UNIVERSITY OF CINCINNATI
May 23, 2008
Date:___________________
Brenda Vose
I, _________________________________________________________,
hereby submit this work as part of the requirements for the degree of:
Doctorate of Philosophy (Ph.D.)
in:
Criminal Justice
It is entitled:
Assessing the Predictive Validity of the Level of Service Inventory-
Revised: Recidivism Among Iowa Parolees and Probationers
This work and its defense approved by:
Francis T. Cullen, Ph.D.
Chair: _______________________________
Christopher Lowenkamp, Ph.D.
_______________________________
Paula Smith, Ph.D.
_______________________________
Melissa Moon, Ph.D.
_______________________________
_______________________________
Assessing the Predictive Validity of the Level of Service Inventory-Revised:
Recidivism Among Iowa Parolees and Probationers
A Dissertation Submitted to the:
Division of Research and Advanced Studies
Of the University of Cincinnati
In Partial Fulfillment of the
Requirements for the Degree of
Doctorate of Philosophy (Ph.D.)
In the Division of Criminal Justice
Of the College of Education, Criminal Justice, and Human Services
2008
by
Brenda Vose, M.A.
B.A., University of Northern Iowa, 1998
M.A. Wichita State University, 2000
Dissertation Committee: Francis T. Cullen, Ph.D. (Chair)
Christopher T. Lowenkamp, Ph.D.
Melissa M. Moon, Ph.D.
Paula Smith, Ph.D.
ABSTRACT
As of 2005, there were approximately 5 million offenders on probation or parole (Glaze
& Bonczar, 2006). Correctional agencies are responsible for managing and treating the
individual risks and needs of the offenders under their supervision. In this context, researchers
have developed classification instruments to aid in the identification of offender risks and needs.
The Level of Service Inventory-Revised (LSI-R) is perhaps the most prominent among these
instruments.
Using a sample of 2,849 probationers and parolees from the state of Iowa, the current
dissertation attempts to contribute to the research on the LSI-R in three ways. First, this study
examines the predictive validity of the LSI-R at time 1 and time 2. Second, this dissertation will
consider the impact that change in LSI-R total score from time 1 to time 2 has on the predictive
validity of the instrument. Third, the predictive validity of the ten domains of LSI-R will be
tested to determine which domain is the most powerful predictor of recidivism and whether a
change score in one domain is more or less important than a change score in the other domains.
The results indicate that the LSI-R is a valid predictor of recidivism at time 1 and time 2.
Moreover, change in total LSI-R score does impact the ability of the instrument to predict
recidivism. Finally, no single domain is more or less important than the others. Policy
implications and direction for future research are discussed.
iii
iv
ACKNOWLEDGEMENTS
I would like to take this opportunity to thank a number of people who have helped see me
through this academic adventure.
First and foremost, thank you to my boss, mentor, and sherpa, Dr. Francis T. Cullen. I truly
appreciate the opportunities, support, motivation, and advice you have given me. I am forever
grateful.
I would like to thank my committee Dr. Francis T. Cullen, Dr. Christopher Lowenkamp, Dr.
Paula Smith, and Dr. Melissa Moon for their valuable comments and direction on this project. I
would like to especially acknowledge Dr. Christopher “Data King” Lowenkamp; without whom
this project would not have been possible.
Special thanks to Dr. James Frank, Dr. Edward Latessa, Dr. Lawrence Travis, Dr. Quint
Thurman, and Dr. Delores Craig-Moreland. You have each been instrumental in shaping my
view on what it means to be an academic. I respect and admire you all.
Thank you to Kristie Blevins, Luahna Winningham Carter, Jim Rauch, Fawn Mitchell and all
distance learning facilitators past and present. Working with you has been a wonderful and I
wish you all the best.
Thanks to my UC friends and colleagues Shamir Ratansi, Dave Carter, Georgia Spiropoulos,
John Schwartz, Denise Nation, Marie Skubak Tillyer, Rob Tillyer, Rebecca Schnupp, Emily
Salisbury, Brian Lovins, and Lori Lovins for sharing in this crazy and rewarding experience.
Jeff Tymony, thank you for taking me under your wing in Wichita and for changing the way I
think about criminal justice and public policy.
Laura Bainbridge, thank you for accidentally introducing me to the research process and to Judy
McDowell for making research fun. Who knew that a simple summer job would turn into this?
Thank you to the Girl Scouts of Shining Trail Council for hosting a mystery event when I was in
middle school that first sparked my interest in crime and criminal justice.
Thank you to my friends, relatives, and the town of Mediapolis. Your support is appreciated.
Last but certainly not least, thank you to my parents for always encouraging education and for
being there every step of the way. I share this accomplishment with you.
v
TABLE OF CONTENTS
TABLE OF CONTENTS ........................................................................................................... VI
CHAPTER 1 .................................................................................................................................. 1
USING THE LEVEL OF SERVICE INVENTORY-REVISED TO ASSESS OFFENDERS
......................................................................................................................................................... 1
PROBATION AND PAROLE IN THE UNITED STATES ..................................................... 3
THE ORIGINS OF COMMUNITY SUPERVISION: THE RISE OF REHABILITATION .............................. 3
THE ATTACK ON REHABILITATION .............................................................................................. 7
THE RISE AND LIMITS OF THE COMMUNITY CONTROL MODEL .................................................. 10
THE REVIVAL OF REHABILITATION ............................................................................... 13
EMPIRICAL SUPPORT FOR REHABILITATION: RESPONDING TO MARTINSON ............................... 14
PRINCIPLES OF EFFECTIVE CORRECTIONAL TREATMENT ........................................................... 19
OFFENDER CLASSIFICATION ............................................................................................. 24
GENERATIONS OF RISK ASSESSMENTS ....................................................................................... 25
THE DEVELOPMENT OF THE LSI: PUTTING EFFECTIVE CORRECTIONAL TREATMENT INTO
PRACTICE ................................................................................................................................... 31
PREDICTIVE VALIDITY OF THE LEVEL OF SUPERVISION/LEVEL OF SERVICE INVENTORY ........... 33
RESEARCH STRATEGY ......................................................................................................... 47
CHAPTER 2 ................................................................................................................................ 49
METHODS .................................................................................................................................. 49
SAMPLE...................................................................................................................................... 49
INDEPENDENT VARIABLES ................................................................................................. 50
DEPENDENT VARIABLE........................................................................................................ 51
STATISTICAL TECHNIQUES ................................................................................................ 51
LIMITATIONS OF THE STUDY ............................................................................................ 52
CHAPTER 3 ................................................................................................................................ 56
RESULTS .................................................................................................................................... 56
THE IMPACT OF THE LSI-R ON RECIDIVISM ................................................................ 56
BIVARIATE ANALYSIS TIME 1 AND TIME 2 FOR SAMPLE ............................................................ 56
MULTIVARIATE ANALYSIS TIME 1 AND TIME 2 FOR SAMPLE ..................................................... 58
CHANGE ANALYSIS FOR SAMPLE ............................................................................................... 66
THE IMPACT OF THE LSI-R ON RECIDIVISM BY GROUP .......................................... 74
BIVARIATE ANALYSIS TIME 1 AND TIME 2 GENDER .................................................................. 74
MULTIVARIATE ANALYSIS TIME 1 AND TIME 2 FOR GENDER .................................................... 76
CHANGE ANALYSIS FOR GENDER............................................................................................... 91
vi
BIVARIATE ANALYSIS TIME 1 AND TIME 2 FOR RACE .............................................................. 103
MULTIVARIATE ANALYSIS TIME 1 AND TIME 2 FOR RACE ....................................................... 105
CHANGE ANALYSIS FOR RACE ................................................................................................. 120
BIVARIATE ANALYSIS TIME 1 AND TIME 2 FOR SUPERVISION STATUS..................................... 131
MULTIVARIATE ANALYSIS TIME 1 AND TIME 2 FOR SUPERVISION STATUS ............................. 134
CHANGE ANALYSIS FOR SUPERVISION STATUS ........................................................................ 149
THE IMPACT OF DOMAINS OF THE LSI-R ON RECIDIVISM ....................................... 161
BIVARIATE ANALYSIS TIME 1 AND TIME 2 FOR SAMPLE .......................................................... 161
MULTIVARIATE ANALYSIS TIME 1 AND TIME 2 FOR SAMPLE ................................................... 164
CHANGE ANALYSIS FOR SAMPLE ............................................................................................. 165
THE IMPACT OF THE LSI-R DOMAINS ON RECIDIVISM BY GROUP .................... 173
BIVARIATE ANALYSIS TIME 1 AND TIME 2 FOR GENDER ......................................................... 173
MULTIVARIATE ANALYSIS TIME 1 AND TIME 2 FOR GENDER .................................................. 176
CHANGE ANALYSIS FOR GENDER............................................................................................. 177
BIVARIATE ANALYSIS TIME 1 AND TIME 2 FOR RACE .............................................................. 188
MULTIVARIATE ANALYSIS TIME 1 AND TIME 2 FOR RACE ....................................................... 191
CHANGE ANALYSIS FOR RACE ................................................................................................. 192
BIVARIATE ANALYSIS TIME 1 AND TIME 2 FOR SUPERVISION STATUS..................................... 203
MULTIVARIATE ANALYSIS TIME 1 AND TIME 2 FOR SUPERVISION STATUS ............................. 206
CHANGE ANALYSIS FOR SUPERVISION STATUS ........................................................................ 212
CHAPTER 4 .............................................................................................................................. 220
CONCLUSION: THE FUTURE OF THE LSI-R.................................................................. 220
SUMMARY OF RESULTS ..................................................................................................... 221
IMPLICATIONS FOR THE THEORY OF EFFECTIVE................................................... 225
CORRECTIONAL INTERVENTION ................................................................................... 225
POLICY IMPLICATIONS...................................................................................................... 226
FUTURE RESEARCH............................................................................................................. 229
REFERENCES.......................................................................................................................... 232
vii
TABLE OF TABLES
TABLE 1.1 SUMMARY FINDINGS FROM PREVIOUS LSI RESEARCH ...............................................................................35
TABLE 1.2 PREDICTIVE VALIDITY ACROSS CATEGORIES .............................................................................................40
TABLE 1.3 MEASURES OF RECIDIVISM ACROSS LSI STUDIES.......................................................................................45
TABLE 1.4 TYPES OF LSI INSTRUMENTS .......................................................................................................................46
TABLE 2.1 SAMPLE CHARACTERISTICS ........................................................................................................................54
TABLE 2.2 DOMAIN TOTALS .........................................................................................................................................55
TABLE 3.1 BIVARIATE CORRELATIONS TIME 1 AND 2 FOR SAMPLE..............................................................................57
TABLE 3.2 RISK CATEGORY AND RECIDIVISM TIME 1 FOR SAMPLE .............................................................................59
TABLE 3.3 RISK CATEGORY AND RECIDIVISM TIME 2 FOR SAMPLE .............................................................................60
TABLE 3.4 MULTIVARIATE TIME 1 FOR SAMPLE ..........................................................................................................61
TABLE 3.5 MULTIVARIATE TIME 2 FOR SAMPLE ..........................................................................................................62
TABLE 3.6 RISK CLASSIFICATION AND RECIDIVISM TIME 2..........................................................................................69
TABLE 3.7 DESCRIPTIVES ON PERCENT AND RAW CHANGE FOR SAMPLE .....................................................................70
TABLE 3.8 MULTIVARIATE SAMPLE PERCENT CHANGE ...............................................................................................71
TABLE 3.9 MULTIVARIATE SAMPLE RAW CHANGE ......................................................................................................72
TABLE 3.10 BIVARIATE CORRELATIONS TIME 1 AND TIME 2 FOR GENDER ..................................................................75
TABLE 3.11 RISK CATEGORY TIME 1 AND RECIDIVISM TIME 1 FOR MALES .................................................................77
TABLE 3.12 RISK CATEGORY TIME 1 AND RECIDIVISM TIME 1 FOR FEMALES..............................................................78
TABLE 3.13 RISK CATEGORY TIME 2 AND RECIDIVISM TIME 2 FOR MALES .................................................................79
TABLE 3.14 RISK CATEGORY TIME 2 AND RECIDIVISM TIME 2 FOR FEMALES..............................................................80
TABLE 3.15 MULTIVARIATE TIME 1 FOR MALES ..........................................................................................................83
TABLE 3.16 MULTIVARIATE TIME 1 FOR FEMALES.......................................................................................................84
TABLE 3.17 MULTIVARIATE TIME 2 FOR MALES ..........................................................................................................85
TABLE 3.18 MULTIVARIATE TIME 2 FOR FEMALES.......................................................................................................86
TABLE 3.19 RISK CLASSIFICATION TIME 1 AND TIME 2 AND RECIDIVISM TIME 2 FOR MALES .....................................93
TABLE 3.20 RISK CLASSIFICATION TIME 1 AND TIME 2 AND RECIDIVISM TIME 2 FOR FEMALES..................................94
TABLE 3.21 DESCRIPTIVES ON PERCENT AND RAW CHANGE FOR GENDER ..................................................................95
TABLE 3.22 MULTIVARIATE PERCENT CHANGE FOR MALES ........................................................................................96
TABLE 3.23 MULTIVARIATE PERCENT CHANGE FOR FEMALES ....................................................................................97
TABLE 3.24 MULTIVARIATE RAW CHANGE FOR MALES...............................................................................................98
TABLE 3.25 MULTIVARIATE RAW CHANGE FOR FEMALES ...........................................................................................99
TABLE 3.26 BIVARIATE CORRELATIONS TIME 1 AND TIME 2 FOR RACE.....................................................................104
TABLE 3.27 RISK CATEGORY TIME 1 AND RECIDIVISM TIME 1 FOR BLACKS .............................................................106
TABLE 3.28 RISK CATEGORY TIME 1 AND RECIDIVISM TIME 1 FOR WHITES ..............................................................107
TABLE 3.29 RISK CATEGORY TIME 2 AND RECIDIVISM TIME 2 FOR BLACKS .............................................................108
TABLE 3.30 RISK CATEGORY TIME 2 AND RECIDIVISM TIME 2 FOR WHITES ..............................................................109
TABLE 3.31 MULTIVARIATE TIME 1 FOR BLACKS ......................................................................................................112
TABLE 3.32 MULTIVARIATE TIME 1 FOR WHITES .......................................................................................................113
TABLE 3.33 MULTIVARIATE TIME 2 FOR BLACKS ......................................................................................................114
TABLE 3.34 MULTIVARIATE TIME 2 FOR WHITES .......................................................................................................115
TABLE 3.35 RISK CLASSIFICATION AND RECIDIVISM TIME 2 FOR BLACKS .................................................................122
TABLE 3.36 RISK CLASSIFICATION AND RECIDIVISM TIME 2 FOR WHITES .................................................................123
TABLE 3.37 DESCRIPTIVES ON PERCENT AND RAW CHANGE FOR RACE.....................................................................124
TABLE 3.38 MULTIVARIATE PERCENT CHANGE FOR BLACKS ....................................................................................125
TABLE 3.39 MULTIVARIATE PERCENT CHANGE FOR WHITES.....................................................................................126
TABLE 3.40 MULTIVARIATE RAW CHANGE FOR BLACKS ...........................................................................................127
TABLE 3.41 MULTIVARIATE RAW CHANGE FOR WHITES ...........................................................................................128
TABLE 3.42 BIVARIATE CORRELATIONS FOR SUPERVISION STATUS TIME 1 AND TIME 2 ...........................................132
TABLE 3.43 RISK CATEGORY TIME 1 AND RECIDIVISM TIME 1 FOR PROBATION ........................................................136
TABLE 3.44 RISK CATEGORY TIME 1 AND RECIDIVISM TIME 1 FOR PAROLE ..............................................................137
TABLE 3.45 RISK CATEGORY TIME 2 AND RECIDIVISM TIME 2 FOR PROBATION ........................................................138
TABLE 3.46 RISK CATEGORY TIME 2 AND RECIDIVISM TIME 2 FOR PAROLE ..............................................................139
TABLE 3.47 MULTIVARIATE TIME 1 FOR PROBATION .................................................................................................140
viii
TABLE 3.48 MULTIVARIATE TIME 1 FOR PAROLE .......................................................................................................141
TABLE 3.49 MULTIVARIATE TIME 2 FOR PROBATION .................................................................................................142
TABLE 3.50 MULTIVARIATE TIME 2 FOR PAROLE .......................................................................................................143
TABLE 3.51 RISK CLASSIFICATION AND RECIDIVISM TIME 2 FOR PROBATIONERS .....................................................151
TABLE 3.52 RISK CLASSIFICATION AND RECIDIVISM TIME 2 FOR PAROLEES .............................................................152
TABLE 3.53 DESCRIPTIVES ON PERCENT AND RAW CHANGE FOR SUPERVISION STATUS ...........................................153
TABLE 3.54 MULTIVARIATE PERCENT CHANGE FOR PROBATION ...............................................................................154
TABLE 3.55 MULTIVARIATE PERCENT CHANGE FOR PAROLE.....................................................................................155
TABLE 3.56 MULTIVARIATE RAW CHANGE FOR PROBATION .....................................................................................156
TABLE 3.57 MULTIVARIATE RAW CHANGE FOR PAROLE ...........................................................................................157
TABLE 3.58 BIVARIATE CORRELATIONS DOMAIN TOTALS AND RECIDIVISM TIME 1 FOR SAMPLE .............................162
TABLE 3.59 BIVARIATE CORRELATIONS DOMAIN TOTALS AND RECIDIVISM TIME 2 FOR SAMPLE .............................163
TABLE 3.60 MULTIVARIATE DOMAINS TIME 1 FOR SAMPLE ......................................................................................166
TABLE 3.61 MULTIVARIATE DOMAINS TIME 2 FOR SAMPLE ......................................................................................167
TABLE 3.62 DESCRIPTIVES PERCENT CHANGE FOR SAMPLE ......................................................................................169
TABLE 3.63 MULTIVARIATE DOMAINS PERCENT CHANGE FOR SAMPLE ....................................................................170
TABLE 3.64 DESCRIPTIVES RAW CHANGE DOMAINS FOR SAMPLE .............................................................................171
TABLE 3.65 MULTIVARIATE DOMAINS RAW CHANGE FOR SAMPLE ...........................................................................172
TABLE 3.67 BIVARIATE CORRELATIONS DOMAIN TOTALS AND RECIDIVISM TIME 1 AND TIME 2 FOR FEMALES .......175
TABLE 3.68 MULTIVARIATE DOMAINS TIME 1 FOR MALES ........................................................................................178
TABLE 3.69 MULTIVARIATE DOMAINS TIME 1 FOR FEMALES ....................................................................................179
TABLE 3.70MULTIVARIATE DOMAINS TIME 2 FOR MALES.........................................................................................180
TABLE 3.71 MULTIVARIATE DOMAINS TIME 2 FOR FEMALES ....................................................................................181
TABLE 3.72 DESCRIPTIVES PERCENT CHANGE DOMAINS FOR GENDER ......................................................................182
TABLE 3.73 MULTIVARIATE DOMAINS PERCENT CHANGE FOR MALES ......................................................................183
TABLE 3.74 MULTIVARIATE DOMAINS PERCENT CHANGE FOR FEMALES ..................................................................184
TABLE 3.75 DESCRIPTIVES RAW CHANGE DOMAINS FOR GENDER ............................................................................185
TABLE 3.76 MULTIVARIATE DOMAINS RAW CHANGE FOR MALES ............................................................................186
TABLE 3.77 MULTIVARIATE DOMAINS RAW CHANGE FOR FEMALES .........................................................................187
TABLE 3.78 BIVARIATE CORRELATIONS DOMAIN TOTALS AND RECIDIVISM TIME 1 AND TIME 2 FOR BLACKS .........189
TABLE 3.79 BIVARIATE CORRELATIONS DOMAIN TOTALS AND RECIDIVISM TIME 1 AND TIME 2 FOR WHITES .........190
TABLE 3.80 MULTIVARIATE DOMAINS TIME 1 FOR BLACKS ......................................................................................193
TABLE 3.81 MULTIVARIATE DOMAINS TIME 1 FOR WHITES.......................................................................................194
TABLE 3.82 MULTIVARIATE DOMAINS TIME 2 FOR BLACKS ......................................................................................195
TABLE 3.83 MULTIVARIATE DOMAINS TIME 2 FOR WHITES.......................................................................................196
TABLE 3.84 DESCRIPTIVES PERCENT CHANGE DOMAINS FOR RACE ..........................................................................197
TABLE 3.85 MULTIVARIATE DOMAINS PERCENT CHANGE FOR BLACKS ....................................................................198
TABLE 3.86 MULTIVARIATE DOMAINS PERCENT CHANGE FOR WHITES ....................................................................199
TABLE 3.87 DESCRIPTIVES RAW CHANGE DOMAINS FOR RACE .................................................................................200
TABLE 3.88 MULTIVARIATE DOMAINS RAW CHANGE FOR BLACKS ...........................................................................201
TABLE 3.89 MULTIVARIATE DOMAINS RAW CHANGE FOR WHITES ...........................................................................202
TABLE 3.90 BIVARIATE CORRELATIONS DOMAIN TOTALS AND RECIDIVISM TIME 1 AND TIME 2 FOR PROBATION....204
TABLE 3.91 BIVARIATE CORRELATIONS DOMAIN TOTALS AND RECIDIVISM TIME 1 AND TIME 2 FOR PAROLE .........205
TABLE 3.92 MULTIVARIATE DOMAINS TIME 1 FOR PROBATION .................................................................................208
TABLE 3.93 MULTIVARIATE DOMAINS TIME 1 FOR PAROLE ......................................................................................209
TABLE 3.94 MULTIVARIATE DOMAINS TIME 2 FOR PROBATION .................................................................................210
TABLE 3.95 MULTIVARIATE DOMAINS TIME 2 FOR PAROLE ......................................................................................211
TABLE 3.96 PERCENT CHANGE DOMAINS FOR SUPERVISION STATUS ........................................................................213
TABLE 3.97 MULTIVARIATE DOMAINS PERCENT CHANGE FOR PROBATION...............................................................215
TABLE 3.98 MULTIVARIATE DOMAINS PERCENT CHANGE FOR PAROLE ....................................................................216
TABLE 3.99 RAW CHANGE DOMAINS FOR SUPERVISION STATUS ...............................................................................217
TABLE 3.100 MULTIVARIATE DOMAINS RAW CHANGE FOR PROBATION ...................................................................218
TABLE 3.101 MULTIVARIATE DOMAINS RAW CHANGE FOR PAROLE .........................................................................219
ix
TABLE OF FIGURES
FIGURE 3.1 CHANGE IN ADJUSTED RATE OF RECIDIVISM BY STANDARD DEVIATION AT TIME 1 FOR
SAMPLE…………………………………………………………………………………………...63
FIGURE 3.2 CHANGE IN ADJUSTED RATE OF RECIDIVISM BY STANDARD DEVIATION AT TIME 2 FOR
SAMPLE…………………………………...………………………………………………………64
FIGURE 3.3 CHANGE IN ADJUSTED RECIDIVISM BY RISK LEVEL WHEN ASSUMING A 10 PERCENTAGE
POINT CHANGE IN RISK LEVEL FOR THE SAMPLE…………………………………………………..73
FIGURE 3.4 CHANGE IN ADJUSTED RATE OF RECIDIVISM BY STANDARD DEVIATION AT TIME 1 FOR
MALES…………………………………………………………………………………………….87
FIGURE 3.5 CHANGE IN ADJUSTED RATE OF RECIDIVISM BY STANDARD DEVIATION AT TIME 1 FOR
FEMALES………………………………………………………………………………………….88
FIGURE 3.6 CHANGE IN ADJUSTED RATE OF RECIDIVISM BY STANDARD DEVIATION AT TIME 2 FOR
MALES…………………………………………………………………………………………….89
FIGURE 3.7 CHANGE IN ADJUSTED RATE OF RECIDIVISM BY STANDARD DEVIATION AT TIME 2 FOR
FEMALES………………………………………………………………………………………….90
FIGURE 3.8 CHANGE IN ADJUSTED RECIDIVISM RATE BY RISK LEVEL WHEN ASSUMING A 10
PERCENTAGE POINT CHANGE IN RISK LEVEL FOR MALES………………………………………...101
FIGURE 3.9 CHANGE IN ADJUSTED OF RECIDIVISM RATE BY RISK LEVEL WHEN ASSUMING A 10
PERCENTAGE POINT CHANGE IN RISK LEVEL FOR FEMALES………………………………………102
FIGURE 3.10 CHANGE IN ADJUSTED RATE OF RECIDIVISM BY STANDARD DEVIATION AT TIME 1 FOR
BLACKS…………………………………………………………………………………………116
FIGURE 3.11 CHANGE IN ADJUSTED RATE OF RECIDIVISM BY STANDARD DEVIATION AT TIME 1 FOR
WHITES…………………………………………………………………………………………..117
FIGURE 3.12 CHANG IN ADJUSTED RATE OF RECIDIVISM BY STANDARD DEVIATION AT TIME 2 FOR
BLACKS………………………………………………………………………………………….118
FIGURE 3.13 CHANGE IN ADJUSTED RATE OF RECIDIVISM BY STANDARD DEBIATION AT TIME 2 FOR
WHITES…………………………………………………………………………………………..119
FIGURE 3.14 CHANGE IN ADJUSTED RECIDIVISM RATE BY RISK LEVEL WHEN ASSUMING A 10
PERCENTAGE POINT CHANGE IN RISK LEVEL FOR BLACKS………………………………………..129
FIGURE 3.15 CHANGE IN ADJUSTED RECIDIVISM RATE BY RISK LEVEL WHEN ASSUMING A 10
PERCENTAGE POINT CHANGE IN RISK LEVEL FOR WHITES………………………………………..130
FIGURE 3.16 CHANGE IN ADJUSTED RATE OF RECIDIVISM BY STANDARD DEVIATION AT TIME 1 FOR
PROBATION……………………………………………………………………………………...144
FIGURE 3.17 CHANGE IN ADJUSTED RATE OF RECIDIVISM BY STANDARD DEVIATION AT TIME 1 FOR
PAROLE………………………………………………………………………………………….145
FIGURE 3.18 CHANGE IN ADJUSTED RATE OF RECIDIVISM BY STANDARD DEVIATION AT TIME 2 FOR
PROBATION……………………………………………………………………………………...146
FIGURE 3.19 CHANGE IN ADJUSTED RATE OF RECIDIVISM BY STANDARD DEVIATION AT TIME 2 FOR
PAROLE………………………………………………………………………………………….147
FIGURE 3.20 CHANGE IN ADJUSTED RECIDIVISM RATE BY RISK LEVEL WHEN ASSUMING A 10
PERCENTAGE POINT CHANGE IN RISK LEVEL FOR PROBATION……………………………………159
FIGURE 3.21 CHANGE IN ADJUSTED RECIDIVISM RATE BY RISK LEVEL WHEN ASSUMING A 10
PERCENTAGE POINT CHANGE IN RISK LEVEL FOR PAROLE………………………………………..160
x
CHAPTER 1
USING THE LEVEL OF SERVICE INVENTORY-REVISED TO ASSESS OFFENDERS
Beginning in the early 1970s, the United States embarked on a “get tough” or “penal
harm” movement (Clear, 1994). In the intervening years, much attention has focused on the rise
in the prison population from under 200,000 to over 2 million offenders behind bars on any
given day. Although this focus is well-deserved, it can have the unanticipated consequence of
diverting attention from the corresponding growth in the population of offenders under
correctional supervision in the community. As of 2005, there were roughly 5 million offenders
on probation or parole and more than 7 million adults under some form of correctional control
(Glaze & Bonczar, 2006). It is estimated that over 600,000 inmates are released into the
community each year (Listwan, Cullen, & Latessa, 2006).
With nearly 5 million offenders in the community under the auspices of the state, the
challenge for corrections is how to supervise these offenders effectively. Collectively, these
offenders pose a potential threat to public safety. Individually, they pose differential threats to
the community. Managing and reducing this risk is a special challenge for corrections officials.
In this context, researchers have developed strategies for assessing offenders. The Level
of Service Inventory-Revised — known by its acronym “the LSI-R” — is perhaps the most
prominent among these instruments. The LSI-R is particularly influential because it not only
measures the risk level of offenders but also identifies their “criminogenic needs” (the factors
predicting involvement in crime). As a result, it provides a basis not only for supervision but
also for effective correctional intervention.
The current dissertation attempts to advance the research on the effectiveness of the LSIR. Using a sample of probationers and parolees from the state of Iowa, it examines the
1
predictive validity of the LSI-R. Specifically, this dissertation will consider the impact of
reassessing offenders on the LSI-R after their initial assessment. That is, does reassessing
offenders after their initial assessment improve the predictive ability of the LSI-R? The concept
of assessing and reassessing offenders also provides the opportunity to examine change scores —
that is, the difference in total LSI-R score on the initial assessment compared to the total LSI-R
score on the reassessment. In turn, the changes in score may help criminal justice practitioners to
adjust offender treatment based on the alterations in their risks or needs. Finally, this dissertation
will review scores in each of the ten domains on the LSI-R to determine which domain is the
most powerful predictor of recidivism and whether change scores in one domain are more or less
important than change scores in the other domains. This information has the potential to assist
criminal justice practitioners in case management and may also shed light on any particular
strengths and weaknesses of the assessment instrument in predicting recidivism.
To place this research in an appropriate context, this chapter is divided into five sections.
The first section explores the rise and changes in probation and parole in the United States. In
recent years, the Progressives’ model of individualized treatment came under attack in favor of
control-oriented offender supervision strategies. The ineffectiveness of this “get tough”
approach has created space for the revitalization of attempts to rehabilitation offenders more
effectively. The LSI-R is an instrument integral to this effort. The second section reviews the
revitalization of offender treatment in recent years. This movement to reaffirm rehabilitation
(Cullen & Gilbert, 1982) has been buoyed by the growing empirical literature showing the
effectiveness of offender treatment. In particular, there is evidence that recidivism is lowered
markedly when correctional programs adhere to the “principles of effective intervention”
(Andrews & Bonta, 2003). The third section explores the use of instruments to assess the risk
2
that offenders will recidivate. The purpose is to provide a background for understanding the
development of the Level of Service Inventory-Revised and how it relates to the principles of
effective supervision. The fourth section then reviews the existing research on the LSI-R. This
provides a context for understanding how the current project advances the extant literature. The
final section presents the research strategy that will inform this dissertation. Specifically, using a
sample of 2,849 probationers and parolees from the state of Iowa, this project intends to explore
three areas: 1) The effectiveness of the LSI-R in predicting recidivism; 2) the impact that change
scores (the difference in total score from the offender’s initial assessment as compared to their
score from the follow up assessment) may or may not have on the LSI-R’s predictive validity;
and 3) the predictive validity of each of the ten domains on the LSI-R and the impact of change
scores in each of the domains.
PROBATION AND PAROLE IN THE UNITED STATES
The Origins of Community Supervision: The Rise of Rehabilitation
Organized community supervision in the form of probation and parole can be traced to
the work of the Progressives during the early part of the twentieth century. This group of
educated and socially conscious reformers sought to improve the lives of Americans by reducing
poverty, improving neighborhood conditions, and more generally, tending to the needs of
individuals. Given their interests, it is not surprising that the Progressives were instrumental in
the development of community corrections initiatives across the country. Unlike the “new
penology” of today where emphasis is placed on the management of groups of offenders based
3
on the seriousness of the crime committed (Feeley & Simon, 1992), the Progressives believed
that one-size-fits-all correctional treatment was inadequate. Instead, Progressives insisted that
each offender be studied on a case-by-case basis so that a treatment plan could be implemented
to meet the specific needs of each individual offender.
This highly individualized approach required the criminal justice system to allocate
resources to assess, treat, and supervise offenders. Thus, probation officers, parole boards, and
parole officers were introduced. Probation officers were employed to gather background
information on each offender, determine the proper treatment, and to supervise low-risk
offenders in the community. Given the indeterminate nature of treatment proposed by the
Progressives, parole boards were established to review the treatment progress of each individual
offender and determine when the offender was fit to return to the community. Offenders were
not simply released into the community, but were placed under the supervision of a parole officer
whose duty it was to continue to supervise, treat, and aid the offender in his or her effort to
assimilate back into pro-social society (Rothman, 1980; Cullen & Gilbert, 1982).
Prior to the organized reforms of the Progressives, community supervision in the form of
probation was informally introduced in Boston by John Augustus. During an eighteen-year
career spanning from 1841 until 1859, Augustus provided supervision and assistance in pursuing
education and employment opportunities to nearly 2,000 accused and convicted criminals in the
Boston area. Augustus did not take responsibility for any and all offenders who crossed his path.
Instead, he conducted what could be seen as an arbitrary and subjective risk assessment to
carefully select offenders whom he surmised to be amenable to reform (e.g., first-time offenders)
and the aptitude to become pro-social members of society (Latessa & Allen, 2003).
4
Lawmakers from neighboring states were intrigued by the work of Augustus and
considered implementing similar probation practices in their respective states. The idea of
releasing adult offenders in the community was initially met with resistance from state and local
legislators. However, legislators were far more receptive to the idea of saving wayward youths
by placing them on probation instead of in jail or prison with hardened criminals. For that
reason, juvenile probation was enthusiastically adopted by states in the northeast and shortly
thereafter, the entire country.
