise).
Limited health characteristics were also obtained including measures of substance use and
of physical and mental disabilities, injuries, and illnesses. These are shown in Table 5.
64
Table 3
Presentence Investigation report, family characteristics.
Symbol Description Coding
psiFamBiri Parent marital status at birth 1 = Yes, 0 Otherwise
psiFamNowi Parent marital status now 1 = Yes, 0 Otherwise
psiFamNow f bi Father alive at birth 1 = Yes, 0 Otherwise
psiFamInvmi Mother involved in parenting 1 = Yes, 0 Otherwise
psiFamInv fi Father involved in parenting 1 = Yes, 0 Otherwise
psiFamRaismi Raised by single mother 1 = Yes, 0 Otherwise
psiFamRais fi Raised by single father 1 = Yes, 0 Otherwise
psiFamRaix fi Raised by extended family 1 = Yes, 0 Otherwise
psiFamRai f fi Raised by foster family 1 = Yes, 0 Otherwise
psiFamNow f ni Father alive now 1 = Yes, 0 Otherwise
psiFamAlimni Mother alive now 1 = Yes, 0 Otherwise
psiFamCnt fi Sustained contact with father 1 = Yes, 0 Otherwise
psiFamCntmi Sustained contact with mother 1 = Yes, 0 Otherwise
psiFamS ibAli Number, siblings [0;+1)
psiFamS ibbni Number, blood-siblings [0;+1)
psiFamS ibsni Number, step-siblings [0;+1)
psiFamS sni PSI author finds a supportive social network 1 = Yes, 0 Otherwise
psiFamAbuphi Physical abuse 1 = Yes, 0 Otherwise
psiFamAbusxi Sexual abuse 1 = Yes, 0 Otherwise
psiFamAbungi Neglect/abandonment 1 = Yes, 0 Otherwise
65
Table 4
Presentence Investigation report, social characteristics.
Symbol Description Coding
psiS ocMarS tai Marital status a
psiS ocMarDivi Ever divorced 1 = Yes, 0 Otherwise
psiS ocChli Children, same residence b
psiS ocHouFai Lives with relativesc 1 = Yes, 0 Otherwise
psiS ocChlBili Number, biological childrend, same residence [0;+1)
psiS ocChlBini Number, biological childrend [0;+1)
psiS ocChlS tni Number, step-childrend [0;+1)
psiS ocToni Number, children total c [0;+1)
a 5 = Married, living together; 4 = Married, living apart; 3 = Divorced, living apart; 2 = Single,
cohabiting; 1 = Single, living alone; 0 = Widowed. b 0 = No children under 18, 1 = Children
under 18, not all same residence, 2 = Children under 18, all same residence.
d Ages 18 or younger. c i.e., parents, siblings, aunts, uncles, cousins, or grandparents.
Table 5
Presentence Investigation report, health characteristics.
Symbol Description Coding
psiMedMdDisi Number, disabilities [0;+1)
psiMedIn ji Number, injuries [0;+1)
psiMedMdDxi Ever diagnosed with a mental illness 1 = Yes, 0 Otherwise
psiMedMdDri Takes prescribed psychotropic medications 1 = Yes, 0 Otherwise
psiMedMdHoi Number, previous mental health hospitalization [0;+1)
psiMedMdS ui Ever attempted suicide 1 = Yes, 0 Otherwise
psiS ubS Mi Self-medicating with alcohol or drugs 1 = Yes, 0 Otherwise
psiMedWouguni Number, gunshot wounds [0;+1)
psiMedWoustbi Number, stabbing wounds [0;+1)
66
As the range of potential predictors obtained from the PSIs was quite large, that missingness
would compromise analytics was expected at the outset. In contrast to those captured in the RNS,
missingness was more widespread with respect to the PSI data and strategies similar to those taken
for the RNS were used to minimize this problem. 63=118 = 0:53 were missing less than 5 values;
20=118 = 0:17 were missing 5 to 9 values; 18=118 = 0:15, 10 to 14; 9=118 = 0:076, 15 to 20; and
8=118 = 0:068 were missing 20 or more. Missing values were imputed based on nonmissing values
among other individual-level measures. Specifically, 5 imputes were derived from random draws
from the conditional distributions of the nonmissing values of each target measure given the values
across the other variables. This resulted in a 5-length vector of imputes for each probationer-value.
Supervision and Management Automated Record Tracking. In addition to those in the RNSs
and PSIs, data provided by the CSOSA also include information contained throughout its SMART
database, a relational database comprising over 350 individual tables. Nearly every piece of in-
formation pertaining to probationers is contained within SMART. Prominent characteristics in-
clude the beginning beginDti and ending outTrmDti dates of the supervision period, the actual
daysS upActi and expected daysS upExpi number of days supervised, and the supervision level
supLvli (i.e., minimum, medium, maximum, or intensive).
The observation period subsumes both the supervision period?the interval spanning from
the supervision period begin date beginDti to the supervision period termination date outTrmDti22
and the post-supervision period?the interval spanning from outTrmDti to the follow-up close
date. For all sampled o enders, the follow-up close date is December 31, 2006.
What I discuss next are those items bearing on performance, such as drug screening, con-
viction and violation histories, and termination modes.
22For most o enders this is the date on which the supervision period ended in one of three modes: successful,
unsuccessful, or revoked. For absconders, termination date is the date of the first of a series of contact losses. For
o enders that died while supervised, termination date is the date of death.
67
Results from drug testing events include which, if any, of the 7 potentially screened sub-
stances were positive at each J event for the i = 1;2;:::;N probationers. For clarity, these sub-
stances are mapped to the index k as
k =
8>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>><
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>:
1; alcohol;
2; methadone;
3; amphetamine;
4; cocaine;
5; marijuana;
6; opiates; and
7; phencyclidine.
The specific data used here include the date of the jth event for the ith probationer outDrgdt ji and
two sets of indicators. The first, outDrgS cr1 ji; outDrgS cr2 ji;:::; outDrgS crk ji, capture whether
the kth substance was screened during the jth event, and, if so, the second set, outDrgPos1 ji;
outDrgPos2 ji;:::; outDrgPosk ji, capture whether the result was positive. For both, 1 = Yes,
0 Otherwise.
The item outS 1i summarizes these items in capturing whether the ith probationer ever tested
positive, provided a bogus specimen, or failed to appear for a drug testing event. The variables
outDrgTotS cr1i; outDrgTotS cr2i;:::; outDrgTotS cr7i capture the total number of drug screens
for the kth substance and outDrgTotPos1i;outDrgTotPos2i;:::;outDrgTotPoski capture the total
number of positive screens for the kth substance; piPoski capture the proportion of positive screens
for the kth substance.
Conviction data are available in SMART as part of a data sharing agreement with the Su-
perior Court of the District of Columbia, a trial court with general jurisdiction over virtually all
68
local legal matters. This agreement allows the CSOSA to identify its o enders and determine
the outcomes of prosecution, trial, and sentencing processes. The specific measures recovered
include for every J conviction event involving each i = 1;2;:::;N probationers, the conviction
date outCnvdt ji, the date of arrest leading to the conviction outCnvArrdt ji, and the type of charge
on which the conviction is made outCnvCg ji. Here, charge types are broadly classified into one of
five categories (see, Appendix C) and are mapped to the index m as
m =
8>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>><
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
:
4; violent;
3; drug- or alcohol-related;
2; property;
1; public disorder; and
0; other.23
The next measures bear on whether and, if so, how often during the observation period
rearrests resulted in convictions. Criteria include outC1i, which captures whether the ith proba-
tioner was arrested and subsequently convicted for any o ense during the supervision period; and
outC2i, which captures whether the ith probationer was arrested and subsequently convicted for
any o ense during the post-supervision period. For each, 1 = Yes, 0 Otherwise.
Violations and modes of termination data were also obtained from SMART. Following
agency definitions, condition violations are classified as either supervision- or drug-related, where
supervision-related violations include violations of general and special conditions (see,Table D1
and Table D2); drug-related violations include only those specifically involving illegal substances.
The date outViodt ji and broken condition outVioCond ji were recovered for every J violation event
involving the i = 1;2;:::;N probationers. The variable outVioTyp ji summarizes the jth event for
69
the ith probationer as
outVioTyp ji =
8>>
>>>>
>><
>>>>
>>>>
:
1; Drug-related;
0; Supervision-related.
Two criteria were constructed: outV1i, which captures the total number of times the ith probationer
violated a supervision-specific condition and outV2i, which captures the total number of times the
ith probationer violated a drug-related condition.
The termination mode outT1i and the date of termination outT2i are obtained from SMART
for the ith probationer. Termination modes capture the process by which probationers completed
their sentence. Possible modes are detailed in Appendix E. Briefly, these are
outT1i =
8>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>><
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
:
1; death;
2; successful;
3; unsuccessful and terminated;
4; revoked; and
5; absconsion:
However, sparse categories necessitated collapsing levels into either successful or unsuccessful
termination. As such, outT1i was recoded into an indicator of failure (i.e., unsuccessful, revoked,
or absconsion).
Contextual measures
The RNSs, the PSIs, and the information housed within SMART describe characteristics
unique to each probationer. To describe their environments data were also obtained from five other
sources including an agency maintaining local geospatial data,24 the U.S. Census Bureau (Census),
24Among other roles the O ce of the Chief Technology O cer for the District of Columbia maintains for the Dis-
trict of Columbia (DC) point, polyline, and polygon data describing streets and administrative and political boundaries.
70
the local police department, and two agencies responsible for business regulation. These environ-
mental data are discussed next beginning with sociodemographic and economic characteristics
obtained from the Census.
Sociodemographic and economic characteristics. Several items were obtained from the
2000 Census Summary File 3 (SF3) to capture the contextual aspects shown in Chapter 2 to relate
well with NSP, which are, namely, wealth and poverty, race and ethnicity, immigration, employ-
ment, age and family structure, and residential stability. The SF3 comprises sample data from
roughly 1:6 U.S. households receiving the Census 2000 long-form questionnaire. The measures
derived here, summarized for each k = 1;2;:::;436 block-group, are shown in Table 6.
Missingness resulted from an absence of data bearing on areas outside of the DC: 3 of-
fenders lived each within separate census tracts of neighboring Prince George?s county, Maryland;
recovering data for these BGs was uncomplicated as they were all obtainable from the Census.
Objective crime measures. A data sharing agreement between the CSOSA and the MPDC
allows CSOSA access to local arrest data. Locations of arrest events are geocoded thereby coordi-
nating each within x-y space then aggregated by crime category within BG. The obtained measures
include the number of arrests for violent arrViok, property arrProk, drug- and alcohol-related
arrDrgk, public-order arrPubk, and otherwise unclassified arrOthk crimes. I then created density
measures of each type of arrests as arrVio:dk, arrPro:dk, arrDrg:dk, arrPub:dk, and arrOth:dk,
respectfully.
These data will be used in linking environmental-level data with probationer residences. I obtained street-level in-
formation which then provided a means for both coordinating points within x-y space and for aggregating these points
within BGs. Note, a Census block-group (BG) consists of all census blocks having the same first digit of their four-digit
identifying numbers within a census tract and generally contain between 600 and 3,000 residents (see, United States
Census Bureau [Census], 2004, A-8). It is precisely this level that all environmental-level data were summarized.
71
Table 6
U.S. Census, sociodemographic characteristics, 2000, by Census block-group, k = 436.
Symbol Description Coding
piBlk Population, BlackPopulation [0;1]
piDi f Populationa, di erent house in 1995Populationa [0;1]
piEdu Populationb, less than a high school diploma or equivalencyPopulationb [0;1]
piFHH Households, female, no husband presentHouseholds [0;1]
piFor Population, foreign bornPopulation [0;1]
piHis Population,Hispanic or LatinoPopulation [0;1]
piPov Populationc, income in 1999 < poverty levelPopulationc [0;1]
piPub Households, public assistance incomeHouseholds [0;1]
piRnt Housing units, renter occupiedOccupied housing units [0;1]
piUne Populationd, unemployedPopulationd [0;1]
rtAK ChildreneAdultsf [0;1]
popDens Populationland area m2 [0;1)
Note. Data are provided by the U.S. Census Bureau, Summary File 3, 2000.
a Population 5 years and over. b Population ages 25 and older.
c Population for whom poverty status is determined. d Population 16 years and over.
e Population under age 18. f Population ages 18 and older.
Access to objective crime measures was just short of nonexistent for the 3 non-DC BGs
wherein sampled o enders resided. Strategies similar to those taken among the individual-level
measures were used to adjust for missingness, but before doing so as much information as possible
was replaced with data provided by the local police department with arresting jurisdiction.
An arrangement with the Prince George?s County Police Department (PGPD) was estab-
lished which enabled recovering the number of arrests for both violent and property crimes within
72
each of the 3 Maryland BGs.25 The two remaining arrest summaries capturing, respectfully, drug
and alcohol arrDrgi and public order arrPubi crimes were imputed.
Specifically, missing values were imputed based on non-missing values across those re-
maining. Five imputes for each were derived from random draws from the conditional distribu-
tions of the nonmissing values of each target given the values across the other variables. This
resulted in a 5-length vector of imputes for each measure per BG.
Commercialization patterns. Data describing the concentrations of businesses within BGs
were collected from the Department of Consumer and Regulatory A airs for the District of
Columbia (DCRA), the regulatory agency charged with licensing as well as monitoring and en-
forcing compliance with commercial regulations in DC. I obtained the location and type of all
licensees in DC, geocoded their locations, then summarized types within BG. The specific mea-
sures derived from these data capture the total number of licenses for employment services; en-
tertainment services; general businesses; housing; public health; and sales, service, and repair. To
represent commercialization I calculated a summary measure busDensk capturing the density of
all licensees26 per 1,000 residential housing units within BG.
I also obtained data specifically addressing retail alcohol licensees. The Alcoholic Beverage
Regulation Administration for the District of Columbia (ABRA) provides data describing alcohol
retailers in DC. This agency issues licenses as well as monitors and enforces compliance with
regulations among liquor stores, brewpubs, nightclubs, restaurants, taverns, hotels, and other es-
tablishments that manufacture, sell, or serve alcoholic beverages in DC. I summarized the location,
class, and type of every alcohol retail licensee in DC. Licensees were then classified by license
25These values were, for Block Group 2, Census Tract 8019.01: arrVioi = 24 and arrProi = 261; for Block Group
1, Census Tract 8012.05: arrVioi = 37 and arrProi = 442; and for Block Group 4, Census Tract 8012.02: arrVioi = 90
and arrProi = 506.
26excluding retail alcohol outlets
73
type, which were one of either (a) o -premises sale of beer, wine, and spirits; (b) o -premises sale
of beer and wine only; (c) on-premises sale of beer, wine, and spirits; or (d) on-premises sale of
beer and wine only. Densities of each licensee were calculated per square meter of BG land area
for each k = 433 as, respectfully, alcBWS 1k, alcBW1k, alcBWS 2k, and alcBW2k. A summary
measure alcDensk was also calculated representing the density of all licensees per 1,000 square
meter of BG land area.
None of the measures capturing commercialization patterns were available for the 3 Mary-
land BGs and these, too, were imputed using the same procedures as outlined above.
Described thus far are the data sources and the measures derived from these sources. Broadly,
these include (a) the RNS, (b) the PSI, and (c) the SMART database, which are each provided by
the CSOSA; objective crime measures, which are derived from data provided by the local police
agency; densities of local businesses, and, in particular, alcohol retailers, which are derived from
data provided by local regulatory agencies; and various social, economic, and housing summaries,
which are derived from data provided by the Census. Among described measures were predictors,
both individual- and contextual-level, as well as the legal and supervision-specific criteria opera-
tionalizing NSP. In the next few passages I review NSP?as defined by the criteria?while also
describing the specific models, each estimated with various General Linear Models (GLMs) (see,
Dobson, 2001; McCullagh & Nelder, 1989), in which these criteria were central.
Procedure
NSP itself is unmeasurable. It is a simplifying concept used to succinctly describe a broad
set of related characteristics. As described previously, NSP encapsulates behaviors classified into
legal and supervision-specific domains. Criteria in the legal domain capture substance use and
rearrests resulting in convictions; supervision-specific criteria include condition violations and
74
termination modes. Because of its multifarious nature, multiple measures are needed to get an
understanding of NSP in the population and thus, here, separate models are estimated for each
embedded feature. The procedures used in estimating these models are described throughout this
section beginning with those related to substance use.
Model MS1 estimates factors associated with the probabilities that probationers in the pop-
ulation will ever fail a drug-testing event while supervised. The criterion for this model was outS 1.
Models MS2A;MS2B;:::;MS2G estimate factors associated with how often probationers in the
population will test positive for the kth screened substance. The criterion for these models were,
respectfully, outDrgTotPosik.
Models MC1 and MC2 estimate the probabilities that probationers in the population will
be convicted for new crimes given an arrest. Model MC1 does this for convictions during the
supervision period; Model MC2, the post-supervision period. The criteria were, respectfully,
outC1i and outC2i.
Model MV1 estimates factors associated with how often probationers in the population
will violate supervision-related conditions and Model MV2 estimates factors associated with how
often probationers in the population will violate drug-related conditions. The criteria were outV1i
and outV2i, respectfully.
Model MT1 estimates likelihoods probationers in the population will terminate sentences
in one of three unsuccessful modes (i.e., unsuccessful, revoked, or absconsion); Model MT2
estimates how soon either are likely to occur. The criteria were, respectfully, outT1 and outT2.
The procedures used in estimating risk of NSP as operationalized by the criteria involved,
first, developing a general model to predict each criterion using predictors identified in the review
as likely influences. Models were reduced to binary trees using recursive partitioning analysis
(RPA) (Breiman, Friedman, Olshen, & Stone, 1984; Clark & Pregibon, 1992; Therneau & Atkin-
75
son, 1997) 27 and then pruned back to account for replacement optimism based on an AIC-like
pruning scheme (see, Venables & Ripley, 2002; Ciampi, Negassa, & Lou, 1995). A multivariable
GLM was then developed using the pruned-tree predictors with any parameterization and func-
tional form adjustments necessary to normalize marginal distributions. Standard errors were cor-
rected using robust variance estimators following procedures outlined by (Huber, 1967; Rogers,
1993; H. White, 1980; R. L. Williams, 2000).
After models were estimated I turned to evaluation and validation. Model performance was
evaluated in terms of calibration and discrimination and models were validated by bootstrapping
(see, Appendix F).
27Using the R library RPART (Therneau & Atkinson, 1997).
76
RESULTS
Analytic results are presented in this chapter. In the first section I describe the central
characteristics moving, generally, from the individual- to the contextual-level then follow with
presentations of model estimates. Detailed discussions are held o until the following chapter.
Description
Descriptions of measures are presented in this section beginning with characteristics of
the sample in terms of supervision-specific measures and then those bearing on individual- and
contextual-level predictors. Descriptions of the criteria are presented along with the model esti-
mates.
Supervision level supLvli took on one of four values:28 109=199 = 0:54 were supervised
at the maximum level and 47=199 = 0:24 were supervised at the intensive level. The remaining
o enders were supervised at medium 32=199 = 0:16 and minimum 11=199 = 0:055 levels.
The majority of o enders began supervision between calender years 2002 (53=199 = 0:27)
and 2003 (92=199 = 0:46); 31=199 = 0:16 began in 2004; 14=199 = 0:07 began in 2001; and 8=199 = 0:04
began in 2000.29 The shortest supervision period was 2 weeks and 2 days; the longest was 6
years and 29 days. Median length of supervision was 1 year and 47 days; the 0.25 and 0.75
quantiles span from 42 weeks to 2 years. On average, sampled o enders served between 109
and 166 (95%CI) fewer days than expected given the expiration date in the original full term
sentence. As a function of the sampling all o enders terminated sentences during the interval
spanning January 1, 2004 to December, 31, 2004.30 Supervision period begin dates beginDti,
28With the exception of one o ender whose supervision level remained undetermined. Supervision level for this
o ender was recoded from TBD to the sample mode, MAX.
29One o ender began supervision in 1998?1 year and 289 days before the next earliest begin date. This o ender is
excluded from Figures 1 and 2.
30One o ender o cially terminated supervision in 2004, but is treated as though he terminated in June, 2003. Within
three months after beginning supervision in April, 2003 this o ender absconded. A warrant was issued and his super-
77
Table 7
CSOSA Risk-needs Screener, social characteristics, n = 199.
Symbol Description Level 95%CI/Proportion
rnsAge Age at the time of assessment [31:6;34:6]
rnsEdu Educational level 10th or below 48199 0:24
11th 35199 0:18
HS/GED 78199 0:39
Some college 38199 0:19
rnsS S N Significant relationships No relationships 10199 0:0503
Relationship with one person 16199 0:0804
Relationship with two or more people 173199 0:87
rnsRL Recent loss Yes 37199 0:19
rnsS taa Instability Currently/Recently incarcerated/shelter 61199 0:31
Two or fewer changes 77199 0:39
Three or more changes 61199 0:31
a Derived from RNS items rnsEmp and rnsRes.
expected termination dates outTrmDtExpi, termination dates outTrmDti, and termination modes
outTrmModi are shown in Figures 1 and 2.31
The Risk-needs Screener (RNS) items are described in Table 7 through Table 9. Items cap-
turing demographic and social characteristics, including age, education, significant relationships,
recent loss, and instability are shown in Table 7. Those capturing criminal histories are shown
in Table 8. Substance abuse, mental health, and agency responsiveness to o enders?including
impressions of risk and cooperation?are shown in Table 9.
Items obtained from the Presentence Investigation report (PSI) as well as those derived
therein are shown in Tables 10?19. Table 10 summarizes substance use and abuse measures. Crim-
vision status was changed to reflect that monitoring was no longer possible, but it was not until March, 2004, that
his supervision period was o cially terminated. Delayed termination in this case captures more administrative than
behavioral e ects and, thus, I treat the date of termination for this o ender (as well as that for the only other absconder)
as the date of the first of an unending series of contact losses.
31Plots are shown by supervision level supLvli only to ease digestion: no relationship is implied.
78
1
11
Oct
2001
Jan
2004
Dec
2004
Dec
2006
Minimum
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a73
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
1
32
Oct
2001
Jan
2004
Dec
2004
Dec
2006
Medium
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
?
a73
a73
a73
a73
a73
a73
a73
a73
a73
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a78
a78
a78
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
Figure 1. Begin date beginDti (4), expected termination date outTrmDtExpi (+), and termina-
tion date outTrmDti marked by termination mode outTrmModi [absconscion ( ), death ( ), re-
vocation ( ), successful ( ), and unsuccessful ( )], and follow-up close dates (O) by supervision
level for minimum (n = 11) and medium (n = 32) supervision levels.
79
1
108
Apr
2000
Jan
2004
Dec
2004
Dec
2006
Maximum
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
?
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a78
a78
a78
a78
a78
a78
a78
a78
a78
a78
a78
a78
a78
a78
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
1
47
Apr
2000
Jan
2004
Dec
2006
Intensive
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
a76
?
?
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a73
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a71
a78
a78
a78
a78
a78
a78
a78
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a27
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
a77
Figure 2. Begin date beginDti (4), expected termination date outTrmDtExpi (+), and termina-
tion date outTrmDti marked by termination mode outTrmModi [absconscion ( ), death ( ), re-
vocation ( ), successful ( ), and unsuccessful ( )], and follow-up close dates (O) by supervision
level for maximum (n = 109) and intensive (n = 47) supervision levels.
80
Table 8
CSOSA Risk-needs Screener, criminal history characteristics, n = 199.
Symbol Description Level Proportion
rnsPV Prior violent o ense None 124199 0:62
One 41199 0:21
Two or More 34199 0:17
rnsNPA Prior adult arrests Two or less 68199 0:34
Three to Four 45199 0:23
Five 12199 0:0603
Six or more 74199 0:37
rnsPS Prior supervision failures None 120199 0:6
One to Two 66199 0:33
Three or more 13199 0:065
rnsFA Frequency of arrests/year None 84199 0:42
One 78199 0:39
Two to Four 37199 0:19
rnsAF Age at first arrest 15 and younger 28199 0:14
16 to 17 37199 0:19
18 to 25 112199 0:56
Over 26 22199 0:11
rnsPC Prior convictions None 54199 0:27
One to Five 121199 0:61
Six or more 24199 0:12
inal histories are described through Tables 11?14. Treatment histories are described in Table 15.
Family characteristics are shown in Table 16. Social characteristics are shown in Table 17. Ed-
ucational and employment characteristics are shown in Table 18. Health characteristics obtained
from the PSIs are shown in Table 19.
What I move into next is a generalization of the geopolitical characteristics of the areas in
which sampled o enders reside. It begins with a description of the geographical unit, the U.S.
81
Table 9
CSOSA Risk-needs Screener, substance use, mental health, and agency responsivenes, n = 199.
Symbol Description Level Proportion
rnsDrga Substance use rnsHS A=NO 32199 0:16
rnsHS A=YES & rnsCS A=NO 94199 0:47
rnsHS A=YES & rnsCS A=YES 73199 0:37
rnsMHb History or current mental disorder Yes 11199 0:055
rnsImp CSO Impression Low 53199 0:27
Medium 94199 0:47
High 52199 0:26
rnsFullCoopa Fully cooperative Yes 170199 0:85
a Derived from RNS items rnsCS A and rnsHS A. b Derived from RNS items rnsCMD and rnsHMD.
Census Bureau (Census) block-group (BG), then moves into a summary of how the contextual
predictors vary across these units.
The BG was chosen because, unlike the census tracts within which they are nested, they
were intentionally designed to represent near-neighborhoods and, unlike the smaller clusters of
blocks they encompass, their sample data is publicly available from the Census.32 Among the
k = 436 BGs, 122=436 = 0:28 were occupied by sampled o enders: 75=436 = 0:17 were occupied by
only 1, 27=436 = 0:06 by 2, and 20=436 = 0:05 by 3 or more. There were 3 BGs having 5 o ender
residents each.
32There are 433 DC BGs, but, all in all, there are k = 436 BGs included in this discussion. A data integrity slippage
resulted in the inclusion of 3 residents of neighboring Prince George?s county, Maryland, in the sample of Black, male
probationers. Usually, non-DC residents supervised by the CSOSA are done so pursuant to the the Interstate Compact
(Court Services and O ender Supervision for the District of Columbia [CSOSA], 2004) which provides the DC ?may
enter into a compact with any of the United States for the mutual helpfulness in relation to persons convicted of crimes
or o enses who may be on probation or parole.? Non-DC resident o enders are usually classified not as regularly
supervised but rather as interstate compact o enders. As the sample frame for this study included only regularly
supervised probationers, a decision was made to include these 3 non-DC residents in the study sample after weighing
costs associated with data loss in an already-limited sample against those of introducing both artifacts due to their
unlikeness vis-`a-vis the population and di culties in recovering comparable environmental characteristics.
82
Table 10
Presentence Investigation report, ever used and admits to problem with substances, by substance
n = 199.
Symbol Description Level Proportion
psiS ubAlAP Admits to problem, alcohol Yes 17199 0:085
psiS ubAlEU Ever use, alcohol Yes 98199 0:49
psiS ubAmAP Admits to problem, amphetamines Yes 3199 0:015
psiS ubAmEU Ever use, amphetamines Yes 13199 0:065
psiS ubCoAP Admits to problem, cocaine Yes 36199 0:18
psiS ubCoEU Ever use, cocaine Yes 106199 0:53
psiS ubMaAP Admits to problem, marijuana Yes 43199 0:22
psiS ubMaEU Ever use, marijuana Yes 159199 0:8
psiS ubHeAP Admits to problem, opiates Yes 29199 0:15
psiS ubHeEU Ever use, opiates Yes 48199 0:24
psiS ubPcAP Admits to problem, PCP Yes 22199 0:11
psiS ubPcEU Ever use, PCP Yes 79199 0:4
psiS ubEU1a Number ever used, substances out of 7 [0;3) 95199 0:48
3 59199 0:3
[4;7] 45199 0:23
psiS ubEU2 Ever used any of 7 substances Yes 187199 0:94
psiS ubAP1a Number admit to problem, substances out of 6 0 125199 0:63
1 31199 0:16
[2;5] 43199 0:22
psiS ubAP2 Admits to problem with any of 6 substances Yes 74199 0:37
a Discretized for presentation.
83
Table 11
Presentence Investigation report, juvenile and adult o ending histories, n = 199.
Symbol Description Level Proportion
psiCrmJuvCasN Number, juvenile cases None 133199 0:67
One 21199 0:11
Two or more 45199 0:23
psiCrmJuvAd j Number, adjudications None 152199 0:76
One or more 47199 0:24
psiCrmJuvCon Number, juvenile confinements None 182199 0:91
One or more 17199 0:085
psiCrmAdlCasNa Number, adult cases [0;4) 65199 0:33
[4;7) 38199 0:19
[7;13) 50199 0:25
[13;37] 46199 0:23
psiCrmAdlCnvNa Number, convictions [0;2) 79199 0:4
2 29199 0:15
[3;6) 45199 0:23
[6;18] 46199 0:23
a Discretized for presentation.
The DC is the most densely populated of the U.S. with roughly 3,600 people per km2.
The 95%CI around mean block-group population spans [1250;1401]; around mean households,
[536;614]; and around mean family units, [254;285]. Not all geographic units in the DC have
residential populations. Several in fact have populations, households, or family units equal to
zero.33
A majority (347214=578133 = 0:60) of the residents are Black; a small proportion (73904=578133 =
0:13) is foreign born; and an even smaller proportion (45151=578133 = 0:078) is Hispanic. The ratio
33Specifically, there are zero populations in census tracts 54.02, 57.02, 62.02, 89.05; zero housing units in tracts
54.02, 57.02, 62.02, 73.08, 89.05, and 98.09; and zero family units in tracts 2.01, 54.02, 57.02, 62.02, 73.08, 89.05,
95.01, and 98.09. Measures drawn from the Summary File 3 (SF3) that rely on these units are thus incalculable, and,
as imputation would be meaningless, such units are dropped from estimation when necessary.
84
Table 12
Presentence Investigation report, adult conviction histories, by conviction type n = 199.
Symbol Description Level Proportion
psiCrmAdlCnvVio Number, violent convictions None 147199 0:74
One 34199 0:17
Two or more 18199 0:09
psiCrmAdlCnvNsp Number, NSP convictions None 158199 0:79
One 32199 0:16
Two or more 9199 0:045
psiCrmAdlCnvWeaa Number, weapons convictions 0 162199 0:81
[1;9] 37199 0:19
psiCrmAdlCnvPro Number, property convictions None 133199 0:67
One 30199 0:15
Two or more 36199 0:18
psiCrmAdlCnvDrg Number, drugs convictions None 85199 0:43
One 45199 0:23
Two or more 69199 0:35
a Discretized for presentation.
of residents ages 17 and younger to those ages 18 and older is 115634=462499 = 0:25. Less than a
quarter (86071=388982 = 0:22) of the residents ages 25 and older failed to earn a high school diploma.
One-half (272935=545475 = 0:50) of the residents ages 5 and older have been living in the same house
for 5 years or more. The majority (147585=250525 = 0:59) of occupied housing units are occupied by
renters rather than owners. 47784=250745 = 0:19 of the households comprise female householders
with no husband present. O cially, 109837=547312 = 0:20 of the population for whom poverty status
is known are impoverished; only 13683=250745 = 0:055 of the households receive public assistance.
Among the population ages 16 and over in the civilian labor force, 31937=297719 = 0:11 are unem-
ployed. Following Sampson et al. (1997), the items piPov, piPub, piFHH, piUne, rtAK, piBlk,
piHis, piFor, piRnt, and piDi f were summarized with factor scores. Loadings after oblimin
rotation are shown in Table 20.
85
Table 13
Presentence Investigation report, triggering o ense, n = 199.
Symbol Description Level 95%CI/Proportion
psiO f f ense Triggering o ense Drug-related 118199 0:59
Violent 37199 0:19
Property 19199 0:095
Non-violent, weapons 16199 0:0804
Other 9199 0:045
psiS taAgr Agrees with o ense account? Yes 166199 0:83
psiS taBlaD Does the o ender blame drugs? Yes 25199 0:13
psiS taBlaM Blames mental disorder? Yes 1199 0:005
psiS taBlaP Does the o ender blame police? Yes 9199 0:045
psiS taBlaS Does o ender blame self? Yes 51199 0:26
psiS taBlaV Does the o ender blame victim? Yes 11199 0:055
psiO f f In j Someone physically injured? Yes 28199 0:14
psiO f f Mon Police recovered money? Yes 77199 0:39
psiNS PBra Instant o ense is bail-reform Yes 16199 0:0804
psiO f f DrgS us Police suspect influence of drugs Yes 11199 0:055
psiO f f DrgVer Drug use verified Yes 11199 0:055
psiO f f IntS us Police suspect influence of alcohol Yes 4199 0:0201
psiO f f IntVer Intoxication verified Yes 1199 0:005
psiS ubLuPosA At lock-up, positive for alcohol Yes 4199 0:0201
psiS ubLuPosD At lock-up, positive for drugs Yes 43199 0:22
psiCrmAdlS FS Salient Factor Score [5:47;6:28]
psiS R Recommended sentence Probation 109199 0:55
Incarceration 55199 0:28
Split-sentence 35199 0:18
86
Table 14
Presentence Investigation report, sentence and supervision histories, n = 199.
Symbol Description Level Proportion
psiS upS ta Criminal justice status at time of arrest Fugitive 3199 0:015
Probation 42199 0:21
Parole 4199 0:02
Free 150199 0:75
psiCrmAdlS upNPra Number, probation sentences 0 54199 0:27
1 54199 0:27
[2;4) 56199 0:28
[4;9] 35199 0:18
psiCrmAdlS upNPaa Number, post-incarceration parole 0 167199 0:84
[1;5] 32199 0:16
psiCrmAdlIncna Number, confinement sentences > 30 days 0 74199 0:37
1 36199 0:18
[2;5) 48199 0:24
[5;13] 41199 0:21
psiS upRev Number, previous supervision failures None 109199 0:55
One 43199 0:22
Two or more 47199 0:24
psiS upAbs Number, previous supervision absconsions None 193199 0:97
One 5199 0:025
Two or more 1199 0:005
psiS upWar Number, previous violation warrants None 169199 0:85
One 23199 0:12
Two or more 7199 0:035
psiCrmAdlS upRev Number, probation or parole revocations None 120199 0:6
One 44199 0:22
Two or more 35199 0:18
b Incarceration sentence for length greater than 30 days.
87
Table 15
Presentence Investigation report, treatment histories, n = 199.
Symbol Description Level Proportion
psiS upT xDr Number, drug treatment referrals None 123199 0:62
One 43199 0:22
Two or more 33199 0:17
psiS upT xDx Number, detoxification referrals None 178199 0:89
One 19199 0:095
Two or more 2199 0:0101
psiS upT xMh Number, mental health referrals None 188199 0:94
One 10199 0:0503
Two or more 1199 0:005
psiS ubT xAny Any previous drug or alcohol treatments Yes 78199 0:39
There were 19,087 business licensees34 across k = 1;2;:::;433 BGs in the District. Den-
sities within BGs ranged between [132;248] licensees per 1,000 housing units (95%CI). There
were 1,663 retail alcohol licensees across BGs. By far, the majority of these retailers were those
983=1663 = 0:59 licensed for on-premises sales of beer, wine, and spirits. The 95%CI around mean
densities of alcohol licensees per 200,000 m2 of BG land area are shown in Table 21.
The 95%CI around mean arrest rates are shown in Table 22. Rates capture the number of
arrests per 1,000 residents ages 18 and older for violent, property, alcohol and drug, public order,
and unclassified crimes as well as an index of all crimes.
In the next section I describe the models operationalizing negative supervision performance
(NSP). Criteria are tabulated in Table 23, and in each section that follows I describe the criterion
of interest, the steps taken to reduce the pool of potential predictors, and the development of the
model. I then present model estimates and conclude each section with an interpretation of these
estimates. These findings are discussed in greater detail in the chapter that follows.
34Excluding licensed alcohol retailers.
88
Table 16
Presentence Investigation report, family characteristics, n = 199.
Symbol Description Level Proportion
psiFamMarBir Parent marital status at birth Not cohabiting 80199 0:40
Cohabiting, unmarried 27199 0:14
Cohabiting, married 92199 0:46
parentInvolved Either parent uninvolved Yes 94199 0:47
parentAlive Both parents alive Yes 109199 0:55
raisedBy Single mother or father or extended family Yes 108199 0:54
psiFamRaix f Raised by extended family Yes 51199 0:26
psiFamRaiF f Raised by foster family Yes 9199 0:045
parentContact Sustained contacta with both parents Yes 104199 0:52
psiFamS S N Supportive social network Yes 82199 0:41
psiFamS ibAlb Number of siblings [0;3) 67199 0:34
3 37199 0:19
[4;7) 62199 0:31
[7;24] 33199 0:17
anyAbuse Any report of abuse Yes 74199 0:37
psiFamNowFb Father alive at birth Yes 191199 0:96
psiFamS ibBnb Number of blood-siblings [0;3) 74199 0:37
3 34199 0:17
[4;6) 42199 0:21
[6;24] 49199 0:25
a Until instant o ense. If deceased, until time of death. b Discretized for presentation.
Estimation
Substance use
Model MS1 estimates factors associated with population probabilities of ever testing pos-
itive, providing a bogus specimen, or failing to appear for a drug-testing event while supervised
and Models MS2A; MS2B;:::; MS2G estimate factors associated with how often probationers
test positive for the kth screened substance.
