Criminal Justice and Behavior
http://cjb.sagepub.com

Evaluating the Predictive Validity of the Compas Risk and Needs Assessment
System
Tim Brennan, William Dieterich and Beate Ehret
Criminal Justice and Behavior 2009; 36; 21
DOI: 10.1177/0093854808326545
The online version of this article can be found at:
http://cjb.sagepub.com/cgi/content/abstract/36/1/21

Published by:
http://www.sagepublications.com

On behalf of:
International Association for Correctional and Forensic Psychology

Additional services and information for Criminal Justice and Behavior can be found at:
Email Alerts: http://cjb.sagepub.com/cgi/alerts
Subscriptions: http://cjb.sagepub.com/subscriptions
Reprints: http://www.sagepub.com/journalsReprints.nav
Permissions: http://www.sagepub.com/journalsPermissions.nav
Citations http://cjb.sagepub.com/cgi/content/refs/36/1/21

Downloaded from http://cjb.sagepub.com at DENVER UNIV on December 12, 2008

 EVALUATING THE PREDICTIVE
VALIDITY OF THE COMPAS RISK AND
NEEDS ASSESSMENT SYSTEM
TIM BRENNAN
WILLIAM DIETERICH
BEATE EHRET
Northpointe Institute for Public Management Inc.

This study examines the statistical validation of a recently developed, fourth-generation (4G) risk–need assessment system
(Correctional Offender Management Profiling for Alternative Sanctions; COMPAS) that incorporates a range of theoretically
relevant criminogenic factors and key factors emerging from meta-analytic studies of recidivism. COMPAS’s automated scoring provides decision support for correctional agencies for placement decisions, offender management, and treatment planning.
The article describes the basic features of COMPAS and then examines the predictive validity of the COMPAS risk scales by
fitting Cox proportional hazards models to recidivism outcomes in a sample of presentence investigation and probation intake
cases (N = 2,328). Results indicate that the predictive validities for the COMPAS recidivism risk model, as assessed by the area
under the receiver operating characteristic curve (AUC), equal or exceed similar 4G instruments. The AUCs ranged from .66
to .80 for diverse offender subpopulations across three outcome criteria, with a majority of these exceeding .70.
Keywords: COMPAS; predictive validity; survival analysis; risk assessment; probation; area under the curve (AUC); criminogenic needs

I

n a recent review of the state of the art in correctional assessment, Andrews, Bonta, and
Wormith (2006) identified Correctional Offender Management Profiling for Alternative
Sanctions (COMPAS; Northpointe Institute for Public Management, 1996) as an example
of an emerging fourth-generation (4G) approach to correctional assessment. They also
noted that the available 4G approaches, with the exception of the Level of Service/Case
Management Inventory (LS/CMI; Andrews, Bonta, & Wormith, 2004), are relatively new
and that validation evidence is still required for these newer approaches. This article
assesses several key aspects of scale reliability and validity for the COMPAS system.

TRENDS IN CORRECTIONAL ASSESSMENT

The past three decades in correctional practice have seen a progression from first-generation
(1G) to currently emerging 4G assessment approaches (Andrews et al., 2006; Blanchette &
Brown, 2006; Bonta, 1996; Clements, 1996). These developments occurred as successive
generations of assessment and classification methods addressed the more obvious weaknesses
of prior phases. These phases and their main characteristics are described below.
The 1G approach relied on clinical and professional judgment in the absence of any
explicit or objective scoring rules. It dominated corrections for several decades and remains
preferred by many correctional decision makers (Boothby & Clements, 2000; Wormith,
2001). Its weaknesses include excessive subjectivity, inconsistency, bias and potential
CRIMINAL JUSTICE AND BEHAVIOR, Vol. 36 No. 1, January 2009 21-40
DOI: 10.1177/0093854808326545
© 2009 International Association for Correctional and Forensic Psychology

21
Downloaded from http://cjb.sagepub.com at DENVER UNIV on December 12, 2008

 22

CRIMINAL JUSTICE AND BEHAVIOR

stereotyping, legal vulnerability, and lower predictive validity than structured objective
methods (Brennan, 1987; Grove & Meehl, 1996; Hastie & Dawes, 2001).
Second-generation (2G) assessments adopted an empirical approach that mainly relied
on simple additive point scales, often with only a few standardized factors (e.g., Austin,
1983; S. D. Gottfredson, 1987; Hoffman, 1994). These mostly reflected Dawes’s (1979)
description of “improper” linear models (p. 571) because the selected factors and weightings were often established by common sense or professional consensus rather than by statistical methods. These methods primarily focused on risk prediction, brevity, and
efficiency. The main criticisms included lack of theoretical background, limited coverage
of risk and need factors, neglect of dynamic (changeable) risk factors, lack of treatment
implications, weak explanatory value, and questionable relevance for female offenders
(Blanchette & Brown, 2006; Jones, 1996). However, as noted by Dawes, these linear
models are often surprisingly effective in terms of predictive validity and generally outperformed professional judgment or the opinions of trained experts (Grove & Meehl, 1996;
Hastie & Dawes, 2001; Mossman, 1994).
Third-generation (3G) assessments of the late 1970s and 1980s introduced a more
explicit, empirically based, and theory-guided approach and a broader selection of criminogenic factors. In addition, some of these factors were designed to be dynamically sensitive to change. The Level of Service Inventory–Revised (LSI-R; Andrews & Bonta, 1995)
exemplified these trends and perhaps has become the most widely used risk and need
assessment in corrections. However, 3G methods, including the LSI-R, eventually were
criticized for a narrow theoretical focus (mainly social learning theory), a failure to address
gender sensitivity, a dominant focus on risk, and failure to assess offender strengths or protective factors as emphasized in the “good lives” model (Andrews et al., 2006; Blanchette
& Brown, 2006; Bloom, 2000; Reisig, Holtfreter, & Morash, 2006; Ward & Stewart, 2003).
Regarding 4G assessments, Andrews et al. (2006) identified several instruments as representing this category, including the Correctional Assessment and Intervention System
(National Council on Crime and Delinquency, 2006), LS/CMI, and COMPAS. Several general features appear to characterize 4G approaches. These include (a) a broader selection of
explanatory theories, (b) broader range of risk and need factors (content validity), (c) incorporation of the strengths or resiliency perspective, (d) more advanced statistical modeling,
(e) seamless integration of the need or risk domain with the agency management information system, and (f) criminal justice databases and Web-based implementation of assessment technology. Such integration allows users to track offenders from intake to case
closure to support sequential case management monitoring, information feedback, and
decision making. COMPAS has incorporated all of these features; interested readers may
obtain full details in Brennan, Dieterich, and Oliver (2007).
The goals of this article are threefold. First, it describes the general design features and
technical overview of the COMPAS system. Second, it assesses the reliabilities of the
COMPAS scales for both male and female offenders. Third, it assesses the predictive validity of the COMPAS scales for both males and females.

BASIC DESIGN FEATURES OF COMPAS

COMPAS is an automated decision-support software package that integrates risk and
needs assessment with several other domains, including sentencing decisions, treatment
Downloaded from http://cjb.sagepub.com at DENVER UNIV on December 12, 2008

 Brennan et al. / EVALUATING COMPAS

23

and case management, and recidivism outcomes. Documentation of the full software functionality is available at www.northpointeinc.com, and detailed information on the various
risk and need scales is provided in the appendix. Beyond the integration of separate databases, the following design features of COMPAS further advance and support evidencebased practice (EBP) in criminal justice agencies.
THEORY-GUIDED ASSESSMENT

Ideally, explanatory theories of criminality should guide the selection of scale content of
an assessment system. Criminologists have long lamented the lack of theory-guided assessments (Bonta, 2002; Clements, 1996; Jones, 1996). Thus, 4G systems have a strong emphasis on theory-guided assessment. In contrast to the LSI-R, which was designed primarily
around a social learning explanation (Andrews & Bonta, 1998), COMPAS broadens the
theoretical coverage to include key constructs from low self-control theory, strain theory or
social exclusion, social control theory (bonding), routine activities–opportunity theory, subcultural or social learning theories, and a strengths or good lives perspective.
BROADBAND COMPREHENSIVE ASSESSMENT