The success of probation work with juveniles, coupled with the emergence of Progressive
ideology and reforms, meant that probation as a sentencing option for adults could no longer be
ignored. To that end, New York was the first state to pass laws to allow adult probation in 1901
(Lindner & Savarese, 1984). Other jurisdictions followed suit and by 1956 all states had adopted
probation practices for adults and juveniles (Latessa & Allen, 2003). Central to the probation
practices of Augustus and the Progressive era was the importance of the probation officer serving
as a social worker to the offender. The offender was placed in the care of the probation worker
so as that the offender had adequate treatment to prevent him or her from committing crimes in
the future. Caseloads were small to allow for more individualized attention from the probation
officer (Rothman, 1980). Unfortunately, the days of small caseloads and individualized
treatment have passed and today probation is by far the most popular of all criminal justice
sanctions. Petersilia (1997, p. 149) estimates that “probation officers are responsible for
supervising two-thirds of all correctional clientele.” As of 2005, it was estimated that
probationers make up 58% of the correctional population (Glaze & Bonczar, 2006).
Endorsed and largely developed by the Progressives, parole was introduced to the
country in the late 1800s and was to serve a multitude of purposes. For Progressives, parole was
5
a way to ensure individualized treatment and release from prison for offenders who have been
rehabilitated and are now ready to rejoin the general population. For prison wardens, parole was
a tool to encourage obedient behavior by those incarcerated who hope to earn good time credits
resulting in early release from prison. Prison wardens also supported parole initiatives because
they allowed for prisoners who were near the end of their sentence to be released early, making
room for newly sentenced offenders to take their place. Lawmakers supported parole because it
appeased the Progressives as well as criminal justice practitioners, thereby creating a win-win
situation (Rothman, 1980). Given its general appeal, twenty states adopted parole by 1900, and
by 1944 all U.S. states had implemented some form of parole (Latessa & Allen, 2003). The
federal government took note of the early success states were experiencing with parole and
proceeded to implement federal parole in 1910 (Hoffman, 2003).
Parole remained in the good graces of policy makers, criminal justice practitioners, and
the general public for a number of years. Reports first surfaced about the inconsistencies in
parole practices in the 1931 Wickersham Report. However, these minor concerns did not
initially take away from states increased use of this correctional option. It was not until the
events of the 1960s and early 1970s that parole practices were truly taken to task. Change in the
social and political climate, coupled with the concerns of unfair practices in parole selection and
qualifications of parole board members, made states consider abolishing this practice altogether.
To that end, in 1976 Maine was the first state to abolish parole (Latessa & Allen, 2003). In
subsequent years, fifteen other states abolished parole practices, meaning thirty-four states still
utilized some form of parole in 2000 (Hughes, Wilson, & Beck, 2001). While some states have
abolished parole, the importance of parole as a correctional option has not waned as the number
6
of offenders on parole increased in nine of the ten years between 1995 and 2005. As of 2005,
there were 784,408 offenders on state and federal parole (Glaze & Bonczar, 2006).
The Attack on Rehabilitation
As the Progressive movement began to lose momentum, so too did rehabilitation
gradually lose its hold as the guiding philosophy of the criminal justice system. The 1960s and
early 1970s were a time of great social and political unrest. The Civil Rights Movement, the
Deinstitutionalization Movement, Bay of Pigs Invasion, the assassinations of Malcolm X,
President Kennedy, Martin Luther King Jr., and Robert Kennedy, incidents of police brutality at
the Democratic National Convention, the Warren Court, the Hippie Movement, Watergate, the
shootings at Kent State University, the Vietnam War, the Beatles, and riots at New York’s Attica
State Penitentiary — these were but a few of the events that polarized the nation’s major political
parties. Both conservatives and liberals were working to come up with solutions to return the
country to a state of equilibrium.
Conservatives blamed liberals for the civil and political unrest of the day, citing social
welfare programs and lax crime control policies as the cause for growing drug culture, unruliness
of teenagers, and citizens speaking out against the government. Furthermore, conservatives
asserted that rehabilitation efforts were detrimental to the welfare of the country because fewer
offenders in prison reduced the incapacitation and deterrent effects of the system. The
incapacitation effect diminished because fewer people were sent to prison and/or were released
early under the guise of liberal rehabilitation efforts. In turn, more criminals in the community
resulted in an increase in crime and social disorder. Moreover, releasing offenders early also
undermined the system’s potential deterrent effect because criminals were no longer required to
7
serve out their entire sentence — that is, potential criminals had less reason to fear the
consequences of committing a criminal act. Conservatives believed that reduced fear of criminal
sanctions would also lead to an increase in crime and social disorder.
Conservatives blamed a predominantly liberal Supreme Court under the direction of
Chief Justice Earl Warren for rulings that increased individual rights, therefore making it more
difficult for police and other criminal practitioners to catch and punish criminals. Conservatives
argued that the indeterminate sentencing model afforded judges and parole boards too much
discretion in sentencing and releasing criminals from state and federal custody. The
conservatives’ solution to the problem of nationwide social disorder was to control crime
through increased punitive sanctions, determinate sentencing, and reduced discretion for judges
and correctional officials. Conservatives believed that the increase in the severity of criminal
sanctions and implementation of determinate sentencing would reestablish the credibility of the
criminal justice system, bolster the incapacitation and deterrent effects, and restore public order.
Liberals, on the other hand, viewed the events of the 1960s and early 1970s as proof that
the government could not be trusted. With citizens being beaten by police, university students
being shot by the National Guard, and the government sending young men to fight in a war
against their will, liberals concluded that the government did not have the interests of all citizens
at heart. If the government could not be trusted to be fair and equitable to the general population,
then by no means could the government be trusted to provide individualized treatment to the
correctional population. The liberals’ solution to the government’s inability to treat citizens and
offenders in a fair and equitable manner was to reduce the discretion of police, judges, and
correctional officials. Furthermore, liberals argued that judges would use the indeterminate
sentencing model to unfairly punish the poor and minority segments of the offender population.
8
A determinate sentencing model would guarantee similar sanctions for criminals who have
committed the same crime and give offenders a definitive release date so that the government
could not selectively hold offenders for an indefinite period of time. In this sense, liberals
believed that determinate sentencing would prevent the state from exacting excessive
punishment on the criminal population.
In a time of social and political upheaval, both conservatives and liberals agreed that two
fundamental changes were in order for the criminal justice system: 1) a reduction in the
discretionary power of criminal justice practitioners; and 2) that indeterminate sentencing would
be abolished and replaced with determinate sentencing. To reiterate, conservatives were in favor
of reduced discretionary power to ensure criminal justice practitioners did not release criminals
early from custody while liberals were in favor of reduced discretionary power to prevent against
abuse and the misuse of power by state and federal officials. With respect to determinate
sentencing, conservatives were in favor because offenders would be required to serve a definite
amount of time — that is, the offender could not commit additional crime while incarcerated,
and incarcerating the offender would serve to deter others from committing crime. Liberals
favored determinate sentencing to guard against the state and federal government from
victimizing offenders by holding them in custody for an indefinite period of time (Cullen &
Gilbert, 1982).
In 1974, a report released by Robert Martinson on rehabilitation programs garnered
attention from law-makers, criminal justice practitioners, criminologists, and the general public.
Martinson (1974, p. 25) reviewed 231 studies on treatment programs and concluded that, “With
few and isolated exceptions, the rehabilitative efforts that have been reported so far have had no
appreciable effect on recidivism. Later in this same report Martinson questions, “Do all of these
9
studies lead us irrevocably to the conclusion that nothing works, that we haven’t the faintest clue
about how to rehabilitate offenders and reduce recidivism?” (1974, p. 48). The findings from the
Martinson study were not solely responsible for ending the reign of rehabilitation as the guiding
philosophy of the criminal justice system. Rather, the findings validated what the general public
had suspected for some time. The findings from the study, in conjunction with the social and
political climate of the late 1960s and early 1970s set in motion a paradigm shift. The notion of
providing individualized treatment and rehabilitation to offenders was no longer in vogue and
had been replaced by a more punitive, crime control model of criminal justice that would define
the criminal justice system for the next thirty plus years.
The Rise and Limits of the Community Control Model
As was discussed in the previous section, the idea of determinate sentencing and reduced
discretion appealed to both conservatives and liberals in the late 1960s and early 1970s.
However, each group had a very different vision as to what these changes would mean for the
criminal justice system. Conservatives saw this change as a way to control crime. Liberals
thought that these changes would protect offenders from the abuse of power by government or
state officials, thereby affording due process to offenders. Given the social and political climate,
the public embraced the conservatives’ crime control model over the liberals’ due process model.
In doing so, judges began to sentence offenders more harshly for their criminal acts. Instead of
the short and definite sentences that liberals had hoped for, criminals were receiving longer
sentences and parole boards were discouraged from granting early release. Conservatives
believed that imposing longer harsher sentences would teach criminals a lesson and deter them
and potential offenders from committing crime in the future.
10
Throughout the 1980s, the criminal justice system continued to “get tough” on crime, and
Americans saw the correctional population escalate from1,842,100 in 1980 to 4,350,300 in 1990
(Glaze & Bonczar, 2006). As the correctional population grew, so too did the concerns of
lawmakers and corrections officials because prisons were quickly running out of space and
money to house newly sentenced offenders. The dramatic growth in the correctional population
and the financial costs of maintaining penitentiaries required law makers and corrections officials
to explore alternatives to sentencing criminals to prison. To that end, intermediate sanctions
were popularized across the country as a means to administer harsh punishment to offenders
without housing them in penitentiaries (Cullen, Wright, & Applegate, 1996).
Boot camps, electronic monitoring, house arrest, and shock incarceration were but a few
of the intermediate sanctions being implemented throughout the country. These programs held
promise for two reasons. First, the intermediate sanction programs could be operated at a
fraction of the cost of incarcerating an offender. Second, these programs were expected to be
punitive enough to prevent recidivism. The crime control model of intermediate sanctions
suggested that the increased supervision of offenders in the community would deter offenders
and potential offenders from committing crime and also result in low recidivism rates and
technical violations among the offenders being supervised.
Corrections officials were no longer concerned with individual treatment and helping
offenders as was the case in the early 1900s, but rather they were responsible for supervising and
monitoring offenders. While low recidivism and technical violations were the goals, policy
makers did not accurately predict the effect that increased supervision would have on recidivism
and technical violations. Instead of deterring offenders, the increases supervision and scrutiny
offenders received from their probation officers simply resulted in the officers having more
11
opportunities to catch their clients doing something wrong. As such, the increased supervision
resulted in higher recidivism rates and technical violations than were documented with standard
probation practices (Cullen et al., 1996; Fulton, Latessa, Stichman, & Travis, 1997).
As the 1980s progressed, the popularity of intensive supervision programs continued to
grow. Researchers began to study these programs to determine if they did indeed reduce
recidivism and provide for a more cost effective way to punish offenders. Perhaps the most
comprehensive of these intensive supervision program studies was conducted by RAND with
Joan Petersilia and Susan Turner leading the project. The RAND project examined 14 intensive
supervision programs in nine states with a sample of more than 2,000 offenders. Petersilia and
Turner found that after one year, offenders involved in the intensive supervision programs were
more likely to be arrested (37%) than offenders who were assigned to the control group (33%).
Moreover, Petersilia and Turner (1993) found that offenders in intensive supervision programs
also had a higher rate of technical violations (65%) than offenders in the control group (38%).
One surprising finding of the study was that offenders who received treatment during their
intensive supervision were less likely to recidivate than offenders who did not receive treatment.
Recall that there were two goals identified as part of the crime control initiative of
intermediate sanctions. The first goal was to reduce recidivism and technical violations. As
reported by Petersilia and Turner (1993), no such reduction in recidivism or technical violations
took place. The second goal was to punish offenders in an inexpensive manner. To intensively
supervise offenders, the caseloads of the officers assigned to monitor those on intensive
supervision were greatly reduced. In turn, additional probation officers needed to be hired in
order to manage those on intensive supervision as well as to manage the offenders who had
12
previously been supervised by the officer now working intensive supervision. As such, there
was an increase in the money spent to staff probation and intensive supervision officers.
In addition to staffing issues, Petersilia (1998) suggests that intensive supervision
programs served as “net widening” tools where judges sentenced offenders to intensive
supervision who were not likely to go to prison in the first place, but would have ordinarily been
placed on traditional probation. The existence of intensive supervision programs as a sentencing
option therefore, increased the financial cost as judges placed offenders on intensive supervision
that would otherwise have been placed on a less expensive form of probation.
Third, policy makers did not anticipate the number of probationers who would recidivate
or accumulate technical violations. When probationers violated the terms of their intensive
supervision, they were subsequently sentenced to other programs or even prison thereby
increasing the amount of money spent per offender. In total, there were several ways in which
intensive supervision produced unintended consequences surrounding financial expenditures
(Petersilia, 1998).
Unfortunately, intensive supervision programs were not effective at reducing recidivism
nor were they particularly inexpensive to operate. The failures of punitive intensive supervision,
coupled with the expense of operating these programs prompted policy makers to begin
considering other viable options to handle the burgeoning correctional population.
THE REVIVAL OF REHABILITATION
In the wake of the Martinson study, there has been a movement to revive rehabilitation.
This movement has involved two main lines of inquiry: one empirical and one theoretical. The
13
empirical approach has consisted of scholars assessing the existing evaluation studies and, in
turn, to judge the accuracy of Martinson’s “nothing works” claim. This research has used
different methodological and statistical techniques, but has come to the same conclusions:
rehabilitation programs work and some work far better than others.
The second line of inquiry has been theoretical. In this approach, scholars have
developed a set of principles that, if adhered to, differentiate effective from ineffective
treatments. This theory of effective intervention is based on social psychological-learning theory
and has received empirical support. This theory also serves as the foundation for the LSI-R, the
instrument being assessed in this dissertation.
Empirical Support for Rehabilitation: Responding to Martinson
The findings from the Martinson report, coupled with the social and political climate of
the early 1970s, were instrumental in ending rehabilitation’s reign as the guiding philosophy of
the criminal justice system. The Martinson report “proved what everyone already knew:
Rehabilitation did not work” (Cullen & Gendreau, 2000, p. 109). While policy makers and the
general public were quick to embrace the Martinson findings, researchers dutifully reviewed the
methodology and findings of Martinson’s work and continued to conduct research in the area of
rehabilitation and program effectiveness. The following section will consider the three types or
styles of research undertaken in the area of correctional rehabilitation and select research
findings that fall within each type. In the thirty plus years following the publication of the
Martinson report, researchers have gone from challenging the notion that nothing works to
emphatically showing that rehabilitation does work. In fact, researchers not only have proven
14
that rehabilitation works, but also have gone on to identify principles and best practices for
implementing the most effective rehabilitation programs.
In reviewing the research on correctional rehabilitation programs it makes sense to
separate the studies into one of the three following types of research: 1) ballot box; 2) narrative
review; and 3) meta-analysis (Cullen & Gendreau, 2000). These three types of research represent
different approaches researchers have taken in trying to understand the issue of program
effectiveness. Remarkably, rehabilitation programs have been found to be effective across all
three types of research.
The ballot box is a very straightforward approach to research. This strategy is much like
an election where voters cast a vote for their candidate of choice. When the polls close, the votes
are counted and the candidate who received the most votes is declared the winner. In the ballot
box approach, the researcher collects the findings from all research done on a specific topic. In
this case, researchers examine all of the studies conducted on rehabilitation programs. The
researcher reviews the findings from each study to determine whether or not the program was
effective or ineffective. After completing a review of all studies, the researcher goes back and
counts the number of studies that were found to be effective and ineffective. If the number of
studies found to be effective is greater than the number of studies found to be ineffective, the
researcher concludes that rehabilitation programs are effective. Conversely, if the number of
studies found to be ineffective is greater than the number of studies found to be effective, then
the researcher concludes that rehabilitation programs are ineffective.
The ballot box approach is beneficial in that the conclusion reached is easily
understandable to consumers of research. The downside to the ballot box approach is that the
conclusion reached is polarizing. With respect to research on the effectiveness of correctional
15
rehabilitation programs, using the ballot box approach, a researcher would conclude either
programs are effective or programs are not effective. In that sense, the issue is presented as
being black or white, when in fact there are many shades of gray in between the two extremes.
The second type of research is the narrative review. In a narrative review, the researcher
gathers all studies on a particular topic and reviews the findings from each of the studies. The
researcher then determines the best way to convey the findings of the research on a particular
topic. This may result in the researcher individually reviewing each study in a very detailed
manner, providing information about the methodology and findings from each individual study.
Another option is for the researcher to create subcategories of findings (e.g., all studies that look
at gender, race, socioeconomic status) and then report general findings from each subcategory,
accompanied by string citations to represent the various studies from each subcategory.
Narrative reviews are beneficial to the extent that they provide detailed information on
each study conducted on a particular topic. However, there are two disadvantages to the
narrative review style of research. First, the findings of the review are subject to bias because
the researcher selects the studies to include in the narrative review and then determines how to
interpret the findings from all of the individual studies included in the review. The second
disadvantage of the narrative review is that the findings from each study are given the same
amount of worth regardless of sample size or methodology. In that sense, a study with findings
based on 25 participants will count the same as a study with findings based on 1,500 participants.
The third and most recently developed type of research is the meta-analysis. The metaanalysis is quite different than the ballot box or narrative review approaches because the metaanalysis employs statistical techniques. Although this method is more complex than either of the
other approaches, the meta-analysis allows a researcher to determine not only whether or not
16
programs are effective, but also the degree to which the programs are effective or ineffective. In
this regard, the researcher can present the findings of the meta-analysis in a single number
known as the effect size. The effect size is the average of the findings across all studies when
controlling for differences in sample size. Another benefit of the meta-analysis is that the
researcher can weight studies according to sample size — that is, the findings from a study based
on a sample of 25 offenders will be given less value than the findings of a study based on a
sample of 1,500 offenders because the latter study more closely represents the population it is
trying to explain.
There are three potential problems with meta-analyses. First and similar to the ballot box
and narrative review, the researcher determines the studies to include in the analysis. In doing
so, the researcher may include studies that support his or her own ideology. Second, the findings
from the meta-analysis will only be as good as the methodology of the studies included in the
analysis. Because the meta-analysis includes findings from a number of individual studies, any
methodological problems that existed in the individual studies will be carried over and impact
the findings of the meta-analysis. Third, because the meta-analysis is highly statistical, the
average criminal justice practitioner is not likely to understand the techniques used by the
researcher and must accept the findings of the researcher on good faith.
Having provided a brief description of the three types of research in correctional
rehabilitation and the advantages and disadvantages of each, it is now appropriate to discuss
select findings from research representing the ballot box, narrative review, and meta-analytic
approaches to research. The findings from research in each of these areas has helped to “reaffirm
rehabilitation” as a viable and important correctional philosophy (Cullen & Gilbert, 1982).
17
Martinson’s 1974 work is perhaps the most famous example of the ballot box approach
to research in that he divided studies into two categories: those that were effective at reducing
recidivism and those that were not. Having counted the ballots for and against effective
programs, Martinson concluded that “nothing works.” The first of many to challenge
Martinson’s “nothing works” notion was Ted Palmer who in 1975 reanalyzed the very data that
Martinson had used in his famous publication the year prior. Palmer found that 39 of the 82
studies included in the Martinson analysis showed rehabilitation programs to be effective at
reducing recidivism. From a ballot box perspective, Martinson was correct in his conclusion that
rehabilitation programs do not work because the final tally was 43 programs ineffective at
reducing recidivism compared to 39 programs effective. However, Martinson’s suggestion that
“nothing works” grossly misrepresented the effectiveness of rehabilitation programs because in
fact, nearly half (48%) of the programs included in the study did work at reducing recidivism
(Palmer, 1975).
Palmer’s contribution to the field of corrections research did not end with his revelation
that treatment does work. Like the Progressives who preceded him by three-quarters of a
century, Palmer reintroduced the idea that correctional treatment that is tailored to the individual
may be more effective than a one-size-fits-all treatment model where all offenders receive the
same treatment regardless of their individual differences (Palmer, 1975; Cullen, 2005).
Gendreau and Ross have completed two comprehensive narrative reviews about
treatment programs. Their first review of research included 95 studies that had been conducted
between 1973 and 1978. Gendreau and Ross (1979) found that 86% of these studies reported
reductions in recidivism. In 1987 Gendreau and Ross presented a second narrative review that
included 130 studies conducted between 1981 and 1987. Based on the findings from the studies
18
reviewed, Gendreau and Ross concluded that, “it is downright ridiculous to say nothing works”
(Gendreau & Ross, 1987, p. 395).
Mark Lipsey was among the early leaders in conducting meta-analyses of rehabilitation
programs. Lipsey’s review of 443 studies of juvenile treatment programs found that 64% of the
programs were effective at reducing recidivism. Across studies, the treatment programs reduced
recidivism an average of 10% (Lipsey, 1992). Subsequent works by Lipsey (1999) and Lipsey
and Wilson (1993) continued to empirically support the effectiveness of rehabilitation programs.
Losel (1995) conducted the most extensive review of correctional program studies. His review
consisted of 13 meta-analyses that included over 500 studies of correctional rehabilitation
programs. Consistent with Lipsey’s 1992 finding, Losel also found rehabilitation programs to
reduce recidivism by 10%.
Across all three types of research, it is evident that rehabilitation programs are indeed
effective. Interestingly enough, there is considerable inconsistency in the degree to which these
programs are effective at reducing recidivism. Gendreau and Ross found reductions in
recidivism ranging from 30% to 60% (Gendreau & Ross, 1979), whereas Losel’s (1995) review
found reductions in recidivism ranging from 8% to 18%. Although it is clear that rehabilitation
is effective, it was unclear as to why some programs work better or worse than others. To that
end, researchers set out to identify the components common across effective programs.
Principles of Effective Correctional Treatment
The principles of effective intervention were developed by Don Andrews and James
Bonta who, along with their fellow Canadian psychologists Paul Gendreau and Robert Ross have
been instrumental in defining what works in correctional programming. The principles of
19
effective intervention emerged out of the “what works” research conducted in the years
following the Martinson publication. These principles are empirically supported and are
described in a number of publications including, but not limited to the following: Andrews,
Bonta, & Hoge, 1990; Andrews, Zinger, Hoge, Bonta, Gendreau, & Cullen, 1990; Andrews and
Bonta (1998), Cullen & Gendreau (2000), Smith, Gendreau, & Goggin (2004), Gendreau, Smith,
& French (2006). The following section includes an overview of the principles of effective
intervention.
The risk principle is straightforward in that it suggests that offenders who are most at risk
for recidivating should receive more intensive treatment than offenders who are less likely to
recidivate. The risk principle relies on the notion that 1) it is possible to predict future behavior;
and 2) a criminal justice practitioner will be able to properly assess an offender’s risk level and
then assign them to suitable treatment based on their risk level. To that end, it is critical that
offenders are properly classified and that treatment services are directed to high-risk offenders
(Andrews et al.,1990; Andrews & Bonta, 1998).
It is commonly believed that working with high-risk offenders is fruitless because they
cannot be changed. On the contrary, high-risk offenders are not beyond reform and are most
likely to benefit from treatment efforts (Lowenkamp & Latessa, 2005). Moreover, it may be
unnecessary to spend tax payer dollars treating low-risk offenders who are not likely to reoffend
regardless of whether they receive treatment. Finally, it is important that high-risk, medium-risk,
and low-risk offenders not be included in the same treatment program. When low-risk offenders
are mixed in with high-risk offenders, the low-risk offenders tend to pick up on the bad behavior
modeled by the high-risk offenders. Similarly, exposing high-risk offenders to low-risk
offenders allows low-risk offenders the opportunity to broaden their social network of criminal
20
acquaintances. In doing, a low-risk offender has the potential to increase his or her risk level if
placed in a treatment program with high-risk offenders (Bonta, Wallace-Capretta, & Rooney,
2000).
The needs principle takes into consideration an individual’s criminogenic needs. There
are four major criminogenic needs, also known as the “Big Four” that include, “antisocial
attitudes, associates, history, and personality” (Andrews & Bonta, 1998, p.9). To elaborate, on
the “Big Four” antisocial attitudes suggests that the offender has values or beliefs that support
his/her inclination to commit crime. Antisocial associates refer to the presence or absence of
peers or family members who commit crime and/or are supportive of criminal activity. History
refers to an individual’s criminal history wherein the more crimes an individual has committed in
the past, the more likely the individual is to commit crime in the future. Finally, antisocial
personality refers to individual characteristics such as impulsivity and low self-control (Cullen &
Gendreau, 2000). The Big Four have been found to be the best predictors of future offending.
Three of the four (antisocial associates, antisocial peers, and antisocial personality) are dynamic
factors — that is, factors that can change. These dynamic factors or criminogenic needs should
be targeted in treatment programs (Gendreau, Little, & Goggin, 1996).
The responsivity principle attempts to match an offender to the type of treatment that will
provide the best opportunity for improvement and change. Andrews and Bonta describe two
types of responsivity: 1) general; and 2) specific. General responsivity acknowledges that
individuals have the capacity to learn and modify their beliefs and behaviors. Cognitive
behavioral programming focuses on changing attitudes and beliefs of an individual and also
targets for the change the thought process an individual uses in making decisions (Van Voorhis
& Lester, 2004). Role play, modeling positive behavior, and the use of consistent reinforcement
21
are key in cognitive behavioral programs. Furthermore, it is suggested that positive
reinforcement should outweigh punishment by at least a 4:1 ratio as part of the overall treatment
strategy (Gendreau, 1996). Finally, cognitive behavioral programs that emphasize modeling
behavior, problem solving, and the use of positive reinforcement have been found to be the most
effective at reducing recidivism and should therefore, serve as the foundation for offender
treatment (MacKenzie, 2000).
Specific responsivity considers characteristics unique to the individual (e.g., intelligence,
gender, anxiety, sexual abuse). These and other characteristics may inhibit an individual’s
ability to learn or modify behavior in certain treatment setting (Andrews & Bonta, 1998;
Andrews, Bonta, & Wormith, 2006; Hubbard, 2007). For this reason, specific responsivity
factors must be identified through the offender assessment process and taken into consideration
when outlining an offender’s case management plan. Ideally, an offender will be place in a
treatment program that addresses his or her general and specific responsivity needs.
Although the risk, needs, and responsivity principles are the major components of the
principles of effective intervention, there are other principles and program delivery
considerations that also improve program effectiveness. For example, it is best to administer
treatment to the offender in their own community rather than treating them in an institutional
setting. Treating offenders in the community allows individuals the opportunity to maintain
employment, pursue educational goals, and maintain relationships with pro-social family
members and peers. Treatment in the community also provides offenders the opportunity to
practice what they have learned through their cognitive behavioral programming and apply it to
real life scenarios (Gendreau 1996).
22
Also relating to program delivery, treatment must be intense and take place over a period
of time. Administering treatment for a day or week will not be effective. Rather, it is
recommended that treatment services should meet or exceed 100 hours over the course of three
or fourth months. Repetition is fundamental to amending an individual’s thinking errors and
restructuring the way in which information is processed and decisions are made. Treatment
should also be multi-modal, meaning that treatment services must be organized in a way to
address a myriad of issues based on the individual needs of offenders (Gendreau, 1996).
Lastly, there should be emphasis on program development, training and monitoring of
staff members, and aftercare. The program must be theory driven with clearly defined goals,
modes of treatment delivery, and outcome measures. This type of programmatic organization
will allow the program to be tested for effectiveness in meeting its specified treatment goals and
outcomes. With reference to staff members, proper training and oversight is imperative to
ensuring the program is administered as directed by theory on which the program is based.
Finally, programs should include an aftercare component, meaning that services are available to
offenders who have completed treatment to guard against possible relapse and/or direct offenders
to additional services that may be needed.
As was previously mentioned, the identification of risk, needs, and responsivity
characteristics is critical to effective intervention. The Level of Service Inventory Revised (LSIR) is an assessment instrument developed by Andrews and Bonta, designed so that a trained
criminal justice practitioner can identify an offender’s risk, need, and specific responsivity
factors. In turn, this information is used to establish a treatment plan for the offender to
minimize the chances of reoffending through placement in proper treatment programs.
23
Details of the instrument and research on the effectiveness of the LSI-R as a classification
instrument will be discussed in subsequent sections. The instrument’s introduction at this
juncture is to illustrate just how far the field has progressed with respect to effective
programming since Martinson’s “nothing works” proclamation. The LSI-R and principles of
effective intervention identify who is best suited for treatment (high-risk offenders), what
effective treatment entails (cognitive behavioral programming), where treatment is best
administered (the offender’s natural environment), when treatment should be administered (after
identifying risk, need, and responsivity characteristics), and why the criminal justice system
should rehabilitate offenders (because it has been empirically proven to reduce recidivism.)
OFFENDER CLASSIFICATION
The criminal justice system supervises over 7 million offenders of all ages, from diverse
backgrounds, and with a variety of individual needs (Glaze & Bonczar, 2006). The individual
differences across offenders make it imprudent for the criminal justice system to take a one-sizefits-all approach to correctional treatment. Instead, the criminal justice system has the daunting
task of identify the risks and needs of every individual offender in order to determine the
appropriate case management plan that will both protect the general public and effectively treat
the offender so that they will not commit crime when they are released from criminal justice
supervision. The following section will provide an overview of offender classification, including
a discussion of the two general types of assessment (clinical judgment and actuarial), a review of
the four generations of risk assessment, and finally a description of the LSI-R.
24
Generations of Risk Assessments
As was previously mentioned, individual differences exist across offenders. To that end,
criminal justice agencies may use a variety of assessment instruments to identify the risks and
needs of their offender population. Regardless of the type assessment being used, there are four
common elements to all offender classification instruments. First, the assessment is administered
to the offender upon his or her arrival at the program or institution. This is important because it
provides practitioners with a starting point from which the offender’s progress can be measured
when the offender is reassessed during or after he or she has completed treatment. Second, the
assessments allow the offenders to be divided into groups based on their risks and needs. Third,
assessments are administered by criminal justice practitioners who have been properly trained on
the administration of the instrument. Fourth and finally, assessment instruments provide
structure and uniformity to the classification process because they ensure all offenders are
evaluated on identical factors (Van Voorhis, 2000).
To date, scholars have identified four generations of risk assessments (Andrews, Bonta,
& Wormith, 2006). The first generation is the earliest form of risk assessment. Subsequent
generations build upon the strengths and improve upon the weaknesses of the generation prior.
The following traces the development and evolution of the four generations of risk assessments.
First generation risk assessments are based on clinical judgment. In other words,
assessments are based on criminal justice practitioner’s own professional and personal
knowledge of the offender population and their ability to accurately assess an offender’s risk
level. To that end, informal interviews and observations made by criminal justice practitioners
concerning the perceived dangerousness of an offender— that is, of the likelihood the offender
would recidivate. The first generation of risk assessments is unique because there is no
25
standardized set of questions to be asked of each offender. Rather, the criminal justice
practitioner has the discretion to ask the offender as many or as few questions as they would like.
Moreover, classification decisions are based on the factors that each individual practitioner deem
relevant in determining risk. Perhaps not surprisingly, the lack of structure in questioning
offenders results in inconsistent classifications across criminal justice practitioners (Schneider,
Ervin, & Snyder-Joy, 1996). That is, two criminal justice practitioners may assess the same
individual but draw two very different conclusions regarding the risk level of that particular
offender.
Another common problem associated with first generation assessments is that the absence
of a standardized set of questions results in criminal justice practitioners over-classifying
offenders. In other words, the criminal justice practitioners determines the offender to be higher
risk or more likely to recidivate than they actually are (Schneider et al., 1996). The tendency to
over-classify offenders can be traced to agencies holding their criminal justice practitioners
accountable for the assessments they conduct. If a criminal justice practitioner under-classifies
an offender and the offender recidivates, the practitioner who conducted the assessment may be
scrutinized by their supervisor and/or the general public for incorrectly identifying the offender’s
risk level (Schneider et al., 1996). To that end, it is in the best interest of the criminal justice
practitioner to err on the side of safety and over-classify an offender because it protects the
general public from the offender and it protects the criminal justice practitioner from being
scrutinized by their supervisors and the general public (Clear & Gallagher, 1985). If the offender
does not recidivate, then the correctional treatment may be credited for successfully
rehabilitating the offender.
26
The need for a standardized assessment tool was apparent, and so the second generation
of risk assessment was born. The second generation risk assessment is a marked improvement
because it relies on actuarial or mechanical assessment methods. Actuarial assessments are
comprised of a standard set of questions that are asked to all offenders. The questions included
on an actuarial assessment are theory driven and based on empirical support (Gnall & Zajac,
2005). The benefit of second generation actuarial assessments is that the structure of the
instruments ensures that all offenders are being assessed on the same factors or types of
information. This greatly reduces the arbitrariness and potential for individual bias that exists in
the prior generation of risk assessment because all assessments are based on clinical judgment
(Hoge, 2002).
Although the second generation instrument is an improvement over first generation
instruments, there are three major problems with these assessment instruments. The first
problem has to do with the criminal justice practitioners who are expected to administer the
assessments. In a study of probation and parole officers in Oklahoma, only 37% agreed that
“risk/need instruments are appropriate for making decisions about the level of supervision”
(Schneider et al., 1996, p. 118). Criminal justice practitioners are aggravated that they have to
attend training on the instrument to ensure they understand the rationale behind the instrument
and are qualified to administer it to offenders. Practitioners also feel that use of the instrument
severely limits their professional discretion (Clear & Gallagher, 1985). Finally, only 15% of the
officers in the Oklahoma study indicate that they think the assessment instrument will better
predict an offender’s risk level than a probation officer (Schneider et al., 1996).
Contrary to the belief of the majority of the officers in the aforementioned study, the vast
majority of research suggests that actuarial prediction instruments are superior to clinical
27
judgment (Sarbin, 1943; Grove & Meehl, 1996; Gardner, Lidz, Mulvey, & Shaw, 1996; Grove,
Zald, Boyd, Lebow, Snitz, & Nelson, 2000; Bonta, 2002). In a meta-analysis of 136 studies of
human behavior conducted 1966-1988, actuarial assessments were found to be 10% more
accurate than clinical or judgment (Grove et al., 2000). Grove and Meehl (1996, p. 320) assert
that relying on clinical judgment instead of using an actuarial assessment instrument “is not only
unscientific and irrational, it is unethical.”