89
Table 17
Presentence Investigation report, social characteristics, n = 199.
Symbol Description Level Proportion
psiS ocMar Cohabiting Yes 149199 0:75
psiS ocHouFa Lives with extended familya Yes 109199 0:55
psiS ocTonb Number, childrenc total 0 72199 0:36
1 68199 0:34
2 33199 0:17
3 13199 0:0653
[4;12] 13199 0:0653
psiS ocChlBil Number, biological childrenc living with None 161199 0:81
One 23199 0:11
Two or more 15199 0:075
psiS ocChlBin Number, biological children None 53199 0:27
One 60199 0:3
Two 39199 0:2
Three or more 47199 0:24
psiS ocChlS tn Number, step-children None 192199 0:96
One 2199 0:01
Two or more 5199 0:025
a Including parents, siblings, aunts, uncles, cousins, or grandparents, for instance.
b Discretized for presentation. c Ages 18 and younger.
Sampled o enders had, on average, just over one year of drug testing data (M = 14:7;S D =
9:86) with about 5 drug testing events monthly, M = 4:69;S D = 1:57. Not all of the 7 substances
were screened at each event: o enders35 were screened for at least 3 substances and, typically,
for 4 or 5. Nearly all o enders were screened at least once for phencyclidine, cocaine, marijuana,
and opiates (190=199 = 0:95, 194=199 = 0:97, and 194=199 = 0:97, 194=199 = 0:97, respectfully).
67=199 = 0:34 of the sample was screened at least once for methadone, 74=199 = 0:37 for alcohol,
and 77=199 = 0:39 for amphetamines.
35Excluding n = 5 o enders who did not have any drug testing events.
90
Table 18
Presentence Investigation report, educational and employment characteristics, n = 199.
Symbol Description Level Proportion
psiEduGED Earned GEDa Yes 33199 0:17
psiEduGrdAtb Highest grade attempted [0;11) 63199 0:32
11 46199 0:23
12 62199 0:31
[13;16] 28199 0:14
psiEduGrdCmb Highest grade completed [6;10) 53199 0:27
[10;12) 71199 0:36
12 51199 0:26
[13;16] 24199 0:12
psiEmpLess Any jobs of duration < 30 days Yes 24199 0:12
psiEmpMore Any jobs of duration > 30 days Yes 155199 0:78
a If not high school graduate. b Discretized for presentation.
Table 19
Presentence Investigation report, health characteristics, n = 199.
Symbol Description Level Proportion
psiMedDis Number, physical disabilities None 191199 0:96
One or more 8199 0:0402
psiMedMdDx Has the o ender been diagnosed with a mental illness Yes 17199 0:085
psiMedMdDr Does the o ender take any psychotropic medications? Yes 13199 0:065
psiMedMdHo Times previously hospitalized for mental health None 192199 0:96
One or more 7199 0:035
psiMedIn j How many injuries are listed? None 145199 0:73
One 41199 0:21
Two or more 13199 0:065
psiMedWouGun Wounds result from gunshots None 179199 0:90
One or more 20199 0:10
psiMedWouS tb Wounds resulting from stabbings? None 192199 0:96
One or more 7199 0:035
91
Table 20
U.S. Census, subset of sociodemographic characteristics, 3-Factor solution, oblimin, k = 436.
Variable Description Factor/Loading
Concentrated disadvantage
piPov Populationc, income in 1999 < poverty levelPopulationc 0:76
piPub Households, public assistance incomeHouseholds 0:88
piFHH Households, female, no husband presentHouseholds 0:89
piUne Populationd, unemployedPopulationd 0:56
rtAK ChildreneAdultsf 0:86
piBlk Population, BlackPopulation 0:67
Immigrant concentration
piHis Population,Hispanic or LatinoPopulation 0:83
piFor Population, foreign bornPopulation 0:92
Residential stability
piRnt Housing units, renter occupiedOccupied housing units 0:63
piDi f Populationa, di erent house in 1995Populationa 0:87
Table 21
One-sided 95% confidence limit below (LCLM) and above (UCLM) mean densities of alcohol
licensees per 200,000 m2 of BG land area, 2004, k = 433.
Variable Description LCLM UCLM
alcBWS 1 O -premises sale of beer, wine, and spirits 0:38 0:64
alcBW1 O -premises sale of beer and wine only 0:73 1:2
alcBWS 2 On-premises sale of beer, wine, and spirits 0:59 1:4
alcBW2 On-premises sale of beer and wine only 0:024 0:17
alcDens Total licensees 1:9 3:2
92
Table 22
One-sided 95% confidence limit below (LCLM) and above (UCLM) mean block-group arrests
per 1,000 residents ages 18 and older, 2004, k = 436.
Variable Description LCLM UCLM
arrVio Violent 19:5 25:7
arrPro Property 10:9 20:7
arrDrg Drug- and alcohol-related 18:9 42:4
arrPub Public-order 19:3 38:6
arrOth Unclassified 50:6 93:9
arrDens Total Arrests 124 217
Table 23
Description of criteria.
Criteria M S D N
Failed a drug testing event while supervised 0:79 0:41 199
Rate of positive screens for alcohol 0:14 0:26 74
Rate of positive screens for methadone 0:02 0:10 67
Rate of positive screens for amphetamines 0:00 0:02 77
Rate of positive screens for cocaine 0:07 0:14 194
Rate of positive screens for marijuana 0:10 0:17 194
Rate of positive screens for opiates 0:03 0:10 194
Rate of positive screens for phencyclidine 0:03 0:10 190
Convicted of new crime during supervision period 0:26 0:44 199
Convicted of new crime during post-supervision period 0:25 0:43 199
Number of supervision-related violations 1:53 2:47 199
Number of drug-related violations 2:40 3:81 199
Terminated unsuccessfully 0:58 0:50 199
Months until termination 17:24 11:41 199
93
Among those screened, 93=194 = 0:48 failed at least one screen for cocaine, 41=194 = 0:21
failed at least one screen for opiates, 33=190 = 0:17 failed at least one for phencyclidine, 81=194 =
0:42 failed at least one for marijuana, 33=74 = 0:45 failed at least one for alcohol, 5=77 = 0:065
failed at least one for amphetamines, and 4=67 = 0:06 failed at least one screen for methadone.
MS1. The criterion for MS1 is whether the ith probationer ever tested positive, provided a
bogus specimen, or failed to appear for a drug testing event while supervised (if so, outS 1i = 1),
and, indeed, a large fraction (158=199 = 0:79) of the sampled probationers did so. The 95%CI
around mean outS 1 spanned [0:74;0:85]. Interest centers on the predicted probability of the cri-
terion, bPr (outS 1i = 1).
Potential predictors were included in a general model which was then recursively parti-
tioned36 into the binary tree shown in Figure 3a.37 This classification tree, since it was likely too
complex to validate, was trimmed back to that shown in Figure 3b.38
Predictors having the largest e ect on whether probationers ever fail a drug-testing event
while supervised included (a) the number of substances out of 7 the o ender ever used psiS ubEU1i,
(b) the expected number of days of supervision daysS upExpi, and (c) the rate of property-related
arrests within the block-group arrPro:di; remaining predictors did not appear in the model.
The resultant model separated sampled probationers into 4 groups. The groups (and pre-
dicted values) included (a) probationers having used fewer than 1 substance out of 7 (No); (b) pro-
bationers having used 1 or more substances out of 7 and expecting more than 563 days of super-
vision (Yes); (c) probationers having used 1 or more substances out of 7, expecting less than 563
days of supervision, and living within a BG having a rate of property-related arrests below 25.55
36Using the R library RPART (Therneau & Atkinson, 1997).
37The Gini rule was used for splitting, prior probabilities were set proportional to observed frequencies, and altered
priors were used for the loss function.
38Cost-complexity pruning was based on the 1 S E rule among cross-validated data.
94
psiSubEU1< 0.5
daysSupX< 563
arrPro.d>=25.55
psiSubCoEU=No
psiEduGrdCm< 9.5
No 
No 
No Yes
Yes
Yes
(a)
psiSubEU1< 0.5
daysSupX< 563
arrPro.d>=25.55
No 
No Yes
Yes
(b)
Figure 3. Initial (a) and pruned (b) classification trees predicting outS 1i, MS1, n = 199.
(Yes); and (d) probationers having used 1 or more substances out of 7, expecting less than 563
days of supervision, and living within a BG having a rate of property-related arrests above 25.55
(No).
An initial model
bPr (outS 1i = 1jxi) = 1
1 + exp   (xi ) 
= 11 + exp   ( 
0 + 1xi1 + 2xi2 + 3xi3)
 ;
where xi1 = psiS ubEU1i, xi2 = daysS upExpi, and xi3 = arrPro:di, was fitted to the sample data
with the pruned-tree predictors in their original form; parameter estimates are shown in the first
column of Table 24.39
39A robust estimator of variance was used in place of the standard estimator to adjust for k = 122 block-group
clusters (Huber, 1967; Rogers, 1993; H. White, 1980).
95
Table 24
Parameter estimates, logistic regression of outS 1i, MS1, n = 199.
b=z b=z
psiS ubEU1 0:451 
(2:23)
daysS upExp 0:001
(1:95)
arrPro:d  0:011
( 1:95)
psiS ubEU1 0:477  
(2:69)
daysS upExp 1:089  
(2:80)
arrPro:d 3:401
(0:79)
  0:383  3:312
( 0:67) ( 0:88)
Model  2 13:669  15:874  
 p < 0:05;  p < 0:01;   p < 0:001.
The Wald test40 of the hypothesis that all coe cients except the intercept were zero H0 :
 1 =  2 =  3 = 0 was rejected,  2W = 13:67; d f = 3; p = 0:00. The test of H0:  1 = 0 was rejected
( 2W = 4:97; d f = 1; p = 0:03), but neither test of H0:  2 = 0 ( 2W = 3:81; d f = 1; p = 0:05) nor
of H0:  3 = 0 ( 2W = 3:81; d f = 1; p = 0:05) were rejected.
The model correctly classified 0:80 of the sample. Its ability to discriminate probationers
with respect to the criteria is described succinctly by the area under the receiver operating char-
acteristic curve (ROC) curve (AUC). This statistic suggests how likely it is that a probationer
ever failing a drug testing event will have a higher predicted probability than their never failing
40The likelihood-ratio chi-square ( 2
L) test is likely invalid given the robust estimator of variance used to adjust for
k = 122 block-group clusters.
96
counterparts. Hosmer and Lemeshow (2000) suggest interpreting AUC as
AUC = 0:5; Lacks discriminatory power;
0:7 AUC < 0:8; Acceptable discriminatory power;
0:8 AUC < 0:9; Excellent discriminatory power;
0:9 AUC < 1; Outstanding discriminatory power.
In this case the model demonstrated acceptable discrimination, AUC = 0:73.
The model had deviance ( 2 log L) = D = 178:55, Akaike?s (1973) information criterion
(AIC) = 0:937, and Bayesian information criterion (BIC) =  853:643; the Hosmer-Lemeshow
goodness-of-fit statistic (bc) suggested the model was empirically consistent,  2HL = 4:49; d f =
8; p = 0:81.
Focus then turned to refining this preliminary model in terms of parametric relationships
and scale beginning with the relationship between outS 1i and psiS ubEU1i.
A plot of the lowess smoothed logit against linear psiS ubEU1i gave a counterintuitive im-
pression of the influence of psiS ubEU1i on outS 1i. There was an apparent near linear increase
in the log-odds of outS 1i that peaked at psiS ubEU1i = 5 and declined thereafter. This apparent
nonlinearity resulted from the poor behavior in the upper tail of the distribution. Although it poten-
tially ranged within [0;7], only 4=199 = 0:02 of the sampled probationers reported psiS ubEU1 > 5.
To account for this the last three levels of psiS ubEU1i were collapsed as psiS ubEU1 i . Replot-
ting the smoothed logit against psiS ubEU1 i gave a more intuitive impression of the influence of
psiS ubEU1i on outS 1i. In addition, there was no evidence to suggest the between-level spacings
were substantially dissimilar.
Attention then turned to the relationship between daysS upExpi and outS 1i. Although the
e ect of daysS upExpi was not significantly di erent from zero (see, column 1 of Table 24), the
tree in Figure 3b indicated this might be due to incorrect functional form: it suggested a break
97
in daysS upExpi near its median of 548 days. Such a nonlinearity was entirely within reason, as
those probationers expecting relatively shorter sentences might be more willing to abstain than
those facing relatively longer sentences who might be overwhelmed by abstention.
A plot of the smoothed logit against linear daysS upExpi showed a clear nonlinearity with
the logit dropping sharply until roughly the first quartile (0:25Q = 365) then turning sharply
upward and continuing essentially linearly. The logit flattened slightly at just over 1,000 days
and then reassumed the slope at around 2,000 days. Despite the apparent cupping e ect in the
right half of the distribution there was little to be gained in modeling it: only 34=199 = 0:17 of
the sampled o enders had daysS upExpi greater than 1,000 days; only 3=199 = 0:015 had values
greater 2,000. On the other hand, the break at roughly 1 year was theoretically interesting.
I tested whether polynomial terms well-described the relationship, but neither quadratic nor
cubic functions41 were statistically better than a linear term. Nor were first, second, or third degree
fractional polynomials.42 As an alternative I tested three di erent piecewise regressions. The first
allowed one linear e ect for daysS upExp at below 365 days and a di erent linear e ect above.
The next allowed one linear e ect up to 548 days and a di erent e ect thereafter. The last allowed
one linear e ect up to 365 days, a second linear e ect between 365 days and 548 days, and a third
e ect from 548 days onward. None of these splines were significantly better than a linear term.
Last I explored binary splits representing high and low values of daysS upExpi. I tested separate
breakpoints at 548 days and at 365 days. Both represented the data better than a linear term; the
latter outperformed the former. Given this, daysS upExpi was split at 365 days and included in
the model as daysS upExp i .
I then turned to the relationship between arrPro:di and outS 1i. A plot of the smoothed
logit against linear arrPro:di showed a clear negative trend beginning at roughly arrPro:d  
41Higher powers were not explored.
42Higher degrees were not explored.
98
40. The tree in Figure 3b suggests a break in arrPro:d at 25.5. Both apparent e ects are likely
artifacts stemming from poor behavior in the upper tail of the distribution: mean arrPro:d was
roughly equal to the 0:78Q, M = 15:40, S D = 38:60. A normal quantile-comparison plot of
arrPro:di identified several outlying values in the right tail; all 3 were represented by the only
3 non-DC residents in the sample. Although a nonlinear function might reasonably approximate
the apparent relationship, given that the arrest densities outside of the District were inessential
and that removing non-DC residents essentially linearized the relationship, these 3 cases were
temporarily dropped from the regression of outS 1i. A plot of the smoothed logit against arrPro:d i
confirmed this. Still, after making these changes some non-normality was still present and was
corrected using the unconditional Box-Cox method. The maximum-likelihood (ML) normalizing
transformation parameter  in x was estimated as b = 0:178; the predictor was normalized by
applying the transformation arrPro:d i = (arrPro:di + 0:5)0:18.
Following these changes the e ect of arrProd i was not significantly di erent from zero
( 2W = 0:46; d f = 1; p = 0:50). Thus, so as not to reduce potentially predictive informa-
tion with respect to the other two predictors in the model, the 3 non-DC cases were re-included
and the model was refitted. Parameter estimates for the regression of outS 1i on psiS ubEU1 i ,
daysS upExp i , and arrPro:d i
bPr (outS 1i = 1jxi) = 1
1 + exp   (xi ) 
= 11 + exp   ( 
0 + 1xi1 + 2xi2 + 3xi3)
 ;
where xi1 = psiS ubEU1 i , xi2 = daysS upExp i , and xi3 = arrPro:d i , are shown in the second
column of Table 24.43 The Wald44 test of H0 :  1 =  2 =  3 = 0 was rejected,  2W = 15:87; d f =
3; p = 0:00. The test of H0 :  1 = 0 was rejected ( 2W = 7:22; d f = 1; p = 0:01), the test of
43A robust estimator of variance was used in place of the standard estimator to adjust for k = 122 block-group
clusters (Huber, 1967; Rogers, 1993; H. White, 1980).
44The  2
L test is likely invalid given the robust estimator of variance used to adjust for k = 122 block-group clusters.
99
0
.2
.4
.6
.8
1
Observed
0 .2 .4 .6 .8 1
Model implied
Figure 4. Comparison between model-implied probabilities of experiencing outS 1i and a moving
average of the proportion of probationers ever failing a drug testing event, MS1, n = 199.
H0 :  2 = 0 was rejected ( 2W = 7:84; d f = 1; p = 0:01). The test of H0 :  3 = 0 was not rejected
( 2W = 0:62; d f = 1; p = 0:43).
Model calibration and discrimination remained virtually unchanged from the preliminary
model. A visual indication of model calibration is given in the plot in Figure 4 which compares
predicted probabilities b from the regression of outS 1i with a moving average of the proportion
of probationers having at least one positive drug screen. The thick line represents the fraction of
probationers failing a drug-testing event while supervised across levels of predicted probabilities.
Close tracking between this and the diagonal line indicates good calibration. That the thick line in
Figure 4 indeed tracks closely with the diagonal suggests the model is largely well-calibrated.
Among the sample, 0:80 were correctly classified. The model demonstrated acceptable
discrimination, AUC = 0:72. It had D=180.23, AIC=0.946, and BIC= 851.965; the bc suggested
it was empirically consistent,  2HL = 5:74; d f = 8; p = 0:68.
100
Table 25
Lower (LCLM) and upper (UCLM) bias-corrected 95% confidence limits, MS1, n = 199.
LCLM UCLM
psiS ubEU1 0:102 0:852
daysS upExp 0:286 1:893
arrPro:d  6:046 12:847
The regression of outS 1i on psiS ubEU1 i , daysS upExp i , and arrPro:d i shown in the
second column of Table 24 was validated by bootstrapping. Bias-corrected confidence intervals
around the estimated parameters are shown in Table 25. As indicated in the table, the e ect
of arrPro:d i is unlikely to replicate in the population. On the other hand, the e ects of both
psiS ubEU1 i and daysS upExp i are likely to be found in the population; interpretations of the
model are based on these latter two e ects while holding arrPro:d at its mean.
Predicted values b ranged within the interval [:002;:965], with mean b = 0:794;S D =
0:14. The two most important characteristics influencing bPr (outS 1i = 1) are psiS ubEU1 and
daysS upExp . Holding all else equal, having used 5 or more substances compared to none in-
creases the predicted probability b an o ender will ever test positive, provide a bogus specimen,
or fail to appear for a drug testing event while supervised by 0:47, from 0:42 to 0:89. The e ect
of psiS ubEU1 is greatest when moving from none to 1, which increases b by 0:11, from 0:58 to
0:69. b continues to increase with each additional substance ever used at a decreasing rate.
The expected length of supervision also plays a role in whether probationers will ever test
positive, provide a bogus specimen, or fail to appear for a drug testing event while supervised.
Those probationers expecting longer periods of supervision (i.e., > 1 year) are more likely to fail.
Their odds of failing are, in fact, 1.7 times larger. Expecting a supervision period of one year or
more versus less than one year is associated with an increase of 0:17 in b , from 0:71 to 0:88.
101
Figure 5. Predicted probabilities an o ender in the population will test positive, fail to appear, or
provide a bogus specimen for a drug testing event while supervised positive for cocaine, by the
number of substances ever used and length of supervision, MS1, n = 199.
The plot in Figure 5 shows the positive e ect of psiS ubEU1 i on outS 1i and how this e ect
di ers by daysS upExp i . When psiS ubEU1 i is low the e ect of daysS upExp i is relatively
large, but as psiS ubEU1 i increases the e ect shrinks. For instance, among probationers having
used none of the 7 substances, the b is 0:26 higher among those expecting a supervision period of
one year or more versus less than one year. This same di erence among probationers having used
5 or more of the 7 substances is 0:07.
MS2A. The criterion for MS2A is the number of positive tests for alcohol outDrgTotPosi1.
The 95%CI around mean npositivesnscreens is shown in Table 26. As indicated, screens for alcohol were
relatively rare among the sample: just under 3 out of every 8 (74=199 = 0:37) probationers were
screened. One or more screens were positive for about half (33=74 = 0:45) of those screened. As
screens for alcohol were relatively uncommon MS2A was excluded from the present analysis.
102
Table 26
One-sided 95% confidence limit below (LCLM) and above (UCLM) mean rate of positive
screens, by substance.
Criterion Description LCLM UCLM nscreened
outDrgTotPosi1
outDrgTotS cri1 Alcohol 0:0755 0:198 74
outDrgTotPosi2
outDrgTotS cri2 Methadone 0:0 0:0444 67
outDrgTotPosi3
outDrgTotS cri3 Amphetamine 0:0 0:00655 77
outDrgTotPosi4
outDrgTotS cri4 Cocaine 0:0519 0:0908 194
outDrgTotPosi5
outDrgTotS cri5 Marijuana 0:0724 0:121 194
outDrgTotPosi6
outDrgTotS cri6 Opiates 0:0187 0:0475 194
outDrgTotPosi7
outDrgTotS cri7 Phencyclidine 0:0153 0:0427 190
MS2B. The criterion for MS2B is the number of positive tests for methadone outDrgTotPosi2.
Screens for methadone were relatively rare among the sample with only 67=199 = 0:34 of the sam-
ple being screened at least once. Screens were positive one or more times for only 4=67 = 0:06 of
those probationers screened. The 95%CI around mean npositivesnscreens is shown in Table 26. As screens
for methadone were relatively uncommon MS2B was excluded from analyses.
MS2C. The criterion for MS2C is the number of positive tests for amphetamines outDrgTotPosi3.
Screens and positive screens for amphetamines were rare. Among the 77=199 = 0:39 of the sample
that was screened, only 5=77 = 0:065 had one or more positive results. The 95%CI around mean
npositives
nscreens is shown in Table 26. As screens for amphetamines were relatively uncommon MS2C was
excluded from analyses.
MS2D. The criterion for model MS2D is the number times the ith probationer tested pos-
itive for cocaine outDrgTotPosi4; interest centers on the expected rate of this criterion in the
population b i. Screens for cocaine occurred 5 times a month, on average, throughout the super-
vision period, M = 5:05;S D = 7:2. 194=199 = 0:97 of the sample was screened at least once for
103
cocaine, one or more of which were positive for 93=194 = 0:48 of those screened. The 95%CI
around mean npositivesnscreens is shown in Table 26.
Potential predictors were included in a general model which was then recursively parti-
tioned45 into the binary tree shown in Figure 6a using Poisson-splitting methods (see, Breiman et
al., 1984; Therneau & Atkinson, 1997).46 This classification tree, since it was likely too complex
to validate, was trimmed back to that shown in Figure 6b.47
Predictors having the largest e ect on the rate of screens for cocaine included (a) age at
the time of assessment rnsAgei, (b) the number of substances ever used psiS ubEU1i, (c) the total
number of children younger than age 18 psiS ocToni, and (d) the expected number of days of
supervision daysS upExpi; remaining predictors did not appear in the model.
The resultant model separated sampled probationers into 5 groups. The groups (and pre-
dicted values) included (a) those younger than age 33 and having used 3 or fewer substances
(0.75); (b) those younger than age 33 and having used more than 3 substances (7.7); (c) those age
33 or older, having 3 or fewer children under age 18, and expecting to serve less than 455 days of
supervision (2.3); (d) those age 33 or older, having 3 or fewer children under age 18, and expecting
to serve more than 455 days of supervision (11); and (e) those age 33 or older and having more
than 3 children under age 18 (27).
Before fitting an initial model, several transformations were made. First, age i was calcu-
lated as age i = rnsAgei10  1 to allow for non-linearities in the e ect of age on the rate of positive
screens for cocaine. Next, the last three levels of psiS ubEU1i were collapsed as psiS ubEU1 i to
remedy sparse representation in the upper regions of the predictor. Finally, psiS ocToni was trun-
45Using the R library RPART (Therneau & Atkinson, 1997).
46The Gini rule was used for splitting, prior probabilities were set proportional to observed frequencies, and altered
priors were used for the loss function.
47Cost-complexity pruning was based on the 1 S E rule among cross-validated data.
104
rnsAge< 33.5
psiSubEU1< 3.5
psiFamRaiXF=No
psiSocTon< 3.5
daysSupX< 455.5
outV2< 4rnsAge< 45.5
rtAK>=0.3174
0.411.5
7.7
1.1 6.7
2.2 9
20
27
(a)
rnsAge< 33.5
psiSubEU1< 3.5 psiSocTon< 3.5
daysSupX< 455.5
0.75 7.7
2.3 11
27
(b)
Figure 6. Initial (a) and pruned (b) classification trees predicting outDrgTotPosi4, MS2D, n =
194.
cated at 4 as psiS ocTon i . This, too, was done in response to sparse representation in the upper
levels of the predictor.
An initial model
bPr (outDrgTotPosi4jxi) = exp(  i) 
outDrgTotPosi4
i
outDrgTotPosi4!
where  i = exp(xi ) and xi1 = age i , xi2 = psiS ubEU1 i , xi3 = daysS upExpi, and xi4 =
psiS ocTon i , was fitted to the sample data with the pruned-tree predictors; parameter estimates
are shown in the first column of Table 27.48
After this initial model was fit there was significant evidence that the observations were
overdispersed with respect to the Poisson model, G2 = 1153:370; p = 0:000. Thus, MS2D was
48A robust estimator of variance was used in place of the standard estimator to adjust for k = 122 block-group
clusters (Huber, 1967; Rogers, 1993; H. White, 1980). Also, to account for varying times-at-risk, the logged number
of drug screens log (outDrgTotS cri4) was included and constrained to a coe cient of 1.
105
Table 27
Parameter estimates, Poisson and negative binomial regressions of outDrgTotPosi4, MS2D, n =
194.
b=z b=z b=z
age  7:698    7:464    7:922   
( 5:96) ( 6:72) ( 6:39)
psiS ubEU1 0:175 0:482   0:482   
(1:63) (4:62) (4:76)
psiS ocTon 0:190 0:198 0:198
(1:44) (1:66) (1:75)
daysS upExp 0:000 0:001
(0:17) (1:65)
daysS upExp 0:691
(1:78)
  1:096  2:460    2:370   
( 2:08) ( 4:60) ( 4:28)
Overdispersion  1:051   1:037   
(7:86) (7:44)
Model  2 56:789   81:913   75:001   
 p < 0:05;  p < 0:01;   p < 0:001.
respecified to include a multiplicative disturbance term  i to capture unobserved heterogeneity.
The model
bPr (outDrgTotPosi4jxi) = exp( e i)e ioutDrgTotPosi4
outDrgTotPosi4!
where e i = exp(xi +  i) and xi1 = age i , xi2 = psiS ubEU1 i , xi3 = daysS upExpi, and xi4 =
psiS ocTon i , was then fitted to the sample data. Parameter estimates from this regression are
shown in the second column of Table 27.49
The model had D=771.03, AIC=4.036, and BIC= 219.325.
49A robust estimator of variance was used in place of the standard estimator to adjust for k = 122 block-group
clusters (Huber, 1967; Rogers, 1993; H. White, 1980). Also, to account for varying times-at-risk, the logged number
of drug screens log (outDrgTotS cri4) was included and constrained to a coe cient of 1.
106
The Wald test50 of the hypothesis that all coe cients except the intercept were zero H0 :
 1 =  2 =  3 =  4 = 0 was rejected,  2W = 81:91; d f = 4; p = 0:00. The tests of H0 :  1 = 0
( 2W = 45:16; d f = 1; p = 0:00) and of H0 :  2 = 0 ( 2W = 21:36; d f = 1; p = 0:00) were
rejected, but neither tests of H0 :  3 = 0 ( 2W = 2:73; d f = 1; p = 0:10) nor of H0 :  4 = 0
( 2W = 2:76; d f = 1; p = 0:10) were rejected.
By comparison, the Poisson regression (PR) shown in the first column of Table 27 per-
formed at its worse in predictions of 0 where it greatly underpredicted counts. The negative
binomial regression (NBR) shown in the second column of Table 27 performed at its worse
in predictions of counts of 3 where it, too, underpredicted counts albeit on a relatively smaller
scale. Overall, the NBR, with mean absolute di erence between predicted and observed values
j b o b pj= 0:008, outperformed the PR, with itsj b o b pj= 0:078.
Before moving on, focus turned to refining the model in terms of parametric relationships
and scale beginning with the relationship between outDrgTotPosi4 and psiS ubEU1 i . In nei-
ther plot of lowess smoothed rates of positive screens against linear psiS ubEU1 i nor of linear
psiS ocTon i were there substantial departures from linearity. These two predictors were thus left
unchanged. On the other hand, a plot of the lowess smoothed rates of positive screens against lin-
ear daysS upExpi did suggest a non-linearity. I explored several alternatives to linearity, however,
neither quadratic nor cubic functions51 were statistically better than a linear term. Nor were first,
second, or third degree fractional polynomials.52 I examined several di erent splines to account
for the non-linearity, but none provided an acceptable fit. Ultimately, daysS upExpi was split at
365 days and included in the model as daysS upExp i .
50The  2
L test is likely invalid given the robust estimator of variance used to adjust for k = 122 block-group clusters.
51Higher powers were not explored.
52Higher degrees were not explored.
107
The model
bPr (outDrgTotPosi4jxi) = exp( e i)e ioutDrgTotPosi4
outDrgTotPosi4!
where e i = exp(xi +  i) and xi1 = age i , xi2 = psiS ubEU1 i , xi3 = daysS upExp i , and xi4 =
psiS ocTon i , was then fitted to the sample data. Parameter estimates from this regression are
shown in the third column of Table 27.53
The model had D=769.19, AIC=4.027, and BIC= 221.168.
The Wald test54 of the hypothesis that all coe cients except the intercept were zero H0 :
 1 =  2 =  3 =  4 = 0 was rejected,  2W = 75:00; d f = 4; p = 0:00. The tests of H0 :  1 = 0
( 2W = 40:87; d f = 1; p = 0:00) and of H0 :  2 = 0 ( 2W = 22:67; d f = 1; p = 0:00) were
rejected, but neither tests of H0 :  3 = 0 ( 2W = 3:18; d f = 1; p = 0:07) nor of H0 :  4 = 0
( 2W = 3:07; d f = 1; p = 0:08) were rejected.
A visual indication of model calibration is given in the plot in Figure 7. This plot shows the
observed and predicted probabilities of counts zero through 9. As indicated, the model-implied
and observed probabilities track closely throughout the distribution, with the poorest fit in counts
of 3.
The NBR of S 2D on age i , psiS ubEU1 i , daysS upExp i , and psiS ocTon i , shown in the
third column of Table 27 was validated by bootstrapping. Bias-corrected confidence intervals
around the estimated parameters are shown in Table 28. As indicated, the e ects of age i and
psiS ubEU1 i were likely to replicate in the population, but neither that of psiS ocTon i nor of
daysS upExp i were.
53A robust estimator of variance was used in place of the standard estimator to adjust for k = 122 block-group
clusters (Huber, 1967; Rogers, 1993; H. White, 1980). Also, to account for varying times-at-risk, the logged number
of drug screens log (outDrgTotS cri4) was included and constrained to a coe cient of 1.
54The  2
L test is likely invalid given the robust estimator of variance used to adjust for k = 122 block-group clusters.
108
Figure 7. Observed and NBR model-implied counts of outDrgTotPosi4, MS2D, n = 194.
The most important characteristics influencing the rate of positive screens for cocaine in-
cluded age at the time of assessment rnsAge i and the number of substances ever used psiS ubEU1 i .
The expected rate of positive screens for cocaine increases with age. Holding all else constant,
the rate of positives for probationers at the lower quartile (0:25Q  23) of age is roughly 0.03.
It increases to 0.06 for those at the median (0:5Q  31) age and to roughly 0.12 for those at the
upper quartile (0:75Q 43). Holding all other predictors constant each additional substance ever
Table 28
Lower (LCLM) and upper (UCLM) bias-corrected 95% confidence limits, MS2D, n = 194.
LCLM UCLM
age i  10:6  5:25
psiS ubEU1 i :27 :694
psiS ocTon i  :0753 :472
daysS upExp i  :0811 1:46
109
Figure 8. Expected rates of positive screens for cocaine, by the number of substances ever used
and age, MS2D, n = 194.
used increases the expected number of positive screens for cocaine by a factor of 1.6 or roughly
62%.
The plot in Figure 8 shows the e ect of age at its 0:25Q, 0:5Q, and 0:75Q on the rate of
positive screens for cocaine across the number of substances ever used. The number of substances
ever used is associated with increases in the rate of positive screens for cocaine and these increases
vary by age. For probationers having used none of the illegal substances age makes little di erence
on the expected rate, but as the number of substances ever used increases the expected rates of
positives are higher for older versus younger o enders. For instance, holding all else constant the
expected rate of positive screens for probationers having used none of the substances is 0.0075 for
probationers age 23, 0.018 for those age 31, and 0.036 for those age 42. Rates for those having
used 3 substances are 0.032 for probationers age 23, 0.077 for those age 31, and 0.15 for those
age 42. For those having used 5 or more substances rates of positives are 0.083 for those age 23,
0.2 for those age 31, and 0.4 for those age 42.
110
MS2E. The criterion for MS2E is the number times the ith probationer tested positive for
marijuana outDrgTotPosi5; interest centers on the expected rate of this criterion in the population
b i. Screens for marijuana occurred roughly 5 times monthly, on average, throughout the supervi-
sion period, M = 4:98;S D = 7:19. 194=199 = 0:97 of the sample was screened at least once for
marijuana; one or more of these were positive for 81=194 = 0:42 of the probationers. The 95%CI
around mean npositivesnscreens is shown in Table 26.
Potential predictors were included in a general model which was then recursively parti-
tioned55 into the binary tree shown in Figure 9a using Poisson-splitting methods (see, Breiman et
al., 1984; Therneau & Atkinson, 1997).56 This classification tree, since it was likely too complex
to validate, was trimmed back to that shown in Figure 9b.57
Predictors having the largest e ect on the rate of positive tests for marijuana included (a) age
at the time of assessment rnsAgei, (b) population density within the block-group popDensi, and
(c) the total number of prior juvenile cases psiCrmJuvCasni; remaining predictors did not appear
in the model.
The resultant model separated sampled probationers into 4 groups. The groups (and pre-
dicted values) included (a) those ages 34 and older (0.2); (b) those younger than age 34 and living
within a block-group with population density equal to or greater than 3.609 (7.8); (c) those younger
than age 34, living within a block-group with population density less than 3.609, and having no
prior juvenile cases (9.1); and (d) those younger than age 34, living within a block-group with
population density less than 3.609, and having 1 or more prior juvenile cases (30).
55Using the R library RPART (Therneau & Atkinson, 1997).
56The Gini rule was used for splitting, prior probabilities were set proportional to observed frequencies, and altered
priors were used for the loss function.
57Cost-complexity pruning was based on the 1 S E rule among cross-validated data.
111
rnsAge>=34.5
popDens>=3.609
psiSocMarSta=Single-cohabiting,Married-living apart
rtAK>=0.6192
psiEmpMore>=0.5
psiEduGrdAt>=10.5
psiFamBir=Married-living together
psiCrmJuvCasn< 0.5
arrPub.d< 8.263
0.2
1.9
1.7
0.767.9
17
20
2 13
30
(a)
rnsAge>=34.5
popDens>=3.609
psiCrmJuvCasn< 0.5
0.2
7.8
9.1 30
(b)
Figure 9. Initial (a) and pruned (b) classification trees predicting outDrgTotPosi5, MS2E, n =
194.
Before fitting an initial model, several transformations were made. First, rnsAgei was
transformed as age i = rnsAgei10 3  36:33. Second, as its levels were sparse in the upper tail,
psiCrmJuvCasni was truncated at 4 as psiCrmJuvCasn i .
An initial model
bPr (outDrgTotPosi5jxi) = exp(  i) 
outDrgTotPosi5
i
outDrgTotPosi5!
where  i = exp(xi ) and xi1 = age i , xi2 = popDensi, and xi3 = psiCrmJuvCasn i , was fitted to
the sample data with the pruned-tree predictors; parameter estimates are shown in the first column
of Table 29.58
58A robust estimator of variance was used in place of the standard estimator to adjust for k = 122 block-group
clusters (Huber, 1967; Rogers, 1993; H. White, 1980). Also, to account for varying times-at-risk, the logged number
of drug screens log (outDrgTotS cri5) was included and constrained to a coe cient of 1.
112
Table 29
Parameter estimates, Poisson and negative binomial regressions of outDrgTotPosi5, MS2E, n =
194.
b=z b=z b=z
age  0:038    0:037   0:037  
( 6:32) ( 3:14) ( 3:13)
popDens  0:053  0:030
( 1:52) ( 1:24)
psiCrmJuvCasn 0:140 0:113 0:107
(1:74) (1:19) (1:14)
popDens  0:465
( 1:11)
  2:527    2:626    2:465   
( 10:43) ( 10:93) ( 6:92)
Overdispersion  1:167   1:165   
(6:04) (6:00)
Model  2 52:995   19:691   17:673   
 p < 0:05;  p < 0:01;   p < 0:001.
After this initial model was fit there was significant evidence that the observations were
overdispersed with respect to the Poisson model, G2 = 1239:181; p = 0:000. Thus, MS2E was
respecified to include a multiplicative disturbance term  i to capture unobserved heterogeneity.