Second, 4G approaches introduce a broader comprehensive coverage of criminogenic
factors to match the theoretical and explanatory complexity of criminal behavior and to
provide sufficient explanatory information to guide case interpretation and intervention
planning. Thus, COMPAS includes both theoretically relevant factors and the critical eight
criminogenic predictive factors that emerged from recent meta-analytic studies (Andrews
et al., 2006; Gendreau, Little, & Goggin, 1996; Lipsey & Derzon, 1998; Lösel, 1995). The
2G approaches generally reflected the opposite trend by minimizing and simplifying
assessment to reduce workload burden on staff, which, not unexpectedly, resulted in
extreme poverty of explanatory components and an almost total lack of treatment guidance
(Austin, 1983; Glaser, 1987; Palmer, 1992).
INTEGRATION OF THE STRENGTH OR RESILIENCY PERSPECTIVE

The strength-based or good lives approach (Andrews et al., 2006; Ward & Brown, 2004)
is a natural extension of the shift toward more comprehensive assessment. In their review,
Andrews et al. (2006) suggested that measures of strengths and well-being are “highly relevant” (p. 23) for correctional assessments. To address this issue, COMPAS includes a
number of strength and protective factors that have shown empirical support for potential
risk reduction and protecting offenders from the full impact of criminogenic needs. These
include job and educational skills, history of successful employment, adequate finances,
safe housing, family bonds, social and emotional support, noncriminal parents and friends,
and so on.
MORE ADVANCED STATISTICAL MODELS

4G assessments, in contrast to earlier approaches, are beginning to use more advanced
statistical methods for predictive modeling and classification. Although Burgess-type,
equally weighted, linear models have performed reasonably well (S. D. Gottfredson, 1987;
Mossman, 1994), powerful multivariate, model-averaging, mixed-model ensemble methods
Downloaded from http://cjb.sagepub.com at DENVER UNIV on December 12, 2008

 24

CRIMINAL JUSTICE AND BEHAVIOR

and artificial intelligence are now entering correctional assessment approaches. For example,
the COMPAS risk and classification models use logistic regression, survival analysis, and
bootstrap classification methods in a broad repertoire of prediction and classification procedures (Brennan, Breitenbach, & Dieterich, 2008). The present article specifically examines its
predictive models for recidivism based on survival analysis (Cox regression).
INTEGRATION WITH CRIMINAL JUSTICE DATABASES TO FACILITATE EBP

Another feature of 4G methods, including COMPAS, is the seamless integration of the
risk and needs domain with separate domains of sentencing decisions, institutional processing and placement decisions, case management decisions, treatments given (type and
amount), and various outcomes (across time). This integration provides support for correctional agencies in implementing EBP studies (see Andrews et al., 2006; Brennan, Wells, &
Alexander, 2004).
The COMPAS system includes two additional design features of some note: a treatmentexplanatory classification to support staff with specific responsivity decisions and gendersensitive calibration, described below.
TREATMENT-EXPLANATORY CLASSIFICATION TO ADDRESS SPECIFIC RESPONSIVITY

Andrews et al. (2006) suggested that specific responsivity of offenders is the least
explored of their risk–need–responsivity principles. Yet specific responsivity is a critical
and recurrent challenge for treatment providers in matching individual offenders to appropriate treatment regimes (Brennan, 2008a; Meier, 2002; Millon & Davis, 1997; Warren &
Hindelang, 1979). COMPAS addresses specific responsivity and client–treatment matching
using two well-known approaches. First, it provides a person-centered assessment chart of
decile scores for each risk and need scale. Second, following the lineage of Marguerite
Warren, Ted Palmer, and others (also see Harris & Jones, 1999; Van Voorhis, 1994), COMPAS provides a treatment-relevant typology that integrates risk and need. This explanatory
typology identifies and demarcates several specific pathways that may guide differential
targeting and programming for diverse offender types who belong in one particular pathway. Although the present article does not address this treatment-relevant taxonomy,
detailed descriptions of these pattern-seeking methods are presented in Brennan et al.
(2008) and Brennan (2008b).
GENDER-SENSITIVE ASSESSMENT

A major criticism of 2G and 3G approaches is that they largely based their assessment
and classification methods on dominantly male samples and then mechanically applied
these to female offenders (Blanchette & Brown, 2006; Bloom, 2000; Brennan, 2008b; Farr,
2000; Hannah-Moffatt & Shaw, 2001). However, compelling arguments are now advanced
for systematic validation of instruments on separate female and male offender samples
(Hardyman & Van Voorhis, 2004; Holtfreter & Cupp, 2007). COMPAS addresses this issue,
first by using separate samples of males and females to develop gender-specific calibrations
of all risk and need factors and second by evaluating its predictive and classification models
on separate male and female samples (see below). Plans are also under way to incorporate
additional gender-specific factors from recent research on gender-sensitive risk and need
factors into COMPAS (Salisbury, Van Voorhis, & Wright, 2006).
Downloaded from http://cjb.sagepub.com at DENVER UNIV on December 12, 2008

 Brennan et al. / EVALUATING COMPAS

25

METHOD
PARTICIPANTS

Participants were 2,328 individuals who were assessed with COMPAS as part of their
processing at entry into probation agencies. All individuals with complete data who were
administered the full COMPAS assessment were included. The sample represented about
15% of all COMPAS assessments conducted at these agencies during this period.
Individuals were excluded if they were missing data (25%) or if they were not administered
the full COMPAS (60%). Women composed 19% of the sample. The ethnic composition of
the sample is 76% White, 15% African American, 7% Latino, and 2% Other. The average
age of participants was 31.9 years (range = 18.0 to 69.7). In the sample overall, 45% of the
presenting offenses were misdemeanors, 48% nonassaultive felonies, and 7% assaultive
felonies. The median number of prior arrests was 3 (range = 0 to 57). Among the probation
cases, 9% were split-sentence cases (jail and probation).
ADMINISTRATION

The assessments were conducted by local probation officers between January 2001 and
December 2004 at 18 county-level probation agencies in an eastern state. Interviews were
conducted at the point of presentence investigation (PSI) or at probation intake (approximately 50% each). Staff and supervisors take a 2-day COMPAS training program that covers relevant interview techniques, response categories, item meanings, and quality
assurance issues. Official criminal records are used to complete the current offense and
criminal history sections of COMPAS prior to the interviews. The interviews typically
require approximately 45 to 60 min, depending on the extent of probing.
MEASURES

COMPAS scales. This study examined the predictive validity of all the COMPAS base
scales (listed in Table 1) and also the main Recidivism Risk Scale. The Recidivism Risk
Scale is a regression model that has been used in COMPAS since 2000. This regression
model was trained to predict new offenses in a probation sample. The system transforms a
linear predictor from the regression model to a decile score. The system calculates a recidivism risk decile score by referencing an appropriate COMPAS norm group. For the current
analyses, a gender-specific composite norm group was used. The composite norm group
(n = 7,381) was constructed from COMPAS assessment data collected in prison (34%), jail
(14%), and probation (53%). The set of base scales included criminal involvement, history
of noncompliance, history of violence, current violence, criminal associates, substance
abuse, financial problems, vocational or educational problems, family criminality, social
environment, leisure, residential instability, social isolation, criminal attitudes, and criminal personality.
Dependent variables: Three outcomes for survival analyses. We matched COMPAS
assessment data with computerized official criminal history records and constructed multiplerecord survival data sets using assessment and event dates in the criminal history data.
These included crime dates, arrest dates, dispositions, disposition dates, sentence type, and
sentence length. The three outcomes selected as dependent variables for this study included
Downloaded from http://cjb.sagepub.com at DENVER UNIV on December 12, 2008

 26

CRIMINAL JUSTICE AND BEHAVIOR

TABLE 1: Alpha Coefficients and Their Differences Between Women and Men, With Pointwise 95%
Confidence Intervals (CIs)

Criminal Involvement
History of Noncompliance
History of Violence
Current Violence
Criminal Associates
Substance Abuse
Financial Problems
Vocational or Educational
Family Criminality
Social Environment
Leisure
Residential Instability
Social Isolation
Criminal Attitudes
Criminal Personality

α

Women

Men

Difference

Lower Bound 95% CI

Upper Bound 95% CI

.87
.68
.73
.59
.68
.79
.73
.71
.63
.74
.82
.65
.81
.82
.76

.85
.62
.70
.62
.68
.81
.75
.73
.64
.70
.81
.61
.82
.82
.76

.87
.67
.73
.59
.68
.78
.72
.71
.63
.74
.83
.65
.81
.82
.76

–.02
–.04
–.03
.03
.00
.03
.03
.02
.02
–.04
–.02
–.05
.01
.01
.00

–.04
–.11
–.07
–.03
–.05
.00
–.02
–.02
–.04
–.09
–.05
–.11
–.02
–.02
–.03

.01
.02
.02
.09
.05
.06
.07
.06
.07
.01
.01
.01
.04
.03
.04

Note. The difference in alphas is significant if the 95% confidence interval does not include zero.