A second problem with second generation risk assessments is that most instruments of
this generation include only static factors. As has been previously discussed, static factors
cannot change (e.g., criminal history, age at first offense, severity of prior convictions) which
makes reassessment efforts futile with second generation instruments. The purpose of
reassessing an offender is to determine whether or not the treatment has decreased the offender’s
risk level (Bonta, 2002). However, second generation risk assessments based on static factors
prevent any conclusions regarding treatment effectiveness from being made because the
offender’s score will be the same or higher than it was on the offender’s initial assessment.
Perhaps the most famous, and still used, second generation risk assessment is the Salient
Factor Score (SFS). The SFS is typically used in making parole decisions. This instrument is
favored because it is quick and easy to administer and score. The instrument itself is made up of
six items, including “prior convictions/adjudications, prior commitments of more than 30 days,
age at current offense/prior commitments, recent commitment-free period, escape status, and
heroin/opiate dependence” (Hoffman, 1983, p. 546). The major drawback of this instrument is
that the items are static in nature. As such, the instrument is very limited and does not lend itself
well to treatment planning or follow up assessments to monitor treatment progress.
28
A third problem is that second generation assessments only consider an offender’s risk
level. Second generation assessments do not take into account responsivity factors. Recall that
responsivity considers the individual characteristics or traits that may make an offender more or
less amenable to change through certain treatment strategies and attempts to match offenders to
treatment programs that will provide the greatest likelihood for success. Factors including, but
not limited to intelligence, gender, anxiety, sexual abuse may limit the likelihood for success in
certain treatment settings (Andrews & Bonta, 1998; Andrews, Bonta, & Wormith, 2006;
Hubbard, 2007).
Third generation risk assessments incorporate the structured set of questions common in
second generation assessment. An important difference between second and third generation
assessment instruments is that third generation risk assessments include select static factors, but
the majority of questions are based on dynamic factors. The inclusion of dynamic factors is
critical in establishing a case management plan for each offender. Moreover, an assessment
instrument based largely on dynamic factors lends itself nicely to offender reassessment – that is,
the assessment results allow a criminal justice practitioner to monitor whether or not treatment is
working for the offender.
If the reassessment suggests the offender’s risk level has been reduced, then criminal
justice practitioners can make the necessary adjustments to the offender’s case management plan.
Similarly, if the reassessment results indicate that treatment is not working, the practitioner may
assign the offender to a different type of treatment program that may provide a better likelihood
for success. Also, if the treatment is not working, it may bring to light potential problems with
the program itself that may warrant further investigation.
29
While the incorporation of dynamic factors into the third generation is a significant
improvement over the previous generation of assessment instruments, a second improvement is
important to note. Third generation risk assessments, such as the LSI-R, take into consideration
responsivity factors. The opportunity to document responsivity factors improves the likelihood
that criminal justice practitioners will properly place offenders in treatment programs to
maximize the offender’s likelihood for success.
A fourth generation of risk assessment instruments has recently emerged. Much like the
third generation of risk assessments, the fourth generation assessment instruments include both
static and dynamic risk factors that have been theoretically and empirically tested. Building on
the knowledge gained through research on the effectiveness of the previous generations of risk
assessment, the fourth generation risk assessments follows the offender through and beyond the
point in time where the individual is released from criminal justice supervision. In a review of
research by Andrews, Bonta, and Wormith (2006) that compares the predictive ability of
assessment instruments from all four generations. Andrews, Bonta, and Wormwith conclude that
the predictive power of assessment instruments has improved with each generation. (For an
introduction to fourth generation assessments and referrals to specific fourth generation
assessment instruments, see Andrews et al., 2006.)
The news that assessment instruments have become increasingly more accurate at
predicting risk of recidivism with each passing generation is certainly encouraging. To date,
fourth generation assessment instruments are still in the early stages and have not been widely
adopted by criminal justice agencies or researched in depth by scholars. To that end, it is
important to focus on the third generation of assessment instruments such as the LSI-R because
30
these instruments are the most commonly used for assessing risks and needs in criminal
offenders in the United States and abroad (Andrews & Bonta, 1995; Hollin & Palmer, 2006).
The Development of the LSI: Putting Effective Correctional Treatment Into Practice
The Level of Supervision Inventory (LSI) was developed in the early 1980s by Canadian
psychologists Don Andrews and James Bonta. In the 1990s, the LSI was updated and renamed
the Level of Service Inventory-Revised (LSI-R). The LSI-R is a third generation risk/needs
assessment instrument based largely on theory and research in the area of social learning. The
assessment instrument includes 54 questions that fall into ten domains or categories. These
include “Criminal History (10), Education/Employment (10), Financial (2), Family/Marital (4),
Accommodation (3), Leisure/Recreation (2), Companions (5), Alcohol/Drug Problems (9),
Emotional/Personal (5), and Attitudes/Orientation (4)” (Andrews & Bonta, 1995, p. 2).
Although the instrument does contain questions that target static factors, the majority of the
questions target dynamic factors that potentially can be changed through treatment.
The LSI-R is a valuable correctional tool and according to Andrews and Bonta (1995, p.
3), the LSI-R is appropriate for use in “identifying treatment targets and monitoring offender risk
while under supervision and/or treatment services, making probation/supervision decisions,
making decisions regarding placement into halfway houses, deciding appropriate security-level
classification within institutions, and assessing the likelihood of recidivism.”
The assessment is designed to be administered by a criminal justice practitioner who has
been trained on the LSI-R instrument. This practitioner administers the instrument in a semistructured interview with the offender that typically takes forty-five minutes to an hour to
complete. The 54 items on the assessment are scored as either Yes or No or on a scale of 0 to 3.
31
The 0 to 3 scale can be translated to the following: “3 = A satisfactory situation with no need for
improvement, 2 = A relatively satisfactory situation with some room for improvement evident, 1
= A relatively unsatisfactory situation with a need for improvement, and 0 = A very
unsatisfactory situation with a very clear and strong need for improvement” (Andrews & Bonta,
1995, p. 5).
Upon completion of the interview, the criminal justice practitioner scores the offender on
the 54 items. One point is awarded per each item that is scored Yes, 1, or 0. The criminal
practitioner tallies up the points based on the offender’s responses to the 54 questions to
determine the total LSI-R score. The score is then compared against the range of scores that fall
within each designated risk level: 0-13 = Low, 14-23 = Low/Moderate, 24-33 = Moderate, 3440 = Medium/High, and 41-54 = High. Based on the risk designation determined by the
offender’s total LSI-R score, the criminal justice practitioner is able to outline a case
management plan most suitable for the offender based in their risk, needs, and responsivity
factors.
The LSI-R is an important tool in promoting effective correctional treatment because it
addresses and overcomes a number of obstacles observed in the assessment instruments of the
second generation and the clinical judgment of the first generation. Unlike the clinical judgment
that represents the first generation of risk assessment and the actuarial, but static nature of
assessments of the second generation, the LSI-R provides structure and dynamic items that have
been empirically proven to be the best predictors of crime. Moreover, the LSI-R is straightforward, easy to administer and score, and allows for the criminal justice practitioner to exercise
professional discretion during the semi-structured interview and scoring. The LSI-R assigns
each offender to a risk category so that an appropriate case management strategy can be put into
32
place. Once the offender has received treatment, then a follow-up LSI-R can be administered to
monitor the offender’s progress and modify the offender’s treatment plan as needed.
In this section, information has been provided outlining the evolution of risk assessments
from the early clinical judgment stage through the actuarial assessments of today that are based
on empirically supported factors critical to identifying risk, need, and responsivity factors for
appropriate and effective treatment. The following section will examine the research conducted
to date on the LSI-R and how the current study will attempt to advance extant literature on the
LSI-R.
Predictive Validity of the Level of Supervision/Level of Service Inventory
Table 1.1 and Table 1.2 present an overview of findings from forty-five studies on the
Level of Supervision/Level of Service Inventory (LSI) conducted between 1982 and 2008. Each
of these studies tests the predictive validity of the LSI and/or various versions of the assessment
instrument. Specifically, the findings describe the degree to which an offender’s total LSI score
can accurately predict the offender’s likelihood to recidivate. The following section is a
discussion of four major conclusions drawn from the review of previous research.
First, the LSI appears to be an empirically supported instrument for predicting recidivism.
As indicated in the Valid Predictor of Recidivism column (Table 1.1), the large majority of
studies (77.5%) report a statistically significant relationship between total LSI score and
recidivism. Although some studies (22.5%) fail to report a significant relationship between the
LSI total score and recidivism, nearly all (97.8%) of the studies report a positive association
between total LSI score and recidivism. That is, the higher the total LSI score, the greater the
33
likelihood that the offender will recidivate. Conversely, the lower the total LSI score, the less
likely the offender will recidivate.
Second, the LSI is a valid predictor of recidivism across groups of offenders. Table 1.2
includes information on the LSI’s predictive validity across categories of age, gender,
correctional placement, and location. Nearly nine in ten studies using adult samples (87.5%)
report the LSI to be a valid predictor of recidivism. The findings from juvenile offender samples
are less favorable, though only five studies of juveniles were included in the review of literature.
Eighty percent of the juvenile samples report a positive association between the LSI and
recidivism and 40% of the juvenile studies report statistically significant findings. This may be
an indication that the instrument does not predict recidivism for juvenile offenders. However,
given the very limited number of studies with juveniles coupled with the small sample sizes, it is
prudent to encourage more research in this area rather than declare the instrument a failure with
juvenile offenders. It should be noted that the most recent and largest study of juveniles to date
found the YLS-CMI to be a statistically significant predictor of recidivism.
The ability of the LSI to predict recidivism for male and female offenders is a topic of
debate among researchers. Some suggest that the LSI may not predict as well for female
offenders as it does for male offenders because the risk factors of female offenders may not be
identical to the risk factors of their male counterparts. These differences may result in female
offenders being misclassified (Reisig, Holtfreter & Morash, 2006; Holtfreter & Culp, 2007).
Despite the potential for differences between male and female offenders (See Table 1.2), the LSI
has been validated for male samples (85.7%), female samples (71.4%), and mixed samples
(93.3%). The present study will also test the ability of the LSI to predict recidivism for
34
Table 1.1 Summary Findings From Previous LSI Research
Author
Year
N
Measure
Strength of Prediction
Measure of Recidivism
Andrews
1982
561
LSI
r = .41
Reconviction
Yes
Bonta & Motiuk
1985
LSI
r = .40 (S1)
r = .32 (S2)
Reincarceration
Reincarceration
Yes
Yes
Andrews et al.
1986
192
LSI
r = .48
Re-arrest
Yes
Motiuk et al.
1986
147
LSI
r = .36
r = .40
Halfway House Outcome
Reincarceration
Yes
Yes
Bonta & Motiuk
1987
108 (S1)
244 (S2)
LSI
r = .58 (S1)
r = .39 (S2)
r = .34 (S1)
r = .31 (S2)
Halfway House Outcome
Halfway House Outcome
Reincarceration
Reincarceration
Yes
Yes
Yes
Yes
Bonta
1989
119
LSI
r = .35 (Natives)
r = .50 (Non-Natives)
r = .51 (Natives)
r = .46 (Non-Natives)
Reincarceration
Reincarceration
Parole Violation
Parole Violation
Yes
Yes
Yes
Yes
Bonta & Motiuk
1990
580
LSI
RIOC = 70%
Reincarceration
NA
Bonta & Motiuk
1992
580
LSI
r = .35
Reincarceration
Yes
Motiuk et al.
1992
97
LSI
RIOC = 38.7%
Reincarceration
NA
75 (S1)
89 (S2)
35
Valid Predictor
of Recidivism
Author
Year
N
Measure
Strength of Prediction
Measure of Recidivism
Shields
1993
162
YO-LSI
r = .563
Reincarceration
Yes
Coulson et al.
1996
526
LSI
r = .51
r = .53
r = .45
New Charges
Parole Violation
Halfway House Outcome
Yes
Yes
Yes
Gendreau et al.
1996
4,579
LSI-R
r = .35
Varies
Yes
Gendreau et al.
1997
2,252
LSI-R
r = .23
Varies
Yes
Kirkpatrick
1998
138
88
31
17
LSI-R
r = .27 (Intake)
r = .40 (3 Months)
r = .29 (9 Months)
r = .60 (12 Months)
Release Outcome
Release Outcome
Release Outcome
Release Outcome
Yes
Yes
No
Yes
O’Keefe et al.
1998
257
LSI
r = .31 (Parole T1)
r = .22 (Parole T2)
r = .08 (CC T1)
r = .11 (CCT2)
Reincarceration
Reincarceration
Reincarceration
Reincarceration
Yes
Yes
No
No
Ilacqua et al.
1999
164
YO-LSI
Risk of recidivating
increased as YO-LSI
scores increased.
New Charges or
Reincarceration
NA
Kirkpatrick
1999
169
LSI-R
r = .41
Release Outcome
Yes
Raynor
2000
948
LSI-R
r = .35
Reconviction
Yes
36
Valid Predictor
of Recidivism
Author
Year
Lowenkamp et al.
2001
Dowdy et al.
N
Measure
Strength of Prediction
Measure of Recidivism
442
LSI-R
r = .26
r = .24
r = .14
Reincarceration
Program Completion
Absconding
Yes
Yes
Yes
2002
140
127
123
LSI
r = .11
r = .14
r = .13
Halfway House Outcome
Re-arrest Any
Re-arrest Felony
No
No
No
Gendreau et al.
2002
7,367
LSI-R
r = .37
Varies
Yes
Austin et al.
2003
985
LSI-R
Risk of recidivating
increased as LSI-R
scores increased.
Re-arrest, Absconding or
Reincarceration
NA
Barnoski & Aos
2003
22,533
LSI-R
r = .29
Reconviction
Yes
Marczyk et al.
2003
95
YLS-CMI
YLS/CMI score did not
predict recidivism.
Re-arrest
No
Mills et al.
2003
209
LSI-R
r = .39
Re-arrest
Yes
Girard & Wormith
2004
630
LSI-OR
r = .39
Reconviction
Yes
Holtfreter et al.
2004
134
LSI-R
r = .16
Re-arrest
No
37
Valid Predictor
of Recidivism
Author
Year
N
Measure
Strength of Prediction
Measure of Recidivism
Miles & Raynor*
2004
1,380
LSI-R
r = .29
Reconviction
Yes
Simourd
2004
129
LSI-R
r = .44
r = .26
r = .31
r = .50
r = .46
Re-arrest
Violent Rearrest
Reconviction
Reincarceration
Supervision Violation
Yes
Yes
Yes
Yes
Yes
Mills et al.
2005
209
LSI-R
r = .39
Re-arrest
Yes
Schmidt et al.
2005
107
YLS-CMI
r = .19
Re-arrest
No
Dahle
2006
307
LSI-R
r = .41
r = .34
r = .29
Reincarceration
NA
Flores et al.
2006
2,030
LSI-R
r = .18
Reincarceration
Yes
Flores et al.
2006
2,107
LSI-R
r = .28
Reincarceration
Yes
Hendricks et al.
2006
200
LSI-R
r = .16
Domestic Violence
No
Hollin & Palmer
2006
216
LSI-R
r = .20
Reconviction
Yes
Holsinger et al.
2006
403
LSI-R
r = .18
Re-arrest
Yes
Mills & Kroner
2006
209
LSI-R
r = .39
Re-arrest
Yes
38
Valid Predictor
of Recidivism
Author
Year
N
Measure
Strength of Prediction
Measure of Recidivism
Reisig et al.
2006
402
Bechtel et al.
2007
Folsom & Atkinson
LSI-R
r = .07
Violation, Re-arrest or
Reconviction
No
4,482
YLS-CMI
r = .196
Reconviction
Yes
2007
100
LSI-R:SR
r = .30
Reconviction
Yes
Palmer & Hollin
2007
96
LSI-R
r = .53
Reconviction
Yes
Lowenkamp & Bechtel
2007
1,145
LSI-R
r = .25
Re-arrest
Yes
Schlager & Simourd
2007
446
LSI-R
r = .06
r = .09
Re-arrest
Reconviction
No
No
Lowenkamp et al.
2008
14,737
LSI-R
r = .35
Varies
Yes
* R scores for this study appear in Raynor (2007).
39
Valid Predictor
of Recidivism
Table 1.2 Predictive Validity Across Categories
Number
Percent
Age
Adults
Juveniles
40
5
88.9
11.1
100
80
35
2
87.5
40
Sex
Female
Male
Mixed Sample
Varies
Missing
7
16
17
3
2
17.5
35.5
37.8
6.7
4.4
100
100
100
100
100
5
12
14
3
2
71.4
85.7a
93.3b
100
0
Correctional Placement
Community Corrections
Jails
Juvenile Detention
Prison
24
3
5
12
53.3
6.7
11.1
26.7
100
100
80
100
18
2
2
11
78.3c
100d
50e
100f
Location
Canada
United States
Other
Varies
19
17
5
4
42.2
37.8
11.1
8.9
100
94.1
100
100
16
10
4
4
94.1g
62.5h
100i
100
a
Percent
Positive
Association
Number
Percent Valid
Predictor of
Recidivism
Calculation based on 14 studies instead of 16 because two studies fail to report significance.
Calculation based on 15 studies instead of 17 because two studies fail to report significance.
c
Calculation based on 23 studies instead of 24 because one study failed to report significance.
d
Calculation based on 2 studies instead of 3 because one study failed to report significance.
e
Calculation based on 4 studies instead of 5 because one study failed to report significance.
f
Calculation based on 11 studies instead of 12 because one study failed to report significance.
g
Calculation based on 17 studies instead of 19 because one study failed to report significance.
h
Calculation based on 16 studies instead of 17 because one study failed to report significance.
I
Calculation based on 4 studies instead of 5 because one study failed to report significance.
b
40
the entire sample (males and females) and then test the predictive validity for separate
subsamples of males and females.
The LSI is designed to be a versatile assessment tool, appropriate for use in a variety of
correctional settings (Andrews & Bonta, 1995). For this reason, researchers have tested the
instrument with offenders in prisons, jails, juvenile detention, and community corrections. As
seen in Table 1.2, research on the predictive validity of the LSI with offenders in prison and jails
has received unanimous support (100%). The LSI also performs well in community corrections
settings (78.3%). It appears the LSI is the weakest predictor for offenders in juvenile detention
because only 50% of the studies report a statistically significant relationship. Again, this finding
should be viewed with caution due to the limited number of studies on juveniles in detention
centers.
The LSI has been adopted for use by domestic and foreign correctional systems. To date,
the predictive validity of the LSI has been tested in Canada, Germany, the United Kingdom, the
Island of Jersey, and the United States. Given the instrument’s Canadian roots, it is no surprise
that Canadian researchers have been actively involved in testing the LSI with Canadian
offenders. As seen in Table 1.2, the results indicate that the LSI is a valid predictor of recidivism
in ninety-four percent (94.1%) of Canadian studies. Seventeen studies of the LSI have been
carried out in the United States with just over sixty percent (62.5%) reporting statistically
significant findings. Five studies have been conducted in Europe and each of the studies report
statistically significant findings between LSI total score and recidivism. Regardless of study
location, the majority of the studies empirically support the LSI as a predictor of recidivism.
Third, the LSI appears to be an effective predictor across measures of recidivism. Table
1.3 provides information on the variety of ways that recidivism has been measured in extant
41
literature on the LSI. In the studies reviewed, reincarceration (35.6%) is the single-most popular
measure of recidivism, followed closely by re-arrest (31.1%) and reconviction (26.7%). Halfway
house outcome (8.9%), absconding (4.4%), new charges (4.4%), parole violation (4.4%),
program completion (2.2%), evidence of domestic violence (2.2%), and violation (2.2%) are
used less often. Regardless of the measure of recidivism, a positive association between total
score and recidivism is consistent across studies. Further, the LSI total score is a statistically
significant predictor of recidivism across all eleven measures of recidivism.
Fourth, the LSI has garnered empirical support through three decades of research.
During this time, the LSI has undergone minor modifications resulting in multiple versions of the
instrument. For this reason, the specific type of LSI instrument is identified for each individual
study. As seen in Table 1.4, twenty-seven of the studies (60%) test the Level of Service
Inventory-Revised (LSI-R). Eleven of the forty-five studies (24.4%) test the original LSI. Three
studies (6.7%) test the recently developed Youth Level of Service/Case Management Inventory
(YLS-CMI). The predictive validity of the Youth Level of Service Inventory (YO-LSI) has been
tested twice (4.4%) and the Level of Service Inventory-Revised: Self Report (LSI-R:SR), and
Level of Service Inventory-Ontario Revision (LSI-OR) have received less scrutiny from
researchers to date and respectively represent roughly two percent (2.2%) of the studies included
in this review. The LSI, LSI-R, LSI-R: OR, LSI-R:SR, YO-LSI, and YLS-CMI have all
received empirical support as valid predictors of recidivism.
Fifth and finally, very few studies on the LSI have administered the assessment multiple
times in order to consider how change in an offender’s total LSI score may impact the
instrument’s ability to predict recidivism (O’Keefe, Klebe, & Hromas, 1998; Hollin, Palmer, &
Clark, 2003; Miles & Raynor, 2004; Raynor, 2007). These studies have indicated that an
42
offender’s LSI total scores has the potential to change over time – that is an offender’s total LSI
scores at the time of reassessment may be higher, lower, or the same as the offender’s total LSI
from the initial assessment (Hollin et al., 2003).
O’Keefe et al. (1998) assessed a sample of parolees and a separate sample of offenders
under community supervision at two distinct points in time. The initial assessment and
reassessment took place roughly six months apart. The predictive validity of the LSI for parolees
was statistically significant and time 1 and time 2 but was not statistically significant at time 1 or
time 2 for the sample of offenders under community supervision. Further, the predictive validity
for the sample of parolees was stronger at time 1 (r = .31) than it was at time 2 (r = .22). The
opposite was true with the sample of offenders under community supervision. Although neither
time 1 or time 2 was statistically significant, time 2 (r = .11) was a slightly better predictor than
time 1 (r = .08) for offenders under community supervision. The present study will also consider
the predictive validity of the LSI-R at two points in time with the entire sample and then with
specific subsamples of males, females, blacks, whites, probationers, and parolees.
Finally, researchers have compared how an increase or decrease in total LSI from time 1
to time 2 affects likelihood of recidivism (Miles & Raynor, 2004; Raynor, 2007). Miles and
Raynor assessed offenders who were beginning select treatment program and then reassessed
these offenders upon program completion. The mean total LSI-R score was lower for all
treatment groups upon reassessment. Further, the authors found that individuals whose LSI-R
score was higher upon reassessment were more likely to recidivate than offenders whose total
LSI-R score was lower at reassessment than it was at the time of their initial assessment (Miles
and Raynor, 2004).
43
The current study hopes to build on the foundation of the aforementioned studies by
examining the predictive validity of the LSI-R at time 1 and time 2. In addition, this dissertation
will test how change in total LSI-R score from initial assessment to reassessment impacts the
predictive validity of the instrument. This project will also consider the specific domains of the
LSI-R— that is, is change in Criminal History more or less important than change in
Attitudes/Orientation or Accommodations. To date, changes scores have received limited
attention in the LSI literature. The goal of the current study is to contribute to the LSI research
by testing the predictive validity of the LSI-R at time 1, time 2, to assess the impact of change in
score on the instrument’s ability to predict recidivism, and to consider whether or not change in
certain domains is more or less important than change in other domains.
The preceding section has included a discussion of four conclusions from the past
twenty-plus years of research on the LSI. In sum, the LSI has received empirical and broadbased support as an international predictor of recidivism for adult offenders in a variety of
correctional settings. Given the instrument’s empirical support, theoretical underpinnings, and
the ease with which the instrument can adopted by a correctional agency for use by its officers, it
is no surprise that the instrument has become one of the most popular assessment tools in use
today. To that end, it is important to continue to conduct replication studies on the instrument
and also to consider new ways in which the LSI may be utilized to better serve correctional
agencies and the offender population.
44
Table 1.3 Measures of Recidivism Across LSI Studies
Measure of Recidivism
Reincarceration
Re-arrest
Reconviction
Halfway House Outcome
Absconding
New Charges
Parole Violation
Release Outcome
Program Completion
Domestic Violence
Violation
Number
Percent
Percent
Positive
Association
Number
16
14
12
4
2
2
2
2
1
1
1
35.6
31.1
26.7
8.9
4.4
4.4
4.4
4.4
2.2
2.2
2.2
100
92.9
100
100
100
100
100
100
100
100
100
12
10
10
3
1
1
2
2
1
1
1
a
Percent Valid
Predictor of
Recidivism
100a
76.9b
83.3
75
100c
100d
100
100
100
0
0
Calculation based on 12 instead of 15 because four studies failed to report significance.
Calculation based on 13 instead of 14 because one study failed to report significance.
c
Calculation based on 1 instead of 2 because one study failed to report significance.
d
Calculation based on 1 instead of 2 because one study failed to report significance.
b
45
Table 1.4 Types of LSI Instruments
Number
LSI-R
LSI
LSI-OR
LSI-R:SR
YO-LSI
YLS-CMI
27
11
1
1
2
3
Percent
60
24.4
2.2
2.2
4.4
6.7
a
Percent
Positive
Association
100
100
100
100
100
66.7
Number
Percent Valid
Predictor of
Recidivism
21
9
1
1
1
1
84a
90b
100
100
100c
33.3
Calculation based on 25 instead of 27 because two studies failed to report significance.
Calculation based on 10 instead of 11 because one study failed to report significance.
c
Calculation based on 1 instead of 2 because one study failed to report significance.
b
46
RESEARCH STRATEGY
The LSI-R has emerged as a widely used and important instrument for assessing the risk of
offenders. As we have seen, the existing literature suggests that the LSI-R has predictive validity
and can be used in the assessment of offenders in correctional agencies. However, the current
research is limited in that, with few exceptions, studies have measured the LSI-R's ability to
predict recidivism at one point in time. Although valuable, these investigations stop short of
providing a more definitive test of the LSI-R's utility.
The LSI-R is based on the principles of effective intervention (see Andrews and Bonta,
2003). This theory suggests that changes in dynamic risk factors — also known as criminogenic
needs — will be followed by changes in behavior. As these risk factors are lessened, then
involvement in criminal behavior should lessen. More directly, the theory asserts that treatment
programs that reduce these risk factors will in turn achieve reductions in subsequent recidivism.
It follows from this discussion that it is essential to measure risk levels not only at one point
in time but also over time. Offenders are under correctional supervision. Depending on the
nature of the correctional supervision — for example, it could be control oriented or treatment
oriented — we would expect risk levels possible to decrease, stay the same, or increase. Change
in risk levels thus should be associated with change in the level of criminal involvement or in the
likelihood of recidivating.
To assess change in risk level, it is thus essential to measure it at two points in time. The
LSI-R thus can be used to assess changes (or stability) in risk levels over time. If the LSI-R has
predictive validity, changes in LSI-R scores should be associated with changes in the risk of
recidivism. If this occurs, then it would be a powerful piece of evidence in support of the use of
47
the LSI-R in correctional agencies.
In this context, the current dissertation explores the predictive validity of the LSI-R with a
sample of 2,849 offenders on probation and parole in the state of Iowa. The members of the
sample were given the LSI-R on two occasions. As a result, the data allow for an assessment of
change in scores over time. In this regard, the dissertation explores the predictive validity of the
LSI-R in three ways: at time 1, at time 2, and changes between time 1 and time 2. These
analyses are conducted for the sample as a whole and for subgroups within the sample (males
and females, blacks and whites, probationers and parolees). In this way the dissertation attempts
to contribute the most systematic test of the LSI-R's predictive validity that is currently available.
48
CHAPTER 2
METHODS
The previous chapter traced the emergence, downfall, and gradual resurgence of
rehabilitation as a guiding philosophy of the correctional system in the United States. The
revival of rehabilitation coupled with the dramatic increase in the number of people under some
form of correctional supervision prompted the need to develop a way in which to classify and
manage the ever-growing offender population. Thus, offender classification instruments were
introduced. Over time, these classification instruments have been and continue to be tested,
retooled, and validated across different types of offender populations and in a variety of
countries and correctional settings.
This chapter will provide information concerning the collection of data for the current
project and a description of the sample characteristics. The types of independent and dependent
variables will be discussed before outlining the statistical techniques employed in the current
study. Finally, study limitations will be addressed.
SAMPLE
The data for the current study were obtained from the Iowa Department of Corrections.
The Iowa Department of Corrections collects and maintains records on all offenders under state
supervision in a statewide database known as the Iowa Correctional Offender Network (ICON).
A request for data was made to the Iowa Department of Corrections and granted resulting in the
data for this dissertation. These data were not collected specifically for this project but are part
49
of the standard record keeping system of the state of Iowa. The present sample includes 2,849
adult probationers and parolees from the state of Iowa. Further, the sample includes 1,976
probationers (69.4%) and 873 parolees (30.60%). The total sample is nearly eighty-six (85.9%)
male and eighty-five percent (84.8%) white. Blacks comprise roughly fifteen percent (15.2%) of
the sample. Table 2.1 provides an overview of characteristics from the current sample.
It is important to note that each offender included in the sample was administered the
LSI-R at least twice during the five year study period beginning August of 2000 and ending
September of 2005. Multiple assessment points provide the opportunity to assess the degree to
which an offender’s risk level changes over time. The mean number of days between an
offender’s initial assessment and reassessment is 364.80. The mean offender age is
approximately 40 (39.51) at the time of their first LSI-R assessment during the study period and
41 (40.52) at the time of the offender’s reassessment.
INDEPENDENT VARIABLES
Total LSI-R score at time 1 and time 2 serve as the primary independent variables used in
the analysis. The LSI-R total score can range from 0 to 54. The mean LSI-R score at initial
assessment is 26.95 and the mean score at reassessment is 27.63. Total scores from each of the
ten domains are also utilized to determine which domain is the strongest predictor of recidivism
and also to note whether or not change in a particular domain has an effect on the predictive
validity of the instrument. Similar to the LSI-R total score, the domain totals are tested at time 1
and time 2. Table 2.2 illustrates the number of points possible in each of the ten domains.
Categorical variables, including gender (male = 0, female = 1), race (white = 0, black = 1), and
50
supervision status (probation = 0, parole = 1), are also tested as possible predictors of recidivism.
Gender, race, age, supervisory status, risk category, and time at risk serve as control variables in
the multivariate analysis.
DEPENDENT VARIABLE
Recidivism is the outcome of interest in this study and is measured as any new
misdemeanor or felony conviction. The recidivism variable is dichotomous where 0 = no and 1
= yes. The follow-up period, called time at risk, varies across offenders depending on when the
LSI-R is first administered. Time at risk time 1 ranges from 558 days to 2,258 days with the
mean number of days at risk being 1,384. Time at risk time 2 ranges from 400 days to 2,124
days with the mean number of days at risk being 1,724.
STATISTICAL TECHNIQUES
Univariate, bivariate, and multivariate analysis will be presented in this dissertation.
Univariate statistics are employed to describe categories of a particular variable. The univariate
statistics are reported as raw numbers and percentages for variables such as race, gender, marital
status, age, supervisory status, measures of recidivism, and type of LSI assessment instrument.
Bivariate correlations between LSI-R total score and recidivism at time 1 and time 2 are
reported. Similarly, bivariate correlations between each of the domain totals and recidivism at
time 1 and time 2 are reported for the entire sample and then also reported for various subgroups
51
including gender, race, and supervisory status. Consistent with much of the previous research on
the LSI-R (See Table 1.1) bivariate correlations are reported as Pearson’s r correlation
coefficients.
Finally, multivariate analyses will be performed to determine the impact change scores
have on the instrument’s ability to predict recidivism. Moreover, multivariate analyses will be
used to determine which of the ten domains serves as the best predictor of recidivism. Finally,
analyses will be conducted to explore if change in certain domains is more or less important that
change in other domains. Again, these questions will be considered at time 1 and time 2.
LIMITATIONS OF THE STUDY
There are three study limitations that warrant discussion. First, the data were obtained
from the state of Iowa. The demographic characteristics of Iowa offenders are not representative
of all offenders in the United States. Specifically, the racial composition of U.S. probationers in
2006 was 55% white and 29% percent black whereas the current sample of probationers is 86%
white and 14% black. U.S. parolees in 2006 were 41% white and 39% black. The Iowa sample
of parolees is 82% white and nearly 18% black (Glaze & Bonczar, 2007).
A second limitation is that there is little consistency in when the reassessment was
completed. Although the reassessment took place on average one year later, some offenders were
reassessed as early as one month following their initial assessment. This is problematic because
research suggests that for best results, treatment should be administered over the course of three
or four months with at least one hundred contact hours (Smith, Gendreau, & Goggin, 2007).
52
Given the limited time between initial assessment and reassessment, there is unlikely to be any
change in an offender’s risk level.
A final limitation is that the offenders in the sample did not all receive the same
treatment. Further, some offenders received a variety of treatments. Although correctional
treatment programs are known to reduce the likelihood of recidivism by 10%, research also
suggests that there is considerable heterogeneity in the degree to which a treatment program
works to reduce recidivism (Andrews, Zinger, Hoge, Bonta, Gendreau, & Cullen, 1990; Cullen
& Gendreau, 2000). To that end, the findings from this study cannot be used to determine which
treatment programs are best for reducing recidivism.
This chapter has provided an overview of the sample, variables of interest, statistical
techniques, and limitations of this dissertation. Considering LSI-R total scores and domain
scores at two points in time allow the opportunity to assess the predictive validity of the LSI-R
for the entire sample and also for various subgroups. The impact that change in score has on the
instrument’s ability to predict is also examined. The results from the analysis will be presented
in the next chapter.