The model
bPr (outDrgTotPosi5jxi) = exp( e i)e ioutDrgTotPosi5
outDrgTotPosi5!
where e i = exp(xi +  i) and xi1 = age i , xi2 = popDensi, and xi3 = psiCrmJuvCasn i , was then
fitted to the sample data. Parameter estimates from this regression are shown in the second column
of Table 29.59
59A robust estimator of variance was used in place of the standard estimator to adjust for k = 122 block-group
clusters (Huber, 1967; Rogers, 1993; H. White, 1980). Also, to account for varying times-at-risk, the logged number
of drug screens log (outDrgTotS cri5) was included and constrained to a coe cient of 1.
113
The model had D=805.31, AIC=4.203, and BIC= 190.311.
By comparison, the PR shown in the first column of Table 29 performed at its worse in
predictions of counts of 0 where it underpredicted counts. The NBR shown in the second column
of Table 29 performed at its worse in predictions of 1 where it overpredicted counts. Overall,
the NBR, with mean absolute di erence between predicted and observed valuesj b o b pj= 0:019,
outperformed the PR, withj b o b pj= 0:064.
I then turned to assessing scale beginning with the relationship involving psiCrmJuvCasn i .
A plot of the univariable lowess smooth of outDrgTotPosi5 versus psiCrmJuvCasn i did not sug-
gest substantial nonlinearity of psiCrmJuvCasn i in the rate of positive screens for marijuana.
This predictor was thus left as is. On the other hand, a similar plot of outDrgTotPosi5 ver-
sus popDensi did suggests a non-linearity. To account for this popDensi was replaced with
popDens i = ppopDensi=10.
The model
bPr (outDrgTotPosi5jxi) = exp( e i)e ioutDrgTotPosi5
outDrgTotPosi5!
where e i = exp(xi +  i) and xi1 = age i , xi2 = popDens i , and xi3 = psiCrmJuvCasn i , was then
refitted to the sample data. Parameter estimates from this regression are shown in the third column
of Table 29.60
The model had D=805.44, AIC=4.203, and BIC= 190.187.
The Wald test61 of the hypothesis that all coe cients except the intercept were zero H0 :
 1 =  2 =  3 = 0 was rejected,  2W = 17:67; d f = 3; p = 0:00. The test of H0:  1 = 0 was rejected
60A robust estimator of variance was used in place of the standard estimator to adjust for k = 122 block-group
clusters (Huber, 1967; Rogers, 1993; H. White, 1980). Also, to account for varying times-at-risk, the logged number
of drug screens log (outDrgTotS cri5) was included and constrained to a coe cient of 1.
61The  2
L test is likely invalid given the robust estimator of variance used to adjust for k = 122 block-group clusters.
114
Figure 10. Observed and NBR model-implied counts of outDrgTotPosi5, MS2E, n = 194.
( 2W = 9:77; d f = 1; p = 0:00) but neither test of H0:  2 = 0 ( 2W = 1:22; d f = 1; p = 0:27) nor
of H0:  3 = 0 ( 2W = 1:30; d f = 1; p = 0:25) were rejected.
A visual indication of model calibration is given in the plot in Figure 10. This plot shows
the observed and predicted probabilities of counts zero through 9. As indicated, the model-implied
and observed probabilities do not track well for counts of zero and 1, but track closely thereafter.
The NBR of S 2E on age i , popDens i , and psiCrmJuvCasn i , shown in the third column of
Table 29 was validated by bootstrapping. Bias-corrected confidence intervals around the estimated
parameters are shown in Table 30. As indicated, only the e ect of age i was likely to replicate in
the population.
The single most important characteristic influencing the rate of positive screens for mari-
juana is age at the time of screening. The expected rate of positive screens for marijuana by age is
plotted in Figure 11. As indicated, the expected rate decreases near-linearly with age. Holding all
115
Table 30
Lower (LCLM) and upper (UCLM) bias-corrected 95% confidence limits, MS2E, n = 194.
LCLM UCLM
age i  :0637  :00997
popDens i  1:39 :465
psiCrmJuvCasn i  :0988 :314
else constant, for those at the lower quartile (0:25Q  23) of age the rate of positives is roughly
0.16. It decreases to 0.08 for those at the median (0:5Q 31) age and to roughly 0.02 for those at
the upper quartile (0:75Q 43). The expected rate flattens thereafter.
MS2F. The criterion this model is the number times the ith probationer tested positive for
opiates outDrgTotPosi6; interest centers on the expected rate of this criterion in the population
b i. Screens for opiates occurred roughly 5 times monthly, on average, throughout the supervision
period, M = 5:05;S D = 7:2. 194=199 = 0:97 of the sample was screened at least once for opiates;
one or more of these was positive for 41=194 = 0:21 of the probationers. The 95%CI around mean
npositives
nscreens is shown in Table 26.
Potential predictors were included in a general model which was then recursively parti-
tioned62 into the binary tree shown in Figure 12a using Poisson-splitting methods (see, Breiman et
al., 1984; Therneau & Atkinson, 1997).63 This classification tree, since it was likely too complex
to validate, was trimmed back to that shown in Figure 12b.64
Predictors having the largest e ect on the rate of positive tests for opiates included (a) hav-
ing ever used opiates psiS ubHeEUi, (b) density of arrests for property crimes within the block-
62Using the R library RPART (Therneau & Atkinson, 1997).
63The Gini rule was used for splitting, prior probabilities were set proportional to observed frequencies, and altered
priors were used for the loss function.
64Cost-complexity pruning was based on the 1 S E rule among cross-validated data.
116
Figure 11. Expected rates of positive screens for marijuana by age at time of screening, MS2E,
n = 194.
group arrPro:di, and (c) the number of prior adult convictions psiCrmAdlCnvi; remaining predic-
tors did not appear in the model.
The resultant model separated sampled probationers into 4 groups. The groups (and pre-
dicted values) included (a) those having never before used opiates (0.14); (b) those having used
opiates at least once and living within a block group with rates of arrests for property crimes of
11.46 or higher (0.88); (c) those having used opiates at least once, living within a block group with
rates of arrests for property crimes lower than 11.46, and having 8 or more prior adult convictions
(2.6); and (d) those having used opiates at least once, living within a block group with rates of
arrests for property crimes lower than 11.46, and having fewer than 8 prior adult convictions (13).
117
psiSubHeEU=No
arrPub.d< 63.74arrPro.d>=11.46
psiCrmAdlCnv>=7.5
arrDrg.d< 13.38
0.093 1.1
0.88
2.6
4.5 18
(a)
psiSubHeEU=No
arrPro.d>=11.46
psiCrmAdlCnv>=7.5
0.14
0.88
2.6 13
(b)
Figure 12. Initial (a) and pruned (b) classification trees predicting outDrgTotPosi6, MS2F, n =
194.
An initial model
bPr (outDrgTotPosi6jxi) = exp(  i) 
outDrgTotPosi6
i
outDrgTotPosi6!
where  i = exp(xi ) and xi1 = psiS ubHeEUi, xi2 = arrPro:di, and xi3 = psiCrmAdlCnvi, was
fitted to the sample data with the pruned-tree predictors in their original form; parameter estimates
are shown in the first column of Table 31.65
After this initial model was fit there was significant evidence that the observations were
overdispersed with respect to the Poisson model, G2 = 249:421; p = 0:000. Thus, MS2F was
respecified to include a multiplicative disturbance term  i to capture unobserved heterogeneity.
65A robust estimator of variance was used in place of the standard estimator to adjust for k = 122 block-group
clusters (Huber, 1967; Rogers, 1993; H. White, 1980). Also, to account for varying times-at-risk, the logged number
of drug screens log (outDrgTotS cri7) was included and constrained to a coe cient of 1.
118
Table 31
Parameter estimates, Poisson and negative binomial regressions of outDrgTotPosi6, MS2F, n =
194.
b=z b=z
psiS ubHeEU 4:013   4:106   
(9:40) (10:22)
arrPro:d  0:034  0:019
( 1:64) ( 1:33)
psiCrmAdlCnv 0:008 0:044
(0:11) (0:79)
  5:871    6:192   
( 16:50) ( 20:49)
Overdispersion  0:780 
(2:47)
Model  2 120:397   174:339   
 p < 0:05;  p < 0:01;   p < 0:001.
The model
bPr (outDrgTotPosi6jxi) = exp( e i)e ioutDrgTotPosi6
outDrgTotPosi6!
where e i = exp(xi +  i) and xi1 = psiS ubHeEUi, xi2 = arrPro:di, and xi3 = psiCrmAdlCnvi,
was then fitted to the sample data. Parameter estimates from this regression are shown in the
second column of Table 31.66
The model had D=350.81, AIC=1.860, and BIC= 644.820.
The Wald test67 of the hypothesis that all coe cients except the intercept were zero H0 :
 1 =  2 =  3 = 0 was rejected,  2W = 174:34; d f = 3; p = 0:00. The test of H0 :  1 = 0 was
66A robust estimator of variance was used in place of the standard estimator to adjust for k = 122 block-group
clusters (Huber, 1967; Rogers, 1993; H. White, 1980). Also, to account for varying times-at-risk, the logged number
of drug screens log (outDrgTotS cri7) was included and constrained to a coe cient of 1.
67The  2
L test is likely invalid given the robust estimator of variance used to adjust for k = 122 block-group clusters.
119
rejected ( 2W = 104:35; d f = 1; p = 0:00) but neither test of H0:  2 = 0 ( 2W = 1:76; d f = 1; p =
0:18) nor of H0:  3 = 0 ( 2W = 0:63; d f = 1; p = 0:43) were rejected.
By comparison, the PR shown in the first column of Table 31 performed at its worse in
predictions of zero counts where it overpredicted. The NBR shown in the second column of
Table 31 performed at its worse in predictions of counts of 1 where it too overpredicted counts.
Overall, the NBR, with mean absolute di erence between predicted and observed valuesj b o b pj=
0:010, outperformed the PR, withj b o b pj= 0:022.
I then turned to assessing scale beginning with the relationship involving psiCrmAdlCnvi.
A plot of the univariable lowess smooth of outDrgTotPosi6 against psiCrmAdlCnvi suggested
the relationship was essentially linear; this predictor was left as is. On the other hand, a simi-
lar plot suggested there was substantial non-linearity in the relationship involving arrPro:di and
outDrgTotPosi6. Some of this stemmed from poor behavior in the upper tail of the distribution.
There were several extreme outliers (arrPro:di  163). Several transformations of arrPro:di
were attempted with and without the outlying cases. None, however, adequately represented the
relationship. Ultimately, arrPro:di was left as is and thus the model was unchanged.
A visual indication of model calibration is given in the plot in Figure 13. This plot shows
the observed and predicted probabilities of counts zero through 9. As indicated, the model-implied
and observed probabilities track closely throughout the distribution, with the poorest fit in counts
of 1.
The NBR of S 2F on psiS ubHeEUi, arrPro:di, and psiCrmAdlCnvi, shown in the second
column of Table 31 was validated by bootstrapping. Bias-corrected confidence intervals around
the estimated parameters are shown in Table 32. As indicated, only the e ect of psiS ubHeEUi is
likely to replicate in the population. Those of arrPro:di and psiCrmAdlCnvi are not.
120
Figure 13. Observed and NBR model-implied counts of outDrgTotPosi6, MS2F, n = 194.
Table 32
Lower (LCLM) and upper (UCLM) bias-corrected 95% confidence limits, MS2F, n = 194.
LCLM UCLM
psiS ubHeEU 3:21 4:99
arrPro:d  :0791 :0418
psiCrmAdlCnv  :0756 :163
The single most important characteristic influencing the rate of positive screens for opiates
is having ever used opiates psiS ubHeEUi. The rate of positive screens among those having never
used opiates is .00179. Having used opiates increases the expected rate of positives by 0.11?a
factor of 60.
MS2G. The criterion for model MS2G is the number times the ith probationer tested posi-
tive for phencyclidine outDrgTotPosi7; interest centers on the expected rate of this criterion in the
population b i. Screens for phencyclidine occurred roughly 5 times monthly, on average, through-
121
out the supervision period, M = 4:59;S D = 7. 190=199 = 0:95 of the sample was screened at least
once for phencyclidine; one or more of these were positive for 33=190 = 0:17 of these probationers.
The 95%CI around mean npositivesnscreens is shown in Table 26.
Potential predictors were included in a general model which was then recursively parti-
tioned68 into the binary tree shown in Figure 14a using Poisson-splitting methods (see, Breiman et
al., 1984; Therneau & Atkinson, 1997).69 This classification tree, since it was likely too complex
to validate, was trimmed back to that shown in Figure 14b.70
Predictors having the largest e ect on the rate of positive phencyclidine screens outDrgTotPosi7
included (a) whether the probationer ever used phencyclidine psiS ubPcEUi, (b) the number of
substances ever used psiS ubEU1i, and (c) the highest grade attempted psiEduGrdAti; remaining
predictors did not appear in the model.
The resultant model separated sampled probationers into 4 groups. The groups (and pre-
dicted values) included probationers (a) never having used phencyclidine (0.26); (b) having previ-
ously used phencyclidine, used more than 2 substances, and whose highest grade attempted was
at least grade 11 (0.64); (c) having previously used phencyclidine, used more than 2 substances,
whose highest grade attempted was 10th or less (4.5) (d) having used phencyclidine and used 2 or
fewer substances (13);
An initial model
bPr (outDrgTotPosi7jxi) = exp(  i) 
outDrgTotPosi7
i
outDrgTotPosi7!
68Using the R library RPART (Therneau & Atkinson, 1997).
69The Gini rule was used for splitting, prior probabilities were set proportional to observed frequencies, and altered
priors were used for the loss function.
70Cost-complexity pruning was based on the 1 S E rule among cross-validated data.
122
psiSubPcEU=No
psiTonEmpb=Probation,Incarceration
psiFamNow=Married-living apart,Divorced-living apart,Unmarried-cohabiting,Unmarried-living apartrnsAge>=31.5
psiSubEU1>=2.5
psiEduGrdAt>=10.5
ImmigrantConcentration>=-0.6567daysSupX< 547
0.0150.35
0.082.6
0.332.1
0.136.9
13
(a)
psiSubPcEU=No
psiSubEU1>=2.5
psiEduGrdAt>=10.5
0.26
0.64 4.5
13
(b)
Figure 14. Initial (a) and pruned (b) classification trees predicting outDrgTotPosi7, MS2G, n =
190.
where  i = exp(xi ) and xi1 = psiS ubPcEUi, xi2 = psiS ubEU1i, and xi3 = psiEduGrdAti, was
fitted to the sample data with the pruned-tree predictors in their original form; parameter estimates
are shown in the first column of Table 33.71
After this initial model was fit there was significant evidence that the observations were
overdispersed with respect to the Poisson model, G2 = 612:486; p = 0:000. Thus, MS2G was
respecified to include a multiplicative disturbance term  i to capture unobserved heterogeneity.
The model
bPr (outDrgTotPosi7jxi) = exp( e i)e ioutDrgTotPosi7
outDrgTotPosi7!
71A robust estimator of variance was used in place of the standard estimator to adjust for k = 122 block-group
clusters (Huber, 1967; Rogers, 1993; H. White, 1980). Also, to account for varying times-at-risk, the logged number
of drug screens log (outDrgTotS cri7) was included and constrained to a coe cient of 1.
123
Table 33
Parameter estimates, Poisson and negative binomial regressions of outDrgTotPosi7, MS2G, n =
190.
b=z b=z b=z
psiS ubPCEU 3:596   2:731   2:721   
(4:42) (4:33) (4:03)
psiS ubEU1  0:615  0:297
( 2:15) ( 1:62)
psiEduGrdAt  0:071  0:249
( 0:82) ( 1:69)
psiS ubEU1  0:422
( 1:61)
psiEduGrdAt  0:428 
( 1:98)
  3:590    1:979 0:298
( 3:98) ( 1:10) (0:11)
Overdispersion  2:183   2:149   
(8:63) (7:86)
Model  2 47:091   24:899   23:693   
 p < 0:05;  p < 0:01;   p < 0:001.
where e i = exp(xi +  i) and xi1 = psiS ubPcEUi, xi2 = psiS ubEU1i, and xi3 = psiEduGrdAti,
was then fitted to the sample data. Parameter estimates from this regression are shown in the
second column of Table 33.72
The model had D=356.69, AIC=1.930, and BIC= 614.014.
By comparison, the PR shown in the first column of Table 33 performed at its worse in
predictions of zero where it overpredicted counts. The NBR shown in the second column of
Table 33 performed at its worse in predictions of counts of 8 where it underpredicted counts.
72A robust estimator of variance was used in place of the standard estimator to adjust for k = 122 block-group
clusters (Huber, 1967; Rogers, 1993; H. White, 1980). Also, to account for varying times-at-risk, the logged number
of drug screens log (outDrgTotS cri7) was included and constrained to a coe cient of 1.
124
Overall, the NBR, with mean absolute di erence between predicted and observed valuesj b o b pj=
0:007, outperformed the PR, withj b o b pj= 0:061.
I then turned to assessing scale. To account for sparse representation in its upper levels,
the last three levels of psiS ubEU1i were collapsed as psiS ubEU1 i . As for psiEduGrdAti, only
13=199 = 0:065 of the probationers had values of psiEduGrdAti less than 9; the same proportion
of probationers had values above 13. To accommodate this, grades 0?9 and grades 13?16 were
collapsed to represent highest grade attempted as 9th or below and 13th and higher, respectfully,
as psiEduGrdAt i .
The model
bPr (outDrgTotPosi7jxi) = exp( e i)e ioutDrgTotPosi7
outDrgTotPosi7!
where e i = exp(xi +  i) and xi1 = psiS ubPcEUi, xi2 = psiS ubEU1 i , and xi3 = psiEduGrdAt i ,
was then refitted to the sample data. Parameter estimates from this regression are shown in the
third column of Table 33.73
The model had D=355.44, AIC=1.923, and BIC= 615.255.
The Wald test74 of the hypothesis that all coe cients except the intercept were zero H0 :
 1 =  2 =  3 = 0 was rejected,  2W = 23:69; d f = 3; p = 0:00. The test of H0:  1 = 0 was rejected
( 2W = 16:23; d f = 1; p = 0:00), the test of H0:  2 = 0 was not rejected ( 2W = 2:61; d f = 1; p =
0:11), and the test of H0:  3 = 0 was rejected,  2W = 3:91; d f = 1; p = 0:05.
A visual indication of model calibration is given in the plot in Figure 15. This plot shows
the observed and predicted probabilities of counts zero through 9. As indicated, the model-implied
73A robust estimator of variance was used in place of the standard estimator to adjust for k = 122 block-group
clusters (Huber, 1967; Rogers, 1993; H. White, 1980). Also, to account for varying times-at-risk, the logged number
of drug screens log (outDrgTotS cri7) was included and constrained to a coe cient of 1.
74The  2
L test is likely invalid given the robust estimator of variance used to adjust for k = 122 block-group clusters.
125
Figure 15. Observed and NBR model-implied counts of outDrgTotPosi7, MS2G, n = 190.
Table 34
Lower (LCLM) and upper (UCLM) bias-corrected 95% confidence limits, MS2G, n = 190.
LCLM UCLM
psiS ubPCEU :366 5:08
psiS ubEU1  1:02 :181
psiEduGrdAt  :96 :103
and observed probabilities track closely throughout the distribution, with the poorest fit in counts
of 1.
The NBR of S 2G on xi1 = psiS ubPcEUi, xi2 = psiS ubEU1 i , and xi3 = psiEduGrdAt i ,
shown in the third column of Table 33 was validated by bootstrapping. Bias-corrected confidence
intervals around the estimated parameters are shown in Table 34. As indicated, the e ects of
psiS ubEU1 and psiEduGrdAt are not likely to replicate in the population.
126
The most important characteristic influencing rates of testing positive for phencyclidine is
having ever used phencyclidine. Holding psiS ubEU1 and psiEduGrdAt at their modal values,
having ever used phencyclidine increases the expected rate by a factor of 15, from .0022 to .034.
Arrest-convictions
Models MC1 and MC2 estimate the probabilities that probationers in the population will
be convicted for new crimes given an arrest. Model MC1 focuses on convictions during the
supervision period; Model MC2, the post-supervision period.
MC1. The criterion for MC1 is whether the ith probationer is arrested and subsequently
convicted on new charges during the supervision period outC1i. If so, outC1i = 1. Roughly one-
fourth (51=199 = 0:26) of the sampled probationers were indeed arrested and subsequently convicted
of new crimes during the supervision period. Interest centers on the predicted probability of the
criterion, bPr (outC1i = 1).
Potential predictors of this process were included in a general model which was then re-
cursively partitioned75 into the binary tree shown in Figure 16a.76 This classification tree, since it
was likely too complex to validate, was trimmed back to that shown in Figure 16b.77
Predictors having the largest influence on whether probationers are arrested and subse-
quently convicted on new charges during the supervision period included (a) the Salient Factor
Score (SFS)-98 psiCrmAdlS FS 1i, (b) the recommended sentence psiS Ri, and (c) the number of
adult convictions involving weapons psiCrmAdlCnvWeai; remaining predictors did not appear in
the model.
75Using the R library RPART (Therneau & Atkinson, 1997).
76The Gini rule was used for splitting, prior probabilities were set proportional to observed frequencies, and altered
priors were used for the loss function.
77Cost-complexity pruning was based on the 1 S E rule among cross-validated data.
127
psiCrmAdlSFS1< 3.5
psiTonEmpb=Split sentence,Incarceration
psiCrmAdlCnvWea< 0.5
psiSupRev>=0.5
psiFamSibBN< 1.5
Yes No 
No 
Yes No 
No 
(a)
psiCrmAdlSFS1< 3.5
psiTonEmpb=Split sentence,Incarceration
psiCrmAdlCnvWea< 0.5
Yes No 
No 
No 
(b)
Figure 16. Initial (a) and pruned (b) classification trees predicting outC1i, MC1, n = 199.
The resultant model separated sampled probationers into 4 groups. These groups (and pre-
dicted values) included probationers having (a) an SFS-98 score lower than 4, a recommended
sentence of either split sentence or incarceration, and fewer than 1 adult convictions involving
weapons (Yes); (b) an SFS-98 score lower than 4, a recommended sentence of either split sen-
tence or incarceration, and 1 or more adult convictions involving weapons (No); (c) an SFS-98
score lower than 4 and a recommended sentence of probation, (No); and (d) an SFS-98 score
equal to or greater than 4, (No).
Indicators recS Inc and recS S pl representing, respectfully, recommended sentences of in-
carceration and split sentence were created and an initial model
bPr (outC1i = 1jxi) = 1
1 + exp   (xi ) 
= 11 + exp   ( 
0 + 1xi1 + 2xi2 + 3xi3 + 4xi4)
 ;
128
Table 35
Parameter estimates, MC1, logistic regression of outC1i, n = 199.
b=z b=z b=z b=z
psiCrmAdlS FS 1  0:201   0:190   0:215   0:468   
( 2:82) ( 2:66) ( 3:12) ( 3:43)
recS Inc 1:615   1:555   
(3:67) (3:57)
recS S pl 0:460 0:483
(0:96) (1:01)
psiCrmAdlCnvWea  0:617 
( 2:20)
CnvWeaHi  0:673  0:606  0:754
( 1:36) ( 1:29) ( 1:37)
probation  1:145   3:354   
( 3:03) ( 3:63)
psiCrmAdlS FS 1 probation 0:431  
(2:65)
  0:487  0:555 0:727 1:852  
( 1:04) ( 1:16) (1:79) (2:88)
Model  2 31:237   28:496   20:221   32:811   
 p < 0:05;  p < 0:01;   p < 0:001.
where xi1 = psiCrmAdlS FS 1i, xi2 = recS Inci, xi3 = recS S pli, and xi4 = psiCrmAdlCnvWeai,
was fitted to the sample data; parameter estimates are shown in the first column of Table 35.78
The Wald test79 of H0:  1 =  2 =  3 =  4 = 0 was rejected,  2W = 31:24; d f = 4; p = 0:00.
The test of H0 :  1 = 0 was rejected ( 2W = 7:97; d f = 1; p = 0:00); the test of H0 :  2 =  3 = 0
was rejected ( 2W = 14:01; d f = 2; p = 0:00); and the test of H0 :  4 = 0 was rejected ( 2W =
4:83; d f = 1; p = 0:03).
78A robust estimator of variance was used in place of the standard estimator to adjust for k = 122 block-group
clusters (Huber, 1967; Rogers, 1993; H. White, 1980).
79The  2
L test is likely invalid given the robust estimator of variance used to adjust for k = 122 block-group clusters.
129
The model correctly classified 0:79 of the sample; the AUC suggested it demonstrated ac-
ceptable discrimination, AUC = 0:76. The model had D=189.45, AIC=1.002, and BIC= 837.448;
thebc suggested it was empirically consistent,  2HL = 8:34; d f = 8; p = 0:40.
Focus turned to refining this preliminary model in terms of parametric relationships and
scale beginning with the relationship between outC1i and psiCrmAdlS FS 1i. A plot of the lowess
smoothed logit against linear psiCrmAdlS FS 1i suggested a linear relationship and there was no
evidence that an interval representation introduced loss of information. The predictor psiCrmAdlS FS 1i
was thus left as linear and continuous in the regression of outC1i.
I next turned to psiCrmAdlCnvWeai. The tree in Figure 16b unexpectedly indicates having
fewer than 1 weapons-related convictions is associated with increased probabilities of being ar-
rested and subsequently convicted on new charges during the supervision period, but that having
more than 1 weapons-related convictions is associated with decreased probabilities.
Some of this unanticipated relationship reflected severe non-normality in the predictor. A
plot of the lowess smoothed logit against linear psiCrmAdlCnvWea indicated there was a sharp
spike in the logit between 0 and 1 previous weapons convictions, a flat relationship from 1 to 3,
then a slow, negative e ect thereafter.
While originally scaled as continuous, ranging within [0;+1), the distribution of psiCrmAdlCnvWeai
in the sample showed strong positive skew. The majority (162=199 = 0:81) of probationers did not
have any previous weapons-related convictions, 30=199 = 0:15 had 1, 4=199 = 0:02 had 2, and
3=199 = 0:015 had 3 or more. The predictor was recoded as a binary indicator CnvWeaHii =
psiCrmAdlCnvWeai  1.
130
A refined model was fitted to the sample data replacing psiCrmAdlCnvWeai with its binary
representation; parameter estimates from the model
bPr (outC1i = 1jxi) = 1
1 + exp   (xi ) 
= 11 + exp   ( 
0 + 1xi1 + 2xi2 + 3xi3 + 4xi4)
 ;
where xi1 = psiCrmAdlS FS 1i, xi2 = recS Inci, xi3 = recS S pli, and xi4 = CnvWeaHii, are shown
in the second column of Table 35.80
The Wald test81 of H0:  1 =  2 =  3 =  4 = 0 was rejected,  2W = 28:50; d f = 4; p = 0:00.
The test of H0 :  1 = 0 was rejected ( 2W = 7:09; d f = 1; p = 0:01); the test of H0 :  2 =  3 = 0
was rejected ( 2W = 13:08; d f = 2; p = 0:00); the test of H0 :  4 = 0 was not rejected ( 2W =
1:85; d f = 1; p = 0:17).
The model correctly classified 0:79 of the sample and the AUC suggested it demonstrated
acceptable discrimination, AUC = 0:75. The model had D=192.49, AIC=1.018, and BIC= 834.411;
thebc suggested it was empirically consistent,  2HL = 8:80; d f = 8; p = 0:36.
The tree in Figure 16b suggests the e ect of recS enti is likely isolated to the comparison of
probation versus both incarceration and split sentence. To capture this, an indicator of probation
recommendations versus the remaining categories probationi = (recS enti == ?Probation?) was
created.
Using it in place of recS enti, the model
bPr (outC1i = 1jxi) = 1
1 + exp   (xi ) 
= 11 + exp   ( 
0 + 1xi1 + 2xi2 + 3xi3)
 ;
80A robust estimator of variance was used in place of the standard estimator to adjust for k = 122 block-group
clusters (Huber, 1967; Rogers, 1993; H. White, 1980).
81The  2
L test is likely invalid given the robust estimator of variance used to adjust for k = 122 block-group clusters.
131
where xi1 = psiCrmAdlS FS 1i, xi2 = probationi, and xi3 = CnvWeaHii, was fitted; parameter
estimates are shown in the third column of Table 35.82
The Wald test83 of H0:  1 =  2 =  3 = 0 was rejected,  2W = 20:22; d f = 3; p = 0:00. The
test of H0 :  1 = 0 was rejected ( 2W = 9:18; d f = 1; p = 0:00); the test of H0 :  3 = 0 was not
rejected ( 2W = 1:67; d f = 1; p = 0:20).
The model correctly classified 0:82 of the sample and the the AUC indicated acceptable
discrimination, AUC = 0:74. The model had D=197.33, AIC=1.032, and BIC= 834.870; the bc
suggested the model was empirically consistent,  2HL = 15:97; d f = 8; p = 0:04.
There is also an indication that the e ect of psiCrmAdlS FS 1 depends on the recommended
sentence. These two predictors overlap slightly: the SFS-98 captures elements of criminal history
(see, Appendix B for details) and, quite rightly, PSI authors base much of their sentence recom-
mendation on these same criteria.
I created the product term psiCrmAdlS FS 1i probationi to capture this potential interac-
tion. The model was refitted as
bPr (outC1i = 1jxi) = 1
1 + exp   (xi ) 
= 11 + exp   ( 
0 + 1xi1 + 2xi2 + 3xi3 + 4xi1xi2)
 ;
where xi1 = psiCrmAdlS FS 1i, xi2 = probationi, and xi3 = CnvWeaHii; parameter estimates are
shown in the fourth column of Table 35.84
The model had D=188.51, AIC=0.998, and BIC= 838.391; thebc suggested the model was
empirically consistent,  2HL = 12:48; d f = 8; p = 0:13.
82A robust estimator of variance was used in place of the standard estimator to adjust for k = 122 block-group
clusters (Huber, 1967; Rogers, 1993; H. White, 1980).
83The  2
L test is likely invalid given the robust estimator of variance used to adjust for k = 122 block-group clusters.
84A robust estimator of variance was used in place of the standard estimator to adjust for k = 122 block-group
clusters (Huber, 1967; Rogers, 1993; H. White, 1980).
132
0
.2
.4
.6
.8
1
Observed
0 .2 .4 .6 .8 1
Model implied
Figure 17. Comparison between model-implied probabilities of experiencing outC1i and a mov-
ing average of the proportion of probationers arrested and convicted on new charges during the
supervision period, n = 199.
The Wald test85 of the interaction H0 :  4 = 0 was rejected ( 2W = 7:03; d f = 1; p =
0:01). Aside from the significant interaction term, there was evidence that this change led to an
improvement in model calibration.
The model correctly classified 0:83 of the sample and the the AUC indicated acceptable
discrimination, AUC = 0:73.
A visual indication of model calibration is given in the plot in Figure 17. The thick line
represents the fraction of probationers arrested and subsequently convicted of a new crime within
the supervision period at each level of predicted probabilities, and, here, suggests the model may
lack calibration throughout this range. However, compared to calibration curves86 for the models
shown in Columns 1?3 of Table 35, the model in Column 4 demonstrates visible improvement.
85The  2
L test is likely invalid given the robust estimator of variance used to adjust for k = 122 block-group clusters.
86Not presented.
133
Table 36
Lower (LCLM) and upper (UCLM) bias-corrected 95% confidence limits, MC1, n = 199.
LCLM UCLM
psiCrmAdlS FS 1  :765  :17
probation  5:31  1:39
CnvWeaHi  1:96 :45
psiCrmAdlS FS 1 probation :0808 :781
The regression of outC1i on psiCrmAdlS FS 1i probationi, CnvWeaHii, and psiCrmAdlS FS 1i 
probationi shown in the fourth column of Table 35 was validated by bootstrapping. Bias-corrected
confidence intervals around the estimated parameters are shown in Table 36. As indicated in the
table, the e ects of psiCrmAdlS FS 1i, probation, and their interaction were likely to validate in
the population. On the other hand, the e ect of CnvWeaHi was not. Interpretations of the model
are based on the e ects of psiCrmAdlS FS 1i, probation, and psiCrmAdlS FS 1 probation while
holding CnvWeaHi at its modal value of no previous weapons-related convictions.
Predicted values b ranged within the interval [0:036;0:864], with mean b = 0:256 (S D =
0:20). The two most important characteristics influencing bPr (outC1i = 1) are psiCrmAdlS FS 1i
and probation.
The plot in Figure 18 shows predicted probabilities an o ender will be arrested and sub-
sequently convicted on new charges during the supervision period by psiCrmAdlS FS 1i and rec-
ommended sentence. Lower SFS-98 scores are associated with increased chances an o ender will
be arrested and subsequently convicted on new charges during the supervision period. As indi-
cated, this is especially true among those whose recommended sentence was either incarceration
or split-sentence. For these o enders, the predicted probability of being arrested and subsequently
convicted increases exponentially as SFS-98 decreases. For those whose recommended sentence
134
Figure 18. Predicted probabilities an o ender in the population will be convicted of a new crime
during the supervision period, by SFS-98 score and recommended sentence, MC1, n = 199.
was probation, however, changes in the SFS-98 scores have little impact: predicted probabilities
are essentially flat. This is expected given the potential overlap between the two measures.
MC2. The criterion for MC2 is whether the ith probationer is arrested and subsequently
convicted on new charges during the post-supervision period. If so, outC2i = 1. Indeed, roughly
one-fourth (49=199 = 0:25) of the sampled probationers were arrested and subsequently convicted
of new crimes during this period. Interest centers on the predicted probability of the criterion,
bPr (outC2i = 1).
Potential predictors of outC2i were included in a general model which was then recursively
partitioned87 into the binary tree shown in Figure 19a.88 This classification tree, since it was likely
too complex to validate, was trimmed back to that shown in Figure 19b.89
87Using the R library RPART (Therneau & Atkinson, 1997).
88The Gini rule was used for splitting, prior probabilities were set proportional to observed frequencies, and altered
priors were used for the loss function.
89Cost-complexity pruning was based on the 1 S E rule among crossvalidated data.
135
psiCrmAdlSFS1< 3.5
arrPub.d>=50.19
psiFamSSN=Yes
psiCrmAdlIncn2>=0.5
rnsNPA=2 or less,5
psiSubAlEU=No
Yes
Yes No 
Yes No 
No 
No 
(a)
psiCrmAdlSFS1< 3.5
arrPub.d>=50.19
psiFamSSN=Yes
Yes
Yes No 
No 
(b)
Figure 19. Initial (a) and pruned (b) classification trees predicting outC2i, MC2, n = 199.
Predictors having the largest influence on whether probationers are arrested and subse-
quently convicted on new charges during the post-supervision period included (a) the SFS-98
psiCrmAdlS FS 1, (b) the rate of public order related arrests within the BG arrPub:d, and (c) whether
the PSI author found a supportive social network psiFamS S N; remaining predictors did not ap-
pear in the model.
The resultant model separated sampled probationers into 4 groups. These groups (and as-
sociated predicted values) were (a) an SFS-98 score lower than 4 and a rate of public order related
arrests greater than or equal to 50.19 (Yes); (b) an SFS-98 score lower than 4, a rate of public
order related arrests less than 50.19, and supportive social network (Yes); (c) an SFS-98 score
lower than 4, a rate of public order related arrests less than 50.19, and an absence of a supportive
social network (No); and (d) an SFS-98 score equal to or greater than than 4, (No).
136
Table 37
Parameter estimates, MC2, logistic regression of outC2i, n = 199.
b=z b=z b=z
psiCrmAdlS FS 1  0:268    0:267    0:268   
( 4:32) ( 4:33) ( 4:28)
arrPub:d 0:005
(1:41)
psiFamS S N  0:021  0:043 0:030
( 0:05) ( 0:11) (0:08)
arrPub:d 0:093
(0:43)
arrPub:dHi 0:916  
(3:22)
 0:195 0:158 0:109
(0:31) (0:19) (0:17)
Model  2 19:294   18:822   24:515   
 p < 0:05;  p < 0:01;   p < 0:001.
An initial model
bPr (outC2i = 1jxi) = 1
1 + exp   (xi ) 
= 11 + exp   ( 
0 + 1xi1 + 2xi2 + 3xi3)
 ;
where xi1 = psiCrmAdlS FS 1i, xi2 = arrPub:di, and xi3 = psiFamS S Ni, was fitted to the sample
data; parameter estimates are shown in the first column of Table 37.90
The Wald test91 of H0:  1 =  2 =  3 = 0 was rejected,  2W = 20:13; d f = 3; p = 0:00. The
test of H0 :  1 = 0 was rejected ( 2W = 18:46; d f = 1; p = 0:00), but neither test of H0 :  2 = 0
or H0 :  3 = 0 were rejected,  2W = 3:66; d f = 1; p = 0:06 and  2W = 0:00; d f = 1; p = 0:97,
respectfully.
90A robust estimator of variance was used in place of the standard estimator to adjust for k = 122 block-group
clusters (Huber, 1967; Rogers, 1993; H. White, 1980).
91The  2
L test is likely invalid given the robust estimator of variance used to adjust for k = 122 block-group clusters.
137
The model correctly classified 0:76 of the sample. and the AUC suggested it demon-
strated acceptable discrimination, AUC = 0:71. The model had D=200.37, AIC=1.047, and
BIC= 831.823; the bc suggested the model was empirically consistent,  2HL = 9:70; d f = 8; p =
0:29.