(a) an arrest for any offense, (b) an arrest for a person offense, and (c) an arrest for a felony
offense.
We defined an offense as a finger-printable arrest involving a charge and filing for any
uniform crime reporting (UCR) code. A person offense is a finger-printable arrest involving a charge and filing for any UCR code for murder, voluntary manslaughter, forcible rape,
robbery, aggravated assault, simple assault, burglary (with weapon of an occupied
dwelling), dangerous weapons, sex offenses, extortion, arson, and kidnap. This category
includes misdemeanor and felony offenses.
ANALYSIS

We fitted separate cause-specific Cox proportional hazard models to each recidivism
outcome (Kalbfleisch & Prentice, 2002). Analysis time is the number of days from COMPAS assessment date to first failure or end of study, whichever occurred first. As mentioned,
the assessments were conducted between January 2001 and December 2004. The end of
study is the date of the recidivism outcomes computer match (March 3, 2006). We determined the failure time point from the offense date associated with the recidivism outcome
of interest. Cases remained in the risk set and contributed information to the analyses until
the point of failure or end of study, whichever occurred first. The models controlled for
intermittent periods of incarceration during the follow-up by removing the case from the
risk set during these intermittent gaps. We also removed split-sentence cases from the risk
set during the time they were in jail. We removed PSI cases that received a subsequent jail
or prison sentence from the risk set during the incarceration period. In the felony offense
model, there were 433 cases with a gap, with an average time on gap of 206 days (range =
1 to 1,440 days). The median time at risk in the felony offense model was 759 days (range =
1 to 1,722 days).
First, we fitted a series of univariable Cox survival models in which the hazard for each
of the three recidivism outcomes was regressed on each COMPAS base scale. These
Downloaded from http://cjb.sagepub.com at DENVER UNIV on December 12, 2008

 Brennan et al. / EVALUATING COMPAS

27

models were fitted in three partitions of the sample: the full sample, men only, and women
only. Next, two multivariate models were fitted to each recidivism outcome in each partition. Model I included all the COMPAS base scales. Model II included all the COMPAS
base scales plus age at first arrest. Finally, a model that included only the Recidivism Risk
Scale was fitted to each recidivism outcome in each partition.
To gauge the predictive utility of the above two COMPAS multivariate base scale models
and the Recidivism Risk Scale model, we estimated the area under the receiver operating
characteristic curve (AUC) for all three models. For survival models, the most relevant
measure of predictive discrimination is the concordance index, which is equivalent to the
AUC and is defined as the probability that the predictor values and survival times for a pair
of randomly selected cases are concordant. A pair is concordant if the case with the higher
predictor value has a shorter survival time. The calculation is based on the number of all
possible pairs of nonmissing observations for which survival time can be ordered and the
proportion of relevant pairs for which the predictor and survival time are concordant
(Harrell, Califf, Pryor, Lee, & Rosati, 1982).

RESULTS
RELIABILITY

Table 1 provides reliability coefficients (Cronbach’s α) to indicate the internal consistency of the core COMPAS scales both for the total sample and by gender. Alpha is the
most widely used measure of internal consistency of summative scales. By convention,
alpha coefficients of .70 or higher indicate satisfactory reliability. The table indicates that a
large majority of these alphas are in the satisfactory range of close to or above .70 with only
a few exceptions (current violence, family criminality, and residential instability), and these
are close to an acceptable range. The alpha coefficients were statistically equivalent for
both genders.
PREDICTIVE VALIDITY

Table 2 provides a summary of the survival experience of men and women for the three
outcomes examined in the study. For each offense type, the table shows the number of failures that occurred during each year and the estimated survivor function at the end of each
year. The survivor function is the probability of surviving beyond Time t, given survival up
to Time t. The survivor function is interpreted as the cumulative proportion surviving over
time. Note that only the first 4 years of the follow-up are shown.
Table 3 shows the results from univariable Cox regressions of the hazard for a new
felony offense on each COMPAS base scale. The results indicate that all except three of
these base scales reach significant levels in predicting felony recidivism. The scales that did
not attain significance were current violence, financial problems, and residential instability.
Note that substance abuse has a negative effect on felony recidivism. This may have
resulted from the selected outcome variable (felony recidivism) for this analysis. Drug
offenders (overall) may be at lower risk for new arrests for serious felony offenses.
Table 4 shows the results of the multivariate Cox regression of the hazard of a felony
offense on the COMPAS base scales. In this model, which combines all COMPAS base
Downloaded from http://cjb.sagepub.com at DENVER UNIV on December 12, 2008

 28

CRIMINAL JUSTICE AND BEHAVIOR

TABLE 2: Number Failing Each Year and Survivor Function Through the End of Each Year for Any
Offense, Offenses Against Persons, and Felony Offenses for Men and Women

Women ( n = 449)
Offense Type

Men ( n = 1,879)

Number Failing Each Year

Survivor Function

Number Failing Each Year

Survivor Function

64
32
14
10

.85
.76
.70
.58

308
180
63
41

.81
.69
.62
.52

11
10
5
2

.97
.95
.92
.89

114
90
33
23

.93
.87
.83
.78

16
12
7
1

.96
.93
.89
.88

134
95
31
20

.92
.85
.82
.77

Any
1st year
2nd year
3rd year
4th year
Person
1st year
2nd year
3rd year
4th year
Felony
1st year
2nd year
3rd year
4th year

Note. The description of the survival experience is limited to the first 4 years of the follow-up.

TABLE 3: Results From Univariable Cox Proportional Hazards Models Regressing the Hazard for a
Felony Offense on Each COMPAS Base Scale

Criminal involvement
History of noncompliance
History of violence
Current violence
Criminal associates
Substance abuse
Financial problems
Vocational or educational
Criminal attitudes
Family criminality
Social environment
Leisure
Residential instability
Criminal personality
Social isolation

coeff

exp(coeff)

SE(coeff)

p Value

0.033
0.148
0.108
0.101
0.148
–0.091
0.009
0.118
0.057
0.100
0.257
0.073
0.016
0.048
0.029

1.03
1.16
1.11
1.11
1.16
0.91
1.01
1.13
1.06
1.11
1.29
1.08
1.02
1.05
1.03

0.013
0.020
0.018
0.052
0.022
0.023
0.024
0.014
0.009
0.036
0.032
0.016
0.014
0.007
0.010

.008
.000
.000
.052
.000
.000
.691
.000
.000
.006
.000
.000
.277
.000
.003

scales, the significant predictors of felony recidivism are history of noncompliance, criminal associates, substance abuse, financial problems, vocational or educational, and high
crime (social) environment. As often occurs when multiple predictor variables are used and
when some are correlated, the signs of certain parameters can take unexpected directions.
Table 5 shows the results of the Cox regression of the hazard for a new felony offense
on the levels of the Recidivism Risk Scale. The hazard of a new felony offense for cases
that score high on the Recidivism Risk Scale is 5.66 times the hazard of cases that score
low (p value < .001). The hazard of the high-risk group relative to the medium-risk group
is 1.84 (p value < .001).
Downloaded from http://cjb.sagepub.com at DENVER UNIV on December 12, 2008

 Brennan et al. / EVALUATING COMPAS

29

TABLE 4: Results From Multivariable Cox Proportional Hazards Model Regressing the Hazard for a
Felony Offense on the COMPAS Base Scales

Criminal involvement
History of noncompliance
History of violence
Current violence
Criminal associates
Substance abuse
Financial problems
Vocational or educational
Criminal attitudes
Family criminality
Social environment
Leisure
Residential instability
Criminal personality
Social isolation

coeff

exp(coeff)

SE(coeff)

p Value

0.003
0.124
0.026
0.001
0.072
–0.134
–0.059
0.081
0.014
–0.001
0.117
0.015
–0.002
0.007
–0.003

1.00
1.13
1.03
1.00
1.07
0.87
0.94
1.08
1.01
1.00
1.12
1.01
1.00
1.01
1.00

0.018
0.032
0.024
0.057
0.028
0.027
0.026
0.017
0.011
0.040
0.037
0.021
0.015
0.011
0.012

.854
.000
.276
.982
.009
.000
.025
.000
.212
.978
.002
.476
.880
.555
.771

TABLE 5: Results From Cox Proportional Hazards Model Regressing the Hazard for a Felony Offense on
the Recidivism Risk Scale

Medium risk
High risk

coeff

exp(coeff)

SE(coeff)

p Value

95% Confidence Interval

1.12
1.73

3.07
5.66

0.473
0.843

< .001
< .001

2.27, 4.15
4.22, 7.58

Note. The reference category is the low risk level of the Recidivism Risk Scale.