53
Table 2.1 Sample Characteristics
Variable
Number
Percent
Supervision Status
Probation
Parole
1,976
873
69.4
30.6
Sex
Male
Female
2,448
401
85.6
14.1
Race
White
Black
2,416
433
84.8
15.2
1,175
802
750
64
13
41.2
28.2
26.3
2.2
0.5
13
883
1,435
516
0.5
31.0
50.4
18.1
Marital Status
Single
Divorced
Married
Common Law
Widowed
Age
Under 25
26-35
36-45
46+
Mean Age at LSI-R Assessment 1
Mean Age at LSI-R Assessment 2
39.51
40.52
Mean LSI-R Score Time 1
Mean LSI-R Score Time 2
26.95
27.63
Mean Time Between Assessments
Mean Time at Risk T1
Mean Time at Risk T2
364.80
1,384.56
1,019.23
54
Table 2.2 Domain Totals
Domain
Criminal History l
Education/Employment
Financial
Family/Marital
Accommodation
Leisure/Recreation
Companions
Alcohol/Drug Problem
Emotional/Personal
Attitudes/Orientation
Total Possible
10
10
2
4
3
2
5
9
5
4
55
CHAPTER 3
RESULTS
The previous chapter provided an overview of the data collection, study methodology,
and analysis conducted for this project. The current chapter is divided into three sections. The
first section reports the results of the analysis for the entire sample. The second section includes
a discussion of the results for the gender, race, and supervisory status subgroups. The final
section explores the results from the analysis of the LS-R domains for the sample and subgroups.
THE IMPACT OF THE LSI-R ON RECIDIVISM
Bivariate Analysis Time 1 and Time 2 for Sample
The predictive validity of the LSI-R at time 1 is tested by calculating the correlation
between total LSI-R score at time 1 and recidivism. Similarly, the predictive validity of the LSIR at time 2 is tested by calculating the correlation between total LSI-R score at time 2 and
recidivism time 2. Table 3.1 depicts a Pearson correlation of .137 between total LSI-R score and
recidivism at time 1 for the sample and .193 between total LSI-R score and recidivism at time 2
for the sample. These relationships are statistically significant at the .01 level and indicate that
the higher the LSI-R total score, the more likely the offender is to recidivate. Moreover, these
findings suggest that the LSI-R is a valid predictor of recidivism at time 1 and time 2.
Confidence intervals are calculated around the Pearson correlation scores to determine
whether or not the correlation at time 1 is significantly different than the correlation at time 2.
The 95% confidence intervals for the correlation at time 1 are .10 to .17 compared to intervals of
56
Table 3.1 Bivariate Correlations Time 1 and 2 for Sample
Pearson Correlation
Sample Time 1
Sample Time 2
.137
.193
N
Sig
2849
2849
57
p<.01
p<.01
CI 95%
Lower
Upper
.10
.16
.17
.23
.16 to .23 for the correlation at time 2. The overlap in the range of correlations generated with
data from time 1 and time 2 indicates that there is no significant difference between the time 1
and time 2 correlations for the sample.
Given that risk category and recidivism are nominal variables, it is appropriate to
calculate Chi-square to test if offender risk category is related to offender likelihood to
recidivate. Tables 3.2 and 3.3 present chi-square results between risk category and recidivism at
time 1 and time 2 for the sample. At time 1, X2 (4, N = 2849) = 49.883, p = .000. At time 2, X2
(4, N=2849) = 104.580, p = .000. These findings suggest that risk category is a statistically
significant predictor of recidivism for the sample at time 1 and time 2.
Multivariate Analysis Time 1 and Time 2 for Sample
Although the bivariate correlation between total LSI-R Score and recidivism at time 1
and time 2 support the predictive validity of the LSI-R, it is important to include multivariate
analysis in order to control for variables such as race, age, gender, and supervisory status
(Hirschi & Gottfredson, 1983; Whitehead, 1991; Cohen, 1995; Gendreau, Little, & Goggin,
1996; Minor, Wells & Sims, 2003; Lowenkamp & Bechtel, 2007). Logistic regression models
are estimated at time 1 and time 2 using the above control variables and a measure of change on
the LSI-R (both the percent change and raw score change). The multivariate tables in this
section report regression coefficients, standard errors, Wald statistics, degrees of freedom,
significance values, exponent (B) values, and 95% confidence intervals for exponent (B).
Table 3.4 presents the multivariate model at time 1 for the entire sample at time 1. Time
at risk – the number of days between initial LSI-R assessment and record check – and total LSIR score at time 1 are statistically significant predictors. The Exp (B) values suggest that for a
58
Table 3.2 Risk Category and Recidivism Time 1 for Sample
Recidivism T1
No
Yes
Total
Low
135
78.9%
36
21.1%
171
100.0%
Low/Moderate
494
64.8%
268
35.2%
762
100.0%
Moderate
746
59.6%
506
40.4%
1252
100.0%
Medium/High
309
55.1%
252
44.9%
561
100.0%
High
45
43.7%
58
56.3%
103
100.0%
Total
1729
60.7%
1120
39.3%
2849
100.0%
59
Table 3.3 Risk Category and Recidivism Time 2 for Sample
Recidivism T2
No
Yes
Total
Low
182
91.9%
16
8.1%
198
100.0%
Low/Moderate
592
80.9%
140
19.1%
732
100.0%
Moderate
740
68.7%
337
31.3%
1077
100.0%
Medium/High
421
64.6%
231
35.4%
652
100.0%
High
114
60.0%
76
40.0%
190
100.0%
Total
2049
71.9%
800
28.1%
2849
100.0%
60
Table 3.4 Multivariate Time 1 for Sample
B
S.E.
Wald
df
Sig.
Exp(B)
Race
Age
Gender
Supervisory Status
Time at Risk T1
Total LSI-R Score T1
Constant
.085
-.008
.100
.044
.000
.039
-1.814
.109
.005
.112
.085
.000
.005
.327
.609
2.302
.806
.272
14.518
63.031
30.703
1
1
1
1
1
1
1
.435
.129
.369
.602
.000
.000
.000
1.088
.992
1.105
1.045
1.000
1.040
.163
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
74.937 (6)
3740.560
.026
.035
61
95.0% C.I.for EXP(B)
Lower
Upper
.880
.981
.888
.885
1.000
1.030
1.347
1.002
1.376
1.234
1.001
1.050
Table 3.5 Multivariate Time 2 for Sample
B
S.E.
Wald
df
Sig.
Exp(B)
Race
Age
Gender
Supervisory Status
Time at Risk T2
Total LSI-R Score T2
Constant
.098
-.009
.177
.036
.001
.059
-3.411
.120
.006
.122
.094
.000
.005
.340
.674
2.208
2.097
.148
70.395
129.321
100.783
1
1
1
1
1
1
1
.412
.137
.148
.701
.000
.000
.000
1.103
.991
1.193
1.037
1.001
1.061
.033
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
189.693 (6)
3190.053
.064
.093
62
95.0% C.I.for EXP(B)
Lower
Upper
.872
.979
.940
.862
1.001
1.050
1.395
1.003
1.515
1.247
1.001
1.072
one unit change, total LSI-R score at time 1 (1.040) is a slightly stronger predictor than time at
risk (1.000) at time 1. Race, age, gender, and supervisory status are not significant predictors of
recidivism at time 1.
Figure 3.1 presents the adjusted rate of recidivism by standard deviation for the sample at
time 1. The mean LSI-R score for the sample at time 1 is 27 and the corresponding rate of
recidivism is 27%. Offenders whose LSI-R total score at time 1 falls 1, 2, or 3 standard
deviations from the mean have a greater likelihood of recidivating as compared to offenders
whose LSI-R score at time 1 falls -1, -2, or -3 standard deviations from the mean. It is important
to note that the increases in the rate of recidivism are more dramatic with higher total LSI-R
score compared to the decreases in the rate of recidivism associated with lower total LSI-R
scores. Specifically, the rate of recidivism for offenders 3 standard deviations from mean (49%)
minus the rate of recidivism for offenders 2 standard deviations from the mean (41%) is 8%.
The rate of recidivism for offenders -2 standard deviations from the mean (16%) minus the rate
of recidivism for offenders -3 standard deviations from the mean (12%) is 4%.
The results from the multivariate model at time 2 for the sample are outlined in Table 3.5.
This model includes the following variables: race, age, gender, supervisory status, time at risk
time 2, and total LSI-R score time 2. Much like the multivariate model at time 1, the time 2
model reports time at risk time 2 and total LSI-R score time 2 as statistically significant
predictors. The Exp (B) values indicate that for a one unit change, total LSI-R score time 2
(1.061) is a stronger predictor than time at risk time 2 (1.001). Race, age, gender, and
supervisory status are not significant predictors of recidivism at time 2.
Figure 3.2 presents the adjusted rate of recidivism by standard deviation for the sample at
time 2. The mean LSI-R score for the sample at time 2 is 28 and the corresponding rate of
63
Figure 3.1 Change in Adjusted Rate of Recidivism by Standard Deviation at Time 1 for Sample
60
49
50
41
Total LSI-R Score
40
34
30
27
21
20
16
12
10
0
2 (-3SD)
11 (-2SD)
19 (-1SD)
27 (0)
35 (+1SD)
Standard Deviation
AdjustedRate of Recidivism
64
43 (+2SD)
52 (+3SD)
Figure 3.2 Change in Adjusted Rate of Recidivism by Standard Deviation at Time 2 for Sample
70
63
60
51
Total LSI-R Score
50
38
40
30
26
18
20
11
10
7
0
1 (-3SD)
10 (-2SD)
19 (-1SD)
28 (0)
37 (+1SD)
Standard Deviation
AdjustedRate of Recidivism
65
46 (+2SD)
54 (+3SD)
recidivism is 26%. Offenders whose LSI-R total score at time 2 falls 1, 2, or 3 standard
deviations from the mean have a greater likelihood of recidivating as compared to offenders
whose LSI-R score at time 1 falls -1, -2, or -3 standard deviations from the mean. It is important
to note that the increases in the rate of recidivism are more dramatic with higher total LSI-R
score compared to the decreases in the rate of recidivism associated with lower total LSI-R
scores. Specifically, the rate of recidivism for offenders 3 standard deviations from mean (63%)
minus the rate of recidivism for offenders 2 standard deviations from the mean (51%) is 12%.
The rate of recidivism for offenders -2 standard deviations from the mean (11%) minus the rate
of recidivism for offenders -3 standard deviations from the mean (7%) is 4%.
Change Analysis for Sample
This section discusses the results from the change analysis for the entire sample. Table
3.6 outlines the raw number and percent of offenders by risk category who recidivate after their
time 2 LSI-R assessment. The risk categories of the LSI-R at time 1 assessment are represented
by rows. The risk categories at time 2 assessment are represented by columns. Simply put, the
findings in the table indicate that high-risk offenders are more likely to recidivate than low-risk
offenders. Moreover, a change in risk level from time 1 assessment to time 2 assessment has an
impact on rate of recidivism. For example, offenders who moderate risk at time 1 assessment
and medium/high risk at time 2 have a 34.1% chance of failure. Offenders who are moderate
risk at time 1 and then low/moderate at time 2 have a 21.20% likelihood of recidivism. These
findings suggest that change in risk level does impact rate of recidivism for the sample.
Specifically, an increase in risk level results in higher failure rates and a decrease in risk level
results in lower failure rates.
66
The current project examines both percent change and raw change. Percent change is
meaningful when discussing increases and decreases in rates of recidivism and can be interpreted
by individuals who are not familiar with the LSI-R instrument. One drawback of using raw
change is that it requires the reader to be familiar with the scoring system of the LSI-R.
However, raw change is more descriptive than percent change because the risk categories of the
LSI-R are based on raw scores and change in raw score provides insight on any change in the
offender’s risk level. This is important because offender risk level is regarded as an important
factor when determining program placement. Specifically, high-risk offenders require more
intensive treatment and supervision than low-risk offenders (Andrews & Bonta, 1998; Gendreau,
1996; Marlowe et al., 2006).
Table 3.7 provides descriptive statistics regarding percent change and raw change for the
sample that contribute to the multivariate change analysis. Table 3.8 and Table 3.9 present the
results from the multivariate analysis of percent change and raw change. Time at risk time 2 and
risk category at time 1 are statistically significant in both models. Percent change and raw
change are also significant in their respective models. Finally, the interaction terms for each
model (risk category time 1 and percent change) and (risk category time 1 and raw change) are
significant predictors. The Exp (B) values reveal that for a one unit change, risk category at time
1 (1.620) and (1.584) is the strongest of the three significant predictors and percent change (.988)
and (.937) is the weakest of the three significant predictors in both models. Race, age, gender,
and supervisory status fail to be significant predictors in either of the multivariate change
models.
Figure 3.3 illustrates the impact change in offender risk level can have on likelihood of
recidivism. Forty-eight percent of offenders classified as high-risk recidivated during the study
67
period. A 10% reduction in offender risk level reduces the likelihood of recidivism for high-risk
offenders to 42%. A 10% reduction in risk level for offenders classified as medium/high risk
would reduce recidivism from 36% to 32%. The same reduction in risk level for moderate risk
offenders would result in a 7% drop in recidivism while the recidivism rate for low/moderate and
low-risk offenders would drop 2% and 1% respectively. It is important to note that the effect of
a 10% change in risk level varies across categories of risk. Specifically, reducing the risk of a
high-risk offender by 10% has a greater impact on rate of recidivism than reducing the risk of a
low-risk offender by 10%.
68
Table 3.6 Risk Classification and Recidivism Time 2
Reassessment Risk Category
Low/Moderate
Moderate
Medium/High
Initial Risk Category
Low
Low
6.9%
7/101
16.7%
9/54
50%
8/16
---
---
Low/Moderate
8.5%
7/82
17.9%
72/402
34.3%
79/230
39.5%
17/43
40%
2/5
Moderate
13.3%
2/15
21.1%
53/251
31.3%
207/661
34.1%
98/287
36.8%
14/38
Medium/High
---
24%
6/25
26%
40/154
34.6%
99/286
41.7%
40/96
High
---
---
18.8%
3/16
47.2%
17/36
39.2%
20/51
69
High
Table 3.7 Descriptives on Percent and Raw Change for Sample
Sample Raw Change
Sample Percent Change
Sample Time at Risk 2
N
Range
Minimum
Maximum
2849
2849
2849
46
571
1724
-24
-500
400
22
71
2124
70
Mean
-.67
-5.78
1019.23
Std. Deviation
6.325
33.775
346.266
Table 3.8 Multivariate Sample Percent Change
Race
Age
Gender
Supervisory Status
Time at Risk T2
Risk Category T1
Percent Change
Risk Category T1* Percent
Change
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
df
Sig.
Exp(B)
.074
-.008
.162
.032
.001
.476
-.009
-.004
.120
.006
.122
.094
.000
.052
.002
.383
1.803
1.771
.118
64.391
84.643
21.090
1
1
1
1
1
1
1
.536
.179
.183
.731
.000
.000
.000
1.077
.992
1.176
1.033
1.001
1.610
.991
.852
.980
.926
.859
1.001
1.455
.988
1.362
1.004
1.494
1.242
1.001
1.782
.995
.001
7.019
1
.008
.996
.994
.999
-2.695
.316
72.675
1
.000
.068
182.127 (8)
3197.619
.062
.089
71
95.0% C.I.for EXP(B)
Lower
Upper
Table 3.9 Multivariate Sample Raw Change
B
S.E.
Wald
df
Sig.
Exp(B)
.074
-.009
.163
.031
.001
.479
-.101
.120
.006
.122
.094
.000
.051
.016
.385
2.174
1.785
.106
66.190
87.007
42.056
1
1
1
1
1
1
1
.535
.140
.182
.745
.000
.000
.000
1.077
.991
1.177
1.031
1.001
1.615
.904
.852
.979
.927
.857
1.001
1.460
.876
1.362
1.003
1.495
1.240
1.001
1.786
.932
1.005
1.034
Race
Age
Gender
Supervisory Status
Time at Risk T2
Risk Category T1
Raw Change
Risk Category T1*Raw
Change
Constant
.020
.007
7.062
1
.008
1.020
-2.703
.316
73.048
1
.000
.067
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
187.493 (8)
3192.253
.064
.092
72
95.0% C.I.for EXP(B)
Lower
Upper
Figure 3.3 Change in Adjusted Recidivism by Risk Level When Assuming a 10
Percentage Point Change in Risk Level for the Sample
60
48
50
Adjusted Recidivism Rate
42
40
36
32
30
26
23
18
20
12
16
11
10
0
Low
Low/Moderate
Moderate
Medium/High
Risk Level
No Change
10% Decrease in LSI-R
73
High
THE IMPACT OF THE LSI-R ON RECIDIVISM BY GROUP
Bivariate Analysis Time 1 and Time 2 Gender
The predictive validity of the LSI-R at time 1 for males and females is tested by
calculating the correlation between total LSI-R score at time 1 and recidivism at time 1.
Similarly, the predictive validity of the LSI-R at time 2 for males and females is tested by
calculating the correlation between total LSI-R score at time 2 and recidivism at time 2. Table
3.10 presents the Pearson correlations at time 1 and time 2 for males and females. The
correlations for males are .141 at time 1 and .191 at time 2. The correlations for females are .112
at time 1 and .203 at time 2. The time 1 and 2 correlations for males are statistically significant
at the .01 level. For females, the correlation at time 1 is statistically significant at the .05 level
and the time 2 correlation is significant at the .01 level. The LSI-R is a valid predictor of
recidivism at time 1 and time 2 for males and females.
Confidence intervals are calculated around the Pearson correlation scores to determine
whether or not the LSI-R is a statistically better predictor for one gender over the other. For
males, the 95% confidence intervals for the correlation at time 1 are .10 to .18 as compared to .01
to .21 for females. The overlap in the range indicates that there is no statistically significant
difference in the predictive validity of the LSI-R for males or females at time 1. The same
comparison is examined for the confidence intervals at time 2. The range for males at time 2 is
.15 to .23 and the range for females is .11 to .29. Again, the overlap in the range indicates that
there is no statistically significant difference in the predictive validity of the LSI-R for males or
females at time 2.
74
Table 3.10 Bivariate Correlations Time 1 and Time 2 for Gender
Males Time 1
Females Time 1
Males Time 2
Females Time 2
Pearson Correlation
N
Sig.
.141
.112
.191
.203
2448
401
2448
401
P<.01
P<.05
P<.01
P<.01
75
CI 95%
Lower
Upper
.10
.01
.15
.11
.18
.21
.23
.29
Confidence intervals are also compared to determine if the LSI-R is a better predictor at
time 1 or time 2. For males, the time 1 range is .10 to .18 compared to the time 2 range of .15 to
.23. The overlap in range suggests there is no statistically significant difference in the predictive
validity of the LSI-R at time 1 compared to time 2 for males. The same analysis is carried out
for females. The range for females at time one is .01 to .21 compared to .11 to .29 at time 2.
Again, the overlap in range indicates there is no statistically significant difference in the
predictive validity of the LSI-R at time 1 compared to time 2 for females.
Chi-square is calculated for male and female offenders at time 1 and time 2 to test if
offender risk category is related to offender likelihood to recidivate. Tables 3.11 and 3.12
present chi-square results between risk category time 1 and recidivism time 1 for males X2 (4, N
= 2448) = 44.258, p = .000 and females X2 (4, N = 401) = 6.799, p = .147. Tables 3.13 and 3.14
present chi-square results between risk category time 2 and recidivism time 2 for males X2 (4,
N=2448) = 84.385, p = .000 and females X2 (4, N = 401) = 21.240, p = .000. These findings
suggest that risk category is a statistically significant predictor of recidivism for the males at time
1 and time 2. Risk category is not related to recidivism for females at time 1 but risk category is
a significant predictor of recidivism for females at time 2.
Multivariate Analysis Time 1 and Time 2 for Gender
Although the bivariate correlation between total LSI-R Score and recidivism at time 1
and time 2 for male and females supported the predictive validity of the LSI-R, it is important to
include multivariate analysis in order to control for variables such as race, age, and supervisory
status (Hirschi & Gottfredson, 1983; Whitehead, 1991; Cohen, 1995; Gendreau, Little, &
76
Table 3.11 Risk Category Time 1 and Recidivism Time 1 for Males
Recidivism T1
No
Yes
Total
Low
109
80.1%
27
19.9%
136
100.0%
Low/Moderate
430
65.2%
230
34.8%
660
100.0%
Moderate
650
60.1%
432
39.9%
1082
100.0%
Medium/High
264
54.8%
218
45.2%
482
100.0%
High
39
44.3%
49
55.7%
88
100.0%
Total
1492
60.9%
956
39.1%
2448
100.0%
77
Table 3.12 Risk Category Time 1 and Recidivism Time 1 for Females
Recidivism T1
No
Yes
Total
Low
26
74.3%
9
25.7%
35
100.0%
Low/Moderate
64
62.7%
38
37.3%
102
100.0%
Moderate
96
56.5%
74
43.5%
170
100.0%
Medium/High
45
57.0%
34
43.0%
79
100.0%
High
6
40.0%
9
60.0%
15
100.0%
Total
237
59.1%
164
40.9%
401
100.0%
78
Table 3.13 Risk Category Time 2 and Recidivism Time 2 for Males
Recidivism T2
No
Yes
Total
Low
146
92.4%
12
7.6%
158
100.0%
Low/Moderate
523
81.1%
122
18.9%
645
100.0%
Moderate
637
69.1%
285
30.9%
922
100.0%
Medium/High
369
65.1%
198
34.9%
567
100.0%
High
97
62.2%
59
37.8%
156
100.0%
Total
1772
72.4%
676
27.6%
2448
100.0%
79
Table 3.14 Risk Category Time 2 and Recidivism Time 2 for Females
Recidivism T2
No
Yes
Total
Low
36
90.0%
4
10.0%
40
100.0%
Low/Moderate
69
79.3%
18
20.7%
87
100.0%
Moderate
103
66.5%
52
33.5%
155
100.0%
Medium/High
52
61.2%
33
38.8%
85
100.0%
High
17
50.0%
17
50.0%
34
100.0%
Total
277
69.1%
124
30.9%
401
100.0%
80
Goggin, 1996; Minor, Wells & Sims, 2003; Lowenkamp & Bechtel, 2007). Logistic regression
models were estimated at time 1 and time 2 using the above control variables and a measure of
change on the LSI-R (both the percent change and raw score change). The multivariate tables in
this section report regression coefficients, standard errors, Wald statistics, degrees of freedom,
significance values, exponent (B) values, and 95% confidence intervals for exponent (B).
Tables 3.15 and 3.16 present the multivariate models at time 1 for males and females.
Time at risk and total LSI-R score at time 1 are statistically significant predictors for males.
Total LSI-R score at time 1 is the only significant predictor for females at time 1. The Exp (B)
values suggest that for a one unit change, total LSI-R score at time 1 (1.041) is a better predictor
than time at risk (1.000) at time 1 for males. Race, age, and supervisory status are not significant
predictors of recidivism at time 1 for males or females.
Figures 3.4 and 3.5 present the adjusted rate of recidivism by standard deviation for
males and females at time 1. The mean LSI-R score for the males and females at time 1 is 27.
The rate of recidivism for males with a mean LSI-R score is 26% compared to 39% for females.
Offenders whose LSI-R total score at time 1 falls 1, 2, or 3 standard deviations from the mean
have a greater likelihood of recidivating as compared to offenders whose LSI-R score at time 1
falls -1, -2, or -3 standard deviations from the mean. It is important to note that the increases in
the rate of recidivism are more dramatic with higher total LSI-R score compared to the decreases
in the rate of recidivism associated with lower total LSI-R scores. Specifically, the rate of
recidivism for male offenders 3 standard deviations from mean (49%) minus the rate of
recidivism for male offenders 2 standard deviations from the mean (41%) is 8%. Further, the
rate of recidivism for male offenders -2 standard deviations from the mean (16%) minus the rate
of recidivism for male offenders -3 standard deviations from the mean (12%) is 4%.
81
The comparison for female offenders is similar to that of their male counterparts. The
rate of recidivism for female offenders 3 standard deviations from mean (60%) minus the rate of
recidivism for female offenders 2 standard deviations from the mean (53%) is 7%. Further, the
rate of recidivism for female offenders -2 standard deviations from the mean (27%) minus the
rate of recidivism for female offenders -3 standard deviations from the mean (22%) is 5%.
Regardless of gender, the rate of recidivism increases with increases in total LSI-R score and the
rate of recidivism decreases with decreases in total LSI-R score at time 1.
Tables 3.17 and 3.18 present the multivariate models at time 2 for males and females.
Consistent with the results from the time 1 multivariate models for males, time at risk time 2 and
total LSI-R score at time 2 are statistically significant predictors for males. The Exp (B) values
suggest that for a one unit change, total LSI-R score at time 2 (1.061) is a better predictor than
time at risk (1.001) at time 2 for males. Total LSI-R score time 2 is a statistically significant
predictor of recidivism in the time 2 multivariate model for females. Race, age, and supervisory
status are not significant predictors of recidivism at time 2 for males or females.
Figures 3.6 and 3.7 present the adjusted rate of recidivism by standard deviation for
males and females at time 2. The mean LSI-R score for the males at time 2 is 28 and the mean
for females at time 2 is 27. The rate of recidivism for males with a mean LSI-R score is 26%
compared to 36% for females. Offenders whose LSI-R total score at time 1 falls 1, 2, or 3
standard deviations from the mean have a greater likelihood of recidivating as compared to
offenders whose LSI-R score at time 1 falls -1, -2, or -3 standard deviations from the mean. It is
important to note that the increases in the rate of recidivism are more dramatic with higher total
LSI-R score compared to the decreases in the rate of recidivism associated with lower total LSIR scores. Specifically, the time 2 rate of recidivism for male offenders 3 standard deviations
82
Table 3.15 Multivariate Time 1 for Males
B
S.E.
Wald
Df
Sig.
Exp(B)
Race
Age
Supervisory Status
Time at Risk T1
Total LSI-R Score T1
Constant
.017
-.007
-.003
.000
.041
-1.854
.120
.006
.091
.000
.005
.352
.020
1.589
.001
12.243
56.322
27.678
1
1
1
1
1
1
.889
.208
.976
.000
.000
.000
1.017
.993
.997
1.000
1.041
.157
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
65.210 (5)
3207.252
.026
.036
83
95.0% C.I.for EXP(B)
Lower
Upper
.803
.981
.835
1.000
1.030
1.288
1.004
1.191
1.001
1.053
Table 3.16 Multivariate Time 1 for Females
Race
Age
Supervisory Status
Time at Risk T1
Total LSI-R Score T1
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
Df
Sig.
Exp(B)
.365
-.012
.383
.000
.032
-1.583
.259
.015
.242
.000
.012
.893
1.983
.650
2.496
2.542
6.867
3.145
1
1
1
1
1
1
.159
.420
.114
.111
.009
.076
1.440
.988
1.466
1.000
1.033
.205
14.162 (5)
528.379
.035
.047
84
95.0% C.I.for EXP(B)
Lower
Upper
.867
.958
.912
1.000
1.008
2.393
1.018
2.357
1.001
1.058
Table 3.17 Multivariate Time 2 for Males
B
S.E.
Wald
Df
Sig.
Exp(B)
Race
Age
Supervisory Status
Time at Risk T2
Total LSI-R Score T2
Constant
.070
-.008
.023
.001
.060
-3.466
.134
.007
.101
.000
.006
.367
.277
1.502
.052
63.092
109.166
89.096
1
1
1
1
1
1
.599
.220
.820
.000
.000
.000
1.073
.992
1.023
1.001
1.061
.031
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
160.903 (5)
2720.964
.064
.092
85
95.0% C.I.for EXP(B)
Lower
Upper
.825
.979
.840
1.001
1.050
1.394
1.005
1.247
1.001
1.073
Table 3.18 Multivariate Time 2 for Females
B
Race
Age
Supervisory Status
Time at Risk T2
Total LSI-R Score T2
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
.196
-.015
.131
.001
.058
-2.877
S.E.
Wald
Df
Sig.
Exp(B)
.277
.017
.262
.000
.013
.903
.503
.798
.249
7.383
19.360
10.163
1
1
1
1
1
1
.478
.372
.618
.007
.000
.001
1.217
.985
1.140
1.001
1.059
.056
27.562 (5)
468.460
.066
.094
86
95.0% C.I.for EXP(B)
Lower
Upper
.708
.953
.682
1.000
1.032
2.092
1.018
1.904
1.002
1.087
Figure 3.4 Change in Adjusted Rate of Recidivism by Standard Deviation at Time 1 for Males
60
49
50
41
Total LSI-R Score
40
33
30
26
21
20
16
12
10
0
3 (-3SD)
11 (-2SD)
19 (-1SD)
27 (0)
35 (+1SD)
Standard Deviation
AdjustedRate of Recidivism
87
43 (+2SD)
51 (+3SD)
Figure 3.5 Change in Adjusted Rate of Recidivism by Standard Deviation at Time 1 for Females
70
60
60
53
Total LSI-R Score
50
46
39
40
32
30
27
22
20
10
0
0 (-3SD)
9 (-2SD)
17 (-1SD)
27 (0)
35 (+1SD)
Standard Deviation
AdjustedRate of Recidivism
88
44 (+2SD)
53 (+3SD)
Figure 3.6 Change in Adjusted Rate of Recidivism by Standard Deviation at Time 2 for Males
70
63
60
50
Total LSI-R Score
50
38
40
30
26
20
17
11
10
7
0
1 (-3SD)
10 (-2SD)
19 (-1SD)
28 (0)
37 (+1SD)
Standard Deviation
AdjustedRate of Recidivism
89
45 (+2SD)
54 (+3SD)
Figure 3.7 Change in Adjusted Rate of Recidivism by Standard Devation at Time 2 for Females
80
72
70
62
Total LSI-R Score
60
49
50
40
36
30
24
20
15
10
10
0
0 (-3SD)
8 (-2SD)
18 (-1SD)
27 (0)
37 (+1SD)
Standard Deviation
AdjustedRate of Recidivism
90
46 (+2SD)
54 (+3SD)
from mean (63%) minus the rate of recidivism for male offenders 2 standard deviations from the
mean (50%) is 13%. The time 2 rate of recidivism for male offenders -2 standard deviations
from the mean (11%) minus the rate of recidivism for male offenders -3 standard deviations from
the mean (7%) is 4%.
The comparison for female offenders at time 2 is similar to that of their male
counterparts. The time 2 rate of recidivism for female offenders 3 standard deviations from
mean (72%) minus the rate of recidivism for female offenders 2 standard deviations from the
mean (63%) is 10%. The time 2 rate of recidivism for female offenders -2 standard deviations
from the mean (15%) minus the rate of recidivism for female offenders -3 standard deviations
from the mean (10%) is 5%. Regardless of gender, the rate of recidivism increases with
increases in total LSI-R score and the rate of recidivism decreases with decreases in total LSI-R
score at time 2.
Change Analysis for Gender
This section discusses the results from the change analysis for males and females. Table
3.19 and Table 3.20 outline the raw number and percent of male and female offenders by risk
category who recidivate after their time 2 LSI-R assessment. The risk categories of the LSI-R at
time 1 assessment are represented by rows. The risk categories at time 2 assessment are
represented by columns. Regardless of gender, the findings in Table 3.19 and Table 3.20
indicate that high-risk offenders are more likely to recidivate than low-risk offenders. Change in
risk category from time 1 to time 2 has an effect of likelihood of failure. Offenders whose risk
level increases from time 1 to time 2 are more likely to recidivate compared to offenders whose
risk level decreases from time 1 to time 2. For example, males who moderate risk at time 1
91
assessment and medium/high risk at time 2 have a 31.6% chance of failure. Males who were
moderate risk at time 1 and then low/moderate at time 2 have a 20% likelihood of recidivism.
A similar trend is evident with females. Females who moderate risk at time 1 assessment
and medium/high risk at time 2 have a 51.4% chance of failure. Females who were moderate
risk at time 1 and then low/moderate at time 2 have a 27.8% likelihood of recidivism. Note, the
sample size for females is smaller (N = 401) than the sample of males (N = 2448). The small
size can result in small, unstable failure rates. Regardless of gender, increases in risk level
correspond with higher rates of recidivism and decreases in risk level result in lower rates of
recidivism.
The descriptive statistics for percent change and raw change for males and females are
presented in Table 3.21. Table 3.22 and Table 3.23 present the results from the multivariate
analysis of percent change for males and females. Table 3.24 and Table 3.25 show the raw
change for males and females. Time at risk time 2, and risk category at time 1 are statistically
significant in both male multivariate models. Percent change, raw change and the interaction
terms (risk category time 1 and percent change) and (risk category time 1 and raw change) are
also statistically significant predictors for males and females. The Exp (B) values reveal that for
a one unit change, risk category at time 1 (1.617) and (1.593) is the strongest of the three
significant predictors and percent change (.988) and raw change (.936) are the weakest of the
three significant predictors in the percent change and raw change models for males. Time at risk
time 2, risk category at time 1, and percent change are also significant predictors in the percent
change and raw change multivariate models for females. Consistent with the findings of their
male counterparts, the Exp (B) values suggest that for a one unit change, risk category at time 1
(1.626) and (1.524) is the strongest of the three statistically significant predictors for females.