Focus then turned to refining this preliminary model in terms of parametric relationships
and scale beginning with the relationship between outC2i and psiCrmAdlS FS 1i. A plot of the
lowess smoothed logit against linear psiCrmAdlS FS 1 indicated the predictor was well-modeled
as linear and continuous. I next turned to arrPub:di. A plot of the lowess smoothed logit against
linear arrPub:d suggested a fairly complex non-linearity: there was an initial spike in the logit
increasing near linearly from roughly arrPub:d = 20 peaking at roughly arrPub:d = 100. It
cupped at roughly arrPub:d = 150 then flattened thereafter. There was likely very little to make of
the apparent nonlinearities in the right half of the distribution. The predictor was poorly behaving
in the upper tail. In fact, mean arrPub:d was roughly equal to the 0:68Q, M = 25:40, S D = 30:90.
A normal quantile-comparison plot of arrPub:di indicated non-normality, especially in the upper
region, which was corrected using the unconditional Box-Cox method. The ML normalizing
transformation parameter  in x was estimated as b = 0:278; the predictor was normalized by
applying the transformation arrPub:d i = (arrPub:di + 0:5)0:278.
The model
bPr (outC2i = 1jxi) = 1
1 + exp   (xi ) 
= 11 + exp   ( 
0 + 1xi1 + 2xi2 + 3xi3)
 ;
where xi1 = psiCrmAdlS FS 1i, xi2 = arrPub:d i , and xi3 = psiFamS S Ni, was refitted to the
sample data; parameter estimates are shown in the second column of Table 37.92
92A robust estimator of variance was used in place of the standard estimator to adjust for k = 122 block-group
clusters (Huber, 1967; Rogers, 1993; H. White, 1980).
138
The predictor arrPub:d i was split at 50.19 and the model
bPr (outC2i = 1jxi) = 1
1 + exp   (xi ) 
= 11 + exp   ( 
0 + 1xi1 + 2xi2 + 3xi3)
 ;
where xi1 = psiCrmAdlS FS 1i, xi2 = arrPub:d:hii, and xi3 = psiFamS S Ni, was refitted to the
sample data; parameter estimates are shown in the third column of Table 37.93
The Wald test94 of H0 :  1 =  2 =  3 = 0 was rejected,  2W = 24:52; d f = 3; p = 0:00.
The test of H0 :  1 = 0 was rejected ( 2W = 18:33; d f = 1; p = 0:00) and the test of H0 :  2 = 0
was rejected ( 2W = 10:36; d f = 1; p = 0:00); the test of H0 :  3 = 0 was not rejected ( 2W =
0:01; d f = 1; p = 0:94).
The model correctly classified 0:78 of the sample and the AUC suggested it demonstrated
acceptable discrimination, AUC = 0:71. The model had D=198.34, AIC=1.037, and BIC= 833.851;
thebc suggested it was empirically consistent,  2HL = 10:08; d f = 8; p = 0:26.
A visual indication of model calibration is given in the plot in Figure 20 which compares
predicted probabilities from the regression of outC2 with a moving average of the proportion of
probationers arrested and convicted on new charges during the supervision period. The thick line
represents the fraction of probationers arrested and subsequently convicted of a new crime during
the post-supervision period at each level of predicted probabilities. Here, that the thick line tracks
closely with the diagonal indicates the model is well-calibrated.
The regression of outC2i on psiCrmAdlS FS 1i, arrPub:d i , and psiFamS S Ni, shown in
the third column of Table 37 was validated by bootstrapping. Bias-corrected confidence inter-
vals around the estimated parameters are shown in Table 38. As indicated in the table, the
93A robust estimator of variance was used in place of the standard estimator to adjust for k = 122 block-group
clusters (Huber, 1967; Rogers, 1993; H. White, 1980).
94The  2
L test is likely invalid given the robust estimator of variance used to adjust for k = 122 block-group clusters.
139
0
.2
.4
.6
.8
1
Observed
0 .2 .4 .6 .8 1
Model implied
Figure 20. Comparison between model-implied probabilities of experiencing outC2i and a mov-
ing average of the proportion of probationers arrested and convicted on new charges after the
supervision period, n = 199.
Table 38
Lower (LCLM) and upper (UCLM) bias-corrected 95% confidence limits, MC2, n = 199.
LCLM UCLM
psiCrmAdlS FS 1i  :4  :136
arrPub:d i :294 1:54
psiFamS S Ni  :771 :832
e ects of psiCrmAdlS FS 1i and arrPub:d i were likely to validate in the population. The ef-
fect of psiFamS S N, however, was not. Interpretations of the model are based on the e ects of
psiCrmAdlS FS 1i and arrPub:d i while holding psiFamS S N at its modal value of ?No?.
Predicted values b ranged within the interval [:0568;:687], with mean b = 0:246 (S D =
0:15). The two most important characteristics influencing bPr (outC2i = 1) are psiCrmAdlS FS 1i
and arrPub:d i .
140
Figure 21. Predicted probabilities an o ender in the population will be convicted of a new crime
during the post-supervision period, by SFS-98 score and densities of public order arrests within
the block group of residence, MC2, n = 199.
Holding psiCrmAdlS FS 1i at its mean and psiFamS S Ni at its modal value, varying arrPub:d i
from its minimum of zero to its maximum of 1 increases the predicted probability of outC2 by
0.1784, from 0.1876 to 0.3659.
The plot in Figure 21 shows changes in the predicted probability an o ender will be arrested
and subsequently convicted on new charges during the post-supervision period by psiCrmAdlS FS 1i
and arrPub:d i . Predicted probabilities increase linearly as SFS-98 decreases. The rate of change
is nearly identical between o enders living in BGs with high and low rates of public order related
arrests. However, predicted probabilities among those living in areas with higher rates of public
order related arrests are consistently higher than those among their counterparts living in areas
with lower rates of public order related arrests.
141
Table 39
Frequencies of violations, by violation type, n = 199.
Variable Description 0 1 2 [3;5) [5;7) [7;23]
outV1 Supervision-specific 79 49 35 22 8 6
outV2 Drug-related 87 28 25 25 13 21
Violations
Model MV1 estimates factors associated with population rates of supervision-related con-
dition violations. Model MV2 similarly estimates factors associated with rates of drug-related
condition violations. I describe these models next.
The criteria outV1 and outV2 are described in Table 39. 79=199 = 0:40 o enders terminated
sentences without violating supervision-related conditions. 185=199 = 0:93 had fewer than 5. One
o ender had 23. Only 2 others had more than 10. Half of the sample 87=199 = 0:44 terminated
without violating any drug-related conditions. Its maximum reached 21, but, here again, few
o enders (10=199 = 0:05) accumulated more than 10 drug-related violations.
MV1. The criterion for MV1 is the number of supervision-related violations, outV1i.
120=199 = 0:6 of the sample had at least one supervision-related violation and o enders accu-
mulated, on average, roughly 2 throughout the supervision period, M = 1:53, S D = 2:47. The
highest number of supervision-related violations was 23. Interest centers on the predicted rate of
the criterion, b i.
Potential predictors were included in a general model which was then recursively parti-
tioned95 into the binary tree shown in Figure 22a using Poisson-splitting methods (see, Breiman et
95Using the R library RPART (Therneau & Atkinson, 1997).
142
al., 1984; Therneau & Atkinson, 1997).96 This classification tree, since it was likely too complex
to validate, was trimmed back to that shown in Figure 22b.97
Predictors having the largest influence on the rate of supervision-related violations included
(a) the impression of risk on the CSO rnsImpi, (b) the SFS-98 score psiCrmAdlS FS 1i, and (c) the
supervision level supLvli; remaining predictors did not appear in the model.
The resultant model separated sampled probationers into 4 groups. The groups (and pre-
dicted values) included those for whom (a) the impression on the interviewing o cer was either
low or medium and SFS-98 score greater than 8 (0.48); (b) the impression on the interviewing o -
cer was either low or medium and SFS-98 score less than or equal to 8 (1.3); (c) the impression on
the interviewing o cer was high and supervised at either minimum, medium, or maximum levels
(1.8); and (d) the impression on the interviewing o cer was high and supervised at the intensive
level (4.6).
The indicators imprHigi, imprMedi, imprLowi were created to represent rnsImpi and the
indicators mini, medi, maxi, and inti were created to represent supLvli.98 An initial model
Pr (outV1ijxi) = exp(  i) 
outV1i
i
outV1i!
where  i = exp(xi ) and xi1 = psiCrmAdlS FS 1i, xi2 = imprHigi, xi3 = imprLowi, xi4 = mini,
xi5 = medi, and xi6 = inti, was then fitted to the sample data; parameter estimates are shown in the
first column of Table 40.99
After this initial model was fit there was significant evidence that the observations were
overdispersed with respect to the Poisson model, G2 = 186:817; p = 0:000. Thus, MV1 was
96The Gini rule was used for splitting, prior probabilities were set proportional to observed frequencies, and altered
priors were used for the loss function.
97Cost-complexity pruning was based on the 1 S E rule among cross-validated data.
98Indicators representing modal categories were omitted as references.
99A robust estimator of variance was used in place of the standard estimator to adjust for k = 122 block-group
clusters (Huber, 1967; Rogers, 1993; H. White, 1980). Also, to account for varying times-at-risk, the logged number of
days supervised log (daysS upActi) was included and constrained to a coe cient of 1 (not reported in estimates tables).
143
rnsImp=Low,Medium
psiCrmAdlSFS1>=8.5
supLvl=Minimum,MediumarrOth.d>=50.98
ImmigrantConcentration< -0.3148
rtAK>=0.351
psiCrmAdlCnvPro>=0.5
rnsFA=None
psiCrmAdlCnvWea>=0.5
psiCrmJuvCasn< 0.5
psiSubSm=Yes
psiFamAbuNG=No
supLvl=Minimum,Medium,Maximum
ImmigrantConcentration>=-0.5353
rnsAge>=41.5
0.140.72
0.26
0.191
1.7
0.23
0.932.2
1.3
1.83.9
0.661.9
3.3
4.6
(a)
rnsImp=Low,Medium
psiCrmAdlSFS1>=8.5supLvl=Minimum,Medium,Maximum
0.48 1.3
1.8 4.6
(b)
Figure 22. Initial (a) and pruned (b) classification trees predicting outV1i, MV1, n = 199.
respecified to include a multiplicative disturbance term  i to capture unobserved heterogeneity.
The model
Pr (outV1ijxi) = exp( e i)e i
outV1i
outV1i!
where e i = exp(xi + i) and xi1 = psiCrmAdlS FS 1i, xi2 = imprHigi, xi3 = imprLowi, xi4 = mini,
xi5 = medi, and xi6 = inti, was then fitted to the sample data. Parameter estimates from this
regression are shown in the second column of Table 40.100
By comparison, the PR shown in the first column of Table 40 performed at its worse in
predictions of 0 where it greatly underpredicted counts. The NBR shown in the second column of
Table 40 performed at its worse in predictions of counts of 2 where it, too, underpredicted. Pre-
dictions from both models converge near counts of 4; both appear equally capable of predictions
100A robust estimator of variance was used in place of the standard estimator to adjust for k = 122 block-group
clusters (Huber, 1967; Rogers, 1993; H. White, 1980). Also, to account for varying times-at-risk, the logged number of
days supervised log (daysS upActi) was included and constrained to a coe cient of 1 (not reported in estimates tables).
144
Table 40
Parameter estimates, Poisson and negative binomial regressions of outV1i, MV1, n = 199.
b=z b=z
psiCrmAdlS FS 1  0:065  0:108  
( 1:92) ( 2:90)
imprHig 0:813  0:721 
(3:00) (2:37)
imprLow 0:153  0:089
(0:67) ( 0:32)
min  0:581  0:830 
( 1:27) ( 2:23)
med  0:337  0:424
( 1:16) ( 1:13)
int 0:481 0:393
(1:83) (1:35)
  5:917    5:239   
( 23:32) ( 21:16)
Overdispersion  0:235
(1:31)
Model  2 20:904  45:384   
 p < 0:05;  p < 0:01;   p < 0:001.
of counts 5. Overall, the NBR, with mean absolute di erence between predicted and observed
valuesj b o b pj= 0:009, outperformed the PR, withj b o b pj= 0:022.
The Wald test101 of the hypothesis that all coe cients in the regression shown in the second
column of Table 40 except the intercept were zero H0 :  1 =  2 =  3 =  4 =  5 =  6 = 0 was
rejected,  2W = 45:38; d f = 6; p = 0:02. The test of H0 :  1 = 0 was rejected,  2W = 8:40; d f =
1; p = 0:00. The test of H0 :  2 =  3 = 0 was rejected,  2W = 7:39; d f = 2; p = 0:02. The test of
H0:  4 =  5 =  6 = 0 was not rejected,  2W = 7:73; d f = 3; p = 0:05.
The model had D=684.95, AIC=3.522, and BIC= 326.071.
101The  2
L test is likely invalid given the robust estimator of variance used to adjust for k = 122 block-group clusters.
145
Figure 23. Observed and NBR model-implied counts of outV1i, MV1, n = 199.
A visual indication of model calibration is given in the plot in Figure 23. This plot shows
the observed and predicted probabilities of counts zero through 9. As indicated, the model-implied
and observed probabilities track closely throughout the distribution, with the poorest fit in counts
of 2.
Focus then turned to the parametric relationship between outV1i and psiCrmAdlS FS 1i.
Checking the scale of the predictor revealed no substantial non-linearity and thus psiCrmAdlS FS 1i
was parameterized as linear and continuous.
The NBR of outV1 on psiCrmAdlS FS 1i, imprHigi, imprLowi, mini, medi, and inti, shown
in the second column of Table 40 was validated by bootstrapping. Bias-corrected confidence
intervals around the estimated parameters are shown in Table 41. As indicated, the e ects of
psiCrmAdlS FS 1i and imprHig were likely to replicate in the population; interpretations are based
on these e ects.
146
Table 41
Lower (LCLM) and upper (UCLM) bias-corrected 95% confidence limits, MV1, n = 198.
LCLM UCLM
psiCrmAdlS FS 1  :19  :0264
imprHig :107 1:34
imprLow  :668 :49
min  4:37 1:93
med  1:71 :0741
int  :972 :187
The most important characteristics influencing the rate of supervision-related violations in-
cluded the SFS-98 score psiCrmAdlS FS 1i and whether interviewing o cers believed the o ender
represented a high supervision risk imprHig.
Each additional point on the SFS-98 decreases the expected number of violations by a
factor of 0.8974 (10.3%) holding all other predictors constant. Having a high impression versus a
medium impression increases the expected number of violations by a factor of 2.0566 (105.7%).
Being supervised at the minimum level versus maximum level decreases the expected number of
violations by a factor of 0.2945 (70.6%).
MV2. The criterion for MV2 is the number of drug-related violations, outV2i. Among
the sample, 112=199 = 0:56 had at least one drug-related violation; they accumulated, on average,
roughly 2 during the supervision period, M = 2:4, S D = 3:81. The maximum observed number
of drug-related violations was 21. Interest centers on the predicted rate of the criterion, b i.
Potential predictors were included in a general model which was then recursively parti-
tioned102 into the binary tree shown in Figure 24a using Poisson-splitting methods (see, Breiman
102Using the R library RPART (Therneau & Atkinson, 1997).
147
et al., 1984; Therneau & Atkinson, 1997).103 This classification tree, since it was likely too com-
plex to validate, was trimmed back to that shown in Figure 24b.104
Predictors having the largest e ect on the rate of drug-related violations included (a) super-
vision level supLvli, (b) the number of previous weapons-related convictions psiCrmAdlCnvWeai,
and (c) the rates of arrests for drug-related crimes within the block-group arrDrg:di; remaining
predictors did not appear in the model.
The resultant model separated sampled probationers into 4 groups. The groups (and pre-
dicted values) included those (a) supervised at either minimum, medium, or maximum levels
(1.9); (b) supervised at intensive level and having one or more previous adult weapons-related
convictions (0.42); (c) supervised at intensive level, having no previous adult weapons-related
convictions, and living within a block-group having rates of drug-related arrests greater than or
equal to 9.95 (3.3); and (d) supervised at intensive level, having no previous adult weapons-related
convictions, and living within a block-group having rates of drug-related arrests less than to 9.95
(9.5).
Indicators mini, medi, maxi, and inti were created to represent supLvli and indicators cnvWea0i,
cnvWea1i, and cnvWea2i were created to represent the levels of psiCrmAdlCnvWeai.105 An ini-
tial model
Pr (outV2ijxi) = exp(  i) 
outV2i
i
outV2i!
103The Gini rule was used for splitting, prior probabilities were set proportional to observed frequencies, and altered
priors were used for the loss function.
104Cost-complexity pruning was based on the 1 S E rule among cross-validated data.
105Indicators representing modal categories were omitted as references.
148
supLvl=Minimum,Medium,Maximum
rnsDrg=HSA=NO,HSA=YES + CSA=NO
psiOffense=Other,Drugs
psiNspBra=No
arrDens< 168.2
ImmigrantConcentration< 0.9181
arrPro.d>=8.266
arrOth.d>=81.87
psiCrmAdlCnvDrg>=0.5
ImmigrantConcentration>=1.112
psiCrmAdlCnvWea>=0.5
arrDrg.d>=9.947
popDens>=7.978
rtAK< 0.4096
0.241.3
2.4
0.782.3
5
0.44
0.273
6.3
0.42
0.81
1.75.7
9.5
(a)
supLvl=Minimum,Medium,Maximum
psiCrmAdlCnvWea>=0.5
arrDrg.d>=9.947
1.9
0.42
3.3 9.5
(b)
Figure 24. Initial (a) and pruned (b) classification trees predicting outV2i, MV2, n = 199.
where  i = exp(xi ) and xi1 = mini, xi2 = medi, xi3 = inti, xi4 = cnvWea1i, xi5 = cnvWea2i,
and xi6 = arrDrgDi was then fitted to the sample data; parameter estimates are shown in the first
column of Table 42.106
There was significant evidence that the observations were overdispersed with respect to the
Poisson model, G2 = 477:904; p = 0:000. MV2 was thus respecified to include a multiplicative
disturbance term  i to capture unobserved heterogeneity. The NBR model
Pr (outV2ijxi) = exp( e i)e i
outV2i
outV2i!
106A robust estimator of variance was used in place of the standard estimator to adjust for k = 122 block-group
clusters (Huber, 1967; Rogers, 1993; H. White, 1980). Also, to account for varying times-at-risk, the logged number of
days supervised log (daysS upActi) was included and constrained to a coe cient of 1 (not reported in estimates tables).
149
Table 42
Parameter estimates, Poisson and negative binomial regressions of outV2i, MV2, n = 199.
b=z b=z b=z
min  1:900   1:955  2:046 
( 2:90) ( 2:31) ( 2:28)
med  0:224  0:323  0:294
( 0:65) ( 0:83) ( 0:77)
intensive 0:616  0:341 0:356
(2:65) (1:16) (1:22)
cnvWea1  1:423    1:309   1:171  
( 3:87) ( 2:90) ( 2:64)
cnvWea2 0:006  0:457  0:433
(0:02) ( 1:05) ( 0:99)
arrDrgD  0:004  0:001  0:012  
( 0:73) ( 1:97) ( 3:10)
  5:248    4:915    4:697   
( 22:88) ( 25:06) ( 21:49)
Overdispersion  0:773   0:757   
(4:73) (4:51)
Model  2 29:670   20:610  26:237   
 p < 0:05;  p < 0:01;   p < 0:001.
where e i = exp(xi +  i) and xi1 = mini, xi2 = medi, xi3 = inti, xi4 = cnvWea1i, xi5 = cnvWea2i,
and xi6 = arrDrgDi was then fitted to the sample data. Parameter estimates from this regression
are shown in the second column of Table 42.107
The Wald test108 of the hypothesis that all coe cients except the intercept were zero H0 :
 1 =  2 =  3 =  4 =  5 =  6 = 0 was rejected,  2W = 20:61; d f = 6; p = 0:02. The
test of H0 :  1 =  2 =  3 = 0 was rejected,  2W = 9:21; d f = 3; p = 0:03. The test of
107A robust estimator of variance was used in place of the standard estimator to adjust for k = 122 block-group
clusters (Huber, 1967; Rogers, 1993; H. White, 1980). Also, to account for varying times-at-risk, the logged number of
days supervised log (daysS upActi) was included and constrained to a coe cient of 1 (not reported in estimates tables).
108The  2
L test is likely invalid given the robust estimator of variance used to adjust for k = 122 block-group clusters.
150
H0:  4 =  5 = 0 was rejected,  2W = 8:78; d f = 2; p = 0:01. The test of H0:  6 = 0 was rejected,
 2W = 3:87; d f = 1; p = 0:05.
The model had D=791.17, AIC=4.056, and BIC= 219.853.
By comparison, the PR shown in the first column of Table 42 performed at its worse in
predictions of 0 where it greatly underpredicted counts. The NBR shown in the second column of
Table 42 performed at its worse in predictions of 1 where it overpredicted counts. The PR con-
tinued to perform poorly through counts [1,5] where it consistently overpredicted counts. Model
predictions from the two models do not converge until near counts of 7. Overall, the NBR, with
mean absolute di erence between predicted and observed valuesj b o b pjof 0:009, outperformed
the PR, withj b o b pj= 0:059.
Focus then turned to refining the model shown in the second column of Table 42 in terms
of the scale of arrDrg:di. A normal quantile-comparison plot of arrDrg:di indicated, aside from
one extreme value, the distribution of arrDrg:di approximated the normal. With this outlying case
removed, the relationship was approximately normal.
The probationer representing this single outlying value was removed from the analysis and
the model
Pr (outV2ijxi) = exp(  i) 
outV2i
i
outV2i!
where  i = exp(xi ) and xi1 = mini, xi2 = medi, xi3 = inti, xi4 = cnvWea1i, xi5 = cnvWea2i, and
xi6 = arrDrgDi was refitted to the sample data with the outlying probationer excluded; parameter
estimates are shown in the third column of Table 42.109
109A robust estimator of variance was used in place of the standard estimator to adjust for k = 122 block-group
clusters (Huber, 1967; Rogers, 1993; H. White, 1980). Also, to account for varying times-at-risk, the logged number of
days supervised log (daysS upActi) was included and constrained to a coe cient of 1 (not reported in estimates tables).
151
Figure 25. Observed and NBR model-implied counts of outV2i, MV2, n = 198.
The Wald test110 of the hypothesis that all coe cients except the intercept were zero H0 :
 1 =  2 =  3 =  4 =  5 =  6 = 0 was rejected,  2W = 26:24; d f = 6; p = 0:00. The
test of H0 :  1 =  2 =  3 = 0 was rejected,  2W = 9:07; d f = 3; p = 0:03. The test of
H0:  4 =  5 = 0 was rejected,  2W = 7:39; d f = 2; p = 0:02. The test of H0:  6 = 0 was rejected,
 2W = 9:59; d f = 1; p = 0:00.
The model had D=784.46, AIC=4.043, and BIC= 220.313.
A visual indication of model calibration is given in the plot of observed and predicted
probabilities of counts 0 through 9 shown in Figure 25. As indicated, the model-implied and
observed probabilities track closely throughout the distribution, with the poorest fit in counts of 1
and 2.
The NBR of outV2 on mini, medi, inti, cnvWea1i, cnvWea2i, and arrDrgDi shown in the
third column of Table 42 was validated by bootstrapping. Bias-corrected confidence intervals
110The  2
L test is likely invalid given the robust estimator of variance used to adjust for k = 122 block-group clusters.
152
Table 43
Lower (LCLM) and upper (UCLM) bias-corrected 95% confidence limits, MV2, n = 198.
LCLM UCLM
mini  9:1 5:007
medi  1:07 :482
intensivei  :236 :947
cnvWea1i  2:08  :261
cnvWea2i  8:94 8:07
arrDrgDi  :0212  :00275
around the estimated parameters are shown in Table 43. As indicated in the table, the e ects of
cnvWea1i and arrDrgDi were likely to replicate in the population; the others were not.
The most important characteristics influencing the rate of violations included cnvWea1i and
arrDrgDi. Being supervised at the minimum level versus maximum level decreases the expected
number of drug-related violations by a factor of 0.1293 (-87.1%) holding all other predictors
constant. Having 1 previous weapons-related convictions versus none decreases the expected
number of drug-related violations by a factor of 0.6565 (-34.4%) holding all other predictors
constant. For a standard deviation increase in the rate of arrests for drug-related crimes (M =
24:89,S D = 23:36), the expected number of drug-related violations decreases by a factor of 0.7560
(-24.4%).
Modes of termination
Model MT1 estimates likelihoods probationers in the population will terminate sentences
unsuccessfully and Model MT2 estimates how soon such failures are likely to occur.
The criteria outT1i and outT2i are described in Table 44. The two commonest modes of
termination were revocation and successful; the two rarest, absconsion and death. Roughly 0.5 of
the sample terminated within 412 days; 0.75 by 730 days; and 0.9 by 951 days. By type, median
153
Table 44
Median days until termination, by mode of termination, n = 199.
Not Successful
Absconsion Death Unsuccessful Revoked Successful
n 2 2 24 89 82
Days 48 329 463 341 546
days until termination ranged (0.25 and 0.75 quantiles) within [37:5;58:5] days until absconsion,
[237;420] days until death, [231;584] days until revoked, [365;738] days until successful, and
[240;681] days until unsuccessful termination modes.
MT1. The criterion for MT1 is whether probationers terminated their sentence unsuccess-
fully outT1i. If so, outT1i = 1. Roughly three-fifths (115=199 = 0:58) of the sampled probationers
did indeed terminate unsuccessfully. Interest centers on the predicted probability of the criterion,
bPr (outT1i = 1).
Potential predictors of this criterion were included in a general model which was then re-
cursively partitioned111 into the binary tree shown in Figure 16a.112 This classification tree, since
it was likely too complex to validate, was trimmed back to that shown in Figure 26b.113
Predictors with the largest influence on the criteria were (a) recommended sentence psiS Ri,
(b) proportion of positive drug screens for cocaine piPos4, (c) supervision level supLvl, and
(d) age at first arrest rnsAFi; remaining predictors did not appear in the model.
The resultant model separated sampled probationers into 5 groups. These groups (and pre-
dicted values) included those (a) recommended to probation, having less than 0.028 of screens
positive for cocaine, and supervised at either minimum or medium levels (Success). There were
111Using the R library RPART (Therneau & Atkinson, 1997).
112The Gini rule was used for splitting, prior probabilities were set proportional to observed frequencies, and altered
priors were used for the loss function.
113Cost-complexity pruning was based on the 1 S E rule among crossvalidated data.
154
psiSR=Probation
piPos4< 0.02838
supLvl=Minimum,Medium
rnsAF=16 to 17,over 26
rnsDrg=HSA=NO,HSA=YES + CSA=NO
psiEduGrdAt>=11.5
alcDens< 2.707
arrDrg.d>=37.48
psiEmpMore>=0.5
Success
Success
SuccessFailure
Failure
SuccessFailure
SuccessFailure
Failure
(a)
psiSR=Probation
piPos4< 0.02838
supLvl=Minimum,Medium
rnsAF=16 to 17,over 26
Success
SuccessFailure
Failure
Failure
(b)
Figure 26. Initial (a) and pruned (b) classification trees predicting outT1i, MT1, n = 199.
1=25 = 0:04 unsuccessful terminations among this group; (b) recommended to probation, having
less than 0.028 of screens positive for cocaine, supervised at either maximum or intensive levels,
and first arrested at ages 16 to 17 or over age 26 (Success). 0=10 = 0:00 of these o enders termi-
nated unsuccessfully; (c) recommended to probation, having less than 0.028 of screens positive for
cocaine, supervised at either maximum or intensive levels, and first arrested at age 15 and younger
or ages 18 to 25 (Failure). 22=41 = 0:54 of these o enders terminated unsuccessfully; (d) recom-
mended to probation and 0.028 or more of screens positive for cocaine (Failure). 22=33 = 0:67
of these o enders terminated unsuccessfully; and (e) recommended to either split-sentence or
incarceration (Failure). 70=90 = 0:78 of these probationers terminated unsuccessfully.
The most influential predictor was whether the PSI author recommended a sentence of
probation versus either incarceration or split-sentence. The indicator probationi was created to
capture this. In addition, the indicators Mini, Medi, Maxi, and Inti were created to represent
155
the levels of supervision level supLvli. Finally, the indicators AF1i, AF2i, AF3i, and AF4i were
created to represent the levels of age at first arrest rnsAFi.
An initial model
bPr (outT1i = 1jxi) = 1
1 + exp   (xi ) 
= 11 + exp   ( 
0 + 1xi1 + 2xi2 + 3xi3 + 4xi4 + 5xi5 + 6xi6 + 7xi7 + 8xi8)
 ;
where xi1 = probationi, xi2 = piPos4i, xi3 = Mini, xi4 = Medi, xi5 = Inti, xi6 = AF1i, xi7 = AF2i,
and xi8 = AF4i, was then fitted to the sample data; parameter estimates are shown in the first
column of Table 45.114
The Wald test115 of H0 :  1 =  2 =  3 =  4 =  5 =  6 =  7 =  8 = 0 was rejected,  2W =
49:08; d f = 8; p < 0:00. The test of H0:  1 = 0 was rejected ( 2W = 16:85; d f = 1; p = 0:00); the
test of H0:  2 = 0 was rejected ( 2W = 13:62; d f = 1; p = 0:00); the test of H0:  3 =  4 =  5 = 0
was rejected ( 2W = 11:62; d f = 3; p = 0:01); the test of H0 :  6 =  7 =  8 = 0 was not rejected
( 2W = 6:63; d f = 3; p = 0:08).
The model correctly classified 0:74 of the sample and the AUC suggested the it demon-
strated acceptable discrimination, AUC = 0:80. The model had D=208.98, AIC=1.141, and
BIC= 796.746; the bc suggested the model was empirically consistent,  2HL = 9:29; d f = 8; p =
0:32.
Focus then turned to refining this preliminary model in terms of parametric relationships
and scale beginning with the relationship between outT1i and piPos4i. The tree in Figure 26b
suggested a split in piPos4 in the low end of the distribution. It ranged within [0:0;0:69] with
M = 0:07 and S D = 0:14 and near evenly split the sample: just under half 93=199 = 0:47 had
piPosi > 0. As this predictor represents a proportion which, in general, do not respond well
114A robust estimator of variance was used in place of the standard estimator to adjust for k = 122 block-group
clusters (Huber, 1967; Rogers, 1993; H. White, 1980).
115The  2
L test is likely invalid given the robust estimator of variance used to adjust for k = 122 block-group clusters.
156
Table 45
Parameter estimates, MT1, logistic regression of outT1i, n = 199.
b=z b=z
probation  1:584    1:565   
( 4:10) ( 4:06)
piPos4 6:438   
(3:69)
Min  2:402  2:376 
( 2:38) ( 2:37)
Med  0:638  0:622
( 1:68) ( 1:69)
Int 0:520 0:550
(1:24) (1:32)
AF1  0:199  0:118
( 0:34) ( 0:21)
AF2  0:782  0:710
( 1:84) ( 1:72)
AF4  1:168  0:912
( 2:15) ( 1:81)
piPos4 1:029  
(2:75)
 1:282   1:240  
(3:36) (3:26)
Model  2 49:079   40:081   
 p < 0:05;  p < 0:01;   p < 0:001.
157
when values approach the boundaries, power transformations were ine ective. As was an arcsine
transformation. On the other hand, a binary split at zero implied a zero-tolerance policy on drug
use. In practice, however, positive screens for cocaine may be met initially with sanctions aimed at
curtailing the behavior and thus some probationers with non-zero proportions of positive screens
for cocaine may indeed terminate successfully. To account for this piPosi was replaced with the
indicator piPos i = piPosi > 0:02, just under the 0:25Q of the distribution of non-zero values.
The model
bPr (outT1i = 1jxi) = 1
1 + exp   (xi ) 
= 11 + exp   ( 
0 + 1xi1 + 2xi2 + 3xi3 + 4xi4 + 5xi5 + 6xi6 + 7xi7 + 8xi8)
 ;
where xi1 = probationi, xi2 = piPos4 i , xi3 = Mini, xi4 = Medi, xi5 = Inti, xi6 = AF1i, xi7 =
AF2i, and xi8 = AF4i, was refitted to the sample data; parameter estimates are shown in the second
column of Table 45.116
The Wald test117 of H0 :  1 =  2 =  3 =  4 =  5 =  6 =  7 =  8 = 0 was rejected,  2W =
40:08; d f = 8; p < 0:00. The test of H0:  1 = 0 was rejected ( 2W = 16:45; d f = 1; p = 0:00); the
test of H0 :  2 = 0 was rejected ( 2W = 7:54; d f = 1; p = 0:01); the test of H0 :  3 =  4 =  5 = 0
was rejected ( 2W = 12:22; d f = 3; p = 0:01); the test of H0 :  6 =  7 =  8 = 0 was not rejected
( 2W = 5:01; d f = 3; p = 0:17).
The model correctly classified 0:74 of the sample. The AUC suggested the model demon-
strated acceptable discrimination, AUC = 0:78. The model had D=216.21, AIC=1.177, and
BIC= 789.522; thebc suggested the model was empirically consistent,  2HL = 11:32; d f = 8; p =
0:18.
116A robust estimator of variance was used in place of the standard estimator to adjust for k = 122 block-group
clusters (Huber, 1967; Rogers, 1993; H. White, 1980).
117The  2
L test is likely invalid given the robust estimator of variance used to adjust for k = 122 block-group clusters.
158
0
.2
.4
.6
.8
1
Observed
0 .2 .4 .6 .8 1
Model implied
Figure 27. Comparison between model-implied probabilities of experiencing outT1i and a mov-
ing average of the proportion of probationers terminating unsuccessfully, n = 199.
A visual indication of model calibration is given in the plot in Figure 27 which compares
predicted probabilities from the regression of outT1 with a moving average of the proportion of
probationers terminating unsuccessfully.
The regression of outT1i on probationi, piPos4 i , Mini, Medi, Inti, AF1i, AF2i, and AF4i
shown in the second column of Table 45 was validated by bootstrapping. Bias-corrected confi-
dence intervals around the estimated parameters are shown in Table 46.
As indicated in Table 46, the e ects of both probationi and piPos4 i on outT1i are likely
to validate in the population. Also, the Wald test118 of H0 :  3 =  4 =  5 = 0 was rejected
( 2W = 22:00; d f = 3; p = 0:00). On the other hand, the test of H0 :  6 =  7 =  8 = 0
was not rejected ( 2W = 4:16; d f = 3; p = 0:25). Interpretations of the regression of outT1i
focus on the e ects likely to replicate in the population, which are, namely whether the PSI author
recommended probation versus either incarceration or split sentence, whether the proportion of
positive screens for cocaine exceeded 0.02, and supervision level.
118The  2
L test is likely invalid given the robust estimator of variance used to adjust for k = 122 block-group clusters.
159
Table 46
Lower (LCLM) and upper (UCLM) bias-corrected 95% confidence limits, MT1, n = 199.
LCLM UCLM
probation  2:4  :727
piPos :23 1:83
Min  3:55  1:21
Med  1:4 :155
Int  :332 1:43
AF1  1:41 1:18
AF2  1:59 :169
AF4  2:01 :183
Predicted values b ranged within the interval [:026;0:944], with mean b = :578 and S D =
:248. The most important characteristics influencing bPr (outT1i = 1) include the recommended
sentence, the proportion of positive screens for cocaine, and supervision level.
The expected odds of terminating unsuccessfully are roughly 4.8 times larger among those
whose recommended sentence is either incarceration or split-sentence compared to those whose
recommended sentence is probation and roughly 2.8 times larger among those having failed more
than 0.02 of cocaine screens compared to their counterparts, holding all else constant. Compared
to those supervised at the maximum level, the expected odds of terminating unsuccessfully are
roughly 90% smaller among those supervised at the minimum level and roughly 46% smaller
among those supervised at the medium level. The odds of terminating unsuccessfully are roughly
73% larger among those supervised at the intensive level.
The plot in Figure 28 shows predicted probabilities of terminating sentences unsuccessfully
by recommended sentence, supervision level, and whether proportions of positive screens for co-
caine exceed 0.02. As indicated, the predicted probabilities of terminating unsuccessfully increase
with each supervision level and probabilities are consistently higher among those recommended
160
Supervision level
Minumum Medium Maximum Intensive
Predicted probability
0.0
0.2
0.4
0.6
0.8
1.0
Incarceration
Probation
or Split-sentence
Figure 28. Predicted probabilities of terminating sentences unsuccessfully, circles represent pro-
bationers having failed more than 0.02 of screens for cocaine and squares represent their coun-
terparts, hollow symbols represent those whose recommended sentence is probation and solid
symbols represent those whose recommended sentence is either incarceration or split sentence,
MT1, n = 199.
to either incarceration or split sentence and among those whose proportions of positive screens for
cocaine exceed 0.02.
MT2. The criterion for MT2 is the time until unsuccessful termination outT2i. Roughly
three-fifths (115=199 = 0:58) of the sampled probationers did indeed terminate unsuccessfully. The
average followup time ranged within the interval [16;2219]; the 95%CI around median survival
time spanned the interval [516;734].
161
Potential predictors of time until failure were included in a general model which was then
recursively partitioned119 into the binary tree shown in Figure 29a.120 This classification tree,
since it was likely too complex to validate, was trimmed back to that shown in Figure 29b.121
Predictors with the largest influence on the criterion were (a) the number of previous drug-
related convictions psiCrmAdlCnvDrgi, (b) the proportion of positive screens for marijuana piPos5i,
and (c) the expected number of days of supervision daysS upExpi; remaining predictors did not
appear in the model.