Figure 1 shows a plot of the Nelson–Aalen (Aalen, 1978; Nelson, 1972) estimator of the
cumulative hazard function Hˆ (t) within levels of the Recidivism Risk Scale. The estimator
is defined as
Hˆ (t) =
j:t j ≤ t

dj
,
nj

(1)

where nj stands for the number of cases in the risk set just before Time tj and dj is the
number of failures at Time tj. The cumulative hazard is the expected number of failures for
an individual as a function of time, if failure was a repeatable process. Figure 1 also shows
the size of the risk set (nj) at 180-day intervals in each level of the Recidivism Risk Scale.
Although the maximum follow-up time in the data is 1,887 days, the time axis in the plot
is truncated to 1,620 days because the risk set is small and the estimates less precise beyond
this point.
Finally, we assessed the discriminatory power of the three models for predicting the
three criteria of interest (general offenses, offenses against persons, and felony offenses) in
each of the three partitions of the sample. As described previously, Model I includes all
COMPAS base scales, Model II adds age at first arrest to these COMPAS base scales, and
Model III represents the COMPAS Recidivism Risk Scale. Again, because we are fitting
survival models, we estimate Harrell’s concordance index (Harrell et al., 1982) and interpret it as the AUC. A rule of thumb according to several recent articles is that AUCs of .70
Downloaded from http://cjb.sagepub.com at DENVER UNIV on December 12, 2008

 30

CRIMINAL JUSTICE AND BEHAVIOR

Low

.6
Cumulative Hazard

95% CI

Medium
High

.4

.2

0

0

180

Number at Risk
Low Risk 1,048 977
Med. Risk 668 578
High Risk 490 378

360

540

720

900 1,080 1,260 1,440 1,620

Days to First Felony Offense
999
572
368

876
480
298

685
366
234

512
289
185

310
178
134

179
116
87

154
93
56

39
29
11

Figure 1: Nelson-Aalen Plot of the Cumulative Hazard for a Felony Offense in Each Level of the
Recidivism Risk Scale, With Pointwise 95% Confidence Bands

Note. CI = confidence interval.

TABLE 6: Area Under the Curve Values for COMPAS Risk Models Predicting Any Offense, Offenses
Against Persons, and Felony Offenses

Total Sample ( N = 2,328)

Women ( n = 449)

Men (n = 1,879)

Model

Any

Person

Felony

Any

Person

Felony

Any

Person

Felony

COMPAS I
COMPAS II
Recidivism Risk III

.66
.68
.68

.72
.73
.71

.70
.72
.70

.69
.72
.65

.78
.80
.76

.68
.69
.66

.67
.68
.68

.71
.72
.70

.71
.73
.71

Note. COMPAS Model I includes COMPAS base scales; COMPAS II adds age at first arrest to Model I.

or above typically indicate satisfactory predictive accuracy, and measures between .60 and
.70 suggest low to moderate predictive accuracy (Aos & Barnoski, 2003; Jones, 1996;
Quinsey, Harris, Rice, & Cormier, 1998).
Table 6 presents the model-specific AUCs for all of these analyses. The AUC values
range from .66 to .80, with a majority being above .70, which suggests satisfactory predictive validities of these COMPAS risk models for all three recidivism outcomes. The AUCs
for person and felony offenses are a little higher than those for the broader and less precisely defined recidivism category “any offense” (which includes both misdemeanor and
felony offenses). The AUCs for predicting person offenses range from .71 to .80, with high
AUC values for women (.76 to .80). However, for women, although the sample size was
449, there were only 29 women with a person offense, which reduces statistical power and
suggests caution regarding this result.

Downloaded from http://cjb.sagepub.com at DENVER UNIV on December 12, 2008

 Brennan et al. / EVALUATING COMPAS

31

TABLE 7: Area Under the Curve Values for COMPAS Risk Models Predicting Any Offense, Offenses
Against Persons, and Felony Offenses for White and African American Men

White Men (n = 1,412)

African American Men (n = 296)

Model

Any

Person

Felony

Any

Person

Felony

COMPAS I
COMPAS II
Recidivism Risk III

.69
.71
.69

.74
.75
.71

.73
.75
.71

.64
.66
.67

.69
.71
.72

.69
.72
.73

Last, we examined the predictive accuracy of each of the models for African American
men and White men. We do not report results for African American women because the
effective sample sizes for most of the outcomes were too small to calculate separate AUCs
for that group. Table 7 presents the results from each model for the outcomes any arrest,
person offense arrest, and felony arrest for African American and White men. The AUCs
for African American men range from .64 to .73. As was the case in the full sample, the
highest AUCs are obtained for the felony offense and person offense arrest outcomes. The
AUC results for White men are quite similar to the results for African American men,
except that White men have somewhat higher AUCs on the COMPAS base scale models.

DISCUSSION AND CONCLUSIONS

This study examined the reliability (internal consistency) and predictive validity of the
COMPAS risk and needs scales on a large sample of PSI and probation cases. A first general conclusion is that a majority of these scales reach levels of internal consistency and
predictive validity that are within generally acceptable ranges. Second, the separate univariable analyses show that a majority of the specific COMPAS risk and need base scales
were significantly associated with felony recidivism (Table 3). Third, results from survival
analyses demonstrated that the predictive power of the three models tested was comparable
to, and in some cases higher than, similar risk predictive instruments in this field. We now
examine some more detailed specific findings.
Regarding internal consistency, most of the scales have alpha coefficients equal to or
greater than .70, with only three exceptions, and the latter were close to acceptable levels.
These satisfactory results reemerged in both male and female subsamples. We found no significant differences in alpha levels between male and female subsamples, suggesting that
the scales are equally reliable for men and women.
Regarding predictive validity, we must first place the present results in the context of recent
research and accuracy performances for offender recidivism studies. An important point is that
the AUC has become the preferred measure of accuracy largely because of its independence
across base rates and selection ratios that allow it to provide clearer comparisons across different predictive instruments and studies (Flores, Lowenkamp, Smith, & Latessa, 2006;
Quinsey et al., 1998). Another contextual issue concerns interpretations of levels of the AUC
coefficient. As noted previously, AUCs in the .50s are considered to have little or no predictive
accuracy, those in the .60s are considered weak, those approaching or above the .70s are moderate, and those in the .80s are strong (Tape, 2003). However, various authors appear to use
Downloaded from http://cjb.sagepub.com at DENVER UNIV on December 12, 2008

 32

CRIMINAL JUSTICE AND BEHAVIOR

different standards. For example, Flores et al. (2006) described their achieved AUC of .689 for
the LSI-R as valid and robust and as moderate to large. Similarly, Kroner, Stadtland, Eidt, and
Nedopil (2007) describe an AUC of .703 as representing a high predictive accuracy in a study
of the Violence Risk Appraisal Guide (VRAG). Perhaps the relative recency of using the AUC
to investigate offender risk assessment explains variations in evaluative statements. Finally,
with more studies reporting AUCs in this area, it appears that the accuracy levels achieved for
most current instruments, across a variety of samples and outcome variables, generally fall in
the range of .65 to .75, with only a few exceptions (see below).
The present study of COMPAS models produced AUCs mostly in a range of .70 to .80.
Specifically, 16 out of 27 cells examined for AUC reached .70 or above, with a smaller set
of cells in the .66 to .69 range. Note in particular the predictive accuracies of COMPAS for
person offenses in which all 9 of the cells (total sample, males or females, and three
models) had AUCs between .70 and .80. Thus, we may conclude that COMPAS predictive
accuracies are similar to or slightly higher than AUCs obtained by other major instruments
in this field (e.g., Barbaree, Seto, Langton, & Peacock, 2001; Barnoski & Aos, 2003; Dahle,
2006; Flores et al., 2006; Grann, Belfrage, & Tengstrom, 2000; Quinsey et al., 1998).
Furthermore, the present findings are in the same range as found in the initial validation
studies of the COMPAS recidivism risk model for probationers that produced AUCs of .72
and .74 over a 24-month outcome period (Brennan, Dieterich, & Oliver, 2004).
The AUCs of the other main instruments often used for offender risk prediction may further help to contextualize the above findings. Perhaps the best known instruments are the
VRAG (Quinsey et al., 1998), the LSI-R (Andrews & Bonta, 1995), and the Psychopathy
Checklist–Revised (PCL-R; Hare, 1991). The AUC values for these instruments in recent
studies are quite varied according to the specific populations, outcome periods, and dependent variables used in specific studies, as illustrated below.
VRAG