92
Table 3.19 Risk Classification Time 1 and Time 2 and Recidivism Time 2 for Males
Reassessment Risk Category
Low/Moderate
Moderate
Medium/High
Initial Risk Category
Low
Low
5.3%
4/75
18.4%
9/49
41.7%
5/12
---
---
Low/Moderate
8.5%
6/71
18.1%
65/359
33%
63/191
41.2%
14/34
40%
2/5
Moderate
16.7%
2/12
20%
43/215
31.5%
180/572
31.6%
79/250
39.4%
13/33
Medium/High
---
22.7%
5/22
25.6%
34/133
35.5%
89/251
38.2%
29/76
High
---
---
21.4%
3/14
50%
16/32
35.7%
15/42
93
High
Table 3.20 Risk Classification Time 1 and Time 2 and Recidivism Time 2 for Females
Initial Risk Category
Low
Reassessment Risk Category
Low/Moderate
Moderate
Medium/High
High
Low
11.5%
3/26
0%
0/5
75%
¾
---
---
Low/Moderate
9.1%
1/11
16.3%
7/43
41.0%
16/39
33.3%
3/9
---
0%
0/3
27.8%
10/36
30.3%
27/89
51.4%
19/37
20%
1/5
Medium/High
---
33.3
1/3
28.6%
6/21
28.6%
10/35
55%
11/20
High
---
---
0%
0/2
25%
¼
55.6%
5/9
Moderate
94
Table 3.21 Descriptives on Percent and Raw Change for Gender
N
Males Raw Change
Females Raw Change
Males Percent Change
Females Percent Change
Males Time at Risk 2
Females Time at Risk 2
Range
2448
46
401
39
2448 454.76
401 561.54
2448 1724.00
401 1549.00
Minimum
-24
-21
-383.33
-500.00
400.00
435.00
Maximum
22
18
71.43
61.54
2124.00
1984.00
95
Mean
-.63
-.95
-5.4423
-7.8403
1018.9722
1020.8354
Std. Deviation
6.285
6.565
32.66831
39.85417
348.49914
332.72815
Table 3.22 Multivariate Percent Change for Males
Race
Age
Supervisory Status
Time at Risk T2
Risk Category T1
Percent Change
Risk Category T1* Percent
Change
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
Df
Sig.
Exp(B)
.035
-.008
.020
.001
.475
-.008
.134
.007
.101
.000
.056
.002
.069
1.398
.039
58.961
71.027
15.616
1
1
1
1
1
1
.793
.237
.843
.000
.000
.000
1.036
.992
1.020
1.001
1.609
.992
.797
.980
.837
1.001
1.440
.988
1.346
1.005
1.243
1.001
1.797
.996
-.004
.001
8.206
1
.004
.996
.993
.999
-2.725
.340
64.276
1
.000
.066
153.986 (7)
2727.881
.061
.088
96
95.0% C.I.for EXP(B)
Lower
Upper
Table 3.23 Multivariate Percent Change for Females
Race
Age
Supervisory Status
Time at Risk T2
Risk Category T1
Percent Change
Risk Category T1* Percent
Change
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
Df
Sig.
Exp(B)
.242
-.013
.122
.001
.488
-.014
.277
.017
.261
.000
.133
.006
.765
.572
.218
5.733
13.572
6.489
1
1
1
1
1
1
.382
.449
.641
.017
.000
.011
1.274
.987
1.130
1.001
1.630
.986
.741
.955
.677
1.000
1.257
.975
2.191
1.020
1.885
1.001
2.114
.997
.001
.004
.029
1
.864
1.001
.993
1.008
-2.278
.870
6.857
1
.009
.102
28.746 (7)
467.276
.069
.097
97
95.0% C.I.for EXP(B)
Lower
Upper
Table 3.24 Multivariate Raw Change for Males
B
S.E.
Wald
Df
Sig.
Exp(B)
.041
-.008
.018
.001
.484
-.102
.134
.007
.101
.000
.056
.017
.094
1.514
.033
60.509
74.502
35.325
1
1
1
1
1
1
.759
.219
.857
.000
.000
.000
1.042
.992
1.018
1.001
1.622
.903
.802
.979
.836
1.001
1.453
.873
1.354
1.005
1.241
1.001
1.810
.934
1.003
1.036
Race
Age
Supervisory Status
Time at Risk T2
Risk Category T1
Raw Change
Risk Category T1*Raw
Change
Constant
.019
.008
5.704
1
.017
1.019
-2.769
.341
66.011
1
.000
.063
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
161.408 (7)
2720.458
.064
.092
98
95.0% C.I.for EXP(B)
Lower
Upper
Table 3.25 Multivariate Raw Change for Females
Race
Age
Supervisory Status
Time at Risk T2
Risk Category T1
Raw Change
Risk Category T1*Raw
Change
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
Df
Sig.
Exp(B)
.203
-.014
.120
.001
.449
-.099
.276
.017
.261
.000
.129
.038
.538
.713
.212
5.981
12.182
6.766
1
1
1
1
1
1
.463
.398
.645
.014
.000
.009
1.225
.986
1.128
1.001
1.567
.906
.713
.954
.676
1.000
1.218
.841
2.104
1.019
1.880
1.002
2.017
.976
.022
.018
1.522
1
.217
1.022
.987
1.059
-2.142
.862
6.176
1
.013
.117
25.226 (7)
470.796
.061
.086
p<.001
99
95.0% C.I.for EXP(B)
Lower
Upper
Race, age, and supervisory status fail to be significant predictors in either of the
multivariate change models.
Figure 3.8 and Figure 3.9 illustrate the impact change in offender risk level can have on
likelihood of recidivism for males and females. Forty-seven percent of male offenders classified
as high-risk recidivated during the study period. A 10% reduction in offender risk level reduces
the likelihood of recidivism for high-risk male offenders to 41%. A 10% reduction in risk level
for male offenders classified as medium/high risk would reduce recidivism from 36% to 31%.
The same reduction in risk level for moderate risk male offenders would result in a 3% drop in
recidivism while the recidivism rate for low/moderate and low-risk male offenders would drop
2% and 1% respectively.
A similar trend is evident for female offenders. Fifty-five percent of female offenders
classified as high-risk recidivated during the study period. A 10% reduction in offender risk
level reduces the likelihood of recidivism for high-risk female offenders to 52%. A 10%
reduction in risk level for female offenders classified as medium/high-risk would reduce
recidivism from 43% to 40%. The same reduction in risk level for moderate, low/moderate, or
low-risk female offenders would result in a 2% drop in recidivism.
100
Figure 3.8 Change in Adjusted Recidivism Rate by Risk Level When Assuming a 10
Percentage Point Change in Risk Level for Males
60
Adjusted Recidivism Rate
50
47
41
40
36
31
30
26
23
18
20
12
16
11
10
0
Low
Low/Moderate
Moderate
Medium/High
Risk Level
No Change
10% Decrease in LSI-R
101
High
Figure 3.9 Change in Adjusted Recidivism Rate by Risk Level When Assuming a 10
Percentage Point Change in Risk Level for Females
60
55
52
50
Adjusted Recidivism Rate
43
40
40
31
30
22
29
20
20
15
13
10
0
Low
Low/Moderate
Moderate
Medium/High
Risk Level
No Change
10% Decrease in LSI-R
102
High
Bivariate Analysis Time 1 and Time 2 for Race
The predictive validity of the LSI-R at time 1 for blacks and whites is tested by
calculating the correlation between total LSI-R score at time 1 and recidivism at time 1.
Similarly, the predictive validity of the LSI-R at time 2 for blacks and whites is tested by
calculating the correlation between total LSI-R score at time 2 and recidivism at time 2. Table
3.26 presents the Pearson correlations at time 1 and time 2 for blacks and whites. The
correlations for blacks are .128 at time 1 and .232 at time 2. The correlations for whites are .139
at time 1 and .186 at time 2. The time 1 and 2 correlations for blacks and whites are statistically
significant at the .01 level. The LSI-R is a valid predictor of recidivism at time 1 and time 2 for
blacks and whites.
Confidence intervals are calculated around the Pearson correlation scores to determine
whether or not the predictive validity of the LSI-R varies across categories of race. For blacks,
the 95% confidence intervals for the correlation at time 1 are .03 to .22 as compared to .10 to .18
for whites. The overlap in the range indicates that there is no statistically significant difference
in the predictive validity of the LSI-R for blacks or whites at time 1. The same comparison is
examined for the confidence intervals at time 2. The range for blacks at time 2 is .14 to .32 and
the range for whites is .15 to .23. Again, the overlap in the range indicates that there is no
statistically significant difference in the predictive validity of the LSI-R for blacks or whites at
time 2.
Confidence intervals are also compared to determine if the LSI-R is a better predictor at
time 1 or time 2. For blacks, the time 1 range is .03 to .22 compared to the time 2 range of .14 to
.32. The overlap in range suggests there is no statistically significant difference in the predictive
validity of the LSI-R at time 1 compared to time 2 for blacks. The same comparison is examined
103
Table 3.26 Bivariate Correlations Time 1 and Time 2 for Race
Blacks Time 1
Whites Time 1
Blacks Time 2
Whites Time 2
Pearson Correlation
N
Sig
.128
.139
.232
.186
433
2416
433
2416
p<.01
p<.01
p<.01
p<.01
104
CI 95%
Lower
Upper
.03
.10
.14
.15
.22
.18
.32
.23
for whites. The range for whites at time one is .10 to .18 compared to .15 to .23 at time 2.
Again, the overlap in range indicates there is no statistically significant difference in the
predictive validity of the LSI-R at time 1 compared to time 2 for whites.
Chi-square is calculated for black and white offenders at time 1 and time 2 to test if
offender risk category is related to offender likelihood to recidivate. Tables 3.27 and 3.28
present chi-square results between risk category time 1 and recidivism time 1 for blacks X2 (4, N
= 433) = 9.104, p = .059 and whites X2 (4, N = 2416) = 43.864, p = .000. Tables 3.29 and 3.30
present chi-square results between risk category time 2 and recidivism time 2 for blacks X2 (4,
N= 433) = 22.545, p = .000 and whites X2 (4, N = 2416) = 84.580, p = .000. These findings
suggest that risk category is a statistically significant predictor of recidivism for the white
offenders at time 1 and time 2. Risk category is not related to recidivism for black offenders at
time 1 but risk category is a significant predictor of recidivism for blacks at time 2.
Multivariate Analysis Time 1 and Time 2 for Race
Although the bivariate correlation between total LSI-R Score and recidivism at time 1
and time 2 for male and blacks and whites supported the predictive validity of the LSI-R, it is
important to include multivariate analysis in order to control for variables such as age, gender,
and supervisory status (Hirschi & Gottfredson, 1983; Whitehead, 1991; Cohen, 1995; Gendreau,
Little, & Goggin, 1996; Minor, Wells & Sims, 2003; Lowenkamp & Bechtel, 2007). Logistic
regression models were estimated at time 1 and time 2 using the above control variables and a
measure of change on the LSI-R (both the percent change and raw score change). The
multivariate tables in this section report regression coefficients, standard errors, Wald statistics,
105
Table 3.27 Risk Category Time 1 and Recidivism Time 1 for Blacks
Recidivism T1
No
Yes
Total
Low
23
79.3%
6
20.7%
29
100.0%
Low/Moderate
85
59.0%
59
41.0%
144
100.0%
Moderate
100
61.7%
62
38.3%
162
100.0%
Medium/High
46
52.9%
41
47.1%
87
100.0%
High
4
36.4%
7
63.6%
11
100.0%
Total
258
59.6%
175
40.4%
433
100.0%
106
Table 3.28 Risk Category Time 1 and Recidivism Time 1 for Whites
Recidivism T1
No
Yes
Total
Low
112
78.9%
30
21.1%
142
100.0%
Low/Moderate
409
66.2%
209
33.8%
618
100.0%
Moderate
646
59.3%
444
40.7%
1090
100.0%
Medium/High
263
55.5%
211
44.5%
474
100.0%
High
41
44.6%
51
55.4%
92
100.0%
Total
1471
60.9%
945
39.1%
2416
100.0%
107
Table 3.29 Risk Category Time 2 and Recidivism Time 2 for Blacks
Recidivism T2
No
Yes
Total
Low
32
100.0%
0
.0%
32
100.0%
Low/Moderate
99
77.3%
29
22.7%
128
100.0%
Moderate
105
66.0%
54
34.0%
159
100.0%
Medium/High
53
62.4%
32
37.6%
85
100.0%
High
17
58.6%
12
41.4%
29
100.0%
Total
306
70.7%
127
29.3%
433
100.0%
108
Table 3.30 Risk Category Time 2 and Recidivism Time 2 for Whites
Recidivism T2
No
Yes
Total
Low
150
90.4%
16
9.6%
166
100.0%
Low/Moderate
493
81.6%
111
18.4%
604
100.0%
Moderate
635
69.2%
283
30.8%
918
100.0%
Medium/High
368
64.9%
199
35.1%
567
100.0%
High
97
60.2%
64
39.8%
161
100.0%
Total
1743
72.1%
673
27.9%
2416
100.0%
109
degrees of freedom, significance values, exponent (B) values, and 95% confidence intervals for
exponent (B).
Tables 3.31 and 3.32 present the multivariate models at time 1 for blacks and whites.
Time at risk and total LSI-R score at time 1 are statistically significant predictors for whites. The
Exp (B) values suggest that for a one unit change, total LSI-R score at time 1 (1.041) is a better
predictor than time at risk (1.000) at time 1 for whites. Total LSI-R score at time 1 is the only
statistically significant predictor in the time 1 multivariate model for blacks. Age, gender, and
supervisory status are not significant predictors of recidivism at time 1 for blacks or whites.
Figures 3.10 and 3.11 present the adjusted rate of recidivism by standard deviation for
blacks and whites at time 1. The mean LSI-R score for blacks at time 1 is 26 and the mean LSIR score for whites at time 1 is 27. The rate of recidivism for blacks with a mean LSI-R score is
18% compared to 29% for whites. Offenders whose LSI-R total score at time 1 falls 1, 2, or 3
standard deviations from the mean have a greater likelihood of recidivating as compared to
offenders whose LSI-R score at time 1 falls -1, -2, or -3 standard deviations from the mean. It is
important to note that the increases in the rate of recidivism are more dramatic with higher total
LSI-R score compared to the decreases in the rate of recidivism associated with lower total LSIR scores. Specifically, the rate of recidivism for black offenders 3 standard deviations from
mean (34%) minus the rate of recidivism for black offenders 2 standard deviations from the
mean (34%) is 6%. The rate of recidivism for black offenders -2 standard deviations from the
mean (11%) minus the rate of recidivism for black offenders -3 standard deviations from the
mean (9%) is 2%.
The rate of recidivism for white offenders 3 standard deviations from mean (52%) minus
the rate of recidivism for white offenders 2 standard deviations from the mean (43%) is 9%. The
110
rate of recidivism for white offenders -2 standard deviations from the mean (18%) minus the rate
of recidivism for female offenders -3 standard deviations from the mean (13%) is 5%.
Regardless of race, the rate of recidivism increases with increases in total LSI-R score and the
rate of recidivism decreases with decreases in total LSI-R score at time 1.
Tables 3.33 and 3.34 present the multivariate models at time 2 for blacks and whites.
Time at risk time 2 and total LSI-R score at time 2 are statistically significant predictors for
blacks and whites. The Exp (B) values suggest that for a one unit change, total LSI-R score at
time 2 (1.069) is a better predictor than time at risk at time 2 (1.001) for blacks. Similarly, total
LSI-R score time 2 (1.060) is a better predictor than time at risk time 2 (1.001) for whites. Age,
gender, and supervisory status are not significant predictors of recidivism at time 2 for blacks or
whites.
Figures 3.12 and 3.13 present the adjusted rate of recidivism by standard deviation for
blacks and whites at time 2. The mean LSI-R score for blacks at time 2 is 27 and the mean for
whites at time 2 is 28. The rate of recidivism for blacks with a mean LSI-R score is 16%
compared to 29% for whites. Offenders whose LSI-R total score at time 1 falls 1, 2, or 3
standard deviations from the mean have a greater likelihood of recidivating as compared to
offenders whose LSI-R score at time 1 falls -1, -2, or -3 standard deviations from the mean. It is
important to note that the increases in the rate of recidivism are more dramatic with higher total
LSI-R score compared to the decreases in the rate of recidivism associated with lower total LSIR scores. Specifically, the time 2 rate of recidivism for black offenders 3 standard deviations
from mean (53%) minus the rate of recidivism for black offenders 2 standard deviations from the
mean (38%) is 5%. The time 2 rate of recidivism for black offenders -2 standard deviations from
111
Table 3.31 Multivariate Time 1 for Blacks
B
S.E.
Wald
Df
Sig.
Exp(B)
Age
Gender
Supervisory Status
Time at Risk T1
Total LSI-R Score T1
Constant
-.008
.418
-.215
.000
.034
-1.657
.014
.255
.210
.000
.012
.840
.336
2.678
1.050
2.859
8.063
3.885
1
1
1
1
1
1
.562
.102
.306
.091
.005
.049
.992
1.519
.806
1.000
1.035
.191
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
14.515 (5)
568.704
.033
.045
112
95.0% C.I.for EXP(B)
Lower
Upper
.965
.921
.534
1.000
1.011
1.020
2.505
1.217
1.001
1.060
Table 3.32 Multivariate Time 1 for Whites
B
S.E.
Wald
Df
Sig.
Exp(B)
Age
Gender
Supervisory Status
Time at Risk T1
Total LSI-R Score T1
Constant
-.008
.027
.092
.000
.040
-1.861
.006
.125
.093
.000
.005
.356
1.672
.048
.995
12.087
54.501
27.300
1
1
1
1
1
1
.196
.827
.319
.001
.000
.000
.992
1.028
1.097
1.000
1.041
.156
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
64.160 (5)
3167.809
.026
.036
113
95.0% C.I.for EXP(B)
Lower
Upper
.981
.804
.915
1.000
1.030
1.004
1.313
1.315
1.001
1.052
Table 3.33 Multivariate Time 2 for Blacks
Age
Gender
Supervisory Status
Time at Risk T2
Total LSI-R Score T2
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
Df
Sig.
Exp(B)
-.014
.265
-.298
.001
.066
-3.358
.016
.277
.238
.000
.013
.890
.816
.917
1.568
12.924
24.756
14.233
1
1
1
1
1
1
.366
.338
.210
.000
.000
.000
.986
1.304
.742
1.001
1.069
.035
41.950 (5)
481.356
.093
.132
114
95.0% C.I.for EXP(B)
Lower
Upper
.956
.757
.465
1.001
1.041
1.017
2.244
1.184
1.002
1.097
Table 3.34 Multivariate Time 2 for Whites
B
S.E.
Wald
Df
Sig.
Exp(B)
Age
Gender
Supervisory Status
Time at Risk T2
Total LSI-R Score T2
Constant
-.008
.151
.100
.001
.058
-3.408
.007
.137
.102
.000
.006
.367
1.389
1.222
.953
57.609
103.419
86.127
1
1
1
1
1
1
.239
.269
.329
.000
.000
.000
.992
1.163
1.105
1.001
1.060
.033
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
150.493 (5)
2705.501
.060
.087
115
95.0% C.I.for EXP(B)
Lower
Upper
.979
.890
.904
1.001
1.048
1.005
1.520
1.351
1.001
1.071
Figure 3.10 Change in Adjusted Rate of Recidivism by Standard Deviation at Time 1 for Blacks
40
34
35
Total LSI-R Score
30
28
25
22
20
18
14
15
11
10
9
5
0
1 (-3SD)
9 (-2SD)
17 (-1SD)
26 (0)
34 (+1SD)
Standard Deviation
AdjustedRate of Recidivism
116
43 (+2SD)
51 (+3SD)
Table 3.11 Change in Adjusted Rate of Recidivism by Standard Deviation at Time 1 for Whites
60
52
50
43
Total LSI-R Score
40
36
29
30
23
18
20
13
10
0
3 (-3SD)
11 (-2SD)
19 (-1SD)
27 (0)
35 (+1SD)
Standard Deviation
AdjustedRate of Recidivism
117
43 (+2SD)
52 (+3SD)
Figure 3.12 Change in Adjusted Rate of Recidivism by Standard Deviation Time 2 for Blacks
60
53
50
38
Total LSI-R
40
30
25
20
16
9
10
3
5
0
0 (-3SD)
9 (-2SD)
18 (-1SD)
27 (0)
13 (+1SD)
Standard Deviation
AdjustedRate of Recidivism
118
45 (+2SD)
54 (+3SD)
Figure 3.13 Change in Adjusted Rate of Recidivism by Standard Deviation at Time 2 for Whites
70
65
60
54
Total LSI-R Score
50
41
40
29
30
20
20
13
10
8
0
0 (-3SD)
10 (-2SD)
19 (-1SD)
28 (0)
37 (+1SD)
Standard Deviation
AdjustedRate of Recidivism
119
46 (+2SD)
54 (+3SD)
the mean (5%) minus the rate of recidivism for black offenders -3 standard deviations from the
mean (3%) is 2%.
The time 2 rate of recidivism for white offenders 3 standard deviations from mean (65%)
minus the rate of recidivism for white offenders 2 standard deviations from the mean (54%) is
11%. The time 2 rate of recidivism for white offenders -2 standard deviations from the mean
(13%) minus the rate of recidivism for white offenders -3 standard deviations from the mean
(8%) is 5%. Regardless of race, the rate of recidivism increases with increases in total LSI-R
score and the rate of recidivism decreases with decreases in total LSI-R score at time 2.
Change Analysis for Race
This section discusses the results from the change analysis for blacks and whites. Table
3.35 and Table 3.36 outline the raw number and percent of black and white offenders by risk
category who recidivate after their time 2 LSI-R assessment. The risk categories of the LSI-R at
time 1 assessment are represented by rows. The risk categories at time 2 assessment are
represented by columns. Regardless of race, the findings in Table 3.33 and Table 3.34 indicate
that high-risk offenders are more likely to recidivate than low-risk offenders. Offenders whose
risk level increases from time 1 to time 2 are more likely to recidivate compared to offenders
whose risk level decreases from time 1 to time 2. For example, black offenders who are
moderate risk at time 1 assessment and medium/high risk at time 2 have a 36.1% chance of
failure. Blacks who were moderate risk at time 1 and then low/moderate at time 2 have a 22.2%
likelihood of recidivism. Note, the sample size for black offenders (N = 433) is smaller than the
sample of white offenders (N = 2416). The small sample size can result in small, unstable failure
rates.
120
A similar trend is evident with white offenders. Whites who are moderate risk at time 1
assessment and medium/high risk at time 2 have a 33.9% chance of failure. Whites who are
moderate risk at time 1 and then low/moderate at time 2 have a 21% likelihood of recidivism.
Regardless of race, increases in risk level correspond with higher rates of recidivism and
decreases in risk level result in lower rates of recidivism.
The descriptive statistics for percent change and raw change for blacks are whites are
presented in Table 3.37. Table 3.38 and Table 3.39 present the results from the multivariate
analysis of percent change for blacks and whites. Table 3.40 and Table 3.41 report the raw
change for blacks and whites. The multivariate percent change for blacks and whites indicate
risk category at time 1, time at risk 2, and percent change are significant predictors. The
interaction term (risk category time 1 and percent change) is a significant predictor for white
offenders but is not a predictor for black offenders. Notice, the Exp (B) values suggest risk
category at time 1 is the most powerful predictor for blacks (1.670) and whites (1.608) in the
percent change models.
The raw change models for blacks and whites find time at risk time 2, risk category time
1, and raw change to be significant predictors of recidivism. The interaction term (risk category
time 1 and raw change) is a significant predictor for black offenders but is not a predictor for
white offenders. The Exp (B) values show risk category time 1 to be the best predictor for
blacks (1.667) and whites (1.565) while raw change is the weakest significant predictor for
blacks (.928) and (.939) for whites. Age, gender, and supervisory status fail to be significant
predictors in either of the multivariate change models.
121
Table 3.35 Risk Classification and Recidivism Time 2 for Blacks
Reassessment Risk Category
Low/Moderate
Moderate
Medium/High
Initial Risk Category
Low
Low
0%
0/17
9.1%
1/11
0%
0/1
---
---
Low/Moderate
0%
0/12
22.6%
19/84
56.1%
23/41
42.9%
3/7
---
Moderate
0%
0/3
22.2%
6/27
24.4%
21/86
36.1%
13/36
20%
2/10
Medium/High
---
50%
3/6
32.1%
9/28
36.6%
15/41
50%
6/12
High
---
---
33.3%
1/3
100%
1/1
57.1%
4/7
122
High
Table 3.36 Risk Classification and Recidivism Time 2 for Whites
Reassessment Risk Category
Low/Moderate
Moderate
Medium/High
Initial Risk Category
Low
Low
8.3%
7/84
18.6%
8/43
53.3%
8/15
---
---
Low/Moderate
10%
7/70
16.7%
53/318
29.6%
56/189
38.9%
14/36
40%
2/5
16.7%
2/12
21%
47/224
32.3%
186/575
33.9%
85/251
42.9%
12/28
Medium/High
---
15.8%
3/19
24.6%
31/126
34.3%
84/245
40.5%
34/84
High
---
---
15.4%
2/13
45.7%
16/35
36.4%
16/44
Moderate
123
High
Table 3.37 Descriptives on Percent and Raw Change for Race
N
Blacks Raw Change
Whites Raw Change
Blacks Percent Change
Whites Percent Change
Blacks Time at Risk 2
Whites Time at Risk 2
Range
433
41
2416
45
433 366.67
2416 571.43
433 1698.00
2416 1724.00
Minimum
Maximum
-19
-24
-300.00
-500.00
403.00
400.00
22
21
66.67
71.43
2101.00
2124.00
124
Mean
-.91
-.63
-7.3368
-5.5008
1037.0600
1016.0397
Std. Deviation
6.439
6.305
34.99121
33.55188
344.78752
346.50510
Table 3.38 Multivariate Percent Change for Blacks
B
Age
Gender
Supervisory Status
Time at Risk T2
Risk Category T1
Percent Change
Risk Category T1* Percent
Change
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
S.E.
Wald
Df
Sig.
Exp(B)
-.012
.301
-.309
.001
.514
-.006
.016
.276
.237
.000
.132
.004
.583
1.192
1.697
12.089
15.218
2.407
1
1
1
1
1
1
.445
.275
.193
.001
.000
.121
.988
1.352
.734
1.001
1.671
.994
.958
.787
.461
1.000
1.291
.986
1.019
2.322
1.169
1.002
2.163
1.002
-.006
.003
3.700
1
.054
.994
.987
1.000
-2.600
.835
9.696
1
.002
.074
39.251(7)
484.055
.087
.124
125
95.0% C.I.for EXP(B)
Lower
Upper
Table 3.39 Multivariate Percent Change for Whites
Age
Gender
Supervisory Status
Time at Risk T2
Risk Category T1
Percent Change
Risk Category T1* Percent
Change
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
Df
Sig.
Exp(B)
-.007
.124
.097
.001
.469
-.009
.007
.137
.102
.000
.056
.002
1.171
.817
.898
52.723
69.039
18.671
1
1
1
1
1
1
.279
.366
.343
.000
.000
.000
.993
1.132
1.102
1.001
1.598
.991
.980
.865
.902
1.001
1.431
.986
1.006
1.480
1.347
1.001
1.785
.995
-.003
.002
3.830
1
.050
.997
.994
1.000
-2.714
.342
62.986
1
.000
.066
146.604(7)
2709.391
.059
.085
126
95.0% C.I.for EXP(B)
Lower
Upper
Table 3.40 Multivariate Raw Change for Blacks
Age
Gender
Supervisory Status
Time at Risk T2
Risk Category T1
Raw Change
Risk Category T1*Raw
Change
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
df
Sig.
Exp(B)
-.015
.268
-.325
.001
.567
-.155
.016
.277
.239
.000
.134
.040
.878
.935
1.849
11.934
18.030
14.736
1
1
1
1
1
1
.349
.333
.174
.001
.000
.000
.985
1.308
.722
1.001
1.764
.857
.955
.759
.452
1.000
1.357
.792
1.017
2.253
1.154
1.002
2.292
.927
.041
.018
5.280
1
.022
1.042
1.006
1.079
-2.671
.841
10.088
1
.001
.069
45.509 (7)
477.797
.100
.142
127
95.0% C.I.for EXP(B)
Lower
Upper
Table 3.41 Multivariate Raw Change for Whites
B
S.E.
Wald
df
Sig.
Exp(B)
-.008
.133
.098
.001
.463
-.091
.007
.137
.102
.000
.056
.017
1.330
.949
.915
54.297
68.741
28.740
1
1
1
1
1
1
.249
.330
.339
.000
.000
.000
.992
1.142
1.103
1.001
1.588
.913
.980
.874
.902
1.001
1.424
.883
1.005
1.493
1.348
1.001
1.772
.944
.999
1.032
Age
Gender
Supervisory Status
Time at Risk T2
Risk Category T1
Raw Change
Risk Category T1*Raw
Change
Constant
.015
.008
3.571
1
.059
1.015
-2.708
.342
62.802
1
.000
.067
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
146.896 (7)
2709.099
.059
.085
128
95.0% C.I.for EXP(B)
Lower
Upper
Figure 3.14 Change in Adjusted Recidivism Rate by Risk Level When Assuming a 10
Percentage Point Change in Risk Level for Blacks
60
50
50
Adjusted Recidivism Rate
43
38
40
32
30
26
23
18
20
11
16
11
10
0
Low
Low/Moderate
Moderate
Medium/High
Risk Level
No Change
129
10% Decrease in LSI-R
High
Figure 3.15 Change in Adjusted Recidivism Rate by Risk Level When Assuming a 10
Percentage Point Change in Risk Level for Whites
60
48
50
Adjusted Recidivism Rate
43
40
36
32
30
26
18
20
12
24
17
11
10
0
Low
Low/Moderate
Moderate
Medium/High
Risk Level
No Change
130
10% Decrease in LSI-R
High
Figure 3.14 and Figure 3.15 illustrate the impact change in offender risk level can have
on likelihood of recidivism for black and white offenders. Fifty percent of black offenders
classified as high-risk recidivated during the study period. A 10% reduction in offender risk
level reduces the likelihood of recidivism for high-risk black offenders to 43%. A 10% reduction
in risk level for black offenders classified as medium/high risk would reduce recidivism from
38% to 32%. The same reduction in risk level for moderate risk black offenders would result in
a 3% drop in recidivism while the recidivism rate for low/moderate risk would drop 2%. There
is no change in risk rate of recidivism for low-risk black offenders.
A similar trend is evident for white offenders. Forty-eight percent of white offenders
classified as high-risk recidivated during the study period. A 10% reduction in offender risk
level reduces the likelihood of recidivism for high-risk white offenders to 43%. A 10%
reduction in risk level for white offenders classified as medium/high risk would reduce
recidivism from 36% to 32%. The same reduction in risk level for moderate risk offenders
would result in a 2% reduction in recidivism while low/moderate and low-risk white offenders
experience a 1% drop in recidivism.
Bivariate Analysis Time 1 and Time 2 for Supervision Status
The predictive validity of the LSI-R at time 1 for offenders on probation tested by
examining the correlation between total LSI-R score at time 1 and recidivism at time 1.
Similarly, the predictive validity of the LSI-R at time 2 for offenders on parole is tested by
calculating the correlation between total LSI-R score at time 2 and recidivism at time 2. Table
3.42 presents the Pearson correlations at time 1 and time 2 for probation and parole. The
correlations for probation are .112 at time 1 and .183 at time 2. The correlations for parole are
131
Table 3.42 Bivariate Correlations for Supervision Status Time 1 and Time 2
Probation Time 1
Parole Time 1
Probation Time 2
Parole Time 2
Pearson Correlation
N
Sig
.112
.192
.183
.214
1976
873
1976
873
p<.01
p<.01
p<.01
p<.01
132
CI 95%
Lower
Upper
.07
.13
.14
.15
.16
.26
.23
.28
.192 at time 1 and .214 at time 2. The time 1 and 2 correlations for probation and parole are
statistically significant at the .01 level. The LSI-R is a valid predictor of recidivism at time 1 and
time 2 for offenders on probation and parole.
Confidence intervals are calculated around the Pearson correlation scores to determine
whether or not the predictive validity of the LSI-R varies by supervision status. For probation,
the 95% confidence intervals for the correlation at time 1 are .07 to .16 as compared to .13 to .26
for parole. The overlap in the range indicates that there is no statistically significant difference
in the predictive validity of the LSI-R for probation or parole at time 1. The same comparison is
examined for the confidence intervals at time 2. The range for probation at time 2 is .14 to .23
and the range for parole is .15 to .28. Again, the overlap in the range indicates that there is no
statistically significant difference in the predictive validity of the LSI-R for probation or parole at
time 2.
Confidence intervals are also compared to determine if the LSI-R is a better predictor at
time 1 or time 2. For probation, the time 1 range is .07 to .16 compared to the time 2 range of
.14 to .23. The overlap in range suggests there is no statistically significant difference in the
predictive validity of the LSI-R at time 1 compared to time 2 for probation. The same
comparison is examined for parole. The range for paroles at time one is .13 to .26 compared to
.15 to .28 at time 2. Again, the overlap in range indicates there is no statistically significant
difference in the predictive validity of the LSI-R at time 1 compared to time 2 for offenders on
parole.
Chi-square is calculated for offenders on probation and parole at time 1 and time 2 to test
if offender risk category is related to offender likelihood to recidivate. Tables 3.43 and 3.44
present chi-square results between risk category time 1 and recidivism time 1 for probation X2 (4,
133
N = 1976) = 21.098, p = .000 and parole X2 (4, N = 873) = 35.254, p = .000. Tables 3.45 and
3.46 present chi-square results between risk category time 2 and recidivism time 2 for probation
X2 (4, N= 1976) = 64.417, p = .000 and parole X2 (4, N = 873) = 43.561, p = .000. These
findings suggest that risk category is a statistically significant predictor of recidivism for
offenders on probation and parole at time 1 and time 2.