The resultant model separated sampled probationers into 4 groups. These groups (and pre-
dicted values) included those having (a) fewer than 5 previous drug convictions, less than 0.3909
of positive screens for marijuana, and expected to be supervised for more than 487 days (0.67);
(b) fewer than 5 previous drug convictions, less than 0.3909 of positive screens for marijuana, and
expecting to be supervised for less than 487 days (1.3); (c) fewer than 5 previous drug convic-
tions and 0.3909 or more of positive screens for marijuana (2.6); and (d) 5 or more previous drug
convictions (3);
The indicators cnvDrg0i, cnvDrg1i, and cnvDrg2i were created to represent the levels
None, One, and Two or more of psiCrmAdlCnvDrgi and an initial Cox regression model
h(tjxi) = h0(t) exp(xi )
= h0(t) exp( 1xi1 + 2xi2 + 3xi3 + 4xi4)
where xi1 = cnvDrg1i, xi2 = cnvDrg2i, xi3 = piPos5i, and xi4 = daysS upExpi, was then fitted to
the sample data; parameter estimates are shown in the first column of Table 47.122
119Using the R library RPART (Therneau & Atkinson, 1997).
120The Gini rule was used for splitting, prior probabilities were set proportional to observed frequencies, and altered
priors were used for the loss function.
121Cost-complexity pruning was based on the 1 S E rule among cross-validated data.
122A robust estimator of variance was used in place of the standard estimator to adjust for k = 122 block-group
clusters (Huber, 1967; Rogers, 1993; H. White, 1980).
162
psiCrmAdlCnvDrg< 4.5
piPos5< 0.3909
daysSupX>=486.5
arrPro.d< 6.074
piPos5< 0.03184
psiCrmAdlCasN< 11.5
arrVio.d>=35.5
psiSubEU2< 0.5
psiFamSibAL< 5.5
psiOffense=Other,Property,Drugs
0.0720.54
0.98
0.411
0.24
0.782.2
2.8
2.6
3
(a)
psiCrmAdlCnvDrg< 4.5
piPos5< 0.3909
daysSupX>=486.5
0.67 1.3
2.6
3
(b)
Figure 29. Initial (a) and pruned (b) classification trees predicting outT2i, MT2, n = 199.
Table 47
Parameter estimates, MT2, Cox regression of outT2i, n = 199.
b=z
cnvDrg1 0:387
(1:41)
cnvDrg2 0:583 
(2:50)
piPos5 2:215   
(4:10)
daysS upExp  0:001   
( 3:38)
Model  2 37:511   
 p < 0:05;  p < 0:01;   p < 0:001.
163
The Wald test123 of H0 :  1 =  2 =  3 =  4 = 0 was rejected,  2W = 37:56; d f = 4; p <
0:00. The test of H0 :  1 =  2 = 0 was rejected ( 2W = 6:39; d f = 2; p = 0:04), the test of
H0:  3 = 0 was rejected ( 2W = 16:82; d f = 1; p = 0:00), and the test of H0:  4 = 0 was rejected
( 2W = 11:39; d f = 1; p = 0:00).
The model had log L= 493.982, AIC=995.964, and BIC=1,009.137.
An analogue to the bc for BLR models has been proposed by May and Hosmer (1998)
that involves partitioning the covariate space in a manner similar to that proposed by Hosmer
and Lemeshow (1980). Rather than partitioning subjects into groups based on the percentiles of
the predicted probabilities from the fitted logistic regression model, subjects are partitioned into
groups based on their predicted hazard rates from the Cox model (also see, DeMaris, 2004; Parzen
& Lipsitz, 1999). There was no evidence here to suggest the model lacked empirical consistency,
 2MH = 10:89; d f = 9; p = 0:28.
A visual indication of model calibration is given in the plot in Figure 30 which compares
the Nelson-Aalen cumulative hazard function with the Cox-Snell residuals. If a Cox regression
fits the data well then the Cox-Snell residuals should follow a standard exponential distribution
with a hazard function of 1 and the cumulative hazard of these residuals should lie on a straight
45 line (see, Cleves, Gould, & Gutierrez, 2002). Here, that the thick line tracks closely with the
diagonal124 indicates the model is well-calibrated.
Focus then turned to assessing whether the two continuous predictors were linear in the
log hazard and, if not, which transformation would linearize them. I began with the relationship
between outT2i and piPos5i by replacing piPos5i with design variables formed from the Q25,
Q50, and Q75 of the non-zero values. These were, respectfully, 0.0656, 0.188, and 0.346. I refitted
123The  2
L test is likely invalid given the robust estimator of variance used to adjust for k = 122 block-group clusters.
124Cleves et al. (2002) point out that due to the reduced e ective sample caused by failures and censoring, there will
be some variability around the 45 line especially in the upper region.
164
0
1
2
3
4
0 1 2 3 4
Cox?Snell residual
Figure 30. Comparison between the Nelson-Aalen cumulative hazard function and the Cox-Snell
residuals, MT2, n = 199.
the model with the design variables then plotted the estimated coe cients against group midpoints.
The overall trend indicated a positive e ect without substantial departure from linearity. The
assumption of linearity in the log hazard was thus supported.
I then turned to checking the scale of daysS upExpi. Here again design variables were
used. I began by replacing daysS upExpi with design variables formed from Q25, Q50, and Q75.
These were, respectfully, 365, 548, and 730 days. I refitted the model with the design variables
then plotted the estimated coe cients against group midpoints. The overall trend indicated a
negative e ect without substantial departure from linearity which, again, supported the assumption
of linearity in the log hazard.
Given evidence supporting the assumption of linearity in the log hazard, the Cox regression
shown in the first column of Table 47 was evaluated for proportional hazards. All the predictors
were examined for proportionality by fitting a model that included the interaction of each variable
with mean-centered log-time log(t) log(t). The nonsignificant Wald tests125 of H0:  1cnvDrg1 
125The  2
L test is likely invalid given the robust estimator of variance used to adjust for k = 122 block-group clusters.
165
Table 48
Lower (LCLM) and upper (UCLM) bias-corrected 95% confidence limits, MT2, n = 199.
LCLM UCLM
cnvDrg1  :169 :943
cnvDrg2 :105 1:06
piPos5 1:04 3:39
daysS upExp  :00182  :00042
[log(t) log(t)] =  2cnvDrg2  [log(t) log(t)] = 0 ( 2W = 4:50; d f = 2; p < 0:11), H0 :
 3 piPos5 [log(t) log(t)] = 0 ( 2W = 0:27; d f = 1; p = 0:60), and H0 :  4daysS upExp 
[log(t) log(t)] = 0 ( 2W = 0:09; d f = 1; p = 0:77) suggested that the hazard function may be
proportional in each of the predictors.
The Cox regression shown in the first column of Table 47 was validated by bootstrapping.
Bias-corrected confidence intervals around the estimated parameters are shown in Table 48. As
indicated in the table, except for the e ect of cnvDrg1i, all e ects are likely to replicate in the
population.
Controlling the other predictors in the model, the hazard rate is expected to increase by
roughly 25% for each 0.10 increase in the proportion of positive screens for marijuana. On the
other hand, the expected hazard decreases by roughly 18% for each 180 day increase in the ex-
pected length of supervision.
Survival experience does not appear to di er between those with no previous drug-related
convictions and those with one. It does di er between those with two or more and those with
none. Compared to those with no previous drug related convictions, those with two or more are
expected to fail at a rate that is roughly 79% higher. It also di ers between those with two or more
and those with one. Compared to those with only one previous drug-related conviction, those with
two or more are expected to fail at a rate that is roughly 22% higher.
166
DISCUSSION
This chapter describes the findings that were reported in the previous chapter, and the ma-
jority of the discussion is devoted to a summary of the key findings and integrating these findings
with extant theory and research. A discussion of the limitations of the study and suggestions for
future work follows, and I conclude with a few summary statements regarding the project as a
whole.
The findings from this study are largely in accordance with available research convincingly
suggesting a sizeable fraction of o enders enters probation yet fails to comply with conditions
of release (see, BJS, 2000, 2002, 2003; Bonczar, 1997; Bork, 1995; Clear et al., 1992; Glaze &
Bonczar, 2006; Glaze & Palla, 2004; Gray et al., 2001; Langan & Cunni , 1992; Mayzer et al.,
2004; Petersilia et al., 1985; Petersilia, 1985a, 1985b, 1998). For instance, about 4 out of every
5 probationers in the sample tested positive, provided a bogus specimen, or failed to appear for
a drug testing event at least once while supervised. One or more screens were positive for about
1 in every 2 probationers screened for marijuana and cocaine and about 1 in every 5 or 6 pro-
bationers screened for opiates and phencyclidine. As for convictions, roughly one-fourth of the
sampled probationers was arrested and subsequently convicted of new crimes during the super-
vision period; a roughly equal fraction was arrested and subsequently convicted of new crimes
during the post-supervision period. Roughly 5 out of every 8 probationers violated one or more
supervision-related conditions and 4 out of every 8 violated one or more drug-related conditions.
Ultimately, about 3 out of every 5 probationers terminated their sentences unsuccessfully. These
levels of negative supervision performance (NSP) steepen the challenges of o ender management
and reintegration facing the Court Services and O ender Supervision Agency for the District of
Columbia (CSOSA).
167
Choices facing supervision agencies often necessitate judgments about the future behaviors
of those under their charge. Such behaviors include whether supervised o enders will abstain
from illegal substances, discontinue criminal involvement, comply with release conditions, and,
ultimately, successfully complete their sentences and reintegrate with their communities. A small
but growing body of research has begun to identify characteristics associated with these aspects
of supervision performance (see, Albonetti & Hepburn, 1997; Benedict & Hu -Corzine, 1997;
R. L. Cohen, 1995; Gray et al., 2001; Harer, 1994; M. Jones, 1995; Kronick et al., 1998; Langan &
Levin, 2002; MacKenzie et al., 1998; MacKenzie & Li, 2002; MacKenzie et al., 1992; MacKenzie
& Souryal, 1994; MacKenzie, 1991; Mayzer et al., 2004; Minor et al., 2003; Morgan, 1993;
Schwaner, 1997; Silver & Chow-Martin, 2002; Sims & Jones, 1997; F. P. Williams III et al.,
2000). This study has drawn heavily on and adds to this body of research by refining this set of
characteristics with respect to probationers supervised in the the District of Columbia (DC) by the
CSOSA.
The results presented in the previous chapter identified from well over 200 theoretically
plausible predictors a very small set that provide the agency with advance notice of the most
challenging groups of o enders. This set of characteristics, tabulated in Table 49 with relevant
models and domains, includes (a) the age at the time of assessment, (b) the expected length of
supervision, (c) the number of substances ever used, (d) whether the probationer had ever used
opiates or phencyclidine, (e) the number of weapons-related convictions, (f) the SFS-98 score,
(g) the recommended sentence, (h) the impression of recidivism risk on the supervising CSO, and
(i) local rates of arrests for drug-related and public order crimes. It is important to point out that
each of these characteristics are knowable prior to the commencement of supervision; most are
derivable from the PSIs and thus all lend themselves to immediate, risk-anticipated security and
treatment decisions.
168
Table
49
Characteristics
associated
with
increased
lev
els
of
ne
gati
ve
supervision
performance
among
Black
males
supervised
in
the
DC.
Model
Criterion
Description
S1
Probabilities
of
ev
er
testing
positi
ve,
pro
viding
abogus
specimen,
or
failing
to
appear
for
adrug-testing
ev
ent
while
supervised
1.
Probationers
expecting
longer
periods
of
supervision
are
more
lik
ely
to
fail
2.
Probationers
ha
ving
used
agreater
number
of
substances
are
more
lik
ely
to
fail
S2D
Expected
rates
of
positi
ve
screens
for
cocaine
1.
Expected
rates
of
positi
ve
screens
increase
with
number
of
substances
ev
er
used
2.
Expected
rates
of
positi
ve
screens
increase
with
age
S2E
Expected
rates
of
positi
ve
screens
for
marijuana
1.
Expected
rates
of
positi
ve
screens
decrease
with
age
S2F
Expected
rates
of
positi
ve
screens
for
opiates
1.
Expected
rates
of
positi
ve
screens
higher
among
those
ha
ving
ev
er
used
heroin
S2G
Expected
rates
of
positi
ve
screens
for
PCP
1.
Expected
rates
of
positi
ve
screens
higher
among
those
ha
ving
ev
er
used
PCP
C1
Probabilities
of
being
arrested
and
subsequently
con
victed
on
ne
w
char
ges
during
the
supervision
period
1.
Lo
wer
SFS-98
scores
are
associated
with
increased
chances
2.
Those
recommended
to
incarceration
or
split-sentence
ha
ve
higher
chances
C2
Probabilities
of
being
arrested
and
subsequently
con
victed
on
ne
w
char
ges
during
the
post-supervision
period
1.
Lo
wer
SFS-98
scores
are
associated
with
increased
chances
2.
O 
enders
living
in
areas
with
higher
public
order
arrests
ha
ve
higher
chances
V1
Expected
rates
of
supervision-related
violations
1.
Lo
wer
SFS-98
scores
are
associated
with
increased
rates
2.
Ha
ving
ahigh
impression
of
risk
increases
the
expected
rate
V2
Expected
rates
of
drug-related
violations
1.
O 
enders
living
in
areas
with
lower
drug-related
arrests
ha
ve
higher
expected
rates
2.
Ha
ving
fewer
weapons-related
con
victions
is
associated
with
higher
expected
rates
T1
Probabilities
of
terminating
unsuccessfully
1.
Those
supervised
at
more
intense
lev
els
ha
ve
higher
chances
2.
Those
failing
0.02
of
cocaine
tests
ha
ve
higher
chances
3.
Those
recommended
to
incarceration
or
split-sentence
ha
ve
higher
chances
T2
Expected
rates
of
unsuccessful
termination
1.
Expected
hazard
increases
as
proportion
of
positi
ve
marijuana
screens
increases
2.
Expected
hazard
decreases
as
length
of
supervision
increases
169
As indicated by the predictors shown in Table 49, those having the strongest influence on
NSP are largely those bearing on criminal histories and substance use. As such, there are some
obvious theoretical linkages that remain unsupported among the data examined here with respect
to substance use, rearrests resulting in convictions, condition violations, and termination modes.
For instance, while research suggests certain educational characteristics will predict NSP,
they were not influential among these data (cf., Gray et al., 2001; Harer, 1994; Irish, 1989; Landis
et al., 1969; Mayzer et al., 2004; Morgan, 1993; Rhodes, 1986; Roundtree et al., 1984; Silver &
Chow-Martin, 2002; Sims & Jones, 1997). Nor were either residential or employment stabilities
(cf., J. Austin & Litsky, 1982; Gray et al., 2001; Harer, 1994; Irish, 1989; M. Jones, 1995;
Landis et al., 1969; MacKenzie & Li, 2002; Mayzer et al., 2004; Morgan, 1993; Silver & Chow-
Martin, 2002; Sims & Jones, 1997; F. P. Williams III et al., 2000). Childhood and family factors
are also expected to influence NSP, but none of the characteristics observed among these data
were influential (cf., Clarke et al., 1988; Harer, 1994; Landis et al., 1969; MacKenzie & Li,
2002; Morgan, 1993; Petersilia, 1985a; Sims & Jones, 1997). Nor were any of the health-related
characteristics (cf., Bland et al., 1998; Estro et al., 1985; Farrington, 1995; Johnston & O?Malley,
1986; Lin et al., 1996; Link, Andrews, & Cullen, 1992; Link, Cullen, & Wozniak, 1992; Monahan,
1992; Shepherd et al., 2002; Teplin et al., 1996; Teplin, 1990). Research also suggests certain
contextual characteristics will influence NSP, however, none of the sociodemographic predictors,
either singly or in combination, were important among these data (cf., Harer, 1994; Silver &
Chow-Martin, 2002; Sims & Jones, 1997). Nor were either local commercialization levels or
concentrations of alcohol-related businesses (cf., Cochran et al., 1998; Costanza et al., 2001;
D. M. Gorman et al., 2001; D. Gorman, Speer, Labouvie, & Subaiya, 1998; Gyimah-Brempong,
2001; R. Lipton & Gruenewald, 2002; Scribner et al., 1999, 1995; Sherman et al., 1989; Speer et
al., 1998; Stitt & Giacopassi, 1992). And, except for arrest rates for public order and drug-related
170
o enses, objective crime levels within the block-group had little influence on the production of
NSP in these data (cf., Boggs, 1965; Carter & Hill, 1978; Gould, 1969; Kelling & Coles, 1996;
Reppeto, 1974; D. A. Smith, 1986; Stark, 1987; Taylor, 1999; Wilson & Kelling, 1982).
The absence of these seemingly fitting predictors of NSP could reflect a genuine lack of re-
lationship or at least a substantially smaller and di erent than expected relationship among Black
male probationers supervised in the DC. At the same time, their absence could be a consequence
of a few specific methodological limitations. I discuss these limitations later in the chapter. Ul-
timately, if causal relationships were chief among methodological interests?which here they are
not?separate, targeted analyses of specific linkages would be an interesting digression.
I turn for now to an examination of those characteristics that are associated with heightened
risks for NSP and that are likely to replicate in the population beginning with characteristics related
to substance use.
This research examined two sets of criteria bearing on substance use. The first examined
whether o enders in the population will ever test positive, provide a bogus specimen, or fail to
appear for a drug-testing event while supervised. The second examined how often they will test
positive for 7 di erent illegal substances.
The two most important characteristics influencing whether probationers will ever test pos-
itive, provide a bogus specimen, or fail to appear for a drug testing event while supervised are the
number of substances ever used and the expected length of supervision. The likelihood of ever
failing a drug testing event is higher among nonabstainers than their counterparts. It is important
to point out, though, that most of the sample reported having used at least one substance; only
12=199 = 0:06 reported having used none. Hence, while having never used illegal substances in
the past is a good predictor of abstinence while supervised, it is a fairly uncommon characteristic
171
and thus of dubious importance when making immediate, risk-anticipated security and treatment
decisions.
Versatility in the number of di erent substances ever used, on the other hand, was diverse.
For instance, about 2 in every 16 (30=199 = 0:15) sampled probationers had previously used one
illegal substance, 4 in every 16 (53=199 = 0:27) had used two, 5 in every 16 (59=199 = 0:30) had used
three, 2 in every 16 (27=199 = 0:14) had used four, and 1 in every 16 (18=199 = 0:09) had used five
or more. The findings reported here suggest the likelihood of failing a drug testing event while
supervised increases with versatility in the number of di erent substances ever used.
That the criteria was itself a measure of drug use highlights the importance of behavioral
stability. In addition, as was pointed out in the review, research strongly links illegal substances
with contemporaneous o ending (Anglin & Speckart, 1988; Bland et al., 1998; M. R. Chaiken &
Chaiken, 1987; Clayton & Tuchfeld, 1982; Dembo et al., 1995; Dembo, Williams, Getreu, et al.,
1991; Dobinson & Ward, 1986; Elliott et al., 1989; L. Gardner & Shoemaker, 1989; S. D. Got-
tfredson & Gottfredson, 1979; Greenfield & Weisner, 1995; Guze et al., 1968; Inciardi, 1980;
D. C. McBride & McCoy, 1981; McGlothin, 1979; Newcomb & Bentler, 1986; Nurco, 1979;
J. Palmer & Carlson, 1976; Speckart & Anglin, 1985; Stacy & Newcomb, 1995; Stice et al., 1998;
Swanson et al., 1990; Wish & Johnson, 1986). Thus, probationers with the highest chances of fail-
ing drug testing events are likely also to be those that will continue to be criminally active while
serving their sentences in the community (J. Austin & Litsky, 1982; Baird et al., 1984; Benedict
& Hu -Corzine, 1997; Gray et al., 2001; Harer, 1994; MacKenzie et al., 1998; MacKenzie &
Li, 2002; Schmidt & Witte, 1988; Silver & Chow-Martin, 2002). The agency should thus care-
fully examine past substance use when making immediate, risk-anticipated security and treatment
decisions bearing on future use.
172
Probationers serving longer sentences are also more likely to fail drug testing events than
their counterparts. The review highlighted evidence that length of supervision has been consis-
tently linked with supervision performance (Benedict & Hu -Corzine, 1997; Kronick et al., 1998;
MacKenzie et al., 1992; MacKenzie, 1991; Mayzer et al., 2004; Morgan, 1993; Rhodes, 1986;
Roundtree et al., 1984; Sims & Jones, 1997). Those serving longer sentences tend to fair worse.
While the review focused mainly on the linkage between sentence length and either contempo-
raneous o ending or failure while supervised, the implication here is that sentence length is also
related to abstinence while supervised: o enders with longer sentences are less likely to abstain.
Whether this reflects anticipation or artifact in unclear.
With lengthening sentences, nonabstainers may believe their chances of staying drug-free
and sober become too slim to pressure abstention. As they have little expectation of complying
with conditions of community release, some o enders may buy time in the community fully an-
ticipating eventual revocation and incarceration. This could suggest either that the conditions of
community release, particularly the policies on drug abstinence, are not fully understood or that
there is a shared belief that drug testing methods are fallible. On the other hand, those unlikely
to abstain while supervised may be more likely to receive longer sentences in the first place as a
result of sentencing patterns targeting chronic drug o enders. Thus, for perfectly understandable
and predictable reasons, those with longer sentences will probably fail drug testing events and
would probably do so even if their sentences were shorter. What is clear, nevertheless, is that su-
pervision agencies should consider the influence of sentence length when making risk-anticipated
decisions bearing on substance use because those serving relatively longer sentences are likely to
continue using substances and, by extension, to be criminally active while serving their sentence
in the community.
173
In addition to assessing whether they will fail drug testing events, this study also estimated
how often probationers in the population will test positive for 7 di erent substances. Unfortu-
nately, sample screening rates were relatively low for alcohol, methadone, and amphetamines.
It was suspected that this infrequency reflected a non-random selection process. As such, these 3
substances were excluded. On the other hand, screens for cocaine, marijuana, opiates, and phency-
clidine were common across the sample. The findings from the models predicting rates of positive
screens for these substances are discussed next beginning with cocaine and marijuana.
Age is an important predictor of how often probationers will test positive for both cocaine
and marijuana. It is, in fact, the single most important characteristic influencing marijuana posi-
tives: younger o enders test positive at higher rates than their older counterparts. Age is important
in predicting rates of cocaine positives as well?as is the number of substances ever used?but the
relationship is inverted: older o enders and those having used a greater number of substances test
positive for cocaine at higher rates than their counterparts.
The bulk of the evidence suggests that o ending and NSP is concentrated among the youth
(Clarke et al., 1988; Cloninger & Guze, 1973; Dembo et al., 1995; Dembo, Williams, Schmeidler,
et al., 1991; Farrington, 1986; Gendreau, Little, & Goggin, 1996; M. R. Gottfredson & Hirschi,
1986; Harer, 1994; Harrison & Gfroerer, 1992; Ho man & Beck, 1984; Irish, 1989; Matza, 1964;
Morgan, 1993; Osgood et al., 1989; Rhodes, 1986; Sampson & Laub, 1993; Sims & Jones, 1997;
Whitehead, 1991; Wolfgang et al., 1987). We expect drug use, by extension, to have a similar
distribution. The evidence presented here with respect to marijuana is largely consistent with such
a pattern. The evidence with respect to cocaine, however, is not.
This raises the question of whether older o enders are more likely to use cocaine and thus
more likely to test positive more often than younger o enders or rather that older cocaine users are
more likely than younger users to continue using cocaine while supervised. The lack of association
174
Table 50
Frequencies, ever used cocaine and marijuana, by age, n = 199.
Cocaine Marijuana
Age n No Yes No Yes
[19;26) 72 52 20 7 65
[26;39) 61 30 31 10 51
[39;62] 66 11 55 23 43
Total 199 93 106 40 159
between past cocaine use and rates of testing positive suggests not all users of cocaine continue
to use the substance while supervised. This inconsistency might then reflect the drug popularity
trends within age groups that Golub and Johnson (1999) discuss. They argue older o enders are
more likely to favor cocaine whereas younger o enders are more likely to favor marijuana (and
see, H. R. White & Gorman, 2000). The findings presented here are consistent with this.
Frequencies of probationers having ever used cocaine and marijuana are tabulated in Ta-
ble 50 by age tertiles. As indicated, previous use of cocaine was less common across the sample
than was previous use of marijuana: just over half (106=199 = 0:53) of the sample had previously
used cocaine whereas about four-fifths (159=199 = 0:80) had previously used marijuana. By age,
older o enders were more likely to have used cocaine than their younger counterparts: 20=72 = 0:28
of those ages 19 to 26 as compared to 55=66 = 0:83 of those ages 39 to 62 had used cocaine in the
past. On the other hand, younger o enders were more likely to have used marijuana than their
older counterparts. In fact, nearly all (65=72 = 0:9) of those ages 19 to 26 had used marijuana as
compared to 43=66 = 0:65 of those ages 39 to 62.
The implication is that even though research largely suggests younger o enders will be the
most crime involved and the most likely to perform poorest while supervised, older o enders do
not necessarily pose less of a supervision challenge. Here, older o enders?especially those with
175
more extensive drug use histories?have higher chances of continued cocaine use while super-
vised. This is particulary important given research suggesting, compared to marijuana, cocaine is
a less manageable substance and one that is more likely to be implicated in predatory crime (see,
D. McBride, 1981; Wright & Klee, 2001; H. R. White & Gorman, 2000). This caution is relevant
when considering the two other substances examined in this research as well.
Positive screens for opiates and phencyclidine were less common than those for cocaine
and marijuana, and, for these substances, the single most important characteristic influencing pop-
ulation rates is having ever used either. Having done so is associated with large increases in the
expected rates of positive screens: expected rates of positive opiate screens, for instance, are about
60 times higher; those of phencyclidine are about 15 times higher.
This is a clear example of behavioral stability with at least one rather obvious treatment
implication: if programs designed specifically for opiate or phencyclidine users exist, selecting
participants from among those having ever used either substance is an obvious tool because, un-
like the case for cocaine and marijuana, there is a considerable amount of persistence in the use of
both. Aside from the obvious importance of abstinence in determining success while supervised,
these findings also raise a public safety concern. Phencyclidine, like cocaine, is less manageable,
tends largely to evoke erratic behaviors, and is associated with a host of physiological problems
in comparison to other substances (see, D. McBride, 1981; Wright & Klee, 2001; H. R. White
& Gorman, 2000). And while opiates are similar to marijuana in that neither tend to evoke un-
controllable behaviors and neither are associated with long term physiological problems, unlike
marijuana, available research suggests opiate users are more criminally active than users of other
types of drugs.
Thus, when making risk-anticipated security and treatment decisions bearing on substance
use, the agency should consider sentence length, age, the number of substances ever used, and
176
whether the o ender has ever used either opiates or phencyclidine. These characteristics are asso-
ciated with subsequent use while supervised; moreover, there is reason to believe those o enders
that fail to abstain will also continue to be criminally active throughout supervision.
Two additional legal criteria examined here included arrests and subsequent convictions
during and shortly after the supervision period. The SFS-98 is an important predictor of convic-
tions during both periods, with lower values linked with increased likelihoods.126 Additionally,
the recommended sentence well-predicts convictions during the supervision period, where those
recommended to either incarceration or split-sentence as opposed to probation stand a greater
chance of being convicted. The importance of the recommended sentence shrinks in comparison
to the influence of the rate of public-order related arrests within the BG of residence once the
supervision period terminates.
The sole purpose of the SFS-98 is to assess the probability that a federal inmate will reo end
once released. As described in Appendix B, this measure is based on a wide range of attributes
including the number of past adjudications and convictions, the number of past commitments, age
at the time of the instant o ense, the time between this and the most recent commitment, and
criminal justice status at the time of arrest. It has been extensively validated and demonstrates
reasonable predictive accuracy (Ho man, 1994; Ho man & Beck, 1974, 1976, 1984), and, not
surprisingly, proves important in predicting these two legal criteria.
Sentence recommendations are made by PSI authors after having researched the back-
ground of o enders. Recommendations are ultimately based on the DC sentencing guidelines
with qualifications stemming from the extensive information in the PSI. Unlike the SFS-98, the
guidelines emphasize only a limited set of attributes: the severity of the o ense on conviction,127
126The SFS-98 is scored such that lower values are associated with higher recidivism risk.
127The severity of the o ense on conviction is based on a hierarchical scale developed by the DC Sentencing and
Criminal Code Revision Commission.
177
Table 51
Component comparisons of the SFS-98 and the recommended sentence.
SFS-98 Recommended sentence
1 Number of past juvenile adjudications and crimi-
nal convictions
Number and severity of past adjudications and
convictions
2 Time between this and the most recent commit-
ment
Time between this and the most recent sentence
3 NA Severity of the o ense on conviction
4 Criminal justice status at the time of arrest NA
5 Number of previous commitments exceeding 30
days
NA
6 Age at the time of the instant o ense NA
the number and severity of prior convictions and adjudications, and the length of time between
the imposition or the expiration of the last sentence and the commission of the instant o ense.
Those with lengthier criminal histories, those with more serious convictable o enses, and those
with shorter durations between this and the most recent sentence are typically recommended to
incarceration or split-sentence.
The components of the two measures are shown in Table 51 with comparable items tab-
ulated together and incomparable items indicated with NAs. As indicated, while there is some
overlap in that both address past adjudications and convictions and the time between this and the
most recent sentence or commitment, there are some important di erences between them. For
example, while the SFS-98 examines the number of past adjudications and convictions, the rec-
ommended sentence examines their severity as well. The recommended sentence also emphasizes
the nature of the convictable o ense. And, unlike the recommended sentence, the SFS-98 exam-
ines the criminal justice status at the time of the arrest, the number of previous commitments, and
age at the time of the instant o ense.
178
While the SFS-98 is a good predictor of behaviors during both periods, the recommended
sentence is important in predicting behaviors only during the supervision period. The severity of
the o ense on conviction is the central contribution of the recommended sentence over and above
what is represented in the SFS-98. Thus, this key aspect of criminal history should be considered
along with the number of past adjudications and convictions, the number of past commitments,
age at the time of the instant o ense, the time between this and the most recent commitment, and
criminal justice status at the time of arrest when making immediate, risk-anticipated security and
treatment decisions especially as they turn on assessing o enders for risk of continued o ending
while supervised.
Once supervision periods terminate, the added information stemming from the recom-
mended sentence is no longer important. Instead, rates of public order arrests become more pre-
dictive: o enders living in areas with higher rates of public order related arrests have increased
chances of being arrested and subsequently convicted shortly after their sentences are complete.
This might be an indication that o enders supervised in these communities are the main
contributors of the high arrest rates. However, as rates of arrests were measured prior to most of
the supervision activity this explanation is not convincing. At least not with these data. A more
convincing explanation points to the likely direct and indirect influences of neighborhood crime
on the residents. Akers (1998), for instance, suggests the level of crime in an area might influence
subsequent NSP directly by providing learning contexts conducive to criminal behaviors. High
crime rates may alternatively stigmatize communities, degrade the moral standing of its residents,
and potentially undermine strong social ties (Stark, 1987). Why this characteristics is important
only after sentences are complete is unclear. There may be some level of protection against these
forces while sentences are active that decays upon completion. Nevertheless, when making risk-
anticipated security and treatment decisions, especially as they pertain to risk of post-sentence
179
recidivism, agencies should consider the impact of local, particularly public-order related arrests,
along with the number of past adjudications and convictions, the number of past commitments,
age at the time of the instant o ense, the time between this and the most recent commitment,
and criminal justice status at the time of arrest, as these characteristics influence the chances of
sustained law-abiding behavior upon release.
This study also examined characteristics associated with higher expected rates of both su-
pervision and drug-related violations. There are some redundancies among these two criteria and
those that have been discussed so far. For instance, one of the most common types of supervision-
related violations was failing to obey all laws, which is also captured in the first arrest-conviction
criterion albeit at a lower burden of proof. Other common supervision-related violations include
failing to report as directed and failing to carry out CSO instructions, and thus supervision-related
violations is a more sensitive one in that it captures not only law obedience but additional non-
compliance as well. Also, along with the use of illegal substances, drug-related violations capture
failing to comply with drug or alcohol treatment or surveillance programs or procedures; visiting
places where illegal substances are bought, sold, or consumed; and purchasing, possessing, or
selling illegal substances. So, while there is some overlap among this criteria and those bearing
on substance use, this one captures a broader range of drug-related behaviors.
The most important characteristics influencing how often o enders will violate supervision-
related conditions are the impression of risk on the supervising CSO and the SFS-98 score. One is
a subjective measure, the other is objective; neither alone is su cient in predicting which o enders
will violate supervision-related conditions at higher rates. As both assess recidivism risk, it is no
surprise?given the nature of these types of violations?to find either to be important.
When conducting the initial screening assessment the supervising CSO makes a judgment
about the level of risk the given o ender represents vis-`a-vis the other o enders in the CSO?s
180
caseload. Those for whom the CSO has a high impression of recidivism risk are expected to violate
supervision related conditions at a rate roughly twice as high as those for whom the CSO has a
lower impression. This may be an indication of the power of professional judgment: CSOs are in a
perfect position to gauge which of their o enders will likely fail to comply with release conditions.
However, because assessments are made by the supervising CSO one has to wonder whether this
indeed represents accurate prediction or, instead, self-fulfillment. It is possible, for example, that
CSOs make inaccurate predictions, but, because they expect certain o enders to violate conditions
more often, they increase monitoring levels for these o enders and this heightened exposure alone
increases the chances of finding violations and of subsequently making good on their predictions.
This self-fulfilling argument, however, is less compelling given the additional importance
of the number of past adjudications and convictions, the number of past commitments, age at the
time of the instant o ense, the time between this and the most recent commitment, and criminal
justice status at the time of arrest as captured in the SFS-98. Here, each additional point on the
SFS-98 is associated with a roughly 10% decrease in the expected number of supervision-related
violations. Thus, when making immediate, risk-anticipated security and treatment decisions?
especially as they bear on risks of supervision-related violations, the agency should carefully
examine the components of the SFS-98 and the impression of risk on the supervising CSO, as
both are associated with changes in expected supervision-related violation rates.
The most important characteristics influencing the rate of drug-related violations are the
number of previous weapons-related convictions and the rates of arrests for drug-related crimes
within the block-group. The findings from this study suggest the number of weapons-related
convictions is inversely related to rates of drug-related violations. This relationship is unexpected,
as the weight of the evidence suggests instead that having more convictions will be associated
with higher likelihoods of NSP (J. Austin & Litsky, 1982; J. Austin et al., 1989; M. Jones, 1995;
181
Mayzer et al., 2004; Petersilia, 1985a; Quinsey et al., 1998; Schwaner, 1997; Silver & Chow-
Martin, 2002; Whitehead, 1991). Most of the empirical studies, however, have focused on either
technical and criminal behaviors or ultimate failure while supervised; little has been written about
this specific, drug-related aspect of supervision performance.
Contrary to expectations, having more weapons-related convictions is associated with low-
ered expected rates of drug-related violations. Compared to those without any previous weapons-
related convictions, for example, the expected number of drug-related violations will be about
34% lower among those with at least one. This could indicate either that those with more extensive
weapons-related criminal histories are less involved in drug-related behaviors or that while super-
vised in the community this group of o enders eschews non-complying behaviors?especially
those related to illegal substances?as much as possible. While it seems obvious that those with
lengthy weapons related convictions will pose a risk to the public, here, at least as it concerns this
specific, drug-related aspect of NSP, there is a lower associated risk of NSP. Also, that weapons-
related convictions were unrelated to other aspects of NSP, including the substance use criteria,
suggests that this characteristic is related more so to the broader range of drug-related behaviors
captured by this criterion than simply contemporary use.
The relationship between local rates of drug-related arrests and the rate of drug-related
violations is also inconsistent with expectations. The findings presented here suggest as local
rates of arrests for drug-related crimes increase, the expected number of drug-related violations
decreases. This finding is not theoretically supported.
Both Akers (1998) and Stark (1987), for instance, provide plausible explanations for why
rates of drug-related arrests might increase the rate of drug-related violations. Indeed, here we
find just the opposite e ect. Its absence when considering the use of substances as captured by
the previously discussed substance use criteria suggests the e ect of local drug-related crime op-
182
erates primarily on the additional characteristics captured in this criterion (viz., failing to comply
with drug or alcohol treatment or surveillance programs or procedures; visiting places where il-
legal substances are bought, sold, or consumed; and purchasing, possessing, or selling illegal
substances). The inverse e ect found here suggests o enders living within high drug-crime neigh-
borhoods might be more aware of the systemic, drug-related problems within their neighborhoods
and thus more likely to comply with drug-related conditions while supervised. This is speculative
and demands further study.
To the extent choices facing the agency depend on expected rates of drug-related violations,
the agency will benefit from carefully examining past weapons-related convictions and local rates
of drug-related arrests, as both are associated with subsequent rates of violations. However, given
that neither of these predictors behave in a way that is consistent with existing theory and research,
these conclusions should be approached cautiously until they can be examined in greater detail.
This study also examined characteristics associated with whether and, if so, how soon of-
fenders will terminate their sentences unsuccessfully. The most important characteristics influenc-
ing whether probationers will terminate their sentence unsuccessfully include the recommended
sentence, the proportion of positive screens for cocaine, and the level of supervision. The most
important characteristics influencing how fast they will do so include the proportion of positive
screens for marijuana, having a history of drug-related convictions, and the expected length of
supervision.