Quinsey et al. (1998) found an AUC of .76 in a large-scale, multiyear recidivism study.
Barbaree et al. (2001) reported AUCs of .69 in predicting serious reoffending and .77 when
predicting any reoffense for sex offenders. Kroner et al. (2007) obtained an AUC of .703 in
a study of reoffending among mentally ill offenders.
LSI-R

The recent review by Andrews et al. (2006) did not provide AUCs for the LSI-R.
However, Barnoski and Aos (2003) found AUCs of .64 to .66 for the LSI-R in predicting
felony and violent recidivism among Washington State prisoners. Flores et al. (2006) found
an AUC of .689 using the LSI-R to predict reincarceration among federal probationers.
Dahle (2006) reported an AUC of .65 using the LSI-R to predict violent recidivism.
Barnoski (2006) reported an AUC of .65 using the LSI-R to predict felony sex recidivism.
PCL-R

AUC levels again varied across studies. For example, in a Swedish study of mentally ill violent offenders, Grann et al. (2000) found AUC levels of .64 to .75 based on various follow-up
time frames. Barbaree et al. (2001) reported AUCs of .61, .65, and .71 for the PCL-R in predicting various recidivism outcomes among sex offenders.
Downloaded from http://cjb.sagepub.com at DENVER UNIV on December 12, 2008

 Brennan et al. / EVALUATING COMPAS

33

The above findings clearly do not exhaust the full range of studies in this area. As more
studies report AUCs for specific instruments, varying populations, outcome variables, and
time frames, it may become possible to identify which instruments perform well in these
varying conditions.
The present study has several strengths and limitations. Strengths include the large sample (N = 2,328), a multiyear outcome period, and an examination of AUCs for different
COMPAS models, different offense categories, and across genders. The incorporation of
survival modeling also shows not only that these AUCs achieved significant discrimination
between recidivists and nonrecidivists but also that the timing of failure events can be predicted using either the COMPAS base scales model or the overall recidivism risk model.
However, the present study did not systematically address variations in predictive accuracy
by offender subgroups broken down by age, ethnicity, and race; level of addiction; length of
follow-up; and so on. Several large-scale studies of COMPAS that will allow such detailed
examination are in progress. The following issues may still require further research.
PREDICTIVE ACCURACY BY OUTCOME OFFENSE

The present study found minor differences in AUCs for COMPAS risk models in predicting differing outcomes and clearly achieved stronger results for new offenses against
persons and for felony offenses than for the broader category of any new offense. This may
stem from the higher precision of definitions for the first two outcomes as opposed to the
less precise “any” new offense, as this included both misdemeanors and felonies across a
wide range of offenses. Barnoski and Drake (2007) found similar results when they examined the validity of a static risk scale for predicting three outcome categories (violent, property and violent, and felony).
PREDICTIVE ACCURACY BY GENDER

The present study found that the basic COMPAS recidivism model predicts behaviors for
men and women about equally well, again similar to Barnoski and Drake (2007), who
found only small differences by gender. We realize that there are substantial concerns in
this topic, and a search for gender-sensitive risk factors is currently under way. COMPAS
is being augmented by additional risk and need factors of specific importance for female
offenders (Blanchette & Brown, 2006; Brennan, 2008b; Salisbury et al., 2006).
PREDICTIVE ACCURACY BY ETHNICITY

The present study found that the COMPAS recidivism models preformed equally well
for African American and White men at predicting the arrest outcomes. There is only one
previous study of which we are aware that examined the predictive accuracy of the COMPAS for different ethnic groups, and that study reported much weaker results for African
American men (Fass, Heilbrun, DeMatteo, & Fretz, 2008). In predicting rearrest within 1
year of release, Fass et al. (2008) reported AUCs for the COMPAS Recidivism Risk Scale
of .81 for Whites, .67 for Hispanics, .48 for African Americans, and .53 for the total sample assessed with COMPAS (N = 276). However, their study has at least one major weakness that renders its findings unreliable. Their small overall sample size and base rates
resulted in extremely small effective sample sizes for the ethnic groups (African American =
36, Hispanic = 4, White = 1), and this almost ensures unreliable results.
Downloaded from http://cjb.sagepub.com at DENVER UNIV on December 12, 2008

 34

CRIMINAL JUSTICE AND BEHAVIOR

PREDICTIVE ACCURACY ACROSS DIVERSE CRIMINAL JUSTICE POPULATIONS

The present study did not include other specific criminal justice populations from prison,
parole, community corrections, or jails. However, a current, large-scale, parole reentry
study is evaluating predictive performance for parolees released to the community (Zhang,
Farabee, & Roberts, 2007). Preliminary findings reported AUCs of .67 rising to .72 with
only minor adjustments to the COMPAS recidivism risk model.
In conclusion, given that instrument validation is an ongoing process, we acknowledge
that numerous further tests and models could be applied to examine the predictive validity
of COMPAS risk models. The present results, however, are encouraging and suggest that
the COMPAS risk models reach levels of reliability, predictive validity, and generalizability that are at least equal to those of other major instruments in offender risk assessment.

APPENDIX
Scale Content and Selection: Theoretical and Empirical Background

This appendix describes the item content and research background for each Correctional Offender
Management Profiling for Alternative Sanctions (COMPAS) scale listed in Table 1. Full details of
psychometric properties, theoretical justification, and supportive empirical studies for each scale are
given in Brennan, Dieterich, and Oliver (2007). We now describe the background of the COMPAS
scales, along with the main items loading on each scale and factor loadings (in parentheses).
Criminal Involvement. This scale includes items pertaining to number of prior arrests and convictions, frequency of incarceration, and criminal justice involvements. The highest loading item is
total number of prior arrests (.52). Past criminal involvement has been consistently supported by
meta-analytic studies as a major risk factor for predicting ongoing criminal behavior (Andrews &
Bonta, 1998; Gendreau, Little, & Goggin, 1996).
History of Violence. This scale includes official history items reflecting prior arrests and convictions for violent felonies, use of weapons, infractions for fighting, and so on. The highest loading
items are the number of prior assaultive felony convictions (.47) and frequency of injury to victims
(.40). The research literature has indicated that the likelihood of future violence appears to increase
with each instance of a prior violent incident (Farrington, 1991; Lipsey & Derzon, 1998; Parker &
Asher, 1987).
History of Noncompliance. This scale includes official items reflecting failures to appear, failures of
drug tests, failures to comply with sentencing conditions, revocations for technical reasons, and so
forth. High-loading items include the number of probation revocations (.56) and prior failures to appear
(.37). Repeated noncompliance with criminal justice conditions and treatment regimes has emerged as
a predictor of both violent and general recidivism (Stalans, Yarnold, Seng, Olsen, & Repp, 2004).
Criminal Associates. This scale assesses associations with others who are involved in drugs, criminal activity, and gangs. High-loading items include friends who have been arrested (.41) and friends
who have been gang members (.43). This construct is of central theoretical importance in both social
learning and subcultural theories of crime (Cullen & Agnew, 2003; Elliott, Huizinga, & Ageton,
1985). Meta-analytic research has consistently shown that having antisocial associates is a major risk
factor for recidivism (Gendreau et al., 1996).
(continued)