Multivariate Analysis Time 1 and Time 2 for Supervision Status
Although the bivariate correlation between total LSI-R Score and recidivism at time 1
and time 2 for probationers and parolees supported the predictive validity of the LSI-R, it is
important to include multivariate analysis in order to control for variables such as race, age, and
gender (Hirschi & Gottfredson, 1983; Whitehead, 1991; Cohen, 1995; Gendreau, Little, &
Goggin, 1996; Minor, Wells & Sims, 2003; Lowenkamp & Bechtel, 2007). Logistic regression
models were estimated at time 1 and time 2 using the above control variables and a measure of
change on the LSI-R (both the percent change and raw score change). The multivariate tables in
this section report regression coefficients, standard errors, Wald statistics, degrees of freedom,
significance values, exponent (B) values, and 95% confidence intervals for exponent (B).
Tables 3.47 and 3.48 present the multivariate models at time 1 for probationers and
parolees. Time at risk and total LSI-R score at time 1 are statistically significant predictors for
probationers. The Exp (B) values suggest that for a one unit change, total LSI-R score at time 1
(1.036) is a better predictor than time at risk (1.001) at time 1 for probationers. Total LSI-R
score at time 1 is the only statistically significant predictor in the time 1 multivariate model for
parolees. Race, age, and gender are not significant predictors of recidivism at time 1 for
probationers or parolees.
134
Figures 3.16 and 3.17 present the adjusted rate of recidivism by standard deviation for
offenders on probation and parole at time 1. The mean LSI-R score for probation and parole at
time 1 is 27. The rate of recidivism for an offender on probation with a mean LSI-R score is
54% compared to 34% for offenders on parole. Offenders whose LSI-R total score at time 1 falls
1, 2, or 3 standard deviations from the mean have a greater likelihood of recidivating as
compared to offenders whose LSI-R score at time 1 falls -1, -2, or -3 standard deviations from
the mean.
The probation sample at time 1 is interesting because the rate of recidivism increase or
decrease across standard deviations is fairly consistent. That is not the case when looking at
parole. For offenders on parole, the increases in the rate of recidivism are more dramatic with
higher total LSI-R score compared to the decreases in the rate of recidivism associated with
lower total LSI-R scores. Specifically, the rate of recidivism for offenders on parole 3 standard
deviations from the mean (63%) minus the rate of recidivism for parole offenders 2 standard
deviations from the mean (54%) is 9%. Further, the rate of recidivism for parole offenders -2
standard deviations from the mean (18%) minus the rate of recidivism for parole offenders -3
standard deviations from the mean (13%) is 5%. Regardless of supervision status, the rate of
recidivism increases with increases in total LSI-R score and the rate of recidivism decreases with
decreases in total LSI-R score at time 1.
Tables 3.49 and 3.50 present the multivariate models at time 2 for probationers and
parolees. Time at risk time 2 and total LSI-R score at time 2 are statistically significant
predictors for probationers and parolees. The Exp (B) values suggest that for a one unit change,
total LSI-R score at time 2 (1.061) is a better predictor than time at risk at time 2 (1.001) for
probationers. Similarly, total LSI-R score time 2 (1.062) is a better predictor than time at risk
135
Table 3.43 Risk Category Time 1 and Recidivism Time 1 for Probation
Recidivism T1
No
Yes
Total
Low
84
75.0%
28
25.0%
112
100.0%
Low/Moderate
345
65.1%
185
34.9%
530
100.0%
Moderate
521
60.2%
344
39.8%
865
100.0%
Medium/High
224
55.3%
181
44.7%
405
100.0%
High
33
51.6%
31
48.4%
64
100.0%
Total
1207
61.1%
769
38.9%
1976
100.0%
136
Table 3.44 Risk Category Time 1 and Recidivism Time 1 for Parole
Recidivism T1
No
Yes
Total
Low
51
86.4%
8
13.6%
59
100.0%
Low/Moderate
149
64.2%
83
35.8%
232
100.0%
Moderate
225
58.1%
162
41.9%
387
100.0%
Medium/High
85
54.5%
71
45.5%
156
100.0%
High
12
30.8%
27
69.2%
39
100.0%
Total
522
59.8%
351
40.2%
873
100.0%
137
Table 3.45 Risk Category Time 2 and Recidivism Time 2 for Probation
Recidivism T2
No
Yes
Total
Low
115
91.3%
11
8.7%
126
100.0%
Low/Moderate
408
80.8%
97
19.2%
505
100.0%
Moderate
520
69.6%
227
30.4%
747
100.0%
Medium/High
299
64.0%
168
36.0%
467
100.0%
High
83
63.4%
48
36.6%
131
100.0%
Total
1425
72.1%
551
27.9%
1976
100.0%
138
Table 3.46 Risk Category Time 2 and Recidivism Time 2 for Parole
Recidivism T2
No
Yes
Total
Low
67
93.1%
5
6.9%
72
100.0%
Low/Moderate
184
81.1%
43
18.9%
227
100.0%
Moderate
220
66.7%
110
33.3%
330
100.0%
Medium/High
122
65.9%
63
34.1%
185
100.0%
High
31
52.5%
28
47.5%
59
100.0%
Total
624
71.5%
249
28.5%
873
100.0%
139
Table 3.47 Multivariate Time 1 for Probation
B
S.E.
Wald
df
Sig.
Exp(B)
Race
Age
Gender
Time at Risk T1
Total LSI-R Score T1
Constant
.189
-.010
-.011
.001
.035
-1.771
.133
.007
.129
.000
.006
.394
2.019
2.275
.007
15.237
33.688
20.214
1
1
1
1
1
1
.155
.131
.933
.000
.000
.000
1.208
.990
.989
1.001
1.036
.170
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
46.841 (5)
2591.705
.023
.032
140
95.0% C.I.for EXP(B)
Lower
Upper
.931
.977
.768
1.000
1.023
1.569
1.003
1.275
1.001
1.048
Table 3.48 Multivariate Time 1 for Parole
B
S.E.
Wald
df
Sig.
Exp(B)
Race
Age
Gender
Time at Risk T1
Total LSI-R Score T1
Constant
-.122
-.003
.434
.000
.049
-1.891
.189
.010
.226
.000
.009
.580
.414
.118
3.682
.855
31.292
10.633
1
1
1
1
1
1
.520
.731
.055
.355
.000
.001
.885
.997
1.544
1.000
1.051
.151
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
37.707 (5)
1138.816
.042
.057
141
95.0% C.I.for EXP(B)
Lower
Upper
.611
.977
.991
1.000
1.033
1.282
1.016
2.405
1.001
1.069
Table 3.49 Multivariate Time 2 for Probation
B
S.E.
Wald
df
Sig.
Exp(B)
Race
Age
Gender
Time at Risk T2
Total LSI-R Score T2
Constant
.230
-.010
.140
.001
.059
-3.495
.146
.007
.141
.000
.006
.408
2.488
1.674
.979
57.077
86.538
73.370
1
1
1
1
1
1
.115
.196
.322
.000
.000
.000
1.259
.991
1.150
1.001
1.061
.030
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
134.798 (5)
2201.025
.066
.095
142
95.0% C.I.for EXP(B)
Lower
Upper
.946
.976
.872
1.001
1.048
1.675
1.005
1.517
1.001
1.074
Table 3.50 Multivariate Time 2 for Parole
Race
Age
Gender
Time at Risk T2
Total LSI-R Score T2
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
df
Sig.
Exp(B)
-.162
-.008
.290
.001
.060
-3.205
.212
.011
.243
.000
.009
.600
.587
.512
1.428
14.485
42.504
28.554
1
1
1
1
1
1
.444
.474
.232
.000
.000
.000
.850
.992
1.337
1.001
1.062
.041
58.949 (5)
984.843
.065
.094
143
95.0% C.I.for EXP(B)
Lower
Upper
.562
.971
.830
1.000
1.043
1.287
1.014
2.152
1.001
1.082
Figure 3.16 Change in Adjusted Rate of Recidivism by Standard Deviation Time 1 for Probation
73
67
61
Total LSI-R Score
54
47
40
34
3 (-3SD)
11 (-2SD)
19 (-1SD)
27 (0)
35 (+1SD)
Standard Deviation
AdjustedRate of Recidivism
144
43 (+2SD)
51 (+3SD)
Figure 3.17 Change in Adjusted Rate of Recidivism by Standard Deviation Time 1 for Parole
70
63
60
54
Total LSI-R Score
50
43
40
34
30
24
18
20
13
10
0
2 (-3SD)
10 (-2SD)
18 (-1SD)
27 (0)
35 (+1SD)
Standard Deviation
AdjustedRate of Recidivism
145
44 (+2SD)
52 (+3SD)
Figure 3.18 Change in Adjusted Rate of Recidivism by Standard Deviation at Time 2 for Probation
70
58
60
Total LSI-R Score
50
46
40
33
30
23
20
15
10
9
6
0
6 (-3SD)
9 (-2SD)
15 (-1SD)
28 (0)
37 (+1SD)
Standard Deviation
AdjustedRate of Recidivism
146
46 (+2SD)
54 (+3SD)
Figure 3.19 Change in Adjusted Rate of Recidivism by Standard Deviation at Time 2 for Parole
80
68
70
Total LSI-R Score
60
55
50
42
40
30
30
20
20
12
10
8
0
0 (-3SD)
9 (-2SD)
18 (-1SD)
27 (0)
36 (+1SD)
Standard Deviation
AdjustedRate of Recidivism
147
45 (+2SD)
54 (+3SD)
time 2 (1.001) for parolees. Race, age, and gender are not significant predictors of recidivism at
time 2 for probationers or parolees.
Figures 3.18 and 3.19 present the adjusted rate of recidivism by standard deviation for
offenders on probation and parole at time 2. The mean LSI-R score for probation at time 2 is 28
and the mean for parole at time 2 is 27. The rate of recidivism for probation with a mean LSI-R
score is 23% compared to 30% for parole. Offenders whose LSI-R total score at time 1 falls 1, 2,
or 3 standard deviations from the mean have a greater likelihood of recidivating as compared to
offenders whose LSI-R score at time 1 falls -1, -2, or -3 standard deviations from the mean. It is
important to note that the increases in the rate of recidivism are more dramatic with higher total
LSI-R score compared to the decreases in the rate of recidivism associated with lower total LSIR scores. Specifically, the time 2 rate of recidivism for probation offenders 3 standard
deviations from mean (58%) minus the rate of recidivism for probation offenders 2 standard
deviations from the mean (46%) is 12%. The time 2 rate of recidivism for probation offenders -2
standard deviations from the mean (9%) minus the rate of recidivism for probation offenders -3
standard deviations from the mean (6%) is 3%.
The time 2 rate of recidivism for parole offenders 3 standard deviations from mean
(68%) minus the rate of recidivism for white offenders 2 standard deviations from the mean
(55%) is 13%. The time 2 rate of recidivism for white offenders -2 standard deviations from the
mean (12%) minus the rate of recidivism for white offenders -3 standard deviations from the
mean (8%) is 4%. Regardless of supervision status, the rate of recidivism increases with
increases in total LSI-R score and the rate of recidivism decreases with decreases in total LSI-R
score at time 2.
148
Change Analysis for Supervision Status
This section discusses the results from the change analysis for offenders on probation and
parole. Table 3.51 and Table 3.52 outline the raw number and percent of probation and parole
offenders by risk category who recidivate after their time 2 LSI-R assessment. The risk
categories of the LSI-R at time 1 assessment are represented by rows. The risk categories at time
2 assessment are represented by columns. Regardless of supervision status, the findings in Table
3.51 and Table 3.52 indicate that high-risk offenders are more likely to recidivate than low-risk
offenders. Change in risk category from time 1 to time 2 has an effect of likelihood of failure.
Offenders whose risk level increases from time 1 to time 2 are more likely to recidivate
compared to offenders whose risk level decreases from time 1 to time 2. For example, offenders
on probation who moderate risk at time 1 assessment and medium/high risk at time 2 have a
35.4% chance of failure. Offenders on probation who are moderate risk at time 1 and then
low/moderate at time 2 have a 20.7% likelihood of recidivism.
A similar trend is evident with offenders on parole. Parolees who are moderate risk at
time 1 assessment and medium/high risk at time 2 have a 30.8% chance of failure. Offenders on
parole who are moderate risk at time 1 and then low/moderate at time 2 have a 22% likelihood of
recidivism. Note, the sample size for parole is smaller (N = 873) than the sample of offenders on
probation (N = 1976). The small size can result in small, unstable failure rates. Regardless of
supervision status, increases in risk level correspond with higher rates of recidivism and
decreases in risk level result in lower rates of recidivism.
149
The descriptive statistics for percent change and raw change for offenders on probation
and parole are presented in Table 3.53. Table 3.54 and Table 3.55 present the results from the
multivariate analysis of percent change for probationers and parolees. Table 3.56 and Table 3.57
report the raw change for probationers and parolees. The multivariate percent change models for
probationers and parolees indicates risk category at time 1 and time at risk 2 are significant
predictors for both groups. Percent change is a significant predictor for offenders on probation
and parole. However, the interaction term (risk category time 1 and percent change) is only
significant in the probation model. The Exp (B) values suggest that for a one unit change, risk
category at time 1 is the most powerful predictor for probationers (1.592) and parolees (1.682) in
the percent change models.
The raw change models for probationers and parolees find time at risk time 2, risk
category time 1, and raw change to be significant predictors of recidivism. The interaction term
(risk category time 1 and raw change) is significant in the parole model but is not a significant
predictor in the probation model. The Exp (B) values show that for a one unit change, risk
category time 1 to be the best predictor for probationers (1.554) and parolees (1.650) while raw
change is the weakest significant predictor for probationers (.937) and (.937) for parolees. Race,
age, and gender fail to be significant predictors in either of the multivariate change models.
Figure 3.20 and Figure 3.21 illustrate the impact change in offender risk level can have
on likelihood of recidivism for offenders on probation or parole. Forty-four percent of
probationers classified as high-risk recidivated during the study period. A 10% reduction in
offender risk level reduces the likelihood of recidivism for high-risk probationers to 38%. A
10% reduction in risk level for probationers classified as medium/high risk would reduce
recidivism from 34% to 29%. The same reduction in risk level for moderate risk probationers
150
Table 3.51 Risk Classification and Recidivism Time 2 for Probationers
Reassessment Risk Category
Low/Moderate
Moderate
Medium/High
Initial Risk Category
Low
Low
9.4%
6/64
13.2%
5/38
60%
6/10
---
---
Low/Moderate
9.1%
5/55
19.1%
54/283
33.1%
52/157
32.3%
10/31
50%
2/4
0%
0/7
20.7%
35/169
30.5%
139/456
35.4%
74/209
33.3%
8/24
Medium/High
---
20%
3/15
25.7%
29/113
36.3%
74/204
38.4%
28/73
High
---
---
9.1%
1/11
43.5%
10/23
33.3%
10/30
Moderate
151
High
Table 3.52 Risk Classification and Recidivism Time 2 for Parolees
Reassessment Risk Category
Low/Moderate
Moderate
Medium/High
Initial Risk Category
Low
Low
2.7%
1/37
25%
4/16
33.3%
2/6
---
---
Low/Moderate
7.4%
2/27
15.1%
18/119
37%
27/73
58.3%
7/12
0%
0/1
Moderate
25%
2/8
22%
18/82
33.2%
68/205
30.8%
24/78
42.9%
6/14
Medium/High
---
30%
3/10
26.8%
11/41
30.5%
25/82
52.2%
12/23
High
---
---
40%
2/5
53.8%
7/13
47.6%
10/21
152
High
Table 3.53 Descriptives on Percent and Raw Change for Supervision Status
N
Probation Raw Change
Parole Raw Change
Probation Percent Change
Parole Percent Change
Probation Time at Risk 2
Parole Time at Risk 2
Range
1976
44
873
45
1976 562.50
873
454.76
1976 1724.00
873 1650.00
Minimum
Maximum
Mean
-24
-23
-500.00
-383.33
400.00
431.00
20
22
62.50
71.43
2124.00
2081.00
-.75
-.50
-5.8489
-5.6237
1015.5445
1027.5865
153
Std. Deviation
6.249
6.492
32.92772
35.63662
344.71491
349.80787
Table 3.54 Multivariate Percent Change for Probation
Race
Age
Gender
Time at Risk T2
Risk Category T1
Percent Change
Risk Category T1* Percent
Change
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
df
Sig.
Exp(B)
.209
-.008
.124
.001
.457
-.009
.146
.007
.142
.000
.063
.002
2.068
1.305
.772
52.013
52.500
13.981
1
1
1
1
1
1
.150
.253
.380
.000
.000
.000
1.233
.992
1.133
1.001
1.580
.991
.927
.977
.858
1.001
1.396
.987
1.640
1.006
1.495
1.001
1.788
.996
-.004
.002
4.913
1
.027
.996
.993
1.000
-2.756
.379
52.855
1
.000
.064
127.927 (7)
2207.896
.063
.090
154
95.0% C.I.for EXP(B)
Lower
Upper
Table 3.55 Multivariate Percent Change for Parole
Race
Age
Gender
Time at Risk T2
Risk Category T1
Percent Change
Risk Category T1* Percent
Change
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
df
Sig.
Exp(B)
-.193
-.007
.273
.001
.518
-.009
.212
.011
.242
.000
.091
.003
.831
.385
1.274
13.216
32.240
7.473
1
1
1
1
1
1
.362
.535
.259
.000
.000
.006
.825
.993
1.314
1.001
1.679
.991
.545
.972
.818
1.000
1.404
.985
1.248
1.015
2.113
1.001
2.007
.998
-.003
.002
1.995
1
.158
.997
.992
1.001
-2.561
.560
20.953
1
.000
.077
58.683 (7)
985.110
.065
.093
155
95.0% C.I.for EXP(B)
Lower
Upper
Table 3.56 Multivariate Raw Change for Probation
B
S.E.
Wald
df
Sig.
Exp(B)
.210
-.009
.128
.001
.455
-.090
.145
.007
.141
.000
.062
.019
2.091
1.515
.814
53.240
52.946
22.381
1
1
1
1
1
1
.148
.218
.367
.000
.000
.000
1.234
.991
1.136
1.001
1.576
.914
.928
.977
.861
1.001
1.394
.880
1.641
1.005
1.499
1.001
1.781
.949
.996
1.032
Race
Age
Gender
Time at Risk T2
Risk Category T1
Raw Change
Risk Category T1*Raw
Change
Constant
.014
.009
2.280
1
.131
1.014
-2.748
.379
52.609
1
.000
.064
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
128.832 (7)
2206.991
.063
.091
156
95.0% C.I.for EXP(B)
Lower
Upper
Table 3.57 Multivariate Raw Change for Parole
Race
Age
Gender
Time at Risk T2
Risk Category T1
Raw Change
Risk Category T1*Raw
Change
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
df
Sig.
Exp(B)
-.197
-.008
.274
.001
.532
-.125
.212
.011
.243
.000
.091
.028
.859
.579
1.273
13.631
33.964
20.469
1
1
1
1
1
1
.354
.447
.259
.000
.000
.000
.821
.992
1.315
1.001
1.702
.882
95.0% C.I.for EXP(B)
Lower
Upper
.542
1.245
.970
1.013
.817
2.117
1.000
1.001
1.423
2.035
.835
.931
.032
.013
6.224
1
.013
1.033
1.007
-2.592
.562
21.279
1
.000
.075
64.384 (7)
979.408
.071
.102
157
1.060
would result in a 3% drop in recidivism while the recidivism rate for low/moderate risk would
drop 2%. There is no change in risk rate of recidivism for low-risk offenders on probation.
A similar trend is evident for offenders on parole. Fifty-six percent of parolees classified
as high-risk recidivated during the study period. A 10% reduction in offender risk level reduces
the likelihood of recidivism for high-risk parolees to 51%. A 10% reduction in risk level for
parolees classified as medium/high risk would reduce recidivism from 44% to 39%. The same
reduction in risk level for moderate risk offenders would result in a 3% reduction in recidivism
while low/moderate and low-risk white offenders experience a 1% drop in recidivism.
The previous section reviewed the results from the time 1, time 2, percent change, and
race change analysis for the entire sample and subgroups including gender, race, and supervisory
status. The results from the current project indicate that the LSI-R is a valid predictor for the
sample and subgroups at time 1 and time 2. Further, the findings from the percent change and
raw change analyses suggest that change in risk level is also a valid predictor of recidivism.
158
Figure 3.20 Change in Adjusted Recidivism Rate by Risk Level When Assuming a 10
Percentage Point Change in Risk Level for Probation
60
50
Adjusted Recidivism Rate
44
38
40
34
29
30
24
21
20
17
11
15
11
10
0
Low
Low/Moderate
Moderate
Medium/High
Risk Level
No Change
10% Decrease in LSI-R
159
High
Figure 3.21 Change in Adjusted Recidivism by Risk Level When Assuming a 10
Percentage Point Change in Risk Level for Parole
60
56
51
50
Adjusted Recidivism Rate
44
39
40
31
28
30
21
20
20
14
13
10
0
Low
Low/Moderate
Moderate
Medium/High
Risk Level
No Change
10% Decrease in LSI-R
160
High
THE IMPACT OF DOMAINS OF THE LSI-R ON RECIDIVISM
The LSI-R is comprised of ten distinct domains including criminal history,
education/employment, financial, family/marital accommodation, leisure/recreation,
companions, alcohol/drug problem, emotional/personal, and attitudes/orientation. The following
section examines bivariate correlations between each domain and recidivism at time 1 and time
2. Next, multivariate models including each of the domains and appropriate control variables
investigate which domain has the greatest impact on recidivism. Multivariate models estimating
how change (percent and raw) impact the ability of the domains to predict recidivism will also be
presented. Results for the entire sample will be presented first followed by findings for gender,
race, and supervision status.
Bivariate Analysis Time 1 and Time 2 for Sample
The predictive validity of the LSI-R domains at time 1 are tested by calculating the
correlation between the total score from each domain at time 1 and recidivism time 1. Tables
3.58 and 3.59 present the bivariate correlations for each domain of the LSI-R at time 1 and time
2. Nine of ten domains are significant predictors at time 1 for the sample. The only domain that
fails to predict recidivism for the sample at time 1 is emotional/personal.
The predictive validity of the LSI-R domains at time 2 are tested by calculating the
correlation between the total score from each domain at time 2 and recidivism time 2. Nine of
ten domains are significant predictors at time 1 for the sample. The only domain that fails to
predict recidivism for the sample at time 2 is emotional/personal.
161
Table 3.58 Bivariate Correlations Domain Totals and Recidivism Time 1 for Sample
Variables
Criminal History
Education/Employment
Financial
Family/Marital
Accommodation
Leisure/Recreation
Companions
Alcohol/Drug Problem
Emotional/Personal
Attitudes/Orientation
Total LSI-R Score
Pearson Correlation
N
Sig
.097
.107
.048
.048
.083
.052
.084
.072
-.027
.095
.137
2849
2849
2849
2849
2849
2849
2849
2849
2849
2849
2849
p<.01
p<.01
p<.05
p<.05
p<.01
p<.01
p<.01
p<.01
NA
p<.01
p<.01
162
CI 95%
Lower
Upper
.06
.07
.01
.01
.05
.02
.05
.04
-.06
.06
.10
.13
.14
.08
.08
.12
.09
.12
.11
.01
.13
.17
Table 3.59 Bivariate Correlations Domain Totals and Recidivism Time 2 for Sample
Variables
Criminal History
Education/Employment
Financial
Family/Marital
Accommodation
Leisure/Recreation
Companions
Alcohol/Drug Problem
Emotional/Personal
Attitudes/Orientation
Total LSI-R Score
Pearson Correlation
N
Sig
.119
.145
.078
.077
.103
.101
.106
.149
-.003
.141
.193
2849
2849
2849
2849
2849
2849
2849
2849
2849
2849
2849
p<.01
p<.01
p<.01
p<.01
p<.01
p<.01
p<.01
p<.01
NA
p<.01
p<.01
163
CI 95%
Lower
Upper
.08
.11
.04
.04
.07
.06
.07
.11
-.04
.11
.16
.16
.18
.11
.11
.14
.14
.14
.19
.03
.18
.23
Confidence intervals are calculated around the Pearson correlation scores to determine
whether or not the correlations at time 1 are significantly different from the correlations at time
2. Confidence intervals from each of the ten domains at time 1 overlap the confidence intervals
for each domain at time 2. For example, the time 1 confidence interval for criminal history is .06
to .13. The confidence interval for criminal history for time 2 is .08 to .16. The overlap in the
time 1 and time 2 ranges indicates that there is no significant difference between the time 1 and
time 2 criminal history correlations. The finding of no significant difference between time 1 and
time 2 correlations is consistent across all ten domains of the LSI-R for the entire sample.
The confidence intervals for all ten domains are also compared to one another to
determine if one domain is significantly stronger than another. Upon review of the domain
confidence intervals at time nine of the ten domains are significant predictors of recidivism at
time 1 and time 2. The emotional/personal domain is the only domain that fails to predict
recidivism at either time. The confidence intervals of the nine significant domains all overlap
one another which indicate that no single domain emerges as a significantly better predictor than
the others at time 1 or time 2.
Multivariate Analysis Time 1 and Time 2 for Sample
Although the bivariate correlations for nine of the ten domains of the LSI-R -R Score are
able to predict recidivism at time 1 and time 2, it is important to include multivariate analysis in
order to control for variables such as race, age, gender, and supervisory status. Logistic
regression models were estimated at time 1 and time 2 using the above control variables and a
measure of change on the LSI-R (both the percent change and raw score change). The
multivariate tables in this section report regression coefficients, standard errors, Wald statistics,
164
degrees of freedom, significance values, exponent (B) values, and 95% confidence intervals for
exponent (B).
Table 3.60 depicts the multivariate model for all domains of the LSI-R at time 1. Race,
age, gender, supervisory status, and time at risk time 1 are included in the model as control
variables. Time at risk time 1, criminal history time 1, education/employment time 1, and
emotional/personal are statistically significant predictors. The Exp (B) values suggest that for a
one unit change, criminal history time 1 (1.079) appears to be the best domain predictor of
recidivism at time 1. Race, age, gender, and supervisory status are not significant predictors of
recidivism at time 1.
Table 3.61 outlines the multivariate model for all domains of the LSI-R at time 2. Race,
age, gender, supervisory status, and time at risk time 2 are included in the model as control
variables. Time at risk time 2, criminal history time 2, alcohol/drug problem time 2,
attitudes/orientation time 2, and education/employment time 2 are statistically significant
predictors. The Exp (B) values suggest that for a one unit change, criminal history time 2
(1.110) is the best domain predictor of recidivism at time 2. Race, age, gender, and supervisory
status are not significant predictors of recidivism at time 2.
Change Analysis for Sample
The current project examines both percent change and raw change in each of the ten
domains of the LSI-R. Percent change is meaningful when discussing increases and decreases in
rates of recidivism and can be interpreted by individuals who are not familiar with the LSI-R
165
Table 3.60 Multivariate Domains Time 1 for Sample
Race
Age
Gender
Supervisory Status
Time at Risk T1
Criminal History T1
Education/Employment T1
Financial T1
Family/Marital T1
Accommodation T1
Leisure/Recreation T1
Companion T1
Alcohol/Drug Problem T1
Emotional/Personal T1
Attitudes/Orientation T1
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
Df
Sig.
Exp(B)
.121
-.009
.112
.043
.000
.076
.047
.048
-.018
.086
-.010
.050
.038
-.077
.078
-1.818
.110
.006
.113
.085
.000
.018
.016
.065
.035
.042
.062
.038
.016
.029
.032
.335
1.226
2.428
.998
.260
17.212
17.316
8.230
.543
.267
4.090
.026
1.748
5.699
7.240
5.943
29.450
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
.268
.119
.318
.610
.000
.000
.004
.461
.605
.043
.871
.186
.017
.007
.015
.000
1.129
.991
1.119
1.044
1.000
1.079
1.048
1.049
.982
1.089
.990
1.052
1.039
.926
1.081
.162
103.546 (15)
3711.952
.036
.048
166
95.0% C.I.for EXP(B)
Lower
Upper
.911
.981
.897
.884
1.000
1.041
1.015
.924
.917
1.003
.877
.976
1.007
.875
1.015
1.400
1.002
1.395
1.234
1.001
1.119
1.082
1.191
1.052
1.184
1.117
1.133
1.072
.979
1.151
Table 3.61 Multivariate Domains Time 2 for Sample
Race
Age
Gender
Supervisory Status
Time at Risk T2
Criminal History T2
Education/Employment T2
Financial T2
Family/Marital T2
Accommodation T2
Leisure/Recreation T2
Companion T2
Alcohol/Drug Problem T2
Emotional/Personal T2
Attitudes/Orientation T2
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
Df
Sig.
Exp(B)
.126
-.010
.172
.032
.001
.105
.055
.083
-.016
.042
.092
.029
.088
-.079
.117
-3.430
.121
.006
.123
.095
.000
.021
.019
.074
.039
.046
.072
.045
.019
.032
.035
.354
1.080
2.473
1.960
.113
75.683
24.122
8.302
1.269
.170
.821
1.615
.396
21.922
5.983
11.352
94.081
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
.299
.116
.162
.737
.000
.000
.004
.260
.680
.365
.204
.529
.000
.014
.001
.000
1.134
.990
1.188
1.032
1.001
1.110
1.057
1.087
.984
1.042
1.096
1.029
1.092
.924
1.124
.032
222.508 (15)
3157.238
.075
.108
167
95.0% C.I.for EXP(B)
Lower
Upper
.894
.978
.933
.857
1.001
1.065
1.018
.940
.911
.953
.951
.941
1.052
.867
1.050
1.439
1.002
1.512
1.243
1.001
1.158
1.097
1.257
1.063
1.140
1.264
1.125
1.132
.984
1.203
instrument. One drawback of using raw change is that it requires the reader to be familiar with
the scoring system of the LSI-R. However, raw change is more descriptive than percent change
because the risk categories of the LSI-R are based on raw scores and change in raw score
provides insight on any change in the offender’s risk level. This is important because offender
risk level is regarded as an important factor when determining program placement. Specifically,
high-risk offenders require more intensive treatment and supervision than low-risk offenders
(Andrews & Bonta, 1998; Gendreau, 1996; Marlowe et al., 2006).
The descriptive statistics for percent change and raw change for each domain in the
sample are presented in Tables 3.62 and 3.63. Table 3.64 and Table 3.65 present the results from
the multivariate analysis of percent change and raw change. Time at risk time 2, is the only
significant predictor in the percent change model. Time at risk time 2, risk category time 1, raw
change criminal history, and raw change leisure/recreation are statistically significant predictors
in the raw change model. The Exp (B) values in the raw change model reveal that for a one unit
change, risk category at time 1 (1.583) is the strongest of the significant predictors. Raw change
in criminal history is a significantly better predictor than raw change raw change
education/employment, raw change accommodation, raw change emotional/personal, raw change
leisure/recreation, raw change companion, and raw change attitudes/orientation. Race, age,
gender, and supervisory status fail to be significant predictors in either of the multivariate change
models.
168
Table 3.62 Descriptives Percent Change for Sample
N
Sample Percent Change Criminal
History
Sample Percent Change
Education/Employment
Sample Percent Change Financial
Sample Percent Change
Family/Marital
Sample Percent Change
Accommodation
Sample Percent Change
Alcohol/Drug Problem
Sample Percent Change
Emotional/Personal
Sample Percent Change
Leisure/Recreation
Sample Percent Change
Companion
Sample Percent Change
Attitudes/Orientation
Range Minimum
Maximum
Mean
Std.
Deviation
2745
600
-500
100
-10.01
34.963
2715
700
-600
100
-18.43
93.491
2335
200
-100
100
5.78
41.717
2484
400
-300
100
-6.25
47.315
1812
300
-200
100
6.33
68.749
2579
900
-800
100
-15.48
87.904
2405
500
-400
100
-12.21
62.285
2488
200
-100
100
2.17
41.330
2703
400
-300
100
-3.93
36.331
2055
400
-300
100
-7.26
76.840
169
Table 3.63 Multivariate Domains Percent Change for Sample
Race
Age
Gender
Supervisory Status
Time at Risk T2
Risk Category T1
Percent Change Criminal History
Percent Change Education/Employment
Percent Change Financial
Percent Change Family/Marital
Percent Change Accommodation
Percent Change Alcohol/Drug Problem
Percent Change Emotional/Personal
Percent Change Leisure/Recreation
Percent Change Companion
Percent Change Attitudes/Orientation
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
Df
Sig.
Exp(B)
-.156
-.008
.428
.055
.001
.222
-.006
.000
-.003
.000
-.002
-.002
-.002
.001
.000
-.002
-2.360
.214
.010
.204
.161
.000
.108
.003
.001
.002
.002
.001
.001
.001
.002
.002
.001
.581
.531
.614
4.419
.117
32.721
4.263
4.074
.082
1.681
.064
1.826
4.522
1.509
.076
.008
3.814
16.497
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
.466
.433
.036
.733
.000
.039
.044
.775
.195
.801
.177
.033
.219
.783
.929
.051
.000
.856
.992
1.534
1.056
1.001
1.249
.994
1.000
.997
1.000
.998
.998
.998
1.001
1.000
.998
.094
65.996 (16)
1098.013
.070
.097
170
95.0% C.I.for EXP(B)
Lower
Upper
.562
.973
1.029
.771
1.001
1.011
.989
.998
.994
.996
.996
.996
.996
.996
.995
.995
1.302
1.012
2.286
1.448
1.002
1.543
1.000
1.003
1.001
1.003
1.001
1.000
1.001
1.005
1.005
1.000
Table 3.64 Descriptives Raw Change Domains for Sample
N
Sample Raw Change Criminal History
Sample Raw Change Education/Employment
Sample Raw Change Financial
Sample Raw Change Family/Marital
Sample Raw Change Accommodation
Sample Raw Change Alcohol/Drug Problem
Sample Raw Change Emotional/Personal
Sample Raw Change Leisure/Recreation
Sample Raw Change Companion
Sample Raw Change Attitudes/Orientation
2849
2849
2849
2849
2849
2849
2849
2849
2849
2849
Range Minimum Maximum
11
15
4
8
6
16
9
4
8
8
-6
-8
-2
-4
-3
-9
-5
-2
-4
-4
171
5
7
2
4
3
7
4
2
4
4
Mean
-.31
.05
.01
-.07
-.04
.02
-.14
.00
-.02
-.17
Std. Deviation
.840
2.475
.565
.784
1.020
2.130
1.005
.689
.763
1.321
Table 3.65 Multivariate Domains Raw Change for Sample
Race
Age
Gender
Supervisory Status
Time at Risk T2
Risk Category T1
Raw Change Criminal History
Raw Change Education/Employment
Raw Change Financial
Raw Change Family/Marital
Raw Change Accommodation
Raw Change Alcohol/Drug Problem
Raw Change Emotional/Personal
Raw Change Leisure/Recreation
Raw Change Companion
Raw Change Attitudes/Orientation
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
Df
Sig.