Probationers most likely to terminate sentences unsuccessfully are those whom the PSI
author recommended a sentence of either incarceration or split-sentence, those having a proportion
of positive screens for cocaine of 0.02 or more, and those supervised at higher levels of intensity.
Sentence recommendations are made by PSI authors after having researched the back-
ground of o enders. Recommendations are ultimately based on the DC sentencing guidelines
183
with qualifications stemming from the extensive information in the PSI. Unlike the SFS-98, the
guidelines emphasize only a limited set of attributes: the severity of the o ense on conviction,128
the number and severity of prior convictions and adjudications, and the length of time between
the imposition or the expiration of the last sentence and the commission of the instant o ense.
Those with lengthier criminal histories, those with more serious convictable o enses, and those
with shorter durations between this and the most recent sentence are typically recommended to
incarceration or split-sentence. These data suggest o enders recommended to incarceration or
split sentence are likely to perform more poorly than those recommended to probation; their odds
of terminating unsuccessfully are roughly 4.8 times larger.
The proportion of positive screens for cocaine is also associated with whether o enders
ultimately terminate sentences unsuccessfully. The expected odds of failure are roughly 2.8 times
larger among those having failed more than 0.02 of their cocaine screens. This suggests that even
though a small proportion of positives are tolerated, continued use of cocaine is clear reason for
supervision failure. From an earlier model we know o enders having the highest expected rates
of cocaine positives are older particularly drug-involved o enders. This group thus poses higher
chances of terminating unsuccessfully.
Supervision level essentially captures intensity: those supervised at higher levels are seen
more often by supervising CSOs and face stricter penalties in response to noncomplying behaviors.
Here, those supervised at higher levels are more likely to fail while supervised. Compared to those
supervised at the maximum level, for instance, the expected odds of terminating unsuccessfully
are roughly 90% smaller among those supervised at the minimum level, 46% smaller among those
supervised at the medium level, and roughly 73% larger among those supervised at the intensive
level.
128The severity of the o ense on conviction is based on a hierarchical scale developed by the DC Sentencing and
Criminal Code Revision Commission.
184
Although the CSO can increase or decrease supervision level based on professional judge-
ment, it is primarily determined by a linear combination of the RNS items (see, Appendix A).
On one hand this suggests, even though the individual RNS items are themselves irrelevant, the
algorithm determining supervision level successfully identifies o enders posing the highest risk
of NSP?at least, that is, with respect to ultimately terminating unsuccessfully. On the other, it
could suggest either that closely monitoring o enders has a detrimental e ect on performance or
simply that monitoring o enders more closely increases the chances that CSOs will find grounds
for termination. These data do not give any insight into which of these plausible explanations
is driving the variations. They do, however, suggest a linear combination of the RNS items is
a good predictor of whether probationers in the population will ultimately fail or succeed while
supervised.
From this model we know which characteristics influence whether probationers will fail
while supervised. These are, namely, the recommended sentence, the proportion of positive
screens for cocaine, and the level of supervision. The next model examined characteristics in-
fluencing how fast they will do so and, in the end, the most important characteristics include the
proportion of positive screens for marijuana, having a history of drug-related convictions, and the
expected length of supervision.
Failure rates will be highest among persistent marijuana users and those having a higher
number of previous drug-related convictions. These two findings are anticipated by the research.
That lower hazards are expected among those with shorter supervision sentences is simply too
closely tied with supervision time to be meaningful. I therefore exclude this finding in the discus-
sion that follows.
The hazard rate increases by roughly 25% for each 0.10 increase in the proportion of pos-
itive screens for marijuana. This suggests drug abstinence is key in successfully completing sen-
185
tences. Misconduct?especially persistent use of marijuana?is grounds for terminating supervi-
sion, but these data also suggest a modicum of leniency. CSOs may, for instance, initially respond
to continued use by increasing sanctions or introducing directed, drug-related interventions. When
these e orts fail they are nevertheless forced to terminate sentences earlier than anticipated.
It is important to note that these data tell us only about the status at the end of the sentence
and not the details on how this status came about. A sentence may be revoked or terminated
unsuccessfully due to specific, unmeasured processes like contemporaneous o ending or, more
generally, lack of adjustment to supervision requirements. It is likely that early failures result not
only from continued marijuana use but also from behaviors related to continued use. For instance,
MacKenzie and her colleagues found that o ending is highest during months when probationers
were actively using illegal substances (MacKenzie et al., 1998; MacKenzie & Li, 2002). Similarly,
J. Austin and Litsky (1982) found those probationers with higher levels of drug use were more
likely to eventually abscond.
Findings from a previous model suggest those most likely to use marijuana while supervised
are younger o enders. Age was, in fact, the most important predictor of this criterion. Thus, those
at highest risk of early failure while supervised are most likely to be younger o enders. The other
influential predictor of the rate of failure also suggests a linkage between illegal substances and
NSP, but here focus shifts from contemporaneous to past use. Probationers with two or more
drug-related convictions are expected to fail at a rate that is roughly 22% higher than those with
only one and 79% higher than those with none.
As noted throughout the review, past behaviors are among the strongest predictors of future
behaviors: both a lengthy history of convictions and patterned criminality is related to subsequent
NSP (J. Austin & Litsky, 1982; J. Austin et al., 1989; Bartell & Thomas, 1977; Cunni , 1986;
Irish, 1989; M. Jones, 1995; Mayzer et al., 2004; McGaha et al., 1987; Petersilia et al., 1985;
186
Petersilia, 1985a; Quinsey et al., 1998; Schwaner, 1997; Silver & Chow-Martin, 2002; Sims
& Jones, 1997; Vito, 1987; Whitehead, 1991). Also, because the number of convictions for
drug-related o enses is more important than, say, the total number of convictions for all types of
o enses or o enses of a di erent nature, illegal substances are obviously implicated in the rate
failure. Research linking a history of substance use (e.g., Benedict & Hu -Corzine, 1997; Silver &
Chow-Martin, 2002; D. A. Smith & Polsenberg, 1992) and an underlying pattern of abuse (Baird
et al., 1984; Gray et al., 2001; Harer, 1994; Schmidt & Witte, 1988) with later o ending and NSP
anticipates this.
Ultimately, we have a fairly strong indication that both contemporaneous and past drug use
has a particularly negative impact on failure rates. This could be because drug use is a particular
form of misconduct that directly triggers early termination, but how closely the use of drugs and
the ultimate termination are tied is unclear. It could also mean that drug use is an indication of
a larger process including contemporaneous o ending or, more generally, a lack of adjustment to
supervision requirements, and it is these characteristics that speed o enders into early termination.
Out of well over 200 potential predictors included in this study, only a fraction were both
related to NSP and likely to replicate in the population. These characteristics are tabulated in
Table 49; a summary of the findings follows.
The most important characteristics of o enders likely to continue using illegal substances
while supervised are past substance use, age, and length of supervision. Those having used a
greater number of illegal substances and those expecting to serve longer community sentences
are more likely than their counterparts to fail drug testing events. And while younger o enders
are at greater risk of using marijuana while supervised, older o enders?especially those with
more extensive substance use histories?are at greater risk of using cocaine. Also, those o enders
187
having ever used either opiates or phencyclidine are much more likely than their counterparts to
use these substances while supervised.
The most important characteristics of o enders likely to be criminally active while super-
vised include the SFS-98 score, the recommended sentence, and the rate of public-order related
arrests within the BG of residence. Those with lower SFS-98 values are more likely to be arrested
and subsequently convicted. During the supervision period, those o enders that were recom-
mended to either incarceration or split-sentence as opposed to probation are more likely to be
convicted and, afterward, those living in BGs with higher rates of public-order related arrests are
more likely to be convicted.
The most important characteristics of o enders unlikely to comply with release conditions
include the impression of risk on the supervising CSO, the SFS-98 score, the number of previous
weapons-related convictions, and the rates of arrests for drug-related crimes within the BG. Those
o enders for whom CSOs have high impressions of recidivism risk and those having lower SFS-98
scores are expected to violate supervision related conditions at a higher rate than their counterparts.
Those o enders having fewer previous weapons-related convictions and those living in BGs with
lower rates of drug-related arrests are expected to violate drug-related conditions at higher rates
than their counterparts.
The most important characteristics of whether and, if so, how soon probationers will termi-
nate sentences unsuccessfully include the recommended sentence, the expected length of supervi-
sion, the level of supervision, the proportion of positive screens for both cocaine and marijuana,
and the number of previous drug-related convictions. Those o enders recommended to either
incarceration or split-sentence, those failing 0.02 or more of their screens for cocaine, and those
supervised at higher levels of intensity are more likely to fail while supervised; those continuing
188
to use marijuana and those having a higher number of previous drug-related convictions will fail
at the fastest rates.
There is reason for caution when considering the findings from this study. I discuss a few
of these next.
This study examined characteristics increasing the risk that supervised o enders will en-
gage in NSP, which was defined here by two domains of criteria: legal and supervision-specific.
Legal criteria included substance use and continued o ending and supervision-specific criteria
included condition violations and termination modes. These criteria were chosen from common
criteria in criminological risk-assessments, but perhaps a better approach would have been to first
survey the supervision sta regarding their focal supervision concerns and then to construct crite-
ria based on their responses. It is thus possible the criteria studied here do not fully represent their
concerns.
Despite having a wide collection of potential predictors there are many that have been omit-
ted. Those characteristics that have not been included, and there are many, were omitted primarily
because of their unrealistic availability for supervision line sta when making immediate, risk-
anticipated decisions. This includes immediate situational, physiological, or cognitive factors, as
well as unforseen sociological, psychological, and contextual characteristics. Also, among those
measures that were included, there is reason to expect any lack of association with NSP is a conse-
quence of how they were measured and scaled and not of a genuine lack of relationship. I discuss
some of these possibilities next.
Available research suggests educational performance, commitment to educational goals,
and educational attainment are inversely associated with criminal justice involvement (Agnew &
White, 1992; Agnew, 1985, 1989, 1992; Beck et al., 1993; Brezina, 1996; Farnworth & Leiber,
1989; Farrington, 1997; S. D. Gottfredson & Gottfredson, 1979; Harrison & Gfroerer, 1992;
189
Hindelang, 1973; Hirschi, 1969; Horney et al., 1995; Jarjoura, 1996; Kruttschnitt et al., 1986;
Quinsey et al., 1998; Sampson & Laub, 1993; Thornberry et al., 1985a, 1985b; Ward & Tittle,
1994; L. Zhang & Messner, 1996). These characteristics are decisive predictors of NSP (Gray
et al., 2001; Harer, 1994; Irish, 1989; Landis et al., 1969; Mayzer et al., 2004; Morgan, 1993;
Rhodes, 1986; Roundtree et al., 1984; Silver & Chow-Martin, 2002; Sims & Jones, 1997). How-
ever, none of the measures used here (viz., the highest grade attempted and completed and whether
the o ender earned a GED) were associated with the production of NSP. While this suggests these
characteristics are unimportant, it could also mean they do not adequately operationalize the con-
cept. In particular, the measures used here capture attainment exclusively. It is likely, then, that if
measures capturing how well o enders performed while in they were in school and how strongly
they agreed with educational goals were included they would appear important in predicting NSP.
Residential instability is also related to o ending (Kasarda & Janowitz, 1974; Sampson,
1988) and NSP (J. Austin & Litsky, 1982; Mayzer et al., 2004; F. P. Williams III et al., 2000).
None of the measures used here, however, namely the number of residential changes within the
previous year and whether the o ender resides with relatives, were important.129 The lack of
relationship among these measures and NSP could suggest residential stability is unrelated to
supervision performance. On the other hand, it might indicate that these measures do not fully
capture the most important aspects. One of these aspects is likely to be the embeddedness of the
o ender within the community (i.e., Sampson, 1988). Had measures tapping both interpersonal
ties with neighbors and willingness to participate in community activities and organizations been
included, it is likely they would be important predictors of the production of NSP in the population.
Unemployment and job instability are both linked with elevated o ending (Farrington,
1986, 1997; Thornberry & Christenson, 1984; Thornberry & Farnworth, 1982), recidivism (S. D. Got-
129Other measures were intended, including the type of living quarters and whether the o ender rents or owns, but
were excluded due to high levels of missingness.
190
tfredson & Gottfredson, 1979; D. M. Gottfredson et al., 1978; Uggen, 2000), and NSP (J. Austin
& Litsky, 1982; Gray et al., 2001; Harer, 1994; Irish, 1989; M. Jones, 1995; Landis et al., 1969;
MacKenzie & Li, 2002; Mayzer et al., 2004; Morgan, 1993; Silver & Chow-Martin, 2002; Sims
& Jones, 1997; F. P. Williams III et al., 2000); however, none of the employment characteristics
included here were important predictors of NSP. While this suggests these characteristics may be
unimportant in the production of NSP, it could also mean that they fail to fully operationalize the
concept.
For instance, the measures used here capture the number of employment changes within the
past year, whether there are any jobs within the past year lasting less than 30 days, the number of
months at the current job, and employment status at time of instant o ense. These characteristics
may not be observed for long enough periods to fully represent stability patterns. Thus, had all
employment behaviors since, say, age 18, rather than only within the past year or month, been
included it is likely these characteristics would prove important predictors of NSP. In addition,
Sampson and Laub (1993) argue that employment stability is only partially important and that the
quality of employment is also important. It is likely then that had measures of wages, job quality,
and job satisfaction been included, they too would lead to di erent findings.
Early childhood experiences are expected to influence NSP (Cloward & Ohlin, 1960; Far-
rington, 2000, 1997; M. R. Gottfredson & Hirschi, 1990; Gove et al., 1979; Hirschi, 1983, 1994;
Kolvin et al., 1988; McCord, 1979; Merton, 1957; Robins, 1978; Sampson & Laub, 1993; Stark,
1987; Thornberry et al., 1999; Wells & Rankin, 1991; Wilson & Hernnstein, 1985); however,
none of the measures included here (viz., parent marital status at birth, whether parents were in-
volved in their upbringing, whether they have sustained contact with their parents, and whether
they experienced neglect or abuse as a child) were important predictors. While this could mean
191
such early experiences do not have a genuine relationship with later NSP, it might also suggest
these measures do not adequately capture the most important processes.
For instance, while the literature highlights lifelong consequences of lowered economic
conditions and the structure of illegitimate opportunities, neither of these characteristics were
available. It is likely, had they been included, they would have been important predictors. Simi-
larly, while parent and sibling criminality are also linked with later o ending, these characteristics
were excluded due to high levels of missingness. Had they been included, it is likely they too
would have been important and the findings would be di erent.
Immediate family characteristics, such as marital quality and family involvement, are also
associated with o ending and NSP (Clarke et al., 1988; Farrington, 1989; M. R. Gottfredson &
Hirschi, 1990; Harer, 1994; Horney et al., 1995; Landis et al., 1969; Laub et al., 1998; MacKenzie
& Li, 2002; Morgan, 1993; Petersilia, 1985a; Sampson & Laub, 1990, 1993; Sampson et al., 1997;
Sims & Jones, 1997; Warr, 1993, 1998; West, 1982; Wilson & Hernnstein, 1985). However, none
of the characteristics included here were important predictors of NSP. This is likely due to an
incomplete operational definition.
In particular, though marital status was included, evidence suggests it is not merely being
married but also the quality of the relationship that matters (Laub et al., 1998; Sampson & Laub,
1993). The absence of this aspect of marriage may explain why it was unrelated to NSP here
and why it has appeared unimportant in other studies (Gray et al., 2001; Mayzer et al., 2004;
Roundtree et al., 1984). The same holds for family involvement. While the number of children
and the number of children under age 18 the o ender lives with were included, these measures
likely fail in capturing the quality of the relationships. Had measures been included that captured,
for example, having an agreeable marital climate or the extent of parent-child involvement, it is
likely they would have been important in predicting NSP and would have led to di erent findings.
192
Research also suggests the less healthy and more injury prone are involved more so crim-
inally than their counterparts (Farrington, 1995; National Commission on Correctional Health
Care, 2002; Shepherd et al., 2002). However, none of the measures capturing physical health
were important here. These were, namely, the number of disabilities, injuries, gunshot wounds,
and stabbing wounds. While this suggests these aspects of physical health are unrelated to the
production of NSP, it does not suggest these attributes are unrelated to NSP. In particular, other
aspects, such as chronic illness, may be important.130
Research also suggests a linkage between mental health and o ending (see, Bland et al.,
1998; Ditton, 1999; Estro et al., 1985; Johnston & O?Malley, 1986; Lin et al., 1996; Link,
Andrews, & Cullen, 1992; Link, Cullen, & Wozniak, 1992; MacKenzie et al., 1998; Monahan,
1992; Teplin et al., 1996; Teplin, 1990) and thus it was expected that mental health characteristics
would be implicated in the production of NSP. However, none of the measures included here
(viz., whether the o ender had a current or history of mental disorder, whether the o ender was
diagnosed with mental illness, whether the o ender is currently prescribed or taking psychotropic
medications, the number of previous mental health hospitalizations, and whether the o ender ever
attempted suicide) were associated with the production of NSP in the population. This does not,
however, suggest mental health is unrelated. Instead, it might reflect inadequate or incomplete
measurement.
Had broader and more refined measures of, for instance, psychological distress, stress, de-
pression, problems with emotions and items capturing defined classes of mental illness, such as
depression, bipolar disorder, schizophrenia, or obsessive-compulsive disorder, it is likely these
characteristics would indeed demonstrate a relationship with NSP resulting in di erent findings.
130Note, several additional health-related characteristics were included as part of the PSI instrument. Due to extreme
missingness, however, they were excluded from the analysis.
193
Despite research linking environmental conditions with individual behaviors (e.g., Garo-
falo, 1987; Greenberg & Schneider, 1994; Moreno , Sampson, & Raudenbush, 2001; Sampson
& Lauritsen, 1994; W. R. Smith, Frazee, & Davison, 2000; Stark, 1987), except for arrest rates
for public order and drug-related o enses, none of the contextual characteristics were important
predictors of NSP.
This included sociodemographic and economic predictors drawn from the 2000 Census,
such as the proportion of the block-group that was non-white, foreign born, Hispanic or Latino,
living in a di erent house in 1995, having less than a high school diploma or equivalency, income
below the poverty line, and unemployed, as well as the proportion of female-headed households
and households receiving public assistance, the proportion of renter occupied housing units, the
ratio of children to adults, and the population per square meter of block-group land area. None of
these items were associated with NSP. These items were also reduced to a 3-factor solution rep-
resenting concentrated disadvantage, immigrant concentration, and residential stability; however,
these items were also unimportant in the production of NSP in the population. While this might
suggest such sociodemographic characteristics are unimportant in the production of NSP, this lack
of relationship could instead reflect that the block-group is a poor neighborhood-level proxy. It
was chosen for this study as it is the closest approximation to the neighborhood for which Census
data are publicly available. Had measures been constructed for areas that more closely align with
neighborhood boundaries, it is possible the findings reported here would di er.
Research also suggests there are intrinsic, crime generating characteristics in areas wherein
residential units are coexistent with or adjacent to commercial areas (see, Kelling & Coles, 1996;
Reiss, 1986; Sampson & Raudenbush, 2001; Skogan, 1992; Stark, 1987; Taylor, 1999; Wilson &
Kelling, 1982). However, the item constructed here representing commercialization?the density
194
of all licensees131 per 1,000 residential housing units within U.S. Census Bureau (Census) block-
group (BG)?was not an important predictor of NSP. This suggests local commercialization levels
are unimportant in the production of NSP. On the other hand, this lack of relationship could be a
consequence of the aggregational measure. There may very well be multiple dimensions of com-
mercialization each with di ering e ects on the production of NSP. Aggregating these dimensions
into a single measure may thus mask any di erential e ects. Had disaggregated measures of vari-
ous types of commercialization been included, it is possible they would change the findings.
There is also evidence suggesting the concentration of alcohol-related businesses is associ-
ated with criminal o ending (see, Cochran et al., 1998; Costanza et al., 2001; D. M. Gorman et
al., 2001; Gyimah-Brempong, 2001; R. Lipton & Gruenewald, 2002; Scribner et al., 1999, 1995;
Sherman et al., 1989; Speer et al., 1998; Stitt & Giacopassi, 1992). The measures included here
captured densities of retailers licensed for the on- and o -premises sale of beer, wine, and spirits
and of beer and wine only. None, however, were associated with NSP. While this suggests concen-
trations of alcohol-related businesses are unimportant in the production of NSP, it could suggest
that the appropriate processes are not fully captured.
For instance, while these measures capture the densities of alcohol-related businesses, they
tell nothing about the situational and sociocultural environments in which they are located (see,
Bushman, 1997; Fagan, 1990; Gustafson, 1994; Linsky et al., 1986; MacAndrew & Edgerton,
1969; Parker & Auerhahn, 1998; Parker & Rebhun, 1995; Reiss & Roth, 1993; Roncek & Maier,
1991; Skog, 1985; Wiseman, 1991). Had measures of, for instance, perceptions with respect to
acceptable alcohol-related behaviors or the levels of uncharacteristic behaviors in and around these
retailers been included, it is possible the findings would di er.
131Excluding retail alcohol outlets.
195
Ample research indicates criminal activities cluster in place and time (e.g., Boggs, 1965;
Bursik & Grasmick, 1993; Carter & Hill, 1978; Gould, 1969; Reiss, 1986; Reppeto, 1974; Sher-
man et al., 1989; Wilson & Hernnstein, 1985) and there is reason to expect such levels will influ-
ence the production of NSP. Much of this relates to direct influences expected from exposure to
criminal activities as well as indirect influences expected through stigmatization, increased crime
tolerances, and weakened formal and informal controls (Akers, 1998; Kelling & Coles, 1996;
D. A. Smith, 1986; Stark, 1987; Taylor, 1999; Wilson & Kelling, 1982). The measures used here
included rates of violent, property, drug- and alcohol-related, public-order, and otherwise unclassi-
fied arrests within the block-group, yet, only arrest rates for public order and drug-related o enses
were related to NSP. Further, the relationship between drug-related o enses and NSP was unantic-
ipated. While this suggests local crime may be unimportant or misunderstood in the production of
NSP, it might also be that the measures included here do not fully capture the important processes.
As Akers (1998) points out, one of the key processes likely involves the sociocultural tra-
ditions and control systems. Objective crime measures may not fully account for this, as areas
with high levels of crime may not necessarily provide learning environments conducive to crime.
Likewise, Stark (1987) points to the stigmatizing e ect of high crime on communities and how
this degrades the moral standing of its residents and potentially undermines strong social ties.
While local crime levels are obviously implicated, areas with high levels of crime may not be
characterized necessarily by weakened social ties.
Thus, despite the lack of support for some obvious theoretical linkages there is reason for
caution when interpreting the findings from this study, as much of it could be a consequence of
the measures chosen and how they were constructed and not a reflection of a genuine lack of
relationship. There are two additional considerations when interpreting these findings. Both bear
196
on the generalizability of these findings to broader supervision populations. The first relates to the
sample size; the second, the sampling frame. I discuss both next.
There were only 200 o enders in the present sample and, due to extreme missingness, one
was removed. This left a sample comparatively smaller than some suggestions (e.g., P. R. Jones,
1996). The reason for the small sample was to accommodate data extracted from the PSIs. Much
of the analytic work here relied on these rich data, but, as I described previously, they did not
readily lend themselves to analysis. It is possible had a larger sample been developed the findings
would di er. However, incorporating more than a handful of PSIs was unapproachable for this
study.
Another reason for caution bears on the generalizability of these findings to broader su-
pervision populations. Because they represented a large proportion of the CSOSA caseload, only
Black male probationers were included. Thus, characteristics found here to associate with height-
ened risks of NSP may not be universal across other dimensions such as non-Blacks, females, and
those sentenced to parole supervision. It made sense to concentrate on the bulk of the agency
caseload for this study. Expanding it to include these smaller caseloads is an obvious direction for
future research. I highlight some additional suggestions next.
The present study should be seen as part, in fact, the beginnings, of a comprehensive risk
assessment strategy. An instrument characterizing o enders at high risk of NSP was developed.
The next step is putting the instrument into practice and assessing how well it performs.
The implications of the study are clear. Past behaviors were once future behaviors and so
it makes good sense to draw our lessons therein. And we need not look too closely: with just a
handful of characteristics bearing on past criminality and substance use we know where to focus
immediate attention when making risk-anticipated security and treatment decisions. The highest
and lowest risks can be identified with this small set of characteristics, and this speeds appropriate
197
control and therapeutic responses and, ultimately, increases agency accountability, public safety,
and o ender reintegration.
The findings presented here, however, were based largely on data either not currently avail-
able or in a format not lending itself to analytics. Presentence Investigation reports (PSIs) are
prime examples. They provide an unmatched picture of personal and social aspects and the most
comprehensive description available of both the triggering event and the criminal and supervision
histories, they provide insight into the production of NSP in the population, but are, for the most
part, unreachable.
An instrument was developed specifically for this study for extracting the most common
PSI features that have also been shown to vary with NSP. The findings from the analysis suggest
that among these data information required to calculate the SFS-98 and information bearing on
substance use and criminal histories are particularly important in predicting NSP. So, to put this
instrument into practice an obvious, necessary first step involves automating the recovery of at
least those data from the PSIs required to calculate the SFS-98 and those bearing on substance use
and criminal histories. Ideally, more formality would be imposed on the content and structure of
the PSIs themselves thus leading to greater consistency across reports. Also, future reports would
be collected and stored in a format conformable to warehousing and this to future researches.
When considering the findings presented here it is important to heed the lessons learned over
the years with respect to the superiority of statistical assessments vis-`a-vis clinical assessments.
Statistical assessments like this one are characterized by objectivity, formality, and empirical rigor,
and, compared to clinical assessments?such as line o cer risk assessments?when validated
and implemented properly they are more accurate and the instruments on which they are based
demonstrate higher reliabilities (Brennan, 1987; Cocozza & Steadman, 1976; D. M. Gottfredson,
1987; Grove & Meehl, 1996; Grove et al., 2000; P. R. Jones, 1996; Lidz et al., 1993; Meehl, 1954;
198
Menzies et al., 1994; Monahan & Steadman, 1994; Monahan et al., 2001; Monahan, 1981; Morris
& Miller, 1985; Mossman, 1994; Quinsey et al., 1998; Quinsey & Maguire, 1986; Rice & Harris,
1995; Sawyer, 1966; van Voorhis & Brown, 1997).
Line o cer risk assessments are likely ripe with misjudgment. Although there is much to
say in favor of such assessments?especially in light of well-developed professional judgement
and creative insight?without clairvoyant knowledge they are doomed to fail. Historically, for
instance, such assessments are accurate no more than one-third of the time (Monahan, 1981). It is
unfair to expect otherwise.
Line o cers may lack an appreciation of basic statistical properties, such as regression ef-
fects, and basic risk assessment concepts, such as true and false positives and negatives, base rates,
and selection ratios. An understanding of both is critical in making precise judgments. More im-
portantly, because human capacity to deal with a large number of potentially highly intercorrelated
variables is limited, they may be unable to apprehend the complexities giving rise the variations
in the data. Even with a handful of predictors with which to wrestle, often they are inappropri-
ately weighted. This stems often from excessive, ungrounded emphases on illusory correlates or
putative causes of criminality and results in assessments that magnify less important factors and
fail to emphasize more important ones. And because they focus exclusively on individualized
assessments, they largely overlook competing influences such as those stemming from social or
environmental forces.
Objectivity and formality are chief among strengths setting statistical assessments apart
from their less powerful counterparts. It is thus imperative statistical assessments remain free from
the spoils undermining clinical assessments. These include, for instance, complete replacement of
model-implied decisions as well as downward or upward departures from model implications.
199
In his ?disturbing little book,? Meehl (1954) asserts risk-based decisions can be made in
either a clinical or statistical fashion?there being no hybrids?and that decision-makers should
use that method which results in the most accurate predictions. The weight of the evidence from
over 80 years of study clearly identifies statistical approaches as the best tools. Further, the view
here is that these assessments should be seen as tools to supplant rather than merely support line
o cer assessments. Such a view is mildly shared among some (e.g., Monahan et al., 2001) and
vigorously defended by others (Quinsey et al., 1998; Webster, Harris, Rice, Cormier, & Quinsey,
1994). In the end, statistical risk assessments should be viewed not as mere guidance, but rather
gospel.
200
Appendix A
Risk-needs Screener
The Risk-needs Screener (RNS) is an instrument designed by the Community Supervi-
sion Services (CSS) and Community Justice Programs (CJP) o ces of the Court Services and
O ender Supervision Agency for the District of Columbia (CSOSA) to partition caseloads into
supervision level groups (viz., minimum, medium, maximum, or intensive) as a function pri-
marily of NSP risk. In 20 items it captures and weights demographic and social characteristics;
criminal and supervision histories; dependencies, disorders, and disabilities; and Community Su-
pervision O cer (CSO) perceptions of risk posed by the screened o ender. These items, their
levels, and the associated item-level weights are described in this Appendix as is the construction
of rnsTotalS corei, a recommended supervision level classifier based on these weights.
Demographic and social characteristics include (a) age as of the day of the screening
rnsAgei, where values are categorized into
rnsAgei =
8>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>><
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
:
Age 35; 0;
30 < Age 34; 2;
25 < Age 29; 3;
21 < Age 24; 4; and
Age 20; 5;
based on the age of the ith o ender at the time of the assessment; (b) the highest education level
completed rnsEdui, recorded as either 10th grade or below, 11th grade, high school diploma or
201
equivalency, or some college. Values are mapped as
rnsEdui =
8>>
>>>>
>>>>
>>>>
>>>>
>>><
>>>>
>>>>
>>>>
>>>>
>>>>
>:
Some College; 0;
HS/GED; 1;
11th; 3; and
10th or below; 5;
(c) residential changes within the past year rnsResi, which is recorded as either currently incar-
cerated in either an institution or Community Correctional Center (CCC), has been released in
the past 6 months or resides at a shelter, 2 or fewer moves or residing in a residential treatment
program, or 3 or more moves. Values reflect
rnsResi =
8>>
>>>>
>>>>
>>>>
><
>>>>
>>>>
>>>>
>>>:
2 or fewer moves in past year; 0;
Currently/Recently incarcerated/shelter; 3; and
3 or more moves in past year; 5;
(d) employment changes within the past year rnsEmpi, which is recorded as either incarcerated
or shelter if the o ender has been released from incarceration and is currently residing in a CCC,
has been released from incarceration in the past 6 months or is currently incarcerated, 2 or fewer
employment changes, or 3 or more changes. Responses encode
rnsEmpi =
8>>
>>>>
>>>>
>>>>
><
>>>>
>>>>
>>>>
>>>:
2 or fewer jobs; 0;
Currently/Recently incarcerated/shelter; 3; and
3 or more jobs in past year or unemployed; 5;
202
(e) the number and quality of close, supportive, prosocial relationships rnsS S Ni. This is recorded
as either None, Relationships with 1 person, or Relationships with 2 or more and
rnsS S Ni =
8>>
>>>>
>>>>
>>>>
><
>>>>
>>>>
>>>>
>>>:
Relationships with 2 or more; 0;
Relationship with 1 person; 3; and
No Relationships; 5;
and (f) whether the o ender experienced the loss of a significant relationship rnsRLi through, for
instance, divorce, death, or separation within the past 6 months. If so, rnsRLi = 5; Otherwise, 0.
Information regarding criminal and supervision histories includes (a) the age at first arrest
rnsAFi recorded as either 15 or younger, 16?17, 18?25, or over 26. Levels are assigned to rnsAFi
as
rnsAFi =
8>>
>>>>
>>>>
>>>>
>>>>
>>><
>>>>
>>>>
>>>>
>>>>
>>>>
>:
Age at first arrest 26; 0;
18 < Age at first arrest 25; 1;
16 < Age at first arrest 17; 3; and
Age at first arrest 15; 5;
(b) the frequency of arrests in the past year rnsFAi, which is recorded as either none, 1, 2?4, or 5
or more. These values are scored such that
rnsFAi =
8>>
>>>>
>>>>
>>>>
>>>>
>>><
>>>>
>>>>
>>>>
>>>>
>>>>
>:
None; 0;
1; 2;
2 Arrests 4; 3; and
Arrests 5; 5;
203
(c) the number of prior convictions rnsPCi, which is recorded as either none, 1?5, or 6 or more.
The number of prior convictions are mapped to rnsPCi as
rnsPCi =
8>>
>>>>
>>>>
>>>>
><
>>>>
>>>>
>>>>
>>>:
None; 0;
1 Convictions 5; 3; and
Convictions 5; 5
(d) the number of previous violent convictions rnsPVi, which is recorded as either none, 1, or 2 or
more. Values are mapped such that
rnsPVi =
8>>
>>>>
>>>>
>>>>
><
>>>>
>>>>
>>>>
>>>:
None; 0;
1; 3; and
O enses 2; 5;
(e) the initiating o ense rnsOOi, which is either violent, drug-related, or non-violent and mapped
as
rnsOOi =
8>>
>>>>
>>>>
>>>>
><
>>>>
>>>>
>>>>
>>>:
Non-violent; 0;
Drug-related; 3; and
Violent; 5;
(f) the number of prior adult arrests rnsNPAi, which is recorded as either 2 or less, 3?4, 5, or 6 or
more. Values of rnsNPAi are recorded as
rnsNPAi =
8>>
>>>>
>>>>
>>>>
><
>>>>
>>>>
>>>>
>>>:
Arrests 2; 0;
3 Arrests 4; 3; and
Arrests 5; 5;
204
(g) the number of prior supervision failures rnsPS Fi, which is recorded as either none, 1?2, or 3
or more with values mapped such that
rnsPS Fi =
8>>
>>>>
>>>>
>>>>
><
>>>>
>>>>
>>>>
>>>:
none; 0;
1 Failures 2; 3; and
Failures 3; 5;
Items bearing on dependencies, disorders, and disabilities include (a) whether there is evi-
dence of current substance abuse rnsCS Ai such as positive drug tests within the past 60 days or
admission of current substance abuse. If so, rnsCS Ai = 5; Otherwise, 0; (b) a history of substance
abuse rnsHS Ai as indicated by previous positive drug tests, admissions of substance abuse, or
sentences to treatment. If so, rnsHS Ai = 5; Otherwise, 0; (c) a current mental disorder rnsCMDi,
which is indicated by the use of psychotropic medications, admissions of psychological problems,
current treatment for mental disorders, or admissions of homicidal or suicidal thoughts. If so,
rnsCMDi = 5; Otherwise, 0; (d) a history of mental disorder rnsHMDi as indicated by previous
prescriptions for psychotropic medications, previous care by a mental health professional for a
period exceeding three months, previous hospitalizations for mental illness, or previous diagnoses
for mental disorders. If so, rnsHMDi = 5; Otherwise, 0; and (e) physical disabilities or illnesses
potentially disruptive for the current sentence rnsPDi, including hyperglycemia, HIV, migraine
headaches, or chronic pain. If so, rnsPDi = 5; Otherwise, 0.
And, finally, CSO perceptions include (a) the level of cooperation at the time of the in-
terview rnsLOCi. This refers to an o ender?s willingness to cooperate with the CSO and the
conditions of supervision. If an o ender appears willing and is able to comply (e.g., is on time for
appointments, appears for drug testing, or participates in treatment) the CSO scores the response
as fully cooperative. If an o ender is willing and able to comply with some but not all conditions
of supervision the CSO scores the response as noncooperative. On the other hand if an o ender is
205
unwilling or unable to comply with these conditions, the CSO scores the response as restrained.
Values reflect
rnsLOCi =
8>>
>>>>
>>>>
>>>>
><
>>>>
>>>>
>>>>
>>>:
Fully Cooperative; 0;
Restrained; 3; and
Non-cooperative; 5;
and (b) the CSOs impression of risk rnsImpi. This is a subjective measure of risk the o ender
represents to the CSO with respect to recidivism. The CSO scores this response as either low,
medium, or high based on an evaluation of other o enders currently supervised.
rnsImpi =
8>>
>>>>
>>>>
>>>>
><
>>>>
>>>>
>>>>
>>>:
Low; 0;
Medium; 3; and
High; 5.
Item-level weight summations return y i which is then discretized as yi using the cutpoints
 0 = 25,  1 = 50, and  2 = 75 as
rnsTotalS corei =
8>>
>>>>
>>>>
>>>>
>>>>
>>><
>>>>
>>>>
>>>>
>>>>
>>>>
>:
0 y i   0; Minimum;
 0 < y i   1; Medium;
 1 < y i   2; Maximum; and
 2 < y i  100; Intensive.
206
Appendix B
SFS-98
The derivation formula for the Salient Factor Score (SFS) 98 as outlined by the United
States Parole Commission (USPC) and the steps taken to construct this measure from the PSIs
is described in this Appendix. The SFS 98 is an additive scale comprising 6 items (see, USPC,
2003), which are
Item A. Prior convictions/adjudications (adult/juvenile). Scores reflect None = 3; One = 2;
Two or three = 1; and Four or more = 0.
Item B. Prior commitment(s) of more than 30 days (adult/juvenile). Scores reflect None = 2;
One or two = 1; and Three or more = 0.
Item C. Age at current o ense/prior commitments. Scores reflect, for those ages 26 and older at
the time of their current o ense,
Item C =
8>>
>>>>
>>>>
>>>>
><
>>>>
>>>>
>>>>
>>>:
3; Three or fewer prior commitments;
2; Four prior commitments; and
1; Five or more prior commitments.