Downloaded from http://cjb.sagepub.com at DENVER UNIV on December 12, 2008

 Brennan et al. / EVALUATING COMPAS

35

APPENDIX (continued)
Substance Abuse. The central items in this scale reflect the influence of alcohol or drugs on the
current offense (.40), perceived benefit of substance abuse treatment (.41), and prior substance abuse
treatment (.37). Drug use has consistently emerged as a significant risk factor for general criminal
and violent behavior (National Institute of Justice, 1999; National Research Council, 1993) and is a
major risk factor in meta-analytic studies (Gendreau et al., 1996).
Financial Problems and Poverty. This scale includes items such as worry about financial survival
(.53), problems paying bills (.52), and not enough money to get by (.52). Although poverty has shown
only modest predictive power in meta-analytic studies (Gendreau et al., 1996), decades of research has
shown reliable associations between poverty and high crime rates and related risk factors such as unstable residence, family disruption, single-parent families, community disorganization, and substandard
housing (National Research Council, 1993; Sampson & Lauritsen, 1994). Theoretically, poverty is a
key factor in strain or social marginalization and subcultural theories (Cullen & Agnew, 2003).
Occupational and Educational Resources or Human Capital. This scale includes items reflecting
levels of educational and vocational occupational success such as job skills (.43), current unemployment (.52), low wages (.49), and employment history (.39). A social achievement (human capital) scale
was selected for both empirical and theoretical reasons (Coleman, 1990; Gendreau et al., 1996; Hagen,
1998). It is central to strain theory because people with lower social capital have fewer life chances and
more restricted opportunities than do those with greater capital. The scale is dynamic because human
capital can be built or destroyed. Job loss or high school dropout may lower economic and social
opportunities, whereas completing job skills training or obtaining a GED may increase these chances.
Family Crime. The COMPAS scale of family criminality includes items assessing the criminality
and drug use of mother, father, and siblings. The highest loading items include parent ever jailed
(.55), parent has had drug problems (.43), and mother ever arrested (.40). Many empirical studies
have linked delinquency and adult crime to antisocial families (Farrington, Jolliffe, Loeber,
Stouthamer-Loeber, & Kalb, 2001; Farrington & West, 1993; Lykken, 1995). Social learning theory
has also linked deviant behavior to violent or criminal family role models and ineffective parenting
(Lykken, 1995). Heritability theories have included gene-based evolutionary theories (Ellis & Walsh,
1997) and biosocial theories of sociopathy (Mealey, 1995).
High Crime Neighborhood. This scale assesses levels of crime (.44), gang activity (.40), and drug
activity (.39) in the person’s neighborhood. Living in a high-crime neighborhood is an established
correlate of both delinquency and adult crime (Sampson & Lauritsen, 1994; Thornberry, Huizinga,
& Loeber, 1995). It plays a role in social disorganization, social learning, and subcultural theories
(Cullen & Agnew, 2003; Sampson, Raudenbush, & Earls, 1997).
Boredom and Lack of Constructive Leisure Activities (Aimlessness). This scale includes items
from two closely linked themes: boredom proneness and lack of engaging leisure activities.
Dominant items include often bored (.46), nothing to do (.47), restless with current activities (.47),
and scattered attention (.37). Although conceptually different, items from these two themes all load
on a single factor. Theoretically, an absence of constructive leisure activities partially reflects weak
engagement bonds of early social control theory (Hirschi, 1969), and a similar concept (idle hands)
enters routine activities theory (Osgood et al., 1996). Finally, restlessness, distractibility, and attention problems enter M. R. Gottfredson and Hirschi’s (1990) low self-control theory and Hare’s
(1991) related construct of psychopathy.
(continued)
Downloaded from http://cjb.sagepub.com at DENVER UNIV on December 12, 2008

 36

CRIMINAL JUSTICE AND BEHAVIOR

APPENDIX (continued)
Residential Instability. The present scale includes items assessing the number of recent moves
(.39), homelessness (.33), and absence of a verifiable address (.34). The background literature indicates that transience is often associated with poverty, poor housing, social disorganization, and crime
(McNaughton, 2007; National Research Council, 1993). Theoretically, transience and homelessness
may weaken social ties and have been associated with family breakup, social exclusion, and stressful life events (Marris, 1987). Theoretically, it plays a role in both social control theory (weakening
or attenuating social bonds) and strain theory (poverty, personal stress, marginalization). We note
that personal stress or distress has emerged as a risk factor with modest predictive validity for recidivism (Gendreau et al., 1996).
Social Isolation Versus Social Support. This scale captures social isolation at one pole and social
supports at the other. It includes items indicating self-reported loneliness (.33), absence of friends
(.40), feeling left out of things (.33), and no close or best friend (.37). Social support theory suggests
that even in high-risk environments social supports may mediate or buffer the criminogenic effects
of economic and social strain (Bennett & Morabito, 2006; Estroff, Zimmer, Lachicotte, & Benoit,
1994; National Research Council, 1993; Stevenson, 1998). In addition, at prison release and reentry
to society, prisoners with stronger social and family supports are found to have lower recidivism
(Solomon, Johnson, Travis, & McBride, 2004). Theoretically, this factor enters both strain theory
(buffering strain) and social control theory (reflecting social bonds).
Criminal Attitude. This scale assesses antisocial attitudes using items that may justify, excuse, or
minimize damage caused by the offender’s crime. Prominent items include the law does not help average people (.32), minor offenses such as drug use don’t hurt anyone (.24), and things stolen from rich
people won’t be missed (.36). Antisocial attitudes have emerged in meta-analytic studies as a major
risk factor (Andrews, Bonta, & Wormith, 2006; Gendreau et al., 1996). There is less agreement on the
particular attitudes that are most useful or predictive (e.g., minimizing the damage of offenses, tolerance for law violation, etc.; Walters, 1995). In the absence of consensus, COMPAS uses a higher order
scale with items adapted from Bandura, Barbaranelli, Caprara, and Pastorelli (1996).
Antisocial Personality. This scale addresses impulsivity, absence of guilt, selfish narcissism, dominance, risk taking, and anger or hostility. Representative items include short temper (.39), often does
things without thinking (.30), and seen as cold and callous (.32). It is not designed as a comprehensive
coverage of all personality subfactors but as a short, broadband scale using only the first principle component from a larger battery of antisocial personality items adapted from Eysenck and Eysenck (1978)
and Bandura et al. (1996). Important subfactors are impulsivity, risk taking, restlessness or boredom,
absence of guilt (callousness), selfish narcissism, interpersonal dominance, and anger or hostility
(Bandura et al., 1996; Cooke, Forth, & Hare, 1996; Lilienfeld & Andrews, 1996; Marcus, 2003;
Widiger & Lynum, 1998). Empirical support for antisocial personality (variously measured) is found
in Gendreau et al. (1996), Bandura et al. (1996), Blackburn and Coid (1998), and Quinsey, Harris, Rice,
and Cormier (1998). Theoretically, antisocial personality plays an important role in theories of antisocial dispositions (Farrington, 2003; M. R. Gottfredson & Hirschi, 1990; Hare, 1991).

REFERENCES
Aalen, O. O. (1978). Nonparametric inference for a family of counting processes. Annals of Statistics, 6, 701-726.
Andrews, D. A., & Bonta, J. (1995). The Level of Service Inventory–Revised. Toronto: Multi-Health Systems.
Andrews, D. A., & Bonta, J. (1998). The psychology of criminal conduct (2nd ed.). Cincinnati, OH: Anderson.