Exp(B)
.060
-.007
.154
.043
.001
.460
-.271
-.024
-.095
-.086
.034
-.166
-.021
-.077
-.041
-.094
-2.762
.121
.006
.123
.094
.000
.051
.051
.020
.083
.058
.046
.070
.059
.023
.045
.038
.319
.249
1.443
1.570
.204
64.329
80.405
28.624
1.452
1.301
2.219
.569
5.669
.130
11.540
.856
6.243
75.170
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
.618
.230
.210
.651
.000
.000
.000
.228
.254
.136
.450
.017
.718
.001
.355
.012
.000
1.062
.993
1.166
1.044
1.001
1.583
.763
.976
.910
.918
1.035
.847
.979
.926
.959
.910
.063
207.057 (16)
3172.489
.070
.101
172
95.0% C.I.for EXP(B)
Lower
Upper
.839
.981
.917
.867
1.001
1.432
.690
.938
.773
.819
.947
.739
.872
.886
.879
.845
1.345
1.005
1.484
1.256
1.001
1.751
.842
1.015
1.070
1.028
1.132
.971
1.099
.968
1.047
.980
THE IMPACT OF THE LSI-R DOMAINS ON RECIDIVISM BY GROUP
Bivariate Analysis Time 1 and Time 2 for Gender
The predictive validity of the LSI-R domains at time 1 are tested by calculating the
correlation between the total score from each domain at time 1 and recidivism time 1. Tables
3.66 and 3.67 present the bivariate correlations for each domain of the LSI-R at time 1 for males
and females. Nine of the ten domains are significant predictors for males at time 1.
Emotional/personal is the only domain that fails to predict recidivism at time 1 for males. The
overlapping confidence intervals suggest that no single domain predicts better than the others for
males at time 1. For females, criminal history is the only significant predictor of recidivism at
time 1.
The predictive validity of the LSI-R domains at time 2 are tested by calculating the
correlation between the total score from each domain at time 2 and recidivism time 2. Nine of
the ten domains are significant predictors for males at time 2. Emotional/personal is the only
domain that fails to predict recidivism at time 2 for males. The overlapping confidence intervals
suggest that no single domain predicts better than the others for males at time 2. For females,
criminal history, education/employment, family/marital, accommodation, alcohol/drug problem,
and attitudes/orientation are statistically significant predictors at time 2. The overlapping
confidence intervals suggest that no single domain predicts better than the others for females at
time 2.
173
Table 3.66 Bivariate Correlations Domain Totals and Recidivism Time 1 and Time 2 for Males
Variables
Criminal History T1
Criminal History T2
Education/Employment T1
Education/Employment T2
Financial T1
Financial T2
Family/Marital T1
Family/Marital T2
Accommodation T1
Accommodation T2
Leisure/Recreation T1
Leisure/Recreation T2
Companions T1
Companions T2
Alcohol/Drug Problem T1
Alcohol/Drug Problem T2
Emotional/Personal T1
Emotional/Personal T2
Attitudes/Orientation T1
Attitudes/Orientation T2
Total LSI-R Score T1
Total LSI-R Score T2
Pearson
Correlation
N
.087
.106
.116
.151
.048
.082
.046
.069
.080
.096
.059
.109
.101
.121
.075
.147
-.036
-.004
.104
.136
.141
.191
2448
2448
2448
2448
2448
2448
2448
2448
2448
2448
2448
2448
2448
2448
2448
2448
2448
2448
2448
2448
2448
2448
174
Sig
p<.01
p<.01
p<.01
p<.01
p<.05
p<.01
p<.05
p<.01
p<.01
p<.01
p<.01
p<.01
p<.01
p<.01
p<.01
p<.01
NA
NA
p<.01
p<.01
p<.01
p<.01
CI 95%
Lower
Upper
.05
.07
.08
.11
.01
.04
.01
.03
.04
.06
.02
.07
.06
.08
.04
.11
-.08
-.04
.06
.10
.10
.15
.13
.15
.16
.19
.09
.12
.09
.11
.12
.14
.10
.15
.14
.16
.11
.19
0
.04
.14
.18
.18
.23
Table 3.67 Bivariate Correlations Domain Totals and Recidivism Time 1 and Time 2 for Females
Variables
Criminal History T1
Criminal History T2
Education/Employment T1
Education/Employment T2
Financial T1
Financial T2
Family/Marital T1
Family/Marital T2
Accommodation T1
Accommodation T2
Leisure/Recreation T1
Leisure/Recreation T2
Companions T1
Companions T2
Alcohol/Drug Problem T1
Alcohol/Drug Problem T2
Emotional/Personal T1
Emotional/Personal T2
Attitudes/Orientation T1
Attitudes/Orientation T2
Total LSI-R Score T1
Total LSI-R Score T2
Pearson
Correlation
N
.160
.194
.050
.113
.043
.054
.056
.123
.105
.148
.013
.057
-.010
.023
.055
.168
.028
.008
.045
.173
.112
.203
401
401
401
401
401
401
401
401
401
401
401
401
401
401
401
401
401
401
401
401
401
401
175
Sig
p<.01
p<.01
NA
p<.05
NA
NA
NA
p<.05
NA
p<.01
NA
NA
NA
NA
NA
p<.01
NA
NA
NA
p<.01
p<.05
p<.01
CI 95%
Lower
Upper
.06
.10
-.05
.02
-.06
-.04
-.04
.03
.01
.05
-.09
-.04
-.11
-.08
-.04
.07
-.07
-.09
-.05
.08
.01
.11
.26
.29
.15
.21
.14
.15
.15
.22
.20
.25
.11
.16
.09
.12
.15
.27
.13
.11
.14
.27
.21
.29
Finally, it is possible to compare confidence intervals across gender at time 1 and time 2.
Comparing the ten domains for males at time 1 to the ten domains for females at time 1 suggests
that none of the domains emerges as a significantly better predictor for one gender over the
other. The same is true when comparing the domains for males and females at time 2. Overlap in
the confidence intervals for each respective domain regardless of gender suggests that there is no
significant difference in the ability of the domains to predict recidivism for males or females.
Multivariate Analysis Time 1 and Time 2 for Gender
Logistic regression models were estimated at time 1 and time 2 using the above control
variables and a measure of change on the LSI-R (both the percent change and raw score change).
The multivariate tables in this section report regression coefficients, standard errors, Wald
statistics, degrees of freedom, significance values, exponent (B) values, and 95% confidence
intervals for exponent (B).
Tables 3.68 and 3.69 present the multivariate domain models at time 1 for males and
females. Time at risk time 1, criminal history time 1, education/employment,
emotional/personal, and attitudes/orientation are as significant predictors in time 1 model for
males. The Exp (B) values suggest that for a one unit change, time at risk (1.000) is the best
predictor in the model. However, the strongest domain predictor for males is criminal history
time 1 (1.063). Criminal history time 1 is the only significant predictor in the time 1 model for
females.
Tables 3.70 and 3.71 present the multivariate domain models at time 2 for males and
females. Time at risk time 2, criminal history time 2, alcohol/drug problem time 2,
education/employment time 2, and attitude/orientation are statistically significant predictors in
176
the multivariate domain model for males at time 2. The Exp (B) values suggest that for a one
unit change, criminal history time 2 (1.094) is the best domain predictor for males at time 2.
Time at risk 2 and criminal history 2 are significant predictors for females at time 2. Consistent
with the findings from the male domain model at time 2, criminal history 2 is the best domain
predictor at time 2 for females.
Change Analysis for Gender
The descriptive statistics for percent change and raw change for gender are presented in
Table 3.72 and 3.75. Table 3.73 and Table 3.74 present the results from the multivariate analysis
of percent change for males and females. Table 3.76 and Table 3.77 show the raw change for
males and females. Time at risk time 2 is the only significant predictor in the percent change
domain model for males and there are no significant predictors in the percent change domain
model for females. Time at risk time 2, risk category at time 1, raw change criminal history, and
raw change leisure/recreation are significant predictors in the raw change model for males. The
Exp (B) values suggest that for a one unit change, risk category at time 1 (1.558) is the strongest
of the statistically significant predictors for males with raw change in leisure/recreation (.911)
the best domain predictor for males in the raw change model. Time at risk time 2, risk category
time 1, and raw change criminal history are statistically significant predictors in the raw change
domain model for females.
177
Table 3.68 Multivariate Domains Time 1 for Males
Race
Age
Supervisory Status
Time at Risk T1
Criminal History T1
Education/Employment T1
Financial T1
Family/Marital T1
Accommodation T1
Leisure/Recreation T1
Companion T1
Alcohol/Drug Problem T1
Emotional/Personal T1
Attitudes/Orientation T1
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
Df
Sig.
Exp(B)
.053
-.007
-.012
.000
.061
.054
.045
-.024
.071
.000
.088
.038
-.091
.090
-1.846
.122
.006
.091
.000
.020
.018
.069
.038
.046
.067
.042
.017
.031
.035
.361
.190
1.501
.016
13.446
9.502
9.622
.419
.394
2.415
.000
4.541
4.813
8.608
6.753
26.101
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
.663
.221
.899
.000
.002
.002
.518
.530
.120
.994
.033
.028
.003
.009
.000
1.054
.993
.988
1.000
1.063
1.056
1.046
.976
1.074
1.000
1.092
1.039
.913
1.094
.158
94.371 (14)
3178.091
.038
.051
178
95.0% C.I.for EXP(B)
Lower
Upper
.831
.981
.826
1.000
1.023
1.020
.913
.906
.982
.877
1.007
1.004
.859
1.022
1.339
1.004
1.182
1.001
1.105
1.093
1.198
1.052
1.174
1.139
1.185
1.074
.970
1.172
Table 3.69 Multivariate Domains Time 1 for Females
Race
Age
Supervisory Status
Time at Risk T1
Criminal History T1
Education/Employment T1
Financial T1
Family/Marital T1
Accommodation T1
Leisure/Recreation T1
Companion T1
Alcohol/Drug Problem T1
Emotional/Personal T1
Attitudes/Orientation T1
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
Df
Sig.
Exp(B)
.434
-.018
.337
.001
.158
-.003
.052
.015
.182
-.101
-.144
.048
-.007
.014
-1.494
.267
.016
.249
.000
.050
.048
.186
.096
.117
.166
.100
.044
.077
.083
.927
2.639
1.200
1.823
3.735
10.157
.005
.077
.024
2.433
.368
2.063
1.181
.009
.026
2.597
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
.104
.273
.177
.053
.001
.945
.782
.876
.119
.544
.151
.277
.926
.871
.107
1.543
.982
1.400
1.001
1.171
.997
1.053
1.015
1.200
.904
.866
1.049
.993
1.014
.224
26.432 (14)
516.108
.064
.086
p<.05
179
95.0% C.I.for EXP(B)
Lower
Upper
.914
.952
.859
1.000
1.063
.908
.731
.842
.954
.653
.711
.962
.853
.861
2.605
1.014
2.283
1.001
1.291
1.094
1.516
1.224
1.509
1.252
1.054
1.143
1.155
1.194
Table 3.70Multivariate Domains Time 2 for Males
Race
Age
Supervisory Status
Time at Risk T2
Criminal History T2
Education/Employment T2
Financial T2
Family/Marital T2
Accommodation T2
Leisure/Recreation T2
Companion T2
Alcohol/Drug Problem T2
Emotional/Personal T2
Attitudes/Orientation T2
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
Df
Sig.
Exp(B)
.106
-.008
.012
.001
.090
.065
.101
-.037
.024
.123
.081
.085
-.081
.101
-3.535
.135
.007
.102
.000
.023
.021
.080
.043
.049
.079
.049
.020
.035
.038
.383
.613
1.620
.014
65.736
15.288
9.667
1.579
.764
.244
2.435
2.699
17.556
5.313
7.238
85.321
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
.434
.203
.906
.000
.000
.002
.209
.382
.622
.119
.100
.000
.021
.007
.000
1.112
.992
1.012
1.001
1.094
1.067
1.106
.963
1.025
1.131
1.085
1.089
.922
1.106
.029
189.475
2692.392
.075
.108
180
95.0% C.I.for EXP(B)
Lower
Upper
.853
.979
.829
1.001
1.046
1.024
.945
.886
.930
.969
.984
1.046
.861
1.028
1.450
1.005
1.235
1.001
1.145
1.111
1.294
1.048
1.128
1.320
1.195
1.133
.988
1.191
Table 3.71 Multivariate Domains Time 2 for Females
Race
Age
Supervisory Status
Time at Risk T2
Criminal History T2
Education/Employment T2
Financial T2
Family/Marital T2
Accommodation T2
Leisure/Recreation T2
Companion T2
Alcohol/Drug Problem T2
Emotional/Personal T2
Attitudes/Orientation T2
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
Df
Sig.
Exp(B)
.194
-.021
.250
.001
.178
-.019
-.059
.112
.218
-.098
-.263
.105
-.067
.218
-2.623
.290
.018
.276
.000
.058
.052
.206
.105
.132
.194
.123
.050
.087
.093
.966
.448
1.374
.824
10.656
9.479
.132
.083
1.136
2.712
.252
4.608
4.311
.585
5.441
7.368
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
.503
.241
.364
.001
.002
.716
.774
.287
.100
.616
.032
.038
.444
.020
.007
1.215
.980
1.284
1.001
1.195
.981
.943
1.118
1.244
.907
.769
1.110
.936
1.244
.073
48.023 (14)
447.998
.113
.159
181
95.0% C.I.for EXP(B)
Lower
Upper
.687
.946
.748
1.000
1.067
.886
.630
.910
.959
.620
.604
1.006
.789
1.035
2.146
1.014
2.205
1.002
1.339
1.087
1.410
1.374
1.612
1.327
.977
1.226
1.109
1.493
Table 3.72 Descriptives Percent Change Domains for Gender
N
Males Percent Change Criminal
History
Females Percent Change Criminal
History
Males Percent Change
Education/Employment
Females Percent Change
Education/Employment
Males Percent Change Financial
Females Percent Change Financial
Males Percent Change
Family/Marital
Females Percent Change
Family/Marital
Males Percent Change
Accommodation
Females Percent Change
Accommodation
Males Percent Change
Alcohol/Drug Problem
Females Percent Change
Alcohol/Drug Problem
Males Percent Change
Emotional/Personal
Females Percent Change
Emotional/Personal
Males Percent Change
Leisure/Recreation
Females Percent Change
Leisure/Recreation
Males Percent Change Companion
Females Percent Change
Companion
Males Percent Change
Attitudes/Orientation
Females Percent Change
Attitudes/Orientation
Range
Minimum
Maximum
Mean
Std.
Deviation
2357 600.00
-500.00
100.00
-9.7001
34.83162
388 483.33
-400.00
83.33
-11.8793
35.73819
2337 700.00
-600.00
100.00
-18.0287
93.89079
378 600.00
-500.00
100.00
-20.9344
91.06384
1997 200.00
338 200.00
-100.00
-100.00
100.00
100.00
5.9339
4.8817
42.06698
39.63653
2131 400.00
-300.00
100.00
-6.2999
47.58004
353 400.00
-300.00
100.00
-5.9254
45.74830
1559 300.00
-200.00
100.00
6.0936
69.20654
253 300.00
-200.00
100.00
7.7734
65.97662
2219 900.00
-800.00
100.00
-14.8466
86.53267
360 800.00
-700.00
100.00
-19.3773
95.96391
2078 500.00
-400.00
100.00
-12.6307
62.52126
327 400.00
-300.00
100.00
-9.5158
60.78428
2141 200.00
-100.00
100.00
2.4054
41.66574
347 200.00
-100.00
100.00
.7205
39.22368
2332 400.00
-300.00
100.00
-3.9151
35.97428
371 400.00
-300.00
100.00
-4.0431
38.54687
1780 400.00
-300.00
100.00
-7.2378
77.73173
275 400.00
-300.00
100.00
-7.4242
70.92671
182
Table 3.73 Multivariate Domains Percent Change for Males
Race
Age
Supervisory Status
Time at Risk T2
Risk Category T1
Percent Change Criminal History
Percent Change Education/Employment
Percent Change Financial
Percent Change Family/Marital
Percent Change Accommodation
Percent Change Alcohol/Drug Problem
Percent Change Emotional/Personal
Percent Change Leisure/Recreation
Percent Change Companion
Percent Change Attitudes/Orientation
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
Df
Sig.
Exp(B)
-.333
.001
.006
.001
.238
-.005
.001
-.003
.001
-.002
-.002
-.001
.001
.000
-.003
-2.653
.242
.011
.174
.000
.116
.003
.002
.002
.002
.001
.001
.001
.002
.003
.001
.625
1.891
.015
.001
26.430
4.203
3.181
.915
1.841
.245
2.507
5.478
.499
.101
.006
3.881
18.017
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
.169
.903
.971
.000
.040
.074
.339
.175
.621
.113
.019
.480
.750
.938
.049
.000
.717
1.001
1.006
1.001
1.269
.995
1.001
.997
1.001
.998
.998
.999
1.001
1.000
.997
.070
54.143 (15)
933.514
.067
.093
183
95.0% C.I.for EXP(B)
Lower
Upper
.446
.980
.715
1.001
1.011
.989
.998
.993
.997
.995
.996
.996
.996
.995
.995
1.152
1.023
1.416
1.002
1.594
1.001
1.005
1.001
1.004
1.000
1.000
1.002
1.005
1.005
1.000
Table 3.74 Multivariate Domains Percent Change for Females
Race
Age
Supervisory Status
Time at Risk T2
Risk Category T1
Percent Change Criminal History
Percent Change Education/Employment
Percent Change Financial
Percent Change Family/Marital
Percent Change Accommodation
Percent Change Alcohol/Drug Problem
Percent Change Emotional/Personal
Percent Change Leisure/Recreation
Percent Change Companion
Percent Change Attitudes/Orientation
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
Df
Sig.
Exp(B)
.758
-.070
.471
.002
.215
-.008
-.003
-.002
-.015
-.002
.001
-.006
-.002
-.001
.000
-.155
.583
.034
.489
.001
.342
.008
.003
.006
.008
.003
.003
.005
.007
.010
.003
1.903
1.692
4.259
.927
4.800
.396
.959
1.634
.176
3.572
.233
.058
1.680
.106
.021
.006
.007
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
.193
.039
.336
.028
.529
.327
.201
.675
.059
.629
.810
.195
.745
.884
.940
.935
2.134
.932
1.601
1.002
1.240
.992
.997
.998
.985
.998
1.001
.994
.998
.999
1.000
.856
27.856 (15)
145.062
.196
.264
p<.05
184
95.0% C.I.for EXP(B)
Lower
Upper
.681
.872
.614
1.000
.634
.976
.991
.987
.970
.992
.996
.984
.984
.979
.993
6.690
.996
4.177
1.003
2.425
1.008
1.002
1.009
1.001
1.005
1.006
1.003
1.012
1.018
1.006
Table 3.75 Descriptives Raw Change Domains for Gender
N
Males Raw Change Criminal History
Females Raw Change Criminal History
Males Raw Change Education/Employment
Females Raw Change Education/Employment
Males Raw Change Financial
Females Raw Change Financial
Males Raw Change Family/Marital
Females Raw Change Family/Marital
Males Raw Change Accommodation
Females Raw Change Accommodation
Males Raw Change Alcohol/Drug Problem
Females Raw Change Alcohol/Drug Problem
Males Raw Change Emotional/Personal
Females Raw Change Emotional/Personal
Males Raw Change Leisure/Recreation
Females Raw Change Leisure/Recreation
Males Raw Change Companion
Females Raw Change Companion
Males Raw Change Attitudes/Orientation
Females Raw Change Attitudes/Orientation
2448
401
2448
401
2448
401
2448
401
2448
401
2448
401
2448
401
2448
401
2448
401
2448
401
Range Minimum Maximum Mean Std. Deviation
11
11
15
12
4
4
8
6
6
6
15
16
9
7
4
4
8
7
8
8
-6
-6
-8
-6
-2
-2
-4
-3
-3
-3
-8
-9
-5
-4
-2
-2
-4
-4
-4
-4
185
5
5
7
6
2
2
4
3
3
3
7
7
4
3
2
2
4
3
4
4
-.30
-.39
.08
-.14
.01
.02
-.08
-.05
-.04
-.03
.02
-.02
-.15
-.08
.00
-.02
-.02
-.05
-.17
-.18
.814
.984
2.467
2.518
.573
.512
.792
.733
1.016
1.051
2.091
2.358
1.011
.966
.696
.648
.752
.827
1.323
1.309
Table 3.76 Multivariate Domains Raw Change for Males
B
Race
Age
Supervisory Status
Time at Risk T2
Risk Category T1
Raw Change Criminal History
Raw Change Education/Employment
Raw Change Financial
Raw Change Family/Marital
Raw Change Accommodation
Raw Change Alcohol/Drug Problem
Raw Change Emotional/Personal
Raw Change Leisure/Recreation
Raw Change Companion
Raw Change Attitudes/Orientation
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
.040
-.007
.028
.001
.462
-.254
-.021
-.098
-.058
.032
-.159
-.031
-.094
-.033
-.103
-2.782
S.E.
Wald
Df
Sig.
Exp(B)
.134
.007
.101
.000
.056
.056
.022
.089
.062
.050
.075
.065
.025
.048
.041
.343
.088
1.143
.078
57.560
68.362
20.403
.959
1.207
.882
.401
4.452
.221
14.099
.459
6.409
65.940
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
.767
.285
.780
.000
.000
.000
.328
.272
.348
.527
.035
.638
.000
.498
.011
.000
1.041
.993
1.029
1.001
1.588
.775
.979
.907
.944
1.032
.853
.970
.911
.968
.902
.062
176.847 (15)
2705.020
.070
.101
186
95.0% C.I.for EXP(B)
Lower
Upper
.800
.980
.843
1.001
1.423
.694
.938
.761
.836
.936
.736
.853
.867
.880
.833
1.354
1.006
1.255
1.001
1.772
.866
1.022
1.080
1.065
1.138
.989
1.102
.956
1.064
.977
Table 3.77 Multivariate Domains Raw Change for Females
B
Race
Age
Supervisory Status
Time at Risk T2
Risk Category T1
Raw Change Criminal History
Raw Change Education/Employment
Raw Change Financial
Raw Change Family/Marital
Raw Change Accommodation
Raw Change Alcohol/Drug Problem
Raw Change Emotional/Personal
Raw Change Leisure/Recreation
Raw Change Companion
Raw Change Attitudes/Orientation
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
.194
-.010
.214
.001
.447
-.365
-.040
-.132
-.300
.017
-.234
.068
.008
-.085
-.030
-2.491
S.E.
Wald
Df
Sig.
Exp(B)
.287
.017
.270
.000
.131
.122
.051
.236
.169
.117
.191
.147
.054
.122
.104
.887
.459
.352
.628
6.610
11.656
8.957
.612
.313
3.165
.021
1.508
.216
.023
.482
.085
7.888
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
.498
.553
.428
.010
.001
.003
.434
.576
.075
.885
.219
.642
.880
.488
.771
.005
1.215
.990
1.238
1.001
1.563
.694
.961
.876
.741
1.017
.791
1.071
1.008
.919
.970
.083
34.803 (15)
461.219
.083
.119
p<.01
187
95.0% C.I.for EXP(B)
Lower
Upper
.692
.958
.730
1.000
1.210
.547
.870
.551
.532
.809
.545
.803
.907
.723
.791
2.132
1.023
2.100
1.002
2.020
.882
1.062
1.392
1.031
1.278
1.150
1.427
1.120
1.167
1.189
Bivariate Analysis Time 1 and Time 2 for Race
The predictive validity of the LSI-R domains at time 1 are tested by calculating the
correlation between the total score from each domain at time 1 and recidivism time 1. Table
3.78 and Table 3.79 present the bivariate correlations for each domain of the LSI-R at time 1 and
time 2 for black and white offenders. The only significant domain predictor at time 1 for blacks
is criminal history. Nine of ten domain predictors are significant for white offenders at time 1.
The only domain that fails to predict for white offenders at time 1 is emotional/personal.
The predictive validity of the LSI-R domains at time 2 are tested by calculating the
correlation between the total score from each domain at time 2 and recidivism time 2. Seven
domains including criminal history, education/employment, financial, family/marital,
accommodation, alcohol/drug problem, and attitudes/orientation are significant for black
offenders at time 2. For white offenders, all ten domains are significant predictors of recidivism
at time 2.
Confidence intervals are calculated around the Pearson correlation scores to determine
whether or not the correlations at time 1 are significantly different from the correlations at time
2. Confidence intervals from each of the ten domains at time 1 overlap the confidence intervals
for each domain at time 2. For example, the confidence intervals for criminal history of blacks
are .07 to .26. The confidence intervals for black criminal history at time 2 are .10 to .29. The
overlap in the time 1 and time 2 ranges indicates that there is no significant difference between
the time 1 and time 2 criminal history correlations for blacks. The finding of no significant
difference between time 1 and time 2 correlations is consistent across all ten domains of the LSIR for the offenders regardless of race.
188
Table 3.78 Bivariate Correlations Domain Totals and Recidivism Time 1 and Time 2 for Blacks
Variables
Criminal History T1
Criminal History T2
Education/Employment T1
Education/Employment T2
Financial T1
Financial T2
Family/Marital T1
Family/Marital T2
Accommodation T1
Accommodation T2
Leisure/Recreation T1
Leisure/Recreation T2
Companions T1
Companions T2
Alcohol/Drug Problem T1
Alcohol/Drug Problem T2
Emotional/Personal T1
Emotional/Personal T2
Attitudes/Orientation T1
Attitudes/Orientation T2
Total LSI-R Score T1
Total LSI-R Score T2
Pearson
Correlation
N
.163
.193
.091
.207
.069
.106
.079
.102
.077
.120
.011
.057
.073
.065
.033
.166
-.056
.003
.085
.132
.128
.232
433
433
433
433
433
433
433
433
433
433
433
433
433
433
433
433
433
433
433
433
433
433
189
Sig
p<.01
p<.01
NA
p<.01
NA
p<.05
NA
p<.05
NA
p<.05
NA
NA
NA
NA
NA
p<.01
NA
NA
NA
p<.01
p<.01
p<.01
CI 95%
Lower
Upper
.07
.10
.26
.29
.09
.29
.16
.20
.17
.20
.17
.22
11
.15
.17
.16
.13
.26
.04
.10
.18
.23
.22
.32
0
.12
-.03
.01
-.02
.01
-.02
.03
-.08
-.04
-.02
-.03
-.06
.07
-.15
-.09
-.01
.04
.03
.14
Table 3.79 Bivariate Correlations Domain Totals and Recidivism Time 1 and Time 2 for Whites
Variables
Pearson
Correlation
N
Sig
CI 95%
Lower
Criminal History T1
Criminal History T2
Education/Employment T1
Education/Employment T2
Financial T1
Financial T2
Family/Marital T1
Family/Marital T2
Accommodation T1
Accommodation T2
Leisure/Recreation T1
Leisure/Recreation T2
Companions T1
Companions T2
Alcohol/Drug Problem T1
Alcohol/Drug Problem T2
Emotional/Personal T1
Emotional/Personal T2
Attitudes/Orientation T1
Attitudes/Orientation T2
Total LSI-R Score T1
Total LSI-R Score T2
.087
.106
.110
.134
.044
.074
.043
.072
.085
.101
.059
.109
.087
.114
.079
.147
-0.21
-.004
.097
.143
.139
.186
2416
2416
2416
2416
2416
2416
2416
2416
2416
2416
2416
2416
2416
2416
2416
2416
2416
2416
2416
2416
2416
2416
190
p<.01
p<.01
p<.01
p<.01
p<.05
p<.01
p<.05
p<.01
p<.01
p<.01
p<.01
p<.01
p<.01
p<.01
p<.01
p<.01
NA
p<.01
p<.01
p<.01
p<.01
p<.01
.05
.07
.07
.09
0
.03
0
.03
.05
.06
.02
.07
.05
.07
.04
.11
NA
-.04
.06
.10
.10
.15
Upper
.13
.15
.115
.17
.08
.11
.08
.11
.13
.14
.10
.15
.13
.15
.12
.19
.17
.04
.14
.18
.18
.23
The confidence intervals for all ten domains are also compared to one another to
determine if one domain is significantly stronger than another. Upon review of the domain
confidence intervals for blacks at time 1 and time 2, no domain is a significantly better predictor
than the others at time 1 or time 2. Review of the confidence intervals at time 1 and time 2 for
whites suggests that no single domain is a better predictor than the others at time 1 or time 2.
Finally, it is possible to compare confidence intervals across race at time 1 and time 2.
Comparing the ten domains for blacks at time 1 to the ten domains for whites at time 1 suggests
that none of the domains emerges as a significantly better predictor for one race over the other.
The same is true when comparing the domains for blacks and whites at time 2. Overlap in the
confidence intervals for each respective domain regardless of race suggests that there is no
significant difference in the ability of the domains to predict recidivism for blacks or whites.
Multivariate Analysis Time 1 and Time 2 for Race
Logistic regression models were estimated at time 1 and time 2 using the above control
variables and a measure of change on the LSI-R (both the percent change and raw score change).
The multivariate tables in this section report regression coefficients, standard errors, Wald
statistics, degrees of freedom, significance values, exponent (B) values, and 95% confidence
intervals for exponent (B).
Tables 3.80 and 3.81 present the multivariate domain models at time 1 for blacks and
whites. Criminal history time1 is the only significant domain predictor in the time 1 model for
blacks. Time at risk time 1, criminal history time 1, and education/employment are statistically
significant predictors for whites at time 1.
191
Tables 3.82 and 3.83 present the multivariate domain models at time 2 for blacks and
whites. Time at risk time 2 and criminal history time 2 are statistically significant predictors in
the multivariate domain model for blacks at time 2. The Exp (B) values suggest that for a one
unit change, criminal history time 2 (1.243) is the best domain predictor for blacks at time 2.
Time at risk 2, criminal history 2, alcohol/drug problem time 2, and attitudes/orientation time 2
are significant predictors for whites at time 2. Consistent with the findings from the domain
model at time 2 for blacks, criminal history 2 (1.092) is the best domain predictor at time 2 for
whites.
Change Analysis for Race
The descriptive statistics for percent change and raw change for race are presented in
Table.3.84 and 3.87. Table 3.85 and Table 3.86 present the results from the multivariate analysis
of percent change for black and white offenders. Table 3.88 and Table 3.89 show the raw
change for black and white offenders. Gender is the only significant predictor in the percent
change domain model for blacks. Risk category time 1 and percent change attitudes/orientation
are statistically significant predictors in the percent change domain model for whites.
Time at risk time 2 and risk category at time 1 are significant predictors in the raw
change model for blacks. Time at risk 2, risk category time 1, raw change criminal history, raw
change leisure/recreation, and raw change attitudes/orientation are statistically significant
predictors in the raw change domain model for whites. The Exp (B) values suggest that for a one
unit change, risk category at time 1 (1.568) is the strongest of the statistically significant
predictors for whites with raw change in criminal history (.744) the best domain predictor for
whites in the raw change model.
192
Table 3.80 Multivariate Domains Time 1 for Blacks
Age
Gender
Supervisory Status
Time at Risk T1
Criminal History T1
Education/Employment T1
Financial T1
Family/Marital T1
Accommodation T1
Leisure/Recreation T1
Companion T1
Alcohol/Drug Problem T1
Emotional/Personal T1
Attitudes/Orientation T1
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
Df
Sig.
Exp(B)
-.020
.376
-.237
.001
.181
.025
.138
.055
.071
-.190
-.017
.006
-.163
.076
-1.257
.015
.263
.216
.000
.052
.043
.177
.098
.116
.169
.101
.042
.075
.089
.885
1.638
2.037
1.200
4.013
12.323
.338
.612
.314
.382
1.257
.028
.021
4.663
.739
2.017
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
.201
.153
.273
.045
.000
.561
.434
.575
.537
.262
.868
.886
.031
.390
.156
.980
1.456
.789
1.001
1.199
1.026
1.148
1.057
1.074
.827
.983
1.006
.850
1.079
.285
31.391
551.828
.070
.095
p<.01
193
95.0% C.I.for EXP(B)
Lower
Upper
.951
.869
.517
1.000
1.083
.942
.812
.871
.856
.594
.807
.926
.733
.907
1.011
2.438
1.205
1.001
1.327
1.116
1.624
1.282
1.348
1.152
1.198
1.093
.985
1.285
Table 3.81 Multivariate Domains Time 1 for Whites
Age
Gender
Supervisory Status
Time at Risk T1
Criminal History T1
Education/Employment T1
Financial T1
Family/Marital T1
Accommodation T1
Leisure/Recreation T1
Companion T1
Alcohol/Drug Problem T1
Emotional/Personal T1
Attitudes/Orientation T1
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
Df
Sig.