For those ages 22 to 25,
Item C =
8>>
>>>>
>>>>
>>>>
><
>>>>
>>>>
>>>>
>>>:
2; Three or fewer prior commitments;
1; Four prior commitments; and
0; Five or more prior commitments.
For those ages 20 to 21,
Item C =
8>>
>>>>
>><
>>>>
>>>>
:
1; Three or fewer prior commitments; and
0; Four prior commitments.
207
And, for those ages 19 and younger, Item C = 0 for any number of prior commitments:
Item D. Recent commitment free period (three years). Scores reflect there was no prior commit-
ment of more than 30 days (adult or juvenile) or released to the community from last such
commitment in at least 3 years prior to the commencement of the current o ense. If so,
Item D = 1; Otherwise, 0.
Item E. Probation/parole/confinement/escape status violator this time. Scores reflect neither on
probation, parole, confinement, or escape status at the time of the current o ense; nor
committed as a probation, parole, confinement, or escape status violator at this time. If
so, Item E = 1; Otherwise, 0.
Item F. Older o enders. Scores reflect the o ender was ages 41 or older at the commencement
of the current o ense and also that the total score from Items A?E is 9 or less. If so, Item
F = 1; Otherwise, 0.
The value of the SFS 98 is the sum of Items A?F and spans the interval [0;11]. Lower scores
reflect higher recidivism risks.
The SFS 98 was proxied among each i = 1;2;:::;N probationers from items found in the
PSIs. Specifically, S FS 1i is calculated as
S FS 1i = 3 (0 ici < 1) + 2 (1 ici < 2)
+ 1 (2 ici < 4) + 0 (4 ici < +1)
with ici representing the sum of adult convictions psiCrmAdlCnvAlli and juvenile adjudications
psiCrmJuvAd jAlli as an Item A proxy.
208
To proxy Item B, S FS 2i is calculated as
S FS 2i = 3 (0 ci < 1) + 2 (1 ci < 3)
+ 1 (3 ci <1);
with ci representing the sum of the number of juvenile psiCrmJuvConi and adult psiCrmAdlInci
commitments lasting longer than 30 days.
S FS 3i is a proxy for Item C as the rth row and cth column of M, where age at current
o ense ri is encoded into one of four categories as
ri =
8>>
>>>>
>>>>
>>>>
>>>>
>>><
>>>>
>>>>
>>>>
>>>>
>>>>
>:
1; Ages 26 and older;
2; ages 22?25;
3; Ages 20?21; and
4; Ages 19 and under;
the number of commitments ci is encoded into one of three categories as
ci =
8>>
>>>>
>>>>
>>>>
><
>>>>
>>>>
>>>>
>>>:
1; Three or fewer;
2; Four; and
3; Five or more;
and M
4 3
=
3 2 12 1 0
1 0 00 0 0.
S FS 4i proxies Item D as
S FS 4i = 1  (ici  1) (bi rci  1095:75 days) 
where ici represents the sum of the number of juvenile psiCrmJuvConi and adult psiCrmAdlInci
commitments of 30 days or more; bi represents the supervision begin date; and rci represents the
date of release from the most recent commitment psiCrmAdlLstICi.
209
To proxy Item E, S FS 5i is calculated as
S FS 5i = (psiS upS tai = 5)
where criminal justice status at the time of arrest psiS upS tai is coded as 0 = fugitive, 1 = proba-
tion, 2 = parole, 3 = supervised release, 4 = incarcerated, 5 = not under criminal justice sentence.
S FS 6i is calculated as
S FS 6i = (xI  41) 
0B
BBBB
BB@
5X
j=1
S FS ji  9
1C
CCCC
CCA
to proxy Item F, where xi represents the age on the date supervision began psiAgei.
Finally, psiS FS i is calculated as
psiS FS i =
6X
j=1
S FS ji ; for i = 1;2;:::;N
and, like the scale it proxies, takes on values within [0;11].
210
Appendix C
Crime categories
Crime taxonomies can be futile and, at times, especially when blindly adopted, veritably
useless. Multifarious itself, the meaning of crime hinges largely on theoretical, empirical, and
organizational perspectives, goals, and requirements. Any taxonomy is germane only in how it an-
swers particular questions at hand, and as this study aims at developing an instrument for guiding
risk-anticipated security and treatment decisions among the most common o enders supervised
by the CSOSA, its significance here is deeply connected to antecedents and consequences of NSP.
As such, crime is cast along those broad dimensions appearing regularly in the researches and lit-
eratures that were discussed in Chapter 2, which were, namely (a) violent, (b) property, (c) drug-
and alcohol-related, and (d) public order.
Enumerating ad nauseam specific o enses subsumed within these 4 broad dimensions is
excessive, so, instead, representative o enses from each are provided as examples.132 Character-
istic crimes of violence include (a) homicide, (b) rape, (c) robbery, (d) assault, and (e) weapons
o enses. Among property crimes are (a) burglary, (b) theft, (c) arson, and (d) fraud. Those char-
acterizing drug- and alcohol-related crimes include (a) selling or possessing illegal substances and
(b) alcohol-related o enses. And, finally, o enses typical of public order crimes include (a) dis-
orderly conduct, (b) vagrancy, and (c) prostitution.
132A complete listing of specific o enses, the mapping between these and the more comprehensive categories, and
the C code to both extract and classify specific crimes is available upon request.
211
Appendix D
Conditions of supervision
All probationers are subject to the general conditions shown in Table D1; certain probation-
ers are subject to the special conditions shown in Table D2. In both cases, conditions are imposed
by the sentencing body and the CSOSA is charged with ensuring they are obeyed.
Table D1
General probation conditions.
Code Description
GC1 Obey all laws, ordinances and regulations
GC2 Keep all appointments with CSO
GC3 Notify CSO of any change of address within 48 hours and obtain the permission
of Probation O cer if planning to leave the Washington Metropolitan Area for
more than two weeks
GC4 Abstain from the use of hallucinatory or other illegal drugs
GC5 Obtain a job as soon as possible or continue present employment
212
Table D2
Special probation conditions.
Code Description
SC01 Victims of Violent Crime Compensation
SC02 Drug/Alcohol Treatment
SC03 Other Condition
SC04 Drug Testing
SC05 Employment
SC06 Education
SC07 Community Service
SC08 Mental Health
SC09 Restitution
SC10 Stay Away Order
SC11 Anger Management
SC12 Sex O ender Conditions
SC13 Fine
SC14 Halfway House
SC15 Electronic Monitoring/Curfew
SC16 Domestic Violence/Family Violence
SC17 VOTEE
213
Appendix E
Modes of termination
Termination modes take on 1 of 5 categories and are captured in outTrmModi as
outTrmModi =
8>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>><
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
:
1; death;
2; successful;
3; unsuccessful and terminated;
4; revoked; and
5; absconsion:
Aside from absconsion, outTrmModi is derived directly from SMART entries. Entries describe
whether probationers (a) died while supervised, (b) complete their sentences and fulfill imposed
conditions, (c) complete their sentences and fail to fulfill imposed conditions, or (d) fail to com-
plete their sentences and instead have it revoked either outright or followed by incarceration.
Determinations of absconsion, in contrast, rely on several entries in Supervision and Man-
agement Automated Record Tracking (SMART). For clarity, absconders are those o enders that
essentially evade the agency?s scope between the beginning of supervision and the full term
date. Once CSOs become aware of this, through, for example, failure to establish initial con-
tacts, a series of missed appointments, or reports from personal or employer contacts, they no-
tify both the sentencing authority and the Metropolitan Police Department for the District of
Columbia (MPDC). A warrant is issued in most cases. In some of these, the warrant is executed
within the relatively short time it takes to determine the probationer had died, been hospitalized,
been arrested, or had forgotten about scheduled contact appointments. In others, months or years
could pass before the warrant is executed?if it is at all.
Absconders are defined here as those probationers having violated General Condition 2
(see, Table D1), a warrant issued subsequent this violation, and a warrant remaining open for at
214
least 30 days. If so, the triggering event is excluded from the set of violations accumulated to date,
the date of the triggering violation replaces the date of supervision termination outTrmdti, and
outTrmModi is set to outTrmModi = 4 to indicate absconsion.
Probationers having violated General Condition 2 and either not having a warrant issued or
having a warrant issued and expired within 30 days are not considered absconders. These events
are, instead, considered acts of noncompliance.
215
Appendix F
Technical details
Technical details with respect to validation by bootstrapping and recursive partitioning are
included in this appendix. Technical details about bootstrapping are provided in teh first section
and those for recursive partitioning are provided in second.
Bootstrapping
Two of the more common validation techniques include cross-validation and resampling.
Traditionally, replications using new data are compared to original estimates (Farrington & Tar-
ling, 1985b; S. D. Gottfredson & Gottfredson, 1986; Monahan et al., 2001). Recently, though,
more are turning to resampling procedures, like bootstrapping, to estimate validity. Often, instru-
ments are developed using one sample, the construction sample, and prospectively validated using
a di erent sample, the validation sample (P. R. Jones, 1996). Factors associated with events and
outcomes of interest are identified in the construction sample, the persistence of which is exam-
ined in the validation sample (P. R. Jones, 1996). One approach includes dividing samples into
two or more subsamples. The population processes are modeled using the construction sample
then applied to the validation sample(s). The di erence in how well the model performed in the
construction versus the validation samples is considered a measure of shrinkage. This sample frac-
tionation method is typical (S. D. Gottfredson & Gottfredson, 1986), but, while appealing, might
be ill-considered. Fractionation methods are the least desirable (S. D. Gottfredson & Gottfred-
son, 1986). Central drawbacks include their waste of potentially useful information (W. Gardner,
Lidz, Mulvey, & Shaw, 1996) and the consequent reduction in e ective sample size. The observa-
tions held out for later validation could better contribute to the modeling. Because the stability of
model estimates depends largely on the number of cases used in estimation, limiting sample size
leaves fewer cases for model construction, reducing stability, and, in turn, constraining reliability
216
(S. D. Gottfredson & Gottfredson, 1986; P. R. Jones, 1996; Monahan et al., 2001). Thus, while
giving an indication of cross-sample performance, this technique will not necessarily reflect the
expected variability when the model is applied to the population (Farrington & Tarling, 1985b).
A more recent alternative involves estimating the expected variability from multiple, unbi-
ased samples (S. D. Gottfredson & Gottfredson, 1986; Monahan et al., 2001). Bootstrapping is
only one of many methods under the rubric of resampling methods. It is, however, the only method
discussed here. This exclusiveness is reflective only of relevance. With advances in computational
power, similar approaches, such as the jackknife or the delta method, are inferior. Essentially
bootstrapped population parameters are estimated by first repetitively sampling observed data,
with replacement, estimating the parameter on each subsample, then calculating confidence in-
tervals around the statistic by pooling and averaging all of the subsample estimates. Monahan
et al. (2001) recently used this approach when examining n = 939 patients from the MacArthur
Risk Assessment Study. Their criterion was serious violence in the community within 20 weeks
of discharge. So as not to limit the data available for analyses, they bootstrapped parameter esti-
mates. Essentially this entailed constructing 1,000 subsamples from their original data, applying
their model to each subsample, then summarizing the central tendency of these estimates.
To gauge how well they will perform when put into practice, models developed in this
study were validated by bootstrapping using, specifically, random-x, or case, resampling (see,
Fox, 2002). Given the regression of yi on xi1;xi2;:::;xik, R resamples are drawn randomly with
replacement from
z =fyi;xi1;xi2;:::;xikg
for i = 1;2;:::;N of size N. The regression of yi on xi1;xi2;:::;xik is fitted to each R boot-
strap sample. These are then used to estimate what the confidence intervals around coe cients
? 1; ? 2;:::; ? k.
217
Recursive partitioning
Initial, general models to predict each criterion using predictors identified in the review as
likely influences were reduced to binary trees using recursive partitioning analysis (RPA) (Breiman
et al., 1984; Clark & Pregibon, 1992; Therneau & Atkinson, 1997)133 and then pruned back to
account for replacement optimism using an AIC-like pruning scheme (see, Venables & Ripley,
2002; Ciampi et al., 1995).
RPA algorithms typically proceed in two stages. The first involves finding the single most
important predictor with respect to its ability to divide the sample into levels of the criterion. This
process repeats, separately, on the resulting halves of the sample and continues recursively until
subsequent improvements are minimal. Second, a constant model is imposed on resulting par-
titions (Hothorn, Hornik, & Zeileis, 2006; Venables & Ripley, 2002). Historically, such models
have been limited to either classification or regression problems, but recent extensions include
the prediction of both rates and survival probabilities (Ciampi et al., 1995; Therneau & Atkinson,
1997). Splits were chosen based on the Gini index (see, Breiman et al., 1984). This strategy is
attractive for risk assessments. Venables and Ripley (2002) point out RPA can be seen as a form
of variable selection, and this is particularly useful when facing large numbers of predictors?
something RPA algorithms handle well. Because they lack parametric assumptions, RPA algo-
rithms can identify potential interactions and monotonic transformations without concerns over
linearity, independence and normality of the errors, or homoscedasticity.
This does not come cheap, however. RPA algorithms are notorious for overfitting and cap-
italizing on selection bias. Initial trees can and often do over-adapt to data and must be adjusted
for this optimism. This is the so-called cost-complexity pruning Breiman et al. (1984) introduced
133Using the R library RPART (Therneau & Atkinson, 1997).
218
(and see, Ripley, 1996). Here, trees were pruned back to account for replacement optimism based
on an AIC-like pruning scheme (see, Venables & Ripley, 2002; Ciampi et al., 1995).
219
References
Agnew, R. (1985). A revised strain theory of delinquency. Social Forces, 64, 151?167.
Agnew, R. (1989). A longitudinal test of revised strain theory. Journal of Quantitative Criminology, 5,
373?387.
Agnew, R. (1992). Foundation for a general strain theory of crime and delinquency. Criminology, 30,
47?87.
Agnew, R., & White, H. (1992). An empirical test of general strain theory. Criminology, 30, 475?499.
Akers, R. L. (1998). Social learning and social structure: A general theory of crime and deviant behavior.
Boston: Northeastern University Press.
Albonetti, C. A., & Hepburn, J. R. (1997). Probation revocation: A proportional hazards model of the
conditioning e ects of social disadvantage. Social Problems, 44, 124?138.
Allan, E., & Ste ensmeier, D. (1989). Youth, underemployment, and property crime: Di erential e ects
of job availability and job quality on juvenile and young adult arrest rates. American Sociological
Review, 54, 107?123.
Andrews, D., & Bonta, J. (1998). The psychology of criminal conduct (2nd ed.). Cincinnati: Anderson.
Andrews, D. A., Zinger, I., Hoge, R. D., Bonta, J., Gendreau, P., & Cullen, F. T. (1990). Does correctional
treatment work? a clinically relevant and psychologically informed meta-analysis. Criminology, 28,
369?404.
Anglin, M. D., & Speckart, G. (1988). Narcotics use and crime: A multisample, multimethod analysis.
Criminology, 26, 197?233.
Austin, J., & Litsky, P. (1982). Identifying absconders from parole and probation supervision: An evalu-
ation of Nevada?s risk screening instruments (Tech. Rep.). San Francisco: The National Council on
Crime and Delinquency.
Austin, J., Quigley, P., & Cuvelier, S. (1989). Evaluating the impact of Ohio?s community corrections
programs on public safety and costs (Tech. Rep.). San Francisco: The National Council on Crime
and Delinquency.
Austin, R. L., & Allen, M. D. (2000). Racial disparity in arrest rates as an explanation of racial disparity
in commitment to Pennsylvania?s prisons. Journal of Research in Crime and Delinquency, 37, 200?
220.
Bailey, W. C. (1966). Correctional outcome: An evaluation of 100 reports. Journal of Criminal Law,
Criminology and Police Science, 57, 153?160.
Baird, S. C., Storrs, G. M., & Connelly, H. N. (1984). Classification of juveniles in corrections: A model
systems approach (Tech. Rep.). Washington, DC: A. D. Little, Inc.
Bartell, T., & Thomas, W. (1977). Recidivists impacts of di erential sentencing practices for felony
o enders. Criminology, 15, 387?396.
220
Beck, A., Gilliard, D., Greenfeld, L., Harlow, C., Hester, T., Jankowski, L., et al. (1993). Survey of prison
inmates 1991 (NCJ No. 136949). Washington, DC: U.S. Department of Justice, O ce of Justice
Programs, Bureau of Justice Statistics.
Benedict, W. R., & Hu -Corzine, L. (1997). Return to the scene of the punishment: Recidivism of
adult male property o enders on felony probation, 1986?1989. Journal of Research in Crime and
Delinquency, 34, 237?252.
Bernard, T. J., & Ritti, R. R. (1991). The Philadelphia birth cohort and selective incapacitation. Journal of
Research in Crime and Delinquency, 28, 33?54.
Bland, R. C., Newman, S. C., Thompson, A. H., & Dyck, R. J. (1998). Psychiatric disorders in the
population and in prisoners. International Journal of Law and Psychiatry, 21, 273?279.
Blumstein, A., Cohen, J., Roth, J. A., & Visher, C. A. (Eds.). (1986). Criminal careers and ?career
criminals? (Vol. 1). Washington, DC: National Academy Press.
Boggs, S. L. (1965). Urban crime patterns. American Sociological Review, 30, 899?908.
Bonczar, T. P. (1997). Characteristics of adults on probation, 1995 (NCJ No. 164267). Washington, DC:
U.S. Department of Justice, O ce of Justice Programs, Bureau of Justice Statistics.
Bork, M. V. (1995). Five-year review of the United States probation data, 1990?94. Federal Probation, 59,
27?33.
Boyum, D. A., & Kleiman, M. A. R. (1995). Alcohol and other drugs. In J. Q. Wilson & J. Petersilia
(Eds.), Crime and public policy. San Francisco: Institute for Contemporary Studies, ICS Press.
Breiman, L., Friedman, J. H., Olshen, R. A., & Stone, C. J. (1984). Classification and regression trees.
Belmont, California: Wadsworth.
Brennan, T. (1987). Classification for control in jails and prisons. In D. M. Gottfredson & M. Tonry (Eds.),
Prediction and classification: Criminal justice decision-making. Chicago: University of Chicago
Press.
Brezina, T. (1996). Adapting to strain: An examination of delinquent coping mechanisms. Criminology,
34, 39?60.
Brown, S. (2002). The dynamic prediction of criminal recidivism: A three wave prospective study. Ph.D.
dissertation, Queen?s University.
Bruce, A. A., Burgess, E. W., & Harno, A. (1928). A study of the indeterminate sentence and parole in the
state of Illinois. Journal of the American Institute of Criminal Law and Criminology, 19, 284.
Bureau of Justice Statistics. (2000). Correctional populations in the United States, 1997 (NCJ No. 177613).
Washington, DC: U.S. Department of Justice, O ce of Justice Programs, Bureau of Justice Statis-
tics.
Bureau of Justice Statistics. (2002). Correctional populations in the United States, 1998 (NCJ No. 192929).
Washington, DC: U.S. Department of Justice, O ce of Justice Programs, Bureau of Justice Statis-
tics.
221
Bureau of Justice Statistics. (2003). Compendium of federal justice statistics, 2001 (NCJ No. 201627).
Washington, DC: U.S. Department of Justice, O ce of Justice Programs, Bureau of Justice Statis-
tics.
Burgess, E. W. (1928). Factors determining success or failure on parole. In A. A. Bruce (Ed.), The workings
of the indeterminate sentence law in the parole system in Illinois (pp. 203?249). Springfield, IL: State
of Illinois.
Bursik, R. J., Jr. (1986). Ecological stability and the dynamics of delinquency. In A. J. Reiss Jr. & M. Tonry
(Eds.), Communities and crime (Vol. 8). Chicago: University of Chicago Press.
Bursik, R. J., Jr., & Grasmick, H. G. (1993). Neighborhoods and crime: The dimensions of e ective
community control. San Francisco: Jossey-Bass, Inc., Publishers.
Bushman, B. (1997). E ects of alcohol on human aggression: Validity of proposed explanations. In
M. Galanter (Ed.), Alcohol and violence: Epidemiology, neurobiology, psychology, family issues
(Vol. 13, pp. 227?243). New York: Plenum.
Cantor, D., & Land, K. C. (1985). Unemployment and crime rates in post-World War II United States: A
theoretical and empirical analysis. American Sociological Review, 50, 317?332.
Carroll, L., & Jackson, P. (1983). Inequality, opportunity, and crime rates in central cities. Criminology,
21, 178?194.
Carter, R. L., & Hill, K. Q. (1978). Criminal?s and noncriminal?s perceptions of urban crime. Criminology,
16, 353?371.
Chaiken, J. M., Chaiken, M. R., & Rhodes, W. (1994). Predicting violent behavior and classifying violent
o enders. In A. J. Reiss Jr. & J. A. Roth (Eds.), (Vol. 4). Washington, DC: National Academy Press.
Chaiken, M. R., & Chaiken, J. M. (1987, June 1987). Selecting ?career criminals? for priority prosecution
(Tech. Rep.). Cambridge, MA: Abt Associates Inc.
Champion, D. (1994). Measuring o ender risk: A criminal justice sourcebook. Westport: CT: Greenwood
Publishing.
Choldin, H. M., & Roncek, D. W. (1976). Density, population potential and pathology: A block-level
analysis. Public Data Use, 4, 19?30.
Ciampi, A., Negassa, A., & Lou, Z. (1995). Tree-structured prediction for censored survival data and the
Cox model. Journal of Clinical Epidemiology, 48, 675-689.
Clark, L. A., & Pregibon, D. (1992). Tree-based models. In J. M. Chambers & T. J. Hastie (Eds.),
Statistical models in S (chap. 9). Pacific Grove, CA: Wadsworth and Brooks/Cole Advanced Books
and Software.
Clarke, S. H., Lin, Y. W., & Wallace, W. L. (1988). Probationer recidivism in North Carolina: Measure-
ment and classification of risk (Tech. Rep.). Chapel Hill: University of North Carolina Institute of
Government.
222
Clayton, R. R., & Tuchfeld, B. S. (1982). The drug-crime debate: Obstacles to understanding the relation-
ship. Journal of Drug Issues, 12, 153?166.
Clear, T. R., Harris, P. M., & Baird, S. C. (1992). Probationer violations and o cer response. Journal of
Criminal Justice, 20, 1?12.
Cleves, M. A., Gould, W. W., & Gutierrez, R. G. (2002). An introduction to survival analysis using Stata.
College Station, TX: Stata Corporation.
Cloninger, C. R., & Guze, S. B. (1973). Psychiatric disorders and criminal recidivism. Archives of General
Psychiatry, 24, 266?269.
Cloward, R. A., & Ohlin, L. E. (1960). Delinquency and opportunity: A theory of delinquency. New York:
Free Press.
Cochran, J. K., Rowan, K., Blount, W. R., Heide, K., & Sellers, C. S. (1998). Beer joints and badasses: An
aggregate-level assessment of alcohol availability and violent crime. Criminal Justice Policy Review,
9, 465?495.
Cocozza, J., & Steadman, H. (1976). The failure of psychiatric predictions of dangerousness: Clear and
convincing evidence. Rutgers Law Review, 29, 1084?1101.
Cohen, A. K. (1955). Delinquent boys. Glencoe, IL: Free Press.
Cohen, L., & Felson, M. (1979). Social change and crime rate trends: A routine activities approach.
American Sociological Review, 44, 588?607.
Cohen, R. L. (1995). Probation and parole violators in state prison, 1991 (NCJ No. 149076). Washington,
DC: U.S. Department of Justice, O ce of Justice Programs, Bureau of Justice Statistics.
Cordilia, A. (1985). Alcohol and property crime: Exploring the causal nexus. Journal Studies on Alcohol,
46, 161?171.
Costanza, S. E., Bankston, W. B., & Shihadeh, E. (2001). Alcohol availability and violent crime rates: A
spatial analysis. Journal of Crime and Justice, 24, 71?83.
Court Services and O ender Supervision for the District of Columbia. (2004). Interstate compact
(CSOSA Fact Sheet No. 061404). Retrieved May 18, 2006, from http://www.csosa.gov/olipa/
factsheets/interstate compact 061404.pdf. Washington, DC.
Crutchfield, R., Geerkin, M., & Grove, W. (1982). Crime rates and social integratoin: the impact of
metropolitan mobility. Criminology, 20, 467?478.
Cullen, F. T., & Gendreau, P. (2001). From nothing works to what works: Changing professional ideology
in the 21st century. The Prison Journal, 81, 313?338.
Cullen, F. T., & Gilbert, K. E. (1982). Rea rming rehabilitation. Cincinnati, OH: Anderson.
Cunni , M. A. (1986). A sentencing postscript: Felony probationers under supervision in the community
(Tech. Rep.). Washington, DC: National Association of Criminal Justice Planners.
Curtis, L. (1975). Violence, race and culture. Lexington, MA: Heath.
D?Alessio, S., & Stolzenberg, L. (2003). Race and probability of arrest. Social Forces, 81, 1381?1397.
223
Danziger, S., & Wheeler, D. (1975). The economics of crime: Punishment or income redistribution. Review
of Social Economy, 33, 113?131.
Dean, C. W., Brame, B., & Piquero, A. R. (1996). Criminal propensities, discrete groups of o enders, and
persistence in crime. Criminology, 34, 547?574.
DeMaris, A. (2004). Regression with social data : Modeling continuous and limited response variables.
Hoboken, NJ: John Wiley & Sons, Inc.
Dembo, R., Turner, G., Schmeidler, J., Borden, P., Manning, D., & Sue, C. C. (1995). Predictors of
recidivism to a juvenile assessment center. International Journal of the Addictions, 30, 1425?1452.
Dembo, R., Williams, L., Getreu, A., Genung, L., Schmeidler, J., Berry, E., et al. (1991). A longitudinal
study of the relationships among marijuana/hashish use, cocaine use and delinquency in a cohort of
high risk youths. Journal of Drug Issues, 21, 287?312.
Dembo, R., Williams, L., Schmeidler, J., Getreu, A., Berry, E., Gemung, L., et al. (1991). Recidivism
among high risk youths: A 2-year followup of a cohort of juvenile detainees. The International
Journal of the Addictions, 26, 1197?1221.
Ditton, P. M. (1999). Mental health and treatment of inmates and probationers (NCJ No. 174463). Wash-
ington, DC: U.S. Department of Justice, O ce of Justice Programs, Bureau of Justice Statistics.
Dobinson, I., & Ward, P. (1986). Heroin and property crime: An Australian perspective. Journal of Drug
Issues, 16, 249?262.
Dobson, A. (2001). Introduction to generalized linear models (2nd ed.). London: Chapman and Hall/CRC.
Efron, B. (1979). Bootstrap methods: Another look at the jackknife. Annals of Statistics, 7(1), 1?26.
Efron, B. (1982). The jackknife, the bootstrap, and other resampling plans. Philadelphia: SIAM.
Efron, B. (1983). Estimating the error rate of a prediction rule: improvements on cross-validation. Journal
of the American Statistical Association, 78, 316?337.
Efron, B., & Tibshirani, R. (1991). Statistical data analysis in the computer age. Science, 253, 390?395.
Efron, B., & Tibshirani, R. (1993). An introduction to the bootstrap. New York: Chapman and Hall.
Elliott, D. S., Huizinga, D., & Menard, S. (1989). Multiple problem youth: Delinquency, substance use,
and mental health problems. New York: Springer-Verlag.
Estro , T. W., Dackis, C. A., Gold, M. S., & Pottash, A. L. C. (1985). Drug abuse and bipolar disorders.
International Journal of Psychiatry and Medicine, 15, 37-40.
Fagan, J. (1990). Intoxication and aggression. In M. Tonry & J. Q. Wilson (Eds.), Drugs and crime, crime
and justice: A review of the research (Vol. 13, pp. 241?320). Chicago: University of Chicago Press.
Farnworth, M., & Leiber, M. (1989). Strain theory revisited: Economic goals, educational means, and
delinquency. American Sociological Review, 55, 236?279.
Farrington, D. P. (1986). Age and crime. In M. Tonry & N. Morris (Eds.), Crime and justice (Vol. 7).
Chicago: University of Chicago Press.
224
Farrington, D. P. (1989). Early predictors of adolescent aggression and adult violence. Violence and
Victims, 4, 79-100.
Farrington, D. P. (1995). Crime and physical health: Illnesses, injuries, accidents, and o ending in the
Cambridge study. Criminal Behavior and Mental Health, 5, 261?278.
Farrington, D. P. (1997). Predictors, causes, and correlates of male youth violence. In M. Tonry &
M. H. Moore (Eds.), Youth violence, crime and justice (Vol. 24). Chicago: University of Chicago.
Farrington, D. P. (2000). Psychosocial predictors of adult antisocial personality and adult convictions.
Behavioral Sciences and the Law, 18, 605?622.
Farrington, D. P., & Hawkins, J. D. (1991). Predicting participation, early onset and later persistence in
o cially recorded o ending. Criminal Behavior and Mental Health, 1, 1?33.
Farrington, D. P., & Tarling, R. (Eds.). (1985a). Albany: State University of New York.
Farrington, D. P., & Tarling, R. (1985b). Criminological prediction: An introduction. In D. P. Farrington
& R. Tarling (Eds.), Prediction in criminology (pp. 2?33). Albany: State University of New York.
Feeley, M. M., & Simon, J. (1992). The new penology: Notes on the emerging strategy of corrections and
its implications. Criminology, 30, 449?74.
Fischer, C. S. (1978). The spread of violent crime from city to countryside, 1955-1975 (Tech. Rep.).
Berkeley: Institute of Urban and Regional Development, University of California.
Fitzgerald, J. L., & Mulford, H. A. (1993). Alcohol availability, drinking contexts and drinking problems:
the Iowa experience. Journal of studies on alcohol, 54, 320?325.
Fogel, D. (1975). We are the living proof: The justice model for corrections. Cincinnati: Anderson
Publishing Company.
Foucault, M. (1977). Discipline and punish: The birth of the prison (A. Sheridan, Trans.). New York:
Pantheon Books.
Fox, J. (1997). Applied Regression Analysis, Linear Models, and Related Methods. Thousand Oaks, CA:
Sage Publications.
Fox, J. (2002). Bootstrapping regression models: Appendix to an R and S-PLUS companion to applied
regression. Thousand Oaks, CA: Sage Publications.
Freedman, J. L. (1975). Crowding and behavior. San Francisco: W. H. Freeman.
Galle, O. R., Gove, W., & McPherson, J. M. (1972). Population density and pathology: What are the
relations for men? Science, 176, 23?30.
Gardner, L., & Shoemaker, D. J. (1989). Social bonding and delinquency: A comparative analysis. Socio-
logical Quarterly, 30, 481?500.
Gardner, W., Lidz, C., Mulvey, E. P., & Shaw, E. (1996). A comparison of actuarial methods for identifying
repetitively violent patients with mental illness. Law and Human Behavior, 20, 35?48.
Garofalo, J. (1987). Reassessing the lifestyle model of criminal victimization. In M. Gottfredson &
T. Hirschi (Eds.), Positive criminology (pp. 23?42). Newbury Park, NJ: Sage Publications.
225
Gastil, R. D. (1971). Homicide and a regional culture of violence. American Sociological Review, 36,
412?427.
Gendreau, P., Goggin, C., & Paparozzi, M. A. (1996). Principles of e ective assessment for community
corrections. Federal Probation, 60, 64?70.
Gendreau, P., Little, T., & Goggin, C. (1996). A meta-analysis of the predictors of adult o ender recidivism:
What works! Criminology, 34, 575?607.
Gendreau, P., & Ross, R. R. (1979). E ective correctional treatment: Bibliotherapy for cynics. Crime and
Delinquency, 25, 463?89.
Gendreau, P., & Ross, R. R. (1987). Revivification of rehabilitation: Evidence from the 1980s. Justice
Quarterly, 4, 349?407.
Glaser, D. (1987). Classification for risk. In D. M. Gottfredson & M. Tonry (Eds.), Prediction and
classification: Criminal justice decision-making. Chicago: University of Chicago Press.
Glaze, L. E., & Bonczar, T. P. (2006). Probation and parole in the United States, 2005 (NCJ No. 215091).
Washington, DC: U.S. Department of Justice, O ce of Justice Programs, Bureau of Justice Statis-
tics.
Glaze, L. E., & Palla, S. (2004). Probation and parole in the United States, 2003 (NCJ No. 205336). Wash-
ington, DC: U.S. Department of Justice, O ce of Justice Programs, Bureau of Justice Statistics.
Glaze, L. E., & Palla, S. (2005). Probation and parole in the United States, 2004 (NCJ No. 210676). Wash-
ington, DC: U.S. Department of Justice, O ce of Justice Programs, Bureau of Justice Statistics.
Glover, A. J., Nicholson, D. E., Hemmati, T., Bernfeld, G. A., & Quinsey, V. L. (2002). A Comparison
of Predictors of General and Violent Recidivism Among High-Risk Federal O enders. Criminal
Justice and Behavior, 29, 235?249.
Glueck, S., & Glueck, E. (1930). 500 criminal careers. New York, NY: Alfred A. Knopf.
Glueck, S., & Glueck, E. (1950). Unraveling juvenile delinquency. New York, NY: The Commenwealth
Fund.
Golub, A. L., & Johnson, B. D. (1999). Cohort changes in illegal drug use among arrestees in manhattan:
From the heroin injection generation to the blunts generation. Substance Use & Misuse, 34, 1733-
1763.
Gorman, D., Speer, P., Labouvie, E., & Subaiya, A. (1998). Risk of assaultive violence and alcohol
availability in New Jersey. American Journal of Public Health, 88, 97?100.
Gorman, D. M., Speer, P. W., & Gruenewald, P. J. (2001). Spatial dynamics of alcohol availability,
neighborhood structure and violent crime. Journal of Studies on Alcohol, 62, 628?36.
Gottfredson, D. M. (1987). Prediction and classification in the criminal justice system. In D. M. Gottfred-
son & M. Tonry (Eds.), Prediction and classification: Criminal justice decision-making. Chicago:
University of Chicago Press.
Gottfredson, D. M., & Tonry, M. (Eds.). (1987). Chicago: University of Chicago Press.
226
Gottfredson, D. M., Wilkins, L. T., & Ho man, P. B. (1978). Guidelines for parole and sentencing.
Lexington: Lexington Books.
Gottfredson, M. R., & Hirschi, T. (1986). The true value of lambda would appear to be zero: An essay
on career criminals, criminal careers, selective incapacitation, cohort studies, and related topics.
Criminology, 24, 213?234.
Gottfredson, M. R., & Hirschi, T. (1990). A general theory of crime. Stanford: Stanford University Press.
Gottfredson, S. D., & Gottfredson, D. M. (1979). Screening for risk: A comparison of methods (Tech.
Rep.). Washington, DC: U.S. Government Printing O ce.
Gottfredson, S. D., & Gottfredson, D. M. (1985). Screening for risk among parolees: policy, practice, and
method. In D. P. Farrington & R. Tarling (Eds.), Prediction in criminology (pp. 54?77). Albany:
State University of New York.
Gottfredson, S. D., & Gottfredson, D. M. (1986). Accuracy of prediction models. In A. J. Reiss Jr. &
J. A. Roth (Eds.), (Vol. 2). Washington, DC: National Academy Press.
Gottfredson, S. D., & Gottfredson, D. M. (1994). Behavioral prediction and the problem of incapacitation.
Criminology, 32, 441?474.
Gould, L. (1969). The changing structure of property crime in an a uent society. Social Forces, 48, 50?59.
Gove, W. R., Hughes, M. L., & Galle, O. R. (1979). Overcrowding in the home. American Sociological
Review, 44, 55?80.
Gray, M. K., Fields, M., & Maxwell, S. R. (2001). Examining probation violations: Who, what, and when.
Crime and Delinquency, 47, 537?557.
Greenberg, M., & Schneider, D. (1994). Violence in American cities: young black males is the answer, but
what was the question. Social Science and Medicine, 39, 179?187.
Greenfeld, L. A. (1998). Alcohol and crime: An analysis of national data on the prevalence of alcohol
involvement in crime (NCJ No. 168632). Washington, DC: U.S. Department of Justice, O ce of
Justice Programs, Bureau of Justice Statistics. Available from http://www.ojp.usdoj.gov/bjs/
pub/pdf/ac.pdf
Greenfield, T. K., & Weisner, C. (1995). Drinking problems and self-reported criminal behavior, arrests
and convictions: 1990 US alcohol and 1989 county surveys. Addiction, 90, 361?373.
Greenwood, P. W., & Abrahamse, A. F. (1982). Selective incapacitation (Tech. Rep.). Santa Monica, CA:
Rand Corporation.
Grove, W. M., & Meehl, P. E. (1996). Comparative e ciency if informal (subjective, impressionistic) and
formal (mechanical, algorithmic) prediction procedures: The clinical-statistical controversy. Psy-
chology, Public Policy, and Law, 2, 293?323.
Grove, W. M., Zald, D. H., Lebow, B. S., Snitz, B. E., & Nelson, C. (2000). Clinical versus mechanical
prediction: A meta-analysis. Psychological Assessment, 12, 19?30.
227
Gruenewald, P., Ponicki, W., & Holder, H. (1993). The relationship of outlet densities to alcohol consump-
tion: A time series cross-sectional analysis. Alcoholism, Clinical and Experimental Research, 17,
38?47.
Gustafson, R. (1994). Alcohol and aggression. Journal of O ender Rehabilitation, 21, 41?80.
Guze, S., Wolfgram, E., McKinney, J., & Cantwell, D. (1968). Delinquency, social maladjustment and
crime: The role of alcoholism (a study of first-degree relatives of convicted criminals). Disease of
the Nervous System, 219, 238?243.