Downloaded from http://cjb.sagepub.com at DENVER UNIV on December 12, 2008

 Brennan et al. / EVALUATING COMPAS

37

Andrews, D. A., Bonta, J., & Wormith, J. S. (2004). The Level of Service/Case Management Inventory (LS/CMI). Toronto:
Multi-Health Systems.
Andrews, D. A., Bonta, J., & Wormith, J. S. (2006). The recent past and near future of risk and/or needs assessment. Crime
& Delinquency, 52, 7-27.
Aos, S., & Barnoski, R. (2003). Washington’s offender accountability act: An analysis of the Department of Correction’s risk
assessment (Document No. 03-12-1202). Washington, DC: Washington Institute for Public Policy.
Austin, J. (1983). Assessing the new generation of prison classification models. Crime & Delinquency, 29, 561-576.
Bandura, A., Barbaranelli, C., Caprara, G. V., & Pastorelli, C. (1996). Mechanisms of moral disengagement in the exercise
of moral agency. Journal of Personality and Social Psychology, 71, 364-374.
Barbaree, H. E., Seto, M., Langton, C. M., & Peacock, E. J. (2001). Evaluating the predictive accuracy of six risk assessment
instruments for adult sex offenders. Criminal Justice and Behavior, 28, 490-521.
Barnoski, R. (2006). Sex offender sentencing in Washington State: Predicting recidivism based on the LSI-R. Retrieved
January 20, 2008 from http://www.wsipp.wa.gov/rtpfiles/06-02-1201.pdf
Barnoski, R., & Aos, S. (2003). Washington’s offender accountability act: An analysis of the Department of Corrections’ risk
assessment (Document No. 03-12-1202). Olympia: Washington State Institute for Public Policy.
Barnoski, R., & Drake, E. K. (2007). Washington’s Offender Accountability Act: Department of Corrections’ static risk
assessment. Olympia: Washington State Institute for Public Policy.
Bennett, R. R., & Morabito, M. S. (2005, November). Institutional social support and crime: A cross-national investigation.
Paper presented at the annual meeting of the American Society of Criminology, Toronto, Canada.
Blackburn, R., & Coid, J. W. (1998). Psychopathy and the dimensions of personality in violent offenders. Personality and
Individual Differences, 25, 129-145.
Blanchette, K., & Brown, S. L. (2006). The assessment and treatment of women offenders: An integrative perspective. New
York: John Wiley.
Bloom, B. (2000). Beyond recidivism: Perspectives on evaluation of programs for female offenders in community corrections. In M. McMahon (Ed.), Assessment to assistance: Programs for women in community corrections (pp. 107-138).
Latham, MD: American Correctional Association.
Bonta, J. (1996). Risk-needs assessment and treatment. In A. T. Harland (Ed.), Choosing correctional options that work:
Defining the demand and evaluating the supply (pp. 18-32). Thousand Oaks, CA: Sage.
Bonta, J. (2002). Risk needs assessment: Guidelines for selection and use. Criminal Justice and Behavior, 29, 355-379.
Boothby, J. L., & Clements, C. B. (2000). A national survey of correctional psychologists. Criminal Justice and Behavior, 27,
715-731.
Brennan, T. (1987). Classification: An overview of selected methodological issues. In D. M. Gottfredson & M. Tonry (Eds.),
Prediction and classification: Criminal justice decision making (pp. 201-248). Chicago: University of Chicago Press.
Brennan, T. (2008a). Explanatory diversity among female delinquents: Examining taxonomic heterogeneity. In R. Zaplin
(Ed.), Female crime and delinquency: Critical perspectives and effective interventions (pp. 197-232). Boston: Jones and
Bartlett.
Brennan, T. (2008b). Institutional assessment and classification of female offenders: From robust beauty to person-centered
assessment. In R. Zaplin (Ed.), Female crime and delinquency: Critical perspectives and effective interventions (pp. 283322). Boston: Jones and Bartlett.
Brennan, T., Breitenbach, M., & Dieterich, W. (2008). Towards an explanatory taxonomy of adolescent delinquents:
Identifying several social-psychological profiles. Journal of Quantitative Criminology, 24, 179-203.
Brennan, T., Dieterich, W., & Oliver, W. (2004). The COMPAS scales: Normative data for males and females. Community
and incarcerated samples. Traverse City, MI: Northpointe Institute for Public Management.
Brennan, T., Dieterich, W., & Oliver, W. (2007). COMPAS: Correctional offender management for alternative sanctioning.
Technical manual and psychometric report (V. 5.01). Traverse City, MI: Northpointe Institute for Public Management.
Brennan, T., Wells, D., & Alexander, J. (2004). Enhancing prison classification systems: The emerging role of management
information systems. Washington, DC: U.S. Department of Justice, National Institute of Corrections.
Clements, C. B. (1996). Offender classification: Two decades of progress. Criminal Justice and Behavior, 23, 121-143.
Coleman, J. S. (1990). Foundations of social theory. Cambridge, MA: Harvard University Press.
Cooke, D. J., Forth, A. E., & Hare, R. D. (1996). Psychopathy: Theory, research and implications for society. Dordrecht,
Netherlands: NATO Science Series.
Cullen, F., & Agnew, R. (2003). Criminological theory: Past to present. Los Angeles: Roxbury.
Dahle, K. P. (2006). Strengths and limitations of actuarial prediction of criminal re-offence in a German prison sample: A
comparative study of LSI-R, HCR-20 and PCL-R. International Journal of Law and Psychiatry, 29(5), 431-442.
Dawes, R. M. (1979). The robust beauty of improper linear models in decision models. American Psychologist, 34, 571-582.
Elliott, D. S., Huizinga, D., & Ageton, S. S. (1985). Explaining delinquency and drug use. Beverly Hills, CA: Sage.
Ellis, L., & Walsh, A. (1997). Gene-based evolutionary theories in criminology. Criminology, 35, 229-276.
Estroff, S., Zimmer, C., Lachicotte, W., & Benoit, J. (1994). The influence of social networks and social support on violence
by persons with serious mental illness. Hospital and Community Psychiatry, 45, 669-679.

Downloaded from http://cjb.sagepub.com at DENVER UNIV on December 12, 2008

 38

CRIMINAL JUSTICE AND BEHAVIOR

Eysenck, S., & Eysenck, H. (1978). Impulsiveness and venturesomeness: Their position in a dimensional system of personality description. Psychological Reports, 43, 1247-1255.
Farr, K. A. (2000). Classification for female inmates: Moving forward. Crime & Delinquency, 46, 3-17.
Farrington, D. (1991). Childhood aggression and adult violence: Early precursors and later life outcomes. In D. Pepler &
K. Rubin (Eds.), The development and treatment of childhood aggression (pp. 5-29). Hillsdale, NJ: Lawrence Erlbaum.
Farrington, D. P. (2003). Developmental and life course criminology: Key theoretical and empirical issues. Criminology, 41,
221-255.
Farrington, D. P., Jolliffe, D., Loeber, R., Stouthamer-Loeber, M., & Kalb, L. M. (2001). The concentration of offenders in
families, and family criminality in the prediction of boys’ delinquency. Journal of Adolescence, 24, 579-596.
Farrington, D. P., & West, D. J. (1993). Criminal, penal, and life histories of chronic offenders: Risk and protective factors
and early identification. Criminal Behaviour and Mental Health, 3, 492-523.
Fass, T. L., Heilbrun, K., DeMatteo, D., & Fretz, R. (2008). The LSI-R and the COMPAS: Validation on two risk-needs tools.
Criminal Justice and Behavior, 35, 1095-1108.
Flores, A. W., Lowenkamp, C. T., Smith, P., & Latessa, E. J. (2006). Validating the Level of Service Inventory–Revised on a
sample of federal probationers. Federal Probation, 70(2), 44-48.
Gendreau, P., Little, T., & Goggin, C. (1996). A meta-analysis of the predictors of adult offender recidivism: What works!
Criminology, 34, 575-607.
Glaser, D. (1987). Classification for risk. In D. M. Gottfredson & M. Tonry (Eds.), Prediction and classification: Criminal
justice decision making (pp. 249-292). Chicago: University of Chicago Press.
Gottfredson, M. R., & Hirschi, T. (1990). A general theory of crime. Stanford, CA: Stanford University Press.
Gottfredson, S. D. (1987). Prediction: An overview of selected methodological issues. In D. M. Gottfredson & M. Tonry
(Eds.), Prediction and classification: Criminal justice decision making (pp. 21-52). Chicago: University of Chicago Press.
Grann, M., Belfrage, H., & Tengstrom, A. (2000). Actuarial assessment of risk for violence: Predictive validity of the VRAG
and the historical part of the HCR-20. Criminal Justice and Behavior, 27, 97-114.
Grove, W. M., & Meehl, P. E. (1996). Comparative efficiency of informal (subjective, impressionistic) and formal (mechanical,
algorithmic) prediction procedures: The clinical-statistical controversy. Psychology, Public Policy and Law, 2, 293-323.
Hagen, J. (1998). Life course capitalization and adolescent behavioral development. In R. Jessor (Ed.), New perspectives on
adolescent risk behavior. Cambridge, MA: Cambridge University Press.
Hannah-Moffatt, K., & Shaw, M. (2001). Taking risks: Incorporating gender and culture into classification and assessment
of federally sentenced women in Canada. Ottawa, Ontario: Status of Women Canada.
Hardyman, P., & Van Voorhis, P. (2004). Developing gender-specific classification systems for women offenders. Washington,
DC: U.S. Department of Justice, National Institute of Corrections.
Hare, R. D. (1991). The Hare Psychopathy Checklist–Revised. Toronto: Multi-Health Systems.
Harrell, F. E., Califf, R. M., Pryor, D. B., Lee, K. L., & Rosati, R. A. (1982). Evaluating the yield of medical tests. Journal
of the American Medical Association, 247, 2543-2546.
Harris, P. W., & Jones, P. R. (1999). Differentiating delinquent youths for program planning and evaluation. Criminal Justice
and Behavior, 26, 403-434.
Hastie, R., & Dawes, R. M. (2001). Rational choice in an uncertain world: The psychology of judgment and decision-making.
Thousand Oaks, CA: Sage.
Hirschi, T. (1969). Causes of delinquency. Berkeley: University of California Press.
Hoffman, P. B. (1994). Twenty years of operational use of a risk prediction instrument: The United States Parole
Commission’s Salient Factor Score. Journal of Criminal Justice, 22, 477-494.
Holtfreter, K., & Cupp, R. (2007). Gender and risk assessment: The empirical status of the LSI-R for women. Journal of
Contemporary Criminal Justice, 23, 363-382.
Jones, P. R. (1996). Risk prediction in criminal justice. In A. T. Harland (Ed.), Choosing correctional options that work
(pp. 33-68). Thousand Oaks, CA: Sage.
Kalbfleisch, J. D., & Prentice, R. L. (2002). The statistical analysis of failure time data (2nd ed.). New York: John Wiley.
Kroner, D. G., Stadtland, M., Eidt, M., & Nedopil, N. (2007). The validity of the Violence Risk Appraisal Guide (VRAG) in
predicting criminal recidivism. Criminal Behaviour and Mental Health, 17, 89-100.
Lilienfeld, S. O., & Andrews, B. P. (1996). Development and preliminary validation of a self-report measure of psychopathic
personality traits in a non-criminal population. Journal of Personality Assessment, 66, 488-524.
Lipsey, M. W., & Derzon, J. H. (1998). Predictors of violent or serious delinquency in adolescence and early adulthood: A
synthesis of longitudinal research. In R. Loeber & D. P. Farrington (Eds.), Serious and violent juvenile offenders: Risk
factors and successful interventions (pp. 86-105). Thousand Oaks, CA: Sage.
Lösel, F. (1995). The efficacy of correctional treatment: A review and synthesis of meta-evaluations. In J. McGuire (Ed.),
What works: Reducing re-offending (pp. 79-111). Chichester, UK: Wiley.
Lykken, D. T. (1995). The antisocial personalities. Mahwah, NJ: Lawrence Erlbaum.
Marcus, B. (2003). An empirical examination of the construct validity of two alternative self-control measures. Educational
and Psychological Measurement, 63, 674-706.
Marris, P. (1987). Loss and change. New York: Pantheon.