Gyimah-Brempong, K. (2001). Alcohol availability and crime: Evidence from census tract data. Southern
Economic Journal, 68, 2?21.
Hackney, S. (1969). Southern violence. American Historical Review, 74, 906?925.
Harer, M. (1994). Recidivism among federal prisoners released in 1987 (Tech. Rep.). Washington, DC:
Federal Bureau of Prisons.
Harrell, F. E., Jr., & Lee, K. L. (1985). A comparison of the discrimination of discriminant analysis
and logistic regression under multivariate normality. In P. K. Sen (Ed.), Biostatistics: Statistics
in biomedical, public health, and environmental sciences. the Bernard G. Greenberg volume (pp.
333?343). New York: North-Holland.
Harrell, F. E., Jr., Lee, K. L., & Mark, D. B. (1996). Multivariable prognostic models: Issues in developing
models, evaluating assumptions and adequacy, and measuring and reducing errors. Statistics in
Medicine, 15, 361?387.
Harries, K. D. (1974). The geography of crime and justice. New York: McGraw-Hill.
Harrison, L., & Gfroerer, J. (1992). The intersection of drug use and criminal behavior: Results from the
National Household Survey on Drug Abuse. Crime and Delinquency, 38, 422?443.
Hart, A. (1923). Predicting parole success. Journal of Criminal Law and Criminology, 14, 405?413.
Heitgerd, J., & Bursik, R. J., Jr. (1987). Extracommunity dynamics and the ecology of delinquency.
American Journal of Sociology, 92, 775?787.
Hindelang, J. (1973). Causes of delinquency: A partial replication and extension. Social Problems, 20,
471?487.
Hirschi, T. (1969). Causes of delinquency. Berkley: University of California Press.
Hirschi, T. (1983). Crime and the family. In J. Q. Wilson (Ed.), Crime and public policy, (pp. 53?68). San
Francisco: ICS Press.
Hirschi, T. (1994). Family. In T. Hirschi & M. R. Gottfredson (Eds.), The generality of deviance (pp.
47?69). New Brunswick, NJ: Transaction Publishers.
von Hirsch, A. (1976). Doing justice: The choice of punishments. New York: Hill and Wang.
Ho man, P. B. (1994). Twenty years of operational use of a risk prediction instrument: The united states
parole commissions salient factor score. Journal of Criminal Justice, 22, 477?494.
228
Ho man, P. B., & Beck, J. L. (1974). Parole decision-making: A salient factor score. Journal of Criminal
Justice, 2, 195?206.
Ho man, P. B., & Beck, J. L. (1976). Salient factor score validation?a 1972 release cohort. Journal of
Criminal Justice, 4, 69?76.
Ho man, P. B., & Beck, J. L. (1984). Burnout?age ar release from prison and recidivism. Journal of
Criminal Justice, 12, 317?623.
Holmes, M., & Taggart, W. (1990). A comparative analysis of research methods in criminology and
criminal justice. Justice Quarterly, 7, 421?437.
Horney, J., Osgood, D. W., & Marshall, I. H. (1995). Criminal careers in the short-term: Intra-individual
variability in crime and its relation to local life circumstances. American Sociological Review, 60,
655?73.
Hosmer, D. W., & Lemeshow, S. (1980). A goodness-of-fit test for the multiple logistic regression model.
Communications in Statistics, A10, 1043?1069.
Hosmer, D. W., & Lemeshow, S. (2000). Applied Logistic Regression. New York: John Wiley & Sons,
Inc.
Hothorn, T., Hornik, K., & Zeileis, A. (2006). party: A laboratory for recursive part(y)itioning [Computer
software manual]. Available from http://CRAN.R-project.org (R package version 0.9-4)
Huber, P. J. (1967). The behavior of maximum likelihood estimates under nonstandard conditions. In
In proceedings of the fifth berkeley symposium on mathematical statistics and probability (Vol. 1,
p. 221-223). Berkeley, CA: University of California Press.
Huiras, J., Uggen, C., & McMorris, B. (2000). Career jobs, survival jobs, and employee deviance: A social
investment model of workplace misconduct. The Sociological Quarterly, 41, 245?263.
Inciardi, J. A. (1980). Youth, drug and street crime. In F. R. Scarpitti & S. K. Datesman (Eds.), Drugs and
the youth culture (pp. 175?204). Beverly Hills, CA: Sage Publications.
Irish, J. F. (1989). Probation and recidivism: A study of probation adjustment and its relationship to post-
probation outcome for adult criminal o enders in Nassau County, New York (Tech. Rep.). Nassau
County, NY: Nassau County, NY, Probation Department.
Irwin, J., & Cressey, D. (1962). Thieves, convicts and the inmate culture. Social Problems, 10, 142?155.
Jackson, P. I. (1984). Opportunity and crime: A function of city size. Sociology and Social Research, 68,
172?193.
Janus, M. G. (1985). Selective incapacitation: Have we tried it? Does it work? Journal of Criminal Justice,
13, 117?129.
Jarjoura, R. G. (1996). The conditional e ect of social class on the dropout-delinquency relationship.
Journal of Research in Crime and Delinquency, 33, 232?255.
Johnston, L. D., & O?Malley, P. M. (1986). Why do the nation?s students use drugs and alcohol? Self-
reported reasons from nine national surveys. Journal of Drug Issues, 16, 29?66.
229
Jones, M. (1995). Predictors of success and failure on intensive probation supervision. American Journal
of Criminal Justice, 19, 239?254.
Jones, P. R. (1996). Risk prediction in criminal justice. In A. T. Harland (Ed.), Choosing correctional
options that work: Defining the demand and evaluating the supply (pp. 33?67). Thousand Oaks,
CA: Sage Publications, Inc.
Kahneman, D., & Tversky, A. (1973). On the psychology of prediction. Psychological Review, 80, 237?
251.
Kasarda, J. D., & Janowitz, M. (1974). Community attachment in mass society. American Sociological
Review, 39, 328?339.
Kazdin, A. (1995). Conduct disorders in childhood and adolescence (2nd ed.). Thousand Oaks, CA: Sage
Publications, Inc.
Kelling, G. L., & Coles, C. M. (1996). Fixing broken windows: Restoring order and reducing crime in our
communities. New York: Simon and Schuster.
Kolvin, I., Miller, F., Fleeting, M., & Kolvin, P. (1988). Risk/protective factors for o ending with particular
reference to deprivation. In R. M. (Ed.), Studies of psychosocial risk: the power of longitudinal data
(p. 77-95). Billings, Worcester, Great Britain: Cambridge University press.
Kposowa, A. J., Breault, K. D., & Harrison, B. M. (1995). Reassessing the structural covariates of violent
and property crimes in the USA: A county level analysis. British Journal of Sociology, 46, 79?105.
Kronick, R. F., Lambert, D. E., & Lambert, E. W. (1998). Recidivism among adult parolees: What makes
the di erence? Journal of O ender Rehabilitation, 28, 61?69.
Kruttschnitt, C., Heath, L., & Ward, D. A. (1986). Family violence, television viewing habits, and other
adolescent experiences related to violent criminal behavior. Criminology, 24, 235?267.
Kubrin, C. E., & Stewart, E. A. (2006). Predicting who reo ends: The neglected role of neighborhood
context in recidivism studies. Criminology, 44, 165?195.
LaFree, G., & Drass, K. A. (1996). The e ect of changes in intraracial income inequality and educational
attainment on changes in arrest rates for African Americans and whites, 1957 to 1990. American
Sociological Review, 61, 614?634.
Land, K. C., McCall, P. L., & Cohen, L. E. (1990). Structural covariates of homicide rates: Are there any
invariances across time and social space? American Journal of Sociology, 95, 922?963.
Landis, J., Merger, J., & Wol , C. (1969). Success and failure of adult probationers in California. Crime
and Delinquency, 6, 34?40.
Langan, P. A., & Cunni , M. A. (1992). Recidivism of felons on probation, 1986?89 (NCJ No. 134177).
Washington, DC: U.S. Department of Justice, O ce of Justice Programs, Bureau of Justice Statis-
tics.
Langan, P. A., & Levin, D. J. (2002). Recidivism of prisoners released in 1994 (NCJ No. 193427). Wash-
ington, DC: U.S. Department of Justice, O ce of Justice Programs, Bureau of Justice Statistics.
230
Laub, J. H., & Lauritsen, J. L. (1994). The precursors of criminal o ending across the life course. Federal
Probation, 58, 51?57.
Laub, J. H., Nagin, D. S., & Sampson, R. J. (1998). Trajectories of change in criminal o ending: Good
marriages and the desistance process. American Sociological Review, 63, 225?238.
Lidz, C. W., Mulvey, E. P., & Gardner, W. (1993). The accuracy of predictions of violence to others.
Journal of the American Medical Association, 269, 1007?1111.
Lin, N., Eisen, S. A., Scherrer, J. F., Goldberg, J., True, W. R., Lyons, M. J., et al. (1996). The influ-
ence of familial and non-familial factors on the association between major depression and substance
abuse/dependence in 1874 monozygotic male twin pairs. Drug and Alcohol Dependence, 43, 49?55.
Link, B. G., Andrews, H., & Cullen, F. T. (1992). The violent and illegal behavior of mental patients
reconsidered. American Sociological Review, 57, 275?292.
Link, B. G., Cullen, F., & Wozniak, J. (1992). Violent and illegal behavior of current and former mental
patients compared to community controls. American Sociological Review, 57, 275?292.
Linnet, K. (1989). Assessing diagnostic tests by a strictly proper scoring rule. Statistics in Medicine, 8,
609-618.
Linsky, A., Colby, J., Jr., & Straus, M. (1986). Drinking norms and alcohol-related problems in the United
States. Journal of Studies in Alcohol, 47, 384?393.
Lipsey, M. W., Cordray, D. S., & Berger, D. E. (1981). Evaluation of a juvenile diversion program: Using
multiple lines of evidence. Evaluation Review, 5, 283?306.
Lipton, D., Martinson, R., & Wilks, J. (1975). The e ectiveness of correctional treatment: A survey of
treatment evaluation studies. New York: Praeger.
Lipton, R., & Gruenewald, P. (2002). The spatial dynamics of violence and alcohol outlets. Journal of
Studies on Alcohol, 63, 187?195.
Loftin, C., & Hill, R. H. (1974). Regional subculture and homicide: An examination of the Gastil-Hackney
thesis. American Sociological Review, 39, 714?724.
Loftin, C., & Parker, R. N. (1985). An error-in -variable model of the e ect of poverty on urban homicide
rates. Criminology, 23, 269?285.
MacAndrew, C., & Edgerton, R. (1969). Drunken comportment: A social explanation. New York: Aldine.
MacKenzie, D. L. (1991). The parole performance of o enders released from shock incarceration (boot
camp prisons): A survival time analysis. Journal of Quantitative Criminology, 7, 213?236.
MacKenzie, D. L. (2000). Evidence-based corrections: Identifying what works. Crime and Delinquency,
46, 457?471.
MacKenzie, D. L., Browning, K., Priu, H., Skroban, S., & Smith, D. (1998). Probationer compliance with
conditions of probation (NCJ No. 142462). Washington, DC: U.S. Department of Justice, O ce of
Justice Programs, National Institute of Justice.
231
MacKenzie, D. L., & Li, S. D. (2002). The impact of formal and informal social controls on the criminal
activities of probationers. Journal of Research in Crime and Delinquency, 39, 243?276.
MacKenzie, D. L., Shaw, J. W., & Souryal, C. (1992). Characteristics associated with successful adjustment
to supervision: A comparison of parolees, probationers, shock participants, and shock dropouts.
Criminal Justice and Behavior, 19, 437?454.
MacKenzie, D. L., & Souryal, C. (1994). Multisite evaluation of shock incarceration (NCJ No. 142462).
Washington, DC: U.S. Department of Justice, O ce of Justice Programs, National Institute of Jus-
tice.
Maltz, M. D. (1984). Recidivism. Orlando: Academic Press.
Martinson, R. (1974). What works? Questions and answers about prison reform. The Public Interest, 35,
22-54.
Matza, D. (1964). Delinquency and drift. New York: Wiley.
May, S., & Hosmer, D. W. (1998). A simplified method for calculating an overall goodness-of-fit test for
the Cox proportional hazards model. Lifetime Data Analysis, 4, 109?120.
Mayzer, R., Gray, M. K., & Maxwell, S. R. (2004). Probation absconders: A unique risk group? Journal
of Criminal Justice, 32, 137?150.
McBride, D. (1981). Drugs and violence. In J. Inciardi (Ed.), The drug- crime connection (pp. 105?124).
Beverly Hills: Sage Publication.
McBride, D. C., & McCoy, C. B. (1981). Crime and drug-using behavior. Criminology, 19, 281?302.
McCord, J. (1979). Some child-rearing antecedents of criminal behavior in adult men. Journal of Person-
ality and Social Psychology, 37, 1477?1486.
McCullagh, P., & Nelder, J. A. (1989). Generalized linear models. London: Chapman and Hall.
McGaha, J., Fichter, M., & Hirschburg, P. (1987). Felony probation: A reexamination of public risk.
American Journal of Criminal Justice, 11, 1?9.
McGlothin, W. H. (1979). Drugs and crime. (NCJ). Washington, DC: National Institute on Drug Abuse.
McReynolds, V. (1987). Variables related to predicting absconding probationers. Ph.D. dissertation,
University of Texas at Austin.
Meehl, P. E. (1954). Clinical versus statistical prediction: A theoretical analysis and review of the evidence.
Minneapolis: University of Minnesota Press.
Melton, G. B., Petrila, J., Poythress, N. G., & Slobogin, C. (1997). Psychological evaluations for the
courts. New York: Guilford Press.
Menzies, R., Webster, C. D., McMain, S., Staley, S., & Scaglione, R. (1994). The dimensions of dan-
gerousness revisited: Assessing forensic predictions about violence. Law and Human Behavior, 18,
1?28.
Merton, R. K. (1957). Social theory and social structure. New York: Glencoe: IL.
232
Messner, S. F. (1983). Regional and racial e ects on the urban homicide rate: The subculture of violence
revisited. American Journal of Sociology, 88, 997?1007.
Messner, S. F., & Tardi , K. (1986). Economic inequality and levels of homicide: An analysis of urban
neighborhood. Criminology, 24, 297?317.
Miczek, K. A., DeBold, J. F., Haney, M., Tidey, J., Vivian, J., & Weerts, E. M. (1994). Alcohol, drugs
of abuse, aggression, and violence. In A. J. Reiss Jr. & J. A. Roth (Eds.), Understanding and
preventing violence: Panel on the understanding and control of violent behavior (Vol. 3, pp. 377?
570). Washington, DC: National Academy Press.
Minor, K. I., Wells, J. B., & Sims, C. (2003). Recidivism among federal probationers?predicting sentence
violations. Federal Probation, 67, 31?36.
Monahan, J. (1981). Predicting violent behavior. Beverly Hills, CA: Sage Publications, Inc.
Monahan, J. (1992). Mental disorder and violent behavior: perceptions and evidence. American Psycholo-
gist, 47, 511?521.
Monahan, J., Steadman, H., Silver, E., Appelbaum, P. S., Robbins, P. C., Mulvey, E. P., et al. (2001).
Rethinking risk assessment: The MacArthur study of mental disorder and violence. New York:
Oxford University Press.
Monahan, J., & Steadman, H. J. (1994). Violence and mental disorder: Developments in risk assessment.
Chicago: University of Chicago Press.
Moreno , J., Sampson, R. J., & Raudenbush, S. (2001). Neighborhood inequality, collective e cacy, and
the spatial dynamics of urban violence. Criminology, 39, 517-560.
Morgan, K. D. (1993). Factors influencing probation outcome: A review of the literature. Federal Proba-
tion, 57, 23?29.
Morris, N., & Miller, M. (1985). Predictions of dangerousness. In M. Tonry & N. Morris (Eds.), Crime
and justice: An annual review of research (Vol. 6). Chicago: University of Chicago Press.
Mossman, D. (1994). Assessing predictions of violence: Being accurate about accuracy. Journal of
Consulting and Clinical Psychology, 62, 783?792.
Mumola, C. J. (1999). Substance abuse and treatment, state and federal prisoners, 1997 (NCJ No. 172871).
Washington, DC: U.S. Department of Justice, O ce of Justice Programs, Bureau of Justice Statis-
tics.
Mumola, C. J., & Bonczar, T. P. (1998). Substance abuse and treatment of adults on probation, 1995 (NCJ
No. 166611). Washington, DC: U.S. Department of Justice, O ce of Justice Programs, Bureau of
Justice Statistics.
National Commission on Correctional Health Care. (2002). The health status of soon-to-be-released in-
mates: A report to congress (Vol. 1; Tech. Rep.). Washington, DC: National Commission on Cor-
rectional Health Care.
233
Newcomb, M. D., & Bentler, P. M. (1986). Drug use, educational aspirations, and work force involvement:
The transition from adolescence to young adulthood. American Journal of Community Psychology,
14, 303?321.
Nurco, D. N. (1979). Ideological aspects of drug abuse (NCJ). Washington, DC: National Institute on
Drug Abuse.
Osgood, D. W., Johnston, L. D., O?Malley, P. M., & Bachman, J. G. (1989). Time trends and age trends in
arrests and self-reported illegal behavior. Criminology, 27, 389?417.
Palmer, J., & Carlson, P. (1976). Problems with the use of regression analysis in prediction studies. Journal
of Research in Crime and Delinquency, 13, 64?81.
Palmer, T. (1975). Martinson revised. Journal of Research in Crime and Delinquency, 12, 133?152.
Parker, R. N. (1993). Alcohol and theories of homicides. In F. Adler & W. Laufer (Eds.), Advances in
criminological theory (Vol. 4, pp. 113?142). New Brunswick, NJ: Transaction Publishers.
Parker, R. N., & Auerhahn, K. (1998). Alcohol, drugs, and violence. American Sociological Review, 24,
297?311.
Parker, R. N., & Rebhun, L. (1995). Alcohol and homicide: A deadly combination of two american
traditions. Albany: SUNY Press.
Parzen, M., & Lipsitz, S. R. (1999). A global goodness-of-fit statistic for Cox regression models. Biomet-
rics, 55, 580?584.
Patterson, E. B. (1991). Poverty, income inequality, and community crime rates. Criminology, 29, 755-776.
Peduzzi, P., Concato, J., Kemper, E., Holford, T., & Feinstein, A. (1996). A simulation study of the
number of events per variable in logistic regression analysis. Journal of Clinical Epidemiology, 49,
1373-1379.
Petersilia, J. (1985a). Community supervision: Trends and critical issues. Crime and Delinquency, 31,
339?347.
Petersilia, J. (1985b). Probation and felony o enders. Federal Probation, 49, 4?9.
Petersilia, J. (1998). A decade of experimenting with intermediate sanctions: What have we learned?
Federal Probation, 62, 3?9.
Petersilia, J., Turner, S., Kahan, J., & Peterson, J. (1985). Executive summary of Rand?s study, ?granting
felons probation public risks and alternatives?. Crime and Delinquency, 31, 379?392.
Piquero, A. R., & Buka, S. L. (2002). Linking juvenile and adult patterns of criminal activity in the
providence cohort of the national collaborative perinatal project. Journal of Criminal Justice, 30,
259?272.
Piquero, A. R., & Chung, H. (2001). On the relationships between gender, early onset, and the seriousness
of o ending. Journal of Criminal Justice, 29, 189?206.
Pogue, T. F. (1975). E ect of police expenditures on crime rates: Some evidence. Public Finance Quarterly,
3, 14?44.
234
Quinsey, V. L., Coleman, G., Jones, B., & Altrows, I. (1997). Proximal antecedents of eloping and
reo ending among supervised mentally disordered o enders. Journal of Interpersonal Violence, 12,
794?813.
Quinsey, V. L., Harris, G. T., Rice, M. E., & Cormier, C. A. (1998). Violent o enders: Appraising and
managing risk. Washington, D.C.: The American Psychological Association.
Quinsey, V. L., & Maguire, A. (1986). Maximum security psychiatric patents: Actuarial and clinical
prediction of dangerousness. Journal of Interpersonal Violence, 1, 143?71.
Rand, A. (1986). Mobility triangles. In R. M. Figlio, S. Hakim, & G. F. Rengart (Eds.), Metropolitan crime
patterns (pp. 117?126). Monsey, NY: Criminal Justice Press.
Reiss, A. J., Jr. (1986). Why are communities important in understanding crime. In A. J. Reiss Jr. &
M. Tonry (Eds.), Communities and crime (pp. 1?33). Chicago: University of Chicago Press.
Reiss, A. J., Jr., & Roth, J. A. (Eds.). (1986). Criminal careers and ?career criminals? (Vol. 2). Washington,
DC: National Academy Press.
Reiss, A. J., Jr., & Roth, J. A. (1993). Understanding and preventing violence: Panel on the understanding
and control of violent behavior (Vol. 1). Washington, DC: National Academy Press.
Reppeto, T. A. (1974). Residential crime. Cambridge: Ballinger.
Rhodes, W. (1986). A survival model with dependent competing events and right-hand side censoring:
Probation and parole as illustration. Journal of Quantitative Criminology, 2, 113?137.
Rice, M. E., & Harris, G. T. (1995). Violent recidivism: Assessing predictive validity. Journal of Consulting
and Clinical Psychology, 63, 737?48.
Ripley, B. D. (1996). Pattern recognition and nueral networks. Cambridge: Cambridge University Press.
Robins, L. (1978). Sturdy childhood predictors of adult antisocial behavior: Replications from longitudinal
studies. Psychological Medicine, 8, 611?622.
Robinson, J., & Smith, G. (1971). The e ectiveness of correctional programs. Crime and Delinquency, 17,
67?80.
Rogers, W. H. (1993). Regression standard errors in clustered samples. Stata Technical Bulletin, 13,
19?23.
Roncek, D. W., & Maier, P. A. (1991). Bars, blocks, and crimes revisited: Linking the theory of routine
activities to the empiricism of ?hot spots?. Criminology, 29, 725?753.
Room, R., & Collins, G. (1983). Alcohol and disinhibition: Nature and meaning of the link (Research
Monograph No. 12). Washington, DC: National Institute on Alcohol Abuse and Alcoholism.
Rosenfeld, R. (1986). Urban crime rates: E ects of inequality, welfare dependency, region, and race. In
J. Byrne & R. J. Sampson (Eds.), The social ecology of crime (p. 116-130). London: Springer-
Verlag.
Roundtree, G., Edwards, D., & Parker, J. (1984). A study of the personal characteristics of probationers as
related to recidivism. Journal of O ender Counseling, Services, and Rehabilitation, 8, 53?61.
235
Sampson, R. J. (1986). Crime in cities: The e ects of formal and informal social control. In A. J. Reiss Jr.
& M. Tonry (Eds.), Communities and crime (Vol. 8, pp. 271?311). Chicago: University of Chicago
Press.
Sampson, R. J. (1988). Local friendship ties and community attachment in mass society: A multilevel
systemic model. American Sociological Review, 53, 766?779.
Sampson, R. J., & Laub, J. H. (1990). Crime and deviance in the life course. American Sociological
Review, 55, 609?627.
Sampson, R. J., & Laub, J. H. (1993). Crime in the making: Pathways and turning points through life.
Cambridge, MA: Harvard University Press.
Sampson, R. J., & Lauritsen, J. L. (1994). Violent victimization and o ending: Individual-, situational-, and
community-level risk factors. In A. J. Reiss Jr. & J. A. Roth (Eds.), Understanding and preventing
violence (Vol. 3, pp. 1?114). Washington, DC: National Academy Press.
Sampson, R. J., & Raudenbush, S. W. (1999). Systematic social observation of public spaces: A new look
at disorder in urban neighborhoods. American Journal of Sociology, 105, 603?651.
Sampson, R. J., & Raudenbush, S. W. (2001). Disorder in urban neighborhoods?does it lead to crime?
(Research Brief). Washington, DC: U.S. Department of Justice, National Institute of Justice.
Sampson, R. J., & Raudenbush, S. W. (2004). Seeing disorder: Neighborhood stigma and the social
construction of ?broken windows?. Social Psychology Quarterly, 67, 319?342.
Sampson, R. J., Raudenbush, S. W., & Earls, F. (1997). Neighborhoods and violent crime: A multilevel
study of collective e cacy. Science, 277, 918?924.
Sawyer, J. (1966). Measurement and prediction, clinical and statistical. Psychological Bulletin, 66, 178?
200.
Schmidt, P., & Witte, A. D. (1988). Predicting recidivism using survival models. New York, NY: Springer-
Verlang, Inc.
Schmidt, P., & Witte, A. D. (1989). Predicting criminal recidivism using ?split population? survival time
models. Journal of Econometrics, 40, 141?59.
Schuerman, L., & Kobrin, S. (1984). Exposure of community mental health clients to the criminal justice
system: Client criminal or patient prisoner. In L. Teplin (Ed.), Mental health and criminal justice.
Beverly Hills: Sage.
Schwaner, S. L. (1997). They can run, but can they hide? A profile of parole violators at large. Journal of
Crime and Justice, 20, 19?32.
Scribner, R. A., Cohen, D., Kaplan, S., & Allen, S. (1999). Alcohol availability and homicide in New
Orleans: Conceptual considerations for small area analysis of the e ect of alcohol outlet density.
Journal of Studies on Alcohol, 60, 310?316.
Scribner, R. A., MacKinnon, D. P., & Dwyer, J. H. (1995). The risk of assaultive violence and alcohol
availability in Los Angeles county. American Journal of Public Health, 85, 335?338.
236
Sellin, T. (1938). Culture conflict and crime. New York: Social Science Research Company.
Seltzer, A., & Langford, A. (1984). Forensic psychiatric assessments in the northwest territories. Canadian
Journal of Psychiatry, 29, 665?668.
Shepherd, J., Farrington, D. P., & Potts, J. (2002). Relations between o ending, injury and illness. Journal
of The Royal Society of Medicine, 95, 539?544.
Sherman, L., Gartin, P. R., & Buerger, M. E. (1989). Hot spots of predatory crime: Routine activities and
the criminology of place. Criminology, 27, 27?55.
Shupe, L. M. (1954). Alcohol and crime: A study of the urine alcohol concentration found in 882 per-
sons arrested during or immediately after the commission of a felony. Journal of Criminal Law,
Criminology and Police Science, 44, 661?664.
Silberman, C. E. (1978). Criminal justice, criminal violence. New York: Random House.
Silver, E., & Chow-Martin, L. (2002). A multiple models approach to assessing recidivism risk: Implica-
tions for judicial decision making. Criminal Justice and Behavior, 29, 538?568.
Sims, B., & Jones, M. (1997). Predicting success or failure on probation: Factors associated with felony
probation outcomes. Crime and Delinquency, 43, 314?327.
Skog, O. J. (1985). The collectivity of drinking cultures: A theory of the distribution of alcohol consump-
tion. British Journal of Addiction, 80, 83?99.
Skogan, W. G. (1992). Disorder and decline: Crime and the spiral of decay in American neighborhoods.
Berkeley, CA: University of California Press.
Smith, C., & Thornberry, T. P. (1995). The relationship between childhood maltreatment and adolescent
involvement in delinquency. Criminology, 33, 451?477.
Smith, D. A. (1986). The neighborhood context of police behavior. In A. J. Reiss Jr. & M. Tonry (Eds.),
Communities and crime (Vol. 8, pp. 313?341). Chicago: University of Chicago Press.
Smith, D. A., & Jarjoura, R. G. (1988). Social structure and criminal victimization. Journal of Research in
Crime and Delinquency, 25, 27?52.
Smith, D. A., & Polsenberg, C. (1992). Specifying the relationship between arrestee drug test-results and
recidivism. Journal of Criminal Law and Criminology, 83, 364?77.
Smith, D. A., Visher, C. A., & Davidson, L. A. (1984). Equity and discretionary justice: The influence of
race on police arrest decisions. Journal of Criminal Law and Criminology, 75, 234?249.
Smith, M. D., & Parker, R. N. (1980). Deterrence, poverty, and type of homicide. American Journal of
Sociology, 85, 614?624.
Smith, W. R., Frazee, S. G., & Davison, E. L. (2000). Furthering the integration of routine activity and
social disorganization theories: Small units of analysis and the study of street robbery as a di usion
process. Criminology, 38, 489?524.
Speckart, G., & Anglin, M. D. (1985). Narcotics use and crime: An analysis of existing evidence for a
causal relationship. Behavioral Science and the Law, 3, 259?283.
237
Speer, P. W., Gorman, D., Labouvie, E. W., & Ontkush, M. J. (1998). Violent crime and alcohol availability:
Relationships in an urban community. Journal of Public Health Policy, 19, 303?18.
Spohn, C. (2000). O ender race and case outcomes: Do crime seriousness and strength of evidence
matter? Final activities report (NCJ No. 184774). Washington, DC: U.S. Department of Justice,
O ce of Justice Programs, Bureau of Justice Statistics.
Stack, S. (1983). Homicide and property crime: The relationships to anomie. Aggressive Behavior, 9,
339?344.
Stacy, A. W., & Newcomb, M. D. (1995). Long-term social-psychological influences on deviant attitudes
and criminal behavior. In H. Kaplan (Ed.), Drugs, crime, and other deviant adaptations: Longitudi-
nal studies (pp. 99?127). New York: Plenum.
Stark, R. (1987). Deviant places: A theory of the ecology of crime. Criminology, 25, 893?909.
Stice, E., Myers, M. G., & Brown, S. (1998). Relations of delinquency to adolescent substance use and
problem use: A prospective study. Psychology of Addictive Behaviors, 12, 136?146.
Stitt, B. G., & Giacopassi, D. J. (1992). Alcohol availability and alcohol-related crime. Criminal Justice
Review, 17, 268?279.
Substance Abuse and Mental Health Services Administration. (2002). 2000 National Household Survey
on Drug Abuse (Tech. Rep.). Washington, DC: U.S. Department of Health & Human Services,
Substance Abuse & Mental Health Services Administration, O ce of Applied Studies.
Swanson, J., Holzer, C. E., III, Ganju, V., & Jono, R. (1990). Violence and psychiatric disorder in the
community: Evidence from the epidemiologic catchment area surveys. Hospital and Community
Psychiatry, 41, 761?770.
Sykes, G., & Messinger, S. (1960). The inmate social system. In R. A. Cloward & et al. (Eds.), Theoretical
studies in the social organization of the prison (pp. 6?9). New York: Social Science Research
Council.
Taylor, R. B. (1999). Incivilities thesis: Theory, measurement, and policy. In R. Langworthy (Ed.),
Measuring what matters (p. 65-88). Washington, DC: National Institute of Justice.
Taylor, R. B., & Covington, J. (1988). Neighborhood changes in ecology and violence. Criminology, 26,
553?589.
Teplin, L. A. (1990). The prevalence of severe mental disorder among male urban jail detainees: Compar-
ison with the Epidemiological Catchment Area program. American Journal of Public Health, 80,
663?669.
Teplin, L. A., Abram, K. M., & McClelland, G. M. (1996). Prevalence of psychiatric disorders among
incarcerated women: I. pretrial jail detainees. Archives of General Psychiatry, 53, 505-512.
Therneau, T. M., & Atkinson, E. J. (1997). An introduction to recursive partitioning using the RPART
routines (Technical Report Series No. 61). Rochester, MN: Department of Health Science Research,
Mayo Clinic.
238
Thornberry, T. P., & Christenson, R. L. (1984). Unemployment and criminal involvement: An investigation
of reciprocal causal structures. American Sociological Review, 49, 398?411.
Thornberry, T. P., & Farnworth, M. (1982). Social correlates of criminal involvement: Further evidence
on the relationship between social status and criminal behavior. American Sociological Review, 47,
505?518.
Thornberry, T. P., Moore, M., & Christenson, R. (1985a). The e ect of dropping out of high school on
subsequent criminal behavior. Criminology, 23, 3?18.
Thornberry, T. P., Moore, M., & Christenson, R. (1985b). The e ect of dropping out of high school on
subsequent criminal behavior. Criminology, 23, 3?18.
Thornberry, T. P., Smith, C. A., Craig Rivera and, D. H., & Stouthamer-Loeber, M. (1999). Family
disruption and delinquency (NCJ No. 178285). Washington, DC: U.S. Department of Justice, O ce
of Justice Programs, O ce of Juvenile Justice and Delinquency Prevention.
Uggen, C. (2000). Work as a turning point in the life course of criminals: A duration model of age,
employment, and recidivism. American Sociological Review, 65, 529?546.
Uggen, C., & Sta , J. (2001). Work as a turning point for criminal o enders. Corrections Management
Quarterly, 5, 1?16.
United States Census Bureau. (2004). Technical documentation: Census 2000 Summary File 3 (No.
SF3/12[RV]). U.S. Census Bureau.
United States Parole Commission. (2003). Rules and procedures manual [Computer software manual]. Re-
trieved June 15, 2006, from http://www.usdoj.gov/uspc/rules procedures/rulesmanual
.htm.
Venables, W., & Ripley, B. (2002). Modern applied statistics with s-plus (4th ed.). New York: Springer-
Verlag.
Vito, G. F. (1987). Felony probation and recidivism: Replication and response. Federal Probation, 50,
17?25.
van Voorhis, P., & Brown, K. (1997). Risk classification in the 1990s (Tech. Rep.). Washington, DC:
National Institute of Corrections.
Ward, D., & Tittle, C. (1994). IQ and delinquency: A test of two competing explanations. Journal of
Quantitative Criminology, 10, 189?212.
Warner, S. B. (1923). Factors determining parole from the Massachusetts Reformatory. Journal of Criminal
Law, Criminology, and Police Science, 14, 172?207.
Warr, M. (1993). Age, peers, and delinquency. Criminology, 31, 17?40.
Warr, M. (1998). Life-course transitions and desistance from crime. Criminology, 36, 183?216.
Webster, C., Harris, G., Rice, M., Cormier, C., & Quinsey, V. (1994). The violence prediction scheme:
Assessing dangerousness in high risk men. Toronto: Centre of Criminology,University of Toronto.
239
Wells, L., & Rankin, J. (1991). Families and delinquency: A meta-analysis of the impact of broken homes.
Social Problems, 38, 71?83.
West, D. (1982). Delinquency: Its roots, careers, and prospects. Cambridge, MA: Harvard University
Press.
White, H. (1980). A heteroskedasticity-consistent covariance matrix estimator and a direct test for het-
eroskedasticity. Econometrica, 48, 817?830.
White, H. R., & Gorman, D. (2000). Dynamics of the drug-crime relationship. In G. LaFree (Ed.), The
nature of crime: Continuity and change (Vol. 1, pp. 151?218). Washington, DC: U.S. Department
of Justice.
Whitehead, J. (1991). The e ectiveness of felony probation: Results from an eastern state. Justice Quar-
terly, 8, 525?543.
Widom, C. S. (1989). Child abuse, neglect, and violent criminal behavior. Criminology, 27, 251?271.
Williams, F. P., III, McShane, M. D., & Dolny, H. M. (2000). Predicting parole absconders. Prison Journal,
80, 24?38.
Williams, R. L. (2000). A note on robust variance estimation for cluster-correlated data. Biometrics, 56,
645?646.
Wilson, J. Q., & Hernnstein, R. J. (1985). Crime and human nature. New York, NY: Simon and Schuster.
Wilson, J. Q., & Kelling, G. L. (1982). Broken windows: The police and neighborhood safety. Atlantic
Monthly, 211, 29?38.
Wirth, L. (1964). The ghetto. In A. J. Reiss Jr. (Ed.), On cities and social life (pp. 84?98). Chicago:
University of Chicago Press. (Original work published 1956)
Wiseman, J. P. (1991). The other half: Wives of alcoholics and their social-psychological situation. New
York: A. de Gruyter.
Wish, E. D., & Johnson, B. D. (1986). The impact of substance abuse on criminal careers. In A. J. Reiss
Jr. & J. A. Roth (Eds.), (Vol. 2, pp. 52?88). Washington, DC: National Academy Press.
Wolfgang, M. E., & Ferracuti, F. (1967). The subculture of violence: Towards an integrated theory in
criminology. London: Tavistock Publications.
Wolfgang, M. E., Figlio, R. M., & Sellin, T. (1972). Delinquency in a birth cohort. Chicago: Chicago
University Press.
Wolfgang, M. E., Thornberry, T. P., & Figlio, R. (1987). From boy to man, from delinquency to crime.
Chicago: University of Chicago Press.
Wright, S., & Klee, H. (2001). Violent crime, aggression and amphetamine: What are the implications for
drug treatment services? Drugs: Education, Prevention and Policy, 8, 73?90.
Zamble, E., & Quinsey, V. L. (1997). The process of criminal recidivism. Cambridge: Cambridge Univer-
sity Press.
240
Zatz, M. S. (1984). Race, ethnicity, and determinate sentencing: A new dimension to an old controversy.
Criminology, 22, 147?165.
Zatz, M. S., & Hagan, J. (1985). Crime, time, and punishment: An exploration of selection bias in
sentencing research. Journal of Quantitative Criminology, 1, 103?126.
Zhang, L., & Messner, S. (1996). School attachment and o cial delinquency status in the people?s republic
of china. Sociological Forum, 11, 285?303.
Zhang, Q., Loeber, R., & Stouthamer-Loeber, M. (1997). Developmental trends of delinquent attitudes and
behaviors: Replications and synthesis across domains, time, and samples. Journal of Quantitative
Criminology, 13, 181?215.
241