Downloaded from http://cjb.sagepub.com at DENVER UNIV on December 12, 2008

 Brennan et al. / EVALUATING COMPAS

39

McNaughton, C. (2007, September). Life on the edge: Substance abuse and homelessness—Escape, resistance or deviance.
Paper presented at the annual meeting of the British Society of Criminology, Glasgow, UK.
Mealey, L. (1995). The sociobiology of sociopathy: An integrated evolutionary model. Behavioral and Brain Sciences, 18,
523-599.
Meier, S. (2002). Bridging case conceptualization, assessment and intervention. Thousand Oaks, CA: Sage.
Millon, T., & Davis, R. D. (1997). The place of assessment in clinical science. In T. Millon (Ed.), The Millon inventories:
Clinical and personality assessment (pp. 3-22). New York: Guilford.
Mossman, D. (1994). Assessing predictions of violence: Being accurate about accuracy. Journal of Consulting and Clinical
Psychology, 62, 783-792.
National Council on Crime and Delinquency. (2006). Correctional Assessment and Intervention System. Retrieved March 28,
2007, from http://www.nccd-crc.org/nccd/n_index_main.html
National Institute of Justice. (1999). Annual report on drug use among adult and juvenile arrestees (1998). Arrestee Drug
Abuse Monitoring Program (ADAM). Washington, DC: Author.
National Research Council. (1993). Understanding and preventing violence (A. J. Reiss & J. A. Roth, Eds.). Washington, DC:
National Academy Press.
Nelson, W. (1972). Theory and applications of hazard plotting for censored failure data. Technometrics, 14, 945-965.
Northpointe Institute for Public Management. (1996). COMPAS [Computer software]. Traverse City, MI: Author.
Osgood, D., Wayne, J. K., Wilson, J. G., Bachman, P., O’Malley, G., & Johnston, L. D. (1996). Routine activities and individual deviant behavior. American Sociological Review, 61, 635-655.
Palmer, T. (1992). The re-emergence of correctional intervention. Newbury Park, CA: Sage.
Parker, J., & Asher, S. (1987). Peer relations and later personal adjustment: Are low accepted children at risk? Psychological
Bulletin, 102, 357-389.
Quinsey, V. L., Harris, G. T., Rice, M. E., & Cormier, C. A. (1998). Violent offenders: Appraising and managing risk.
Washington, DC: American Psychological Association.
Reisig, M. D., Holtfreter, K., & Morash, M. (2006). Assessing recidivism risk across female pathways to crime. Justice
Quarterly, 23, 384-405.
Salisbury, E. J., Van Voorhis, P., & Wright, E. (2006, November). Construction and validation of a gender responsive
risk/needs instrument for women offenders in Missouri and Maui. Paper presented at the annual conference of the
American Society of Criminology, Los Angeles.
Sampson, R. J., & Lauritsen, J. (1994). Deviant lifestyles proximity to crime and the offender-victim link in personal violence. Journal of Research in Crime and Delinquency, 27, 7-40.
Sampson, R. J., Raudenbush, S., & Earls, F. (1997). Neighborhoods and violent crime: A multilevel study of collective efficacy. Science, 277, 918-924.
Solomon, A., Johnson, K. D., Travis, J., & McBride, E. (2004). From prison to work: The employment dimensions of prisoner reentry. Washington, DC: Urban Institute.
Stalans, L. J., Yarnold, P. R., Seng, M., Olsen, D., & Repp, M. (2004). Identifying three types of violent offenders and predicting their recidivism and performance while on probation: A classification tree analysis. Law and Human Behavior, 26,
253-271.
Stevenson, H. C. (1998). Raising safe villages: Cultural-ecological factors that influence the emotional adjustment of adolescents. Journal of Black Psychology, 24, 44-59.
Tape, T. G. (2003). Interpreting diagnostic tests: The area under the ROC curve. Unpublished report, University of Nebraska
Medical Center, Omaha.
Thornberry, T., Huizinga, D., & Loeber, R. (Eds.). (1995). The prevention of serious delinquency and violence: Implications
from the program of research on the causes and correlates of delinquency. Sourcebook on serious, violent and chronic
juvenile offenders. Thousand Oaks, CA: Sage.
Van Voorhis, P. (1994). Psychological classification of the adult male prison inmate. Albany: State University of New York
Press.
Walters, G. D. (1995). The Psychological Inventory of Criminal Thinking Styles: Part I. Reliability and preliminary validity.
Criminal Justice and Behavior, 22, 307-325.
Ward, T., & Brown, M. (2004). The good lives model and conceptual issues in offender rehabilitation. Psychology, Crime and
Law, 10, 243-257.
Ward, T., & Stewart, C. (2003). Criminogenic needs and human needs: A theoretical model. Psychology, Crime and Law, 9,
125-143.
Warren, M. Q., & Hindelang, M. J. (1979). Differential explanation of offender behavior. In H. Toch (Ed.), Psychology of
crime and criminal justice (pp. 166-182). Prospect Heights, IL: Waveland Press.
Widiger, T. A., & Lynum, D. R. (1998). Psychopathy and the five-factor model of personality. In T. Millon, E. Simonsen,
M. Birkett-Smith, & R. Davis (Eds.), Psychopathy: Antisocial, criminal and violent behavior (pp. 171-186). New York:
Guilford.
Wormith, J. S. (2001, July). Assessing offender assessment: Contributing to effective correctional treatment. The ICCA
Journal, pp. 12-23.

Downloaded from http://cjb.sagepub.com at DENVER UNIV on December 12, 2008

 40

CRIMINAL JUSTICE AND BEHAVIOR

Zhang, S., Farabee, D., & Roberts, R. (2007, October). Predicting parolee risk of recidivism. Paper presented at the 66th semiannual meeting of the Association for Criminal Justice Research, Sacramento, CA.
Tim Brennan, PhD, is a senior research scientist at Northpointe Institute. His main research interests include risk assessment, pattern recognition, classification, and machine learning in the context of crime and delinquency.
William Dieterich, PhD, is director of research at Northpointe Institute. His research interests include developing and testing prognostic models for use in criminal justice agencies.
Beate Ehret, PhD, is a research analyst at Northpointe Institute. Her research interests include gender and crime, juvenile
delinquency, and evidence-based practice in criminal justice.

Downloaded from http://cjb.sagepub.com at DENVER UNIV on December 12, 2008