Exp(B)
-.007
.047
.090
.000
.062
.049
.031
-.027
.087
.012
.060
.043
-.065
.081
-1.869
.006
.126
.093
.000
.020
.018
.070
.038
.046
.067
.041
.017
.031
.034
.364
1.476
.141
.934
13.773
9.759
7.687
.192
.524
3.665
.033
2.096
6.319
4.316
5.491
26.339
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
.224
.708
.334
.000
.002
.006
.661
.469
.056
.856
.148
.012
.038
.019
.000
.993
1.048
1.094
1.000
1.063
1.050
1.031
.973
1.091
1.012
1.062
1.044
.937
1.084
.154
84.751 (14)
3147.217
.034
.047
194
95.0% C.I.for EXP(B)
Lower
Upper
.981
.819
.912
1.000
1.023
1.015
.899
.904
.998
.888
.979
1.010
.882
1.013
1.005
1.342
1.314
1.001
1.105
1.088
1.183
1.048
1.194
1.153
1.151
1.080
.996
1.160
Table 3.82 Multivariate Domains Time 2 for Blacks
Age
Gender
Supervisory Status
Time at Risk T2
Criminal History T2
Education/Employment T2
Financial T2
Family/Marital T2
Accommodation T2
Leisure/Recreation T2
Companion T2
Alcohol/Drug Problem T2
Emotional/Personal T2
Attitudes/Orientation T2
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
Df
Sig.
Exp(B)
-.019
.167
-.424
.001
.218
.111
.112
-.011
.079
-.112
-.175
.112
-.099
.051
-3.127
.017
.288
.248
.000
.059
.051
.197
.108
.125
.191
.119
.049
.086
.093
.953
1.255
.335
2.910
15.027
13.479
4.792
.323
.010
.400
.347
2.167
5.210
1.328
.304
10.776
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
.263
.563
.088
.000
.000
.029
.570
.922
.527
.556
.141
.022
.249
.581
.001
.981
1.181
.655
1.001
1.243
1.118
1.118
.989
1.082
.894
.840
1.119
.906
1.052
.044
58.457 (14)
464.849
.127
.180
195
95.0% C.I.for EXP(B)
Lower
Upper
.948
.672
.402
1.001
1.107
1.012
.760
.801
.847
.615
.666
1.016
.766
.878
1.015
2.076
1.065
1.002
1.396
1.235
1.645
1.222
1.383
1.299
1.060
1.231
1.072
1.262
Table 3.83 Multivariate Domains Time 2 for Whites
Age
Gender
Supervisory Status
Time at Risk T2
Criminal History T2
Education/Employment T2
Financial T2
Family/Marital T2
Accommodation T2
Leisure/Recreation T2
Companion T2
Alcohol/Drug Problem T2
Emotional/Personal T2
Attitudes/Orientation T2
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
Df
Sig.
Exp(B)
-.009
.173
.101
.001
.088
.044
.084
-.024
.036
.125
.064
.084
-.076
.128
-3.453
.007
.138
.103
.000
.023
.021
.080
.043
.050
.079
.050
.020
.035
.038
.383
1.606
1.577
.958
61.325
14.692
4.406
1.088
.311
.524
2.508
1.674
17.223
4.616
11.554
81.271
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
.205
.209
.328
.000
.000
.036
.297
.577
.469
.113
.196
.000
.032
.001
.000
.992
1.189
1.106
1.001
1.092
1.045
1.087
.977
1.037
1.133
1.066
1.088
.927
1.136
.032
178.120 (14)
2677.874
.071
.103
196
95.0% C.I.for EXP(B)
Lower
Upper
.979
.907
.904
1.001
1.044
1.003
.929
.898
.941
.971
.967
1.046
.865
1.056
1.005
1.558
1.354
1.001
1.142
1.089
1.272
1.061
1.142
1.322
1.175
1.132
.993
1.223
Table 3.84 Descriptives Percent Change Domains for Race
Black Percent Change Criminal History
White Percent Change Criminal History
Black Percent Change Education/Employment
White Percent Change Education/Employment
Black Percent Change Financial
White Percent Change Financial
Black Percent Change Family/Marital
White Percent Change Family/Marital
Black Percent Change Accommodation
White Percent Change Accommodation
Black Percent Change Alcohol/Drug Problem
White Percent Change Alcohol/Drug Problem
Black Percent Change Emotional/Personal
White Percent Change Emotional/Personal
Black Percent Change Leisure/Recreation
White Percent Change Leisure/Recreation
Black Percent Change Companion
White Percent Change Companion
Black Percent Change Attitudes/Orientation
White Percent Change Attitudes/Orientation
N
Range
415
2330
414
2301
344
1991
382
2102
261
1551
384
2195
357
2048
387
2101
406
2297
301
1754
350.00
600.00
700.00
700.00
200.00
200.00
400.00
400.00
300.00
300.00
885.71
800.00
500.00
500.00
200.00
200.00
400.00
400.00
400.00
400.00
Minimum Maximum
-250.00
-500.00
-600.00
-600.00
-100.00
-100.00
-300.00
-300.00
-200.00
-200.00
-800.00
-700.00
-400.00
-400.00
-100.00
-100.00
-300.00
-300.00
-300.00
-300.00
197
100.00
100.00
100.00
100.00
100.00
100.00
100.00
100.00
100.00
100.00
85.71
100.00
100.00
100.00
100.00
100.00
100.00
100.00
100.00
100.00
Mean
Std. Deviation
-11.9067
-9.6699
-23.3937
-17.5408
6.3953
5.6755
-6.0864
-6.2758
7.4074
6.1466
-18.6077
-14.9317
-8.6368
-12.8296
.6460
2.4512
-2.4384
-4.1968
-5.8693
-7.5019
34.29068
35.07766
104.13687
91.44073
39.71217
42.06281
40.63515
48.43895
74.95329
67.67391
100.33190
85.55866
59.16892
62.80505
41.10956
41.37446
30.75713
37.22891
77.86605
76.68207
Table 3.85 Multivariate Domains Percent Change for Blacks
Age
Gender
Supervisory Status
Time at Risk T2
Risk Category T1
Percent Change Criminal History
Percent Change Education/Employment
Percent Change Financial
Percent Change Family/Marital
Percent Change Accommodation
Percent Change Alcohol/Drug Problem
Percent Change Emotional/Personal
Percent Change Leisure/Recreation
Percent Change Companion
Percent Change Attitudes/Orientation
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
Df
Sig.
Exp(B)
-.006
1.979
-.600
.002
.518
.010
.000
-.009
.006
-.003
-.004
-.006
.004
.007
.007
-3.869
.032
.626
.516
.001
.342
.009
.005
.006
.007
.004
.003
.006
.007
.012
.004
1.821
.032
10.009
1.353
6.689
2.290
1.060
.000
2.076
.659
.633
2.212
.898
.335
.272
3.227
4.515
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
.859
.002
.245
.010
.130
.303
.995
.150
.417
.426
.137
.343
.563
.602
.072
.034
.994
7.236
.549
1.002
1.679
1.010
1.000
.991
1.006
.997
.996
.994
1.004
1.007
1.007
.021
28.312 (15)
133.756
.196
.275
p<.05
198
95.0% C.I.for EXP(B)
Lower
Upper
.935
2.123
.200
1.000
.858
.991
.991
.980
.992
.990
.991
.983
.991
.982
.999
1.058
24.660
1.508
1.003
3.284
1.029
1.010
1.003
1.019
1.004
1.001
1.006
1.017
1.031
1.014
Table 3.86 Multivariate Domains Percent Change for Whites
Age
Gender
Supervisory Status
Time at Risk T2
Risk Category T1
Percent Change Criminal History
Percent Change Education/Employment
Percent Change Financial
Percent Change Family/Marital
Percent Change Accommodation
Percent Change Alcohol/Drug Problem
Percent Change Emotional/Personal
Percent Change Leisure/Recreation
Percent Change Companion
Percent Change Attitudes/Orientation
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
Df
Sig.
Exp(B)
-.012
.290
.088
.001
.189
-.008
.000
-.002
.000
-.002
-.002
-.001
.000
.000
-.004
-2.058
.011
.226
.174
.000
.116
.003
.001
.002
.002
.001
.001
.001
.002
.003
.001
.624
1.226
1.651
.257
25.951
2.686
5.933
.121
1.114
.021
1.688
2.758
.861
.011
.018
7.530
10.865
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
.268
.199
.612
.000
.101
.015
.728
.291
.884
.194
.097
.353
.918
.894
.006
.001
.988
1.337
1.092
1.001
1.209
.992
1.000
.998
1.000
.998
.998
.999
1.000
1.000
.996
.128
60.960 (15)
940.750
.075
.104
199
95.0% C.I.for EXP(B)
Lower
Upper
.967
.858
.777
1.001
.964
.986
.998
.994
.996
.996
.997
.996
.995
.995
.994
1.009
2.082
1.535
1.002
1.516
.998
1.003
1.002
1.003
1.001
1.000
1.001
1.004
1.005
.999
Table 3.87 Descriptives Raw Change Domains for Race
N
Black Raw Change Criminal History
White Raw Change Criminal History
Black Raw Change Education/Employment
White Raw Change Education/Employment
Black Raw Change Financial
White Raw Change Financial
Black Raw Change Family/Marital
White Raw Change Family/Marital
Black Raw Change Accommodation
White Raw Change Accommodation
Black Raw Change Alcohol/Drug Problem
White Raw Change Alcohol/Drug Problem
Black Raw Change Emotional/Personal
White Raw Change Emotional/Personal
Black Raw Change Leisure/Recreation
White Raw Change Leisure/Recreation
Black Raw Change Companion
White Raw Change Companion
Black Raw Change Attitudes/Orientation
White Raw Change Attitudes/Orientation
433
2416
433
2416
433
2416
433
2416
433
2416
433
2416
433
2416
433
2416
433
2416
433
2416
Range Minimum Maximum Mean Std. Deviation
9
11
12
15
4
4
8
8
6
6
16
15
7
9
4
4
7
8
8
8
-6
-6
-6
-8
-2
-2
-4
-4
-3
-3
-9
-8
-4
-5
-2
-2
-4
-4
-4
-4
200
3
5
6
7
2
2
4
4
3
3
7
7
3
4
2
2
3
4
4
4
-.39
-.30
-.02
.07
.00
.01
-.09
-.07
-.05
-.04
-.03
.02
-.09
-.15
-.01
.00
-.04
-.02
-.19
-.16
.946
.820
2.552
2.462
.544
.568
.727
.794
1.020
1.021
2.238
2.111
.958
1.013
.638
.698
.729
.770
1.314
1.322
Table 3.88 Multivariate Domains Raw Change for Blacks
Age
Gender
Supervisory Status
Time at Risk T2
Risk Category T1
Raw Change Criminal History
Raw Change Education/Employment
Raw Change Financial
Raw Change Family/Marital
Raw Change Accommodation
Raw Change Alcohol/Drug Problem
Raw Change Emotional/Personal
Raw Change Leisure/Recreation
Raw Change Companion
Raw Change Attitudes/Orientation
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
Df
Sig.
Exp(B)
-.017
.309
-.249
.001
.529
-.131
-.107
.083
-.267
-.033
-.168
.099
-.072
-.047
-.016
-2.517
.017
.285
.243
.000
.132
.118
.052
.221
.162
.118
.195
.164
.055
.123
.098
.863
1.099
1.174
1.056
12.668
16.005
1.234
4.281
.142
2.736
.081
.743
.364
1.680
.148
.026
8.499
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
.294
.279
.304
.000
.000
.267
.039
.707
.098
.776
.389
.547
.195
.700
.871
.004
.983
1.362
.779
1.001
1.698
.877
.898
1.087
.766
.967
.845
1.104
.931
.954
.984
.081
44.459 (15)
478.847
.098
.139
201
95.0% C.I.for EXP(B)
Lower
Upper
.951
.779
.484
1.001
1.310
.696
.812
.704
.558
.768
.576
.801
.835
.750
.812
1.015
2.380
1.254
1.002
2.201
1.106
.994
1.678
1.051
1.218
1.239
1.522
1.037
1.213
1.192
Table 3.89 Multivariate Domains Raw Change for Whites
B
Age
Gender
Supervisory Status
Time at Risk T2
Risk Category T1
Raw Change Criminal History
Raw Change Education/Employment
Raw Change Financial
Raw Change Family/Marital
Raw Change Accommodation
Raw Change Alcohol/Drug Problem
Raw Change Emotional/Personal
Raw Change Leisure/Recreation
Raw Change Companion
Raw Change Attitudes/Orientation
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
-.006
.136
.110
.001
.450
-.296
-.009
-.134
-.055
.049
-.173
-.034
-.081
-.035
-.108
-2.774
S.E.
Wald
Df
Sig.
Exp(B)
.007
.138
.103
.000
.056
.057
.022
.090
.062
.050
.075
.064
.025
.048
.041
.344
.945
.976
1.143
51.642
64.797
27.390
.173
2.235
.764
.962
5.282
.286
10.676
.523
6.858
64.943
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
.331
.323
.285
.000
.000
.000
.677
.135
.382
.327
.022
.593
.001
.470
.009
.000
.994
1.146
1.116
1.001
1.568
.744
.991
.874
.947
1.050
.841
.966
.922
.966
.898
.062
173.090 (15)
2682.905
.069
.100
202
95.0% C.I.for EXP(B)
Lower
Upper
.981
.875
.912
1.001
1.405
.666
.949
.733
.838
.952
.726
.852
.878
.878
.828
1.007
1.501
1.366
1.001
1.749
.831
1.034
1.043
1.070
1.158
.975
1.096
.968
1.062
.973
Bivariate Analysis Time 1 and Time 2 for Supervision Status
The predictive validity of the LSI-R domains at time 1 are tested by calculating the
correlation between the total score from each domain at time 1 and recidivism time 1. Tables
3.90 and 3.91 present the bivariate correlations for each domain of the LSI-R at time 1 and time
2 for probationers and parolees. Six domains including criminal history, education/employment,
family/marital, accommodation, companions, and attitudes/orientation are significant for
probationers at time 1. For parolees, seven domains are significant including criminal history,
education/employment, accommodation, leisure/recreation, companions, alcohol/drug problem,
and attitudes/orientation.
The predictive validity of the LSI-R domains at time 2 are tested by calculating the
correlation between the total score from each domain at time 2 and recidivism time 2 for
offenders on probation and parole. Nine of ten domains are significant at time 2 for
probationers. The only domain that fails to predict recidivism for probationers at time 2 is
emotional/personal. The same is true for domain predictors at time 2 for parolees. Nine of ten
domains are significant. The only domain that fails to predict recidivism at time 2 for parolees is
emotional/personal.
Confidence intervals are calculated around the Pearson correlation scores to determine
whether or not the correlations at time 1 are significantly different from the correlations at time
2. Confidence intervals from each of the ten domains at time 1 overlap the confidence intervals
for each domain at time 2. For example, the confidence intervals at time 1 for probationer
criminal history are .02 to .11. The confidence intervals for probationer criminal history at time
203
Table 3.90 Bivariate Correlations Domain Totals and Recidivism Time 1 and Time 2 for Probation
Variables
Pearson
Correlation
N
Sig
CI 95%
Lower
Criminal History T1
Criminal History T2
Education/Employment T1
Education/Employment T2
Financial T1
Financial T2
Family/Marital T1
Family/Marital T2
Accommodation T1
Accommodation T2
Leisure/Recreation T1
Leisure/Recreation T2
Companions T1
Companions T2
Alcohol/Drug Problem T1
Alcohol/Drug Problem T2
Emotional/Personal T1
Emotional/Personal T2
Attitudes/Orientation T1
Attitudes/Orientation T2
Total LSI-R Score T1
Total LSI-R Score T2
.065
.092
.092
.139
.049
.082
.053
.061
.069
.092
.037
.084
.063
.093
.043
.152
-.015
.021
.103
.145
.112
.183
1976
1976
1976
1976
1976
1976
1976
1976
1976
1976
1976
1976
1976
1976
1976
1976
1976
1976
1976
1976
1976
1976
204
p<.01
p<.01
p<.01
p<.01
NA
p<.01
p<.05
p<.01
p<.01
p<.01
NA
p<.01
p<.05
p<.01
NA
p<.01
NA
NA
p<.01
p<.01
p<.01
p<.01
.02
.05
.05
.10
0
.04
.01
.02
.02
.05
-.01
.04
.02
.05
0
.11
-.06
-.02
.06
.10
.07
.14
Upper
.11
.14
.14
.18
.09
.13
.10
.11
.11
.14
..08
.13
.11
.14
.09
.20
.03
.07
.15
.19
.16
.23
Table 3.91 Bivariate Correlations Domain Totals and Recidivism Time 1 and Time 2 for Parole
Variables
Criminal History T1
Criminal History T2
Education/Employment T1
Education/Employment T2
Financial T1
Financial T2
Family/Marital T1
Family/Marital T2
Accommodation T1
Accommodation T2
Leisure/Recreation T1
Leisure/Recreation T2
Companions T1
Companions T2
Alcohol/Drug Problem T1
Alcohol/Drug Problem T2
Emotional/Personal T1
Emotional/Personal T2
Attitudes/Orientation T1
Attitudes/Orientation T2
Total LSI-R Score T1
Total LSI-R Score T2
Pearson
Correlation
N
.169
.177
.139
.160
.046
.070
.039
.113
.115
.129
.085
.140
.131
.134
.136
.144
-.053
-.056
.078
.133
.192
.214
873
873
873
873
873
873
873
873
873
873
873
873
873
873
873
873
873
873
873
873
873
873
205
Sig
p<.01
p<.01
p<.01
p<.01
NA
p<.05
NA
p<.01
p<.01
p<.01
p<.05
p<.01
p<.05
p<.01
p<.01
p<.01
NA
NA
p<.05
p<.01
p<.01
p<.01
CI 95%
Lower
Upper
.10
.11
.07
.09
-.02
0
-.03
.05
.05
.06
.02
.07
.07
.07
.07
.08
-.12
-.12
.01
.07
.13
.15
.24
.25
.21
.23
.11
.14
.11
.18
.18
.20
.15
.21
.20
.20
.20
.21
.01
.01
.14
.20
.26
.28
2 are .11 to .25. The overlap in the time 1 and time 2 ranges indicates that there is no significant
difference between the time 1 and time 2 criminal history correlations for probationers. The only
domain predictor that is a significantly better predictor at time 2 than time 1 is alcohol/drug
problem for probationers. The remaining correlations are not significantly different from time 1
to time 2 for probationers and parolees.
The confidence intervals for all ten domains are also compared to one another to
determine if one domain is significantly stronger than another. Upon review of the domain
confidence intervals for offenders on probation and parole at time 1 and time 2, the
emotional/personal domain is the only domain with confidence intervals that fail to consistently
overlap the confidence intervals for the other nine domains. This finding suggests that the
emotional/personal domain is a significantly poorer predictor as compared to the other domains
for probation and parole at time 1 and time 2.
Finally, it is possible to compare confidence intervals across supervisory status at time 1
and time 2. Comparing the ten domains for probation at time 1 to the ten domains for parole at
time 1 suggests that none of the domains emerges as a significantly better predictor for one
supervisory status over the other. The same is true when comparing the domains for probation
and parole at time 2. Overlap in the confidence intervals for each respective domain regardless
of supervisory status, there is no significant difference in the ability of the domains to predict
recidivism for offenders on probation or parole.
Multivariate Analysis Time 1 and Time 2 for Supervision Status
Logistic regression models were estimated at time 1 and time 2 using the above control
variables and a measure of change on the LSI-R (both the percent change and raw score change).
206
The multivariate tables in this section report regression coefficients, standard errors, Wald
statistics, degrees of freedom, significance values, exponent (B) values, and 95% confidence
intervals for exponent (B).
Tables 3.92 and 3.93 present the multivariate domain models at time 1 for probationers
and parolees. Time at risk time 1 and attitudes/orientation are as significant predictors in time 1
model for probationers. Criminal history time 1 and alcohol/drug problem time 1 are significant
domain predictors for parolees. The Exp (B) values suggest that for a one unit change, criminal
history time 1 (1.140) is the best predictor in the time 1 parolee model.
Tables 3.94 and 3.95 present the multivariate domain models at time 2 for probationers
and parolees. Time at risk time 2, criminal history time 2, education/employment time 2,
alcohol/drug problem time 2, and attitude/orientation are statistically significant predictors in the
multivariate domain model for probationers at time 2. The Exp (B) values suggest that for a one
unit change, attitudes/orientation time 2 (1.157) is the best domain predictor for probationers at
time 2. Time at risk 2, criminal history 2, and attitudes/orientation time 2 are significant
predictors for parolees at time 2. The Exp (B) values suggest that for a one unit change, criminal
history time 2 (1.176) is the best domain predictor at time 2 for offenders on parole.
207
Table 3.92 Multivariate Domains Time 1 for Probation
Race
Age
Gender
Time at Risk T1
Criminal History T1
Education/Employment T1
Financial T1
Family/Marital T1
Accommodation T1
Leisure/Recreation T1
Companion T1
Alcohol/Drug Problem T1
Emotional/Personal T1
Attitudes/Orientation T1
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
Df
Sig.
Exp(B)
.212
-.009
-.016
.001
.049
.040
.041
.006
.076
-.026
.029
.017
-.050
.119
-1.710
.135
.007
.130
.000
.022
.019
.077
.042
.051
.073
.046
.019
.034
.039
.402
2.478
1.979
.015
16.614
4.762
4.203
.287
.022
2.182
.129
.390
.786
2.095
9.481
18.119
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
.115
.159
.903
.000
.029
.040
.592
.882
.140
.720
.532
.375
.148
.002
.000
1.236
.991
.984
1.001
1.050
1.041
1.042
1.006
1.079
.974
1.029
1.017
.951
1.126
.181
61.170 (14)
2577.376
.031
.041
208
95.0% C.I.for EXP(B)
Lower
Upper
.949
.978
.763
1.000
1.005
1.002
.896
.926
.976
.844
.940
.980
.889
1.044
1.609
1.004
1.270
1.001
1.097
1.081
1.212
1.093
1.193
1.124
1.127
1.056
1.018
1.214
Table 3.93 Multivariate Domains Time 1 for Parole
Race
Age
Gender
Time at Risk T1
Criminal History T1
Education/Employment T1
Financial T1
Family/Marital T1
Accommodation T1
Leisure/Recreation T1
Companion T1
Alcohol/Drug Problem T1
Emotional/Personal T1
Attitudes/Orientation T1
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
Df
Sig.
Exp(B)
-.080
-.005
.508
.000
.131
.059
.067
-.081
.101
.036
.102
.093
-.147
.000
-2.082
.194
.010
.233
.000
.033
.031
.122
.065
.076
.117
.069
.030
.053
.059
.607
.169
.276
4.752
1.846
16.044
3.693
.305
1.591
1.766
.094
2.179
9.806
7.561
.000
11.747
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
.681
.599
.029
.174
.000
.055
.581
.207
.184
.759
.140
.002
.006
.998
.001
.924
.995
1.661
1.000
1.140
1.061
1.070
.922
1.107
1.037
1.107
1.098
.863
1.000
.125
66.394 (14)
1110.129
.073
.099
209
95.0% C.I.for EXP(B)
Lower
Upper
.632
.975
1.053
1.000
1.069
.999
.842
.812
.953
.824
.967
1.036
.778
.891
1.349
1.015
2.623
1.001
1.216
1.127
1.358
1.046
1.285
1.305
1.267
1.164
.959
1.122
Table 3.94 Multivariate Domains Time 2 for Probation
Race
Age
Gender
Time at Risk T2
Criminal History T2
Education/Employment T2
Financial T2
Family/Marital T2
Accommodation T2
Leisure/Recreation T2
Companion T2
Alcohol/Drug Problem T2
Emotional/Personal T2
Attitudes/Orientation T2
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
Df
Sig.
Exp(B)
.266
-.009
.135
.001
.078
.060
.087
-.051
.038
.028
.029
.093
-.030
.145
-3.421
.148
.007
.143
.000
.026
.023
.089
.047
.055
.087
.055
.022
.039
.042
.422
3.246
1.605
.902
59.014
8.993
6.810
.958
1.181
.472
.103
.272
17.580
.588
12.130
65.712
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
.072
.205
.342
.000
.003
.009
.328
.277
.492
.748
.602
.000
.443
.000
.000
1.305
.991
1.145
1.001
1.081
1.061
1.091
.950
1.038
1.028
1.029
1.097
.971
1.157
.033
152.983 (14)
2182.841
.075
.107
210
95.0% C.I.for EXP(B)
Lower
Upper
.977
.976
.866
1.001
1.027
1.015
.916
.866
.933
.867
.924
1.051
.900
1.066
1.744
1.005
1.515
1.001
1.138
1.110
1.299
1.042
1.156
1.220
1.146
1.146
1.047
1.255
Table 3.95 Multivariate Domains Time 2 for Parole
Race
Age
Gender
Time at Risk T2
Criminal History T2
Education/Employment T2
Financial T2
Family/Marital T2
Accommodation T2
Leisure/Recreation T2
Companion T2
Alcohol/Drug Problem T2
Emotional/Personal T2
Attitudes/Orientation T2
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
Df
Sig.
Exp(B)
-.158
-.010
.293
.001
.162
.048
.104
.054
.044
.217
.034
.082
-.194
.056
-3.471
.216
.011
.249
.000
.038
.036
.136
.072
.085
.133
.083
.036
.061
.063
.640
.538
.801
1.384
17.088
18.320
1.790
.590
.568
.271
2.681
.165
5.276
10.039
.789
29.375
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
.463
.371
.239
.000
.000
.181
.443
.451
.603
.102
.685
.022
.002
.374
.000
.854
.990
1.340
1.001
1.176
1.050
1.110
1.056
1.045
1.243
1.034
1.085
.824
1.058
.031
87.002 (14)
956.790
.095
.136
211
95.0% C.I.for EXP(B)
Lower
Upper
.559
.968
.823
1.000
1.092
.978
.851
.917
.885
.958
.879
1.012
.731
.935
1.303
1.012
2.182
1.001
1.267
1.127
1.448
1.216
1.234
1.612
1.216
1.164
.929
1.197
Change Analysis for Supervision Status
The descriptive statistics for percent change and raw change for offenders on probation
and parole are presented in Table 3.96 and 3.99. Table 3.97 and Table 3.98 present the results
from the multivariate analysis of percent change for offenders on probation and paroles. Table
3.100 and Table 3.101 show the raw change for probationers and parolees. Time at risk time 2 is
the only significant predictor in the percent change domain model for probationers and there are
no significant predictors in the percent change domain model for parolees. Time at risk time 2,
risk category at time 1, raw change criminal history, and raw change leisure/recreation are
significant predictors in the raw change model for probationers. The Exp (B) values suggest that
for a one unit change, risk category at time 1 (1.542) is the strongest of the statistically
significant predictors for probationers with raw change in leisure/recreation (.909) the best
domain predictor for probationers in the raw change model. Time at risk time 2, risk category
time 1, and raw change criminal history are statistically significant predictors in the raw change
domain model for parolees. Raw change criminal history is the best domain predictor for
parolees.
This section has outlined the results from the domain analysis for the entire sample and
subgroups at time 1, time 2, percent change, and raw change. The next chapter will review
major findings, discuss policy implications, and provide suggestions for future research.
212
Table 3.96 Percent Change Domains for Supervision Status
Probation Percent Change
Criminal History
Parole Percent Change Criminal
History
Probation Percent Change
Education/Employment
Paroles Percent Change
Education/Employment
Probation Percent Change
Financial
Paroles Percent Change Financial
Probation Percent Change
Family/Marital
Paroles Percent Change
Family/Marital
Probation Percent Change
Accommodation
Paroles Percent Change
Accommodation
Probation Percent Change
Alcohol/Drug Problem
Paroles Percent Change
Alcohol/Drug Problem
Probation Percent Change
Emotional/Personal
Paroles Percent Change
Emotional/Personal
Probation Percent Change
Leisure/Recreation
Paroles Percent Change
Leisure/Recreation
Probation Percent Change
Companion
Paroles Percent Change
Companion
Probation Percent Change
Attitudes/Orientation
Paroles Percent Change
Attitudes/Orientation
N
Range
Minimum
Maximum
Mean
Std.
Deviation
1903
600.00
-500.00
100.00
-9.9528
34.50750
842
600.00
-500.00
100.00
-10.1330
35.99122
1883
700.00
-600.00
100.00
-18.7930
93.74748
832
700.00
-600.00
100.00
-17.6192
92.95979
1621
200.00
-100.00
100.00
5.5213
41.78195
714
200.00
-100.00
100.00
6.3725
41.59352
1739
400.00
-300.00
100.00
-7.3270
48.06238
745
400.00
-300.00
100.00
-3.7248
45.45518
1253
300.00
-200.00
100.00
5.0279
68.92470
559
300.00
-200.00
100.00
9.2427
68.32643
1789
800.00
-700.00
100.00
-15.7815
89.01557
790
900.00
-800.00
100.00
-14.7942
85.38547
1670
500.00
-400.00
100.00
-12.8693
64.15539
735
400.00
-300.00
100.00
-10.7029
57.82349
1732
200.00
-100.00
100.00
.3464
41.55966
756
200.00
-100.00
100.00
6.3492
40.51856
1881
400.00
-300.00
100.00
-4.3718
37.74893
822
400.00
-300.00
100.00
-2.9278
32.85586
1430
400.00
-300.00
100.00
-7.7855
73.81618
625
400.00
-300.00
100.00
-6.0667
83.39592
213
214
Table 3.97 Multivariate Domains Percent Change for Probation
Race
Age
Gender
Time at Risk T2
Risk Category T1
Percent Change Criminal History
Percent Change Education/Employment
Percent Change Financial
Percent Change Family/Marital
Percent Change Accommodation
Percent Change Alcohol/Drug Problem
Percent Change Emotional/Personal
Percent Change Leisure/Recreation
Percent Change Companion
Percent Change Attitudes/Orientation
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
Df
Sig.
Exp(B)
-.081
.000
.292
.002
.200
-.006
.001
-.002
.001
-.002
-.003
-.001
.000
.002
-.004
-2.912
.264
.012
.245
.000
.133
.004
.002
.003
.002
.001
.001
.002
.003
.003
.002
.699
.093
.000
1.415
31.168
2.269
2.617
.234
.456
.122
1.343
5.708
.379
.020
.486
6.503
17.348
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
.760
.996
.234
.000
.132
.106
.628
.499
.727
.246
.017
.538
.888
.486
.011
.000
.923
1.000
1.338
1.002
1.222
.994
1.001
.998
1.001
.998
.997
.999
1.000
1.002
.996
.054
56.244 (15)
742.473
.085
.119
215
95.0% C.I.for EXP(B)
Lower
Upper
.549
.977
.828
1.001
.941
.987
.998
.993
.997
.996
.995
.996
.995
.996
.993
1.549
1.023
2.164
1.002
1.585
1.001
1.004
1.003
1.005
1.001
1.000
1.002
1.005
1.008
.999
Table 3.98 Multivariate Domains Percent Change for Parole
B
Race
Age
Gender
Time at Risk T2
Risk Category T1
Percent Change Criminal History
Percent Change Education/Employment
Percent Change Financial
Percent Change Family/Marital
Percent Change Accommodation
Percent Change Alcohol/Drug Problem
Percent Change Emotional/Personal
Percent Change Leisure/Recreation
Percent Change Companion
Percent Change Attitudes/Orientation
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
-.343
-.033
.845
.001
.248
-.003
-.001
-.004
-.002
-.001
.000
-.004
.001
-.005
.000
-.790
S.E.
Wald
.383
.020
.395
.000
.189
.005
.002
.003
.003
.002
.002
.003
.004
.005
.002
1.073
.806
2.709
4.575
2.985
1.721
.399
.219
1.839
.560
.208
.009
1.855
.019
.960
.001
.542
22.781 (15)
341.883
.078
.107
216
Df
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
Sig.
.369
.100
.032
.084
.190
.527
.640
.175
.454
.649
.924
.173
.891
.327
.972
.462
Exp(B)
.709
.967
2.328
1.001
1.282
.997
.999
.996
.998
.999
1.000
.996
1.001
.995
1.000
.454
95.0% C.I.for EXP(B)
Lower
Upper
.335
.929
1.073
1.000
.884
.988
.995
.989
.992
.995
.997
.990
.992
.986
.996
1.501
1.006
5.050
1.001
1.858
1.006
1.003
1.002
1.003
1.003
1.003
1.002
1.009
1.005
1.004
Table 3.99 Raw Change Domains for Supervision Status
N
Probation Raw Change Criminal History
Parole Raw Change Criminal History
Probation Raw Change Education/Employment
Paroles Raw Change Education/Employment
Probation Raw Change Financial
Paroles Raw Change Financial
Probation Raw Change Family/Marital
Paroles Raw Change Family/Marital
Probation Raw Change Accommodation
Paroles Raw Change Accommodation
Probation Raw Change Alcohol/Drug Problem
Paroles Raw Change Alcohol/Drug Problem
Probation Raw Change Emotional/Personal
Paroles Raw Change Emotional/Personal
Probation Raw Change Leisure/Recreation
Paroles Raw Change Leisure/Recreation
Probation Raw Change Companion
Paroles Raw Change Companion
Probation Raw Change Attitudes/Orientation
Paroles Raw Change Attitudes/Orientation
1976
873
1976
873
1976
873
1976
873
1976
873
1976
873
1976
873
1976
873
1976
873
1976
873
Range Minimum Maximum Mean Std. Deviation
11
8
15
12
4
4
8
8
6
6
16
15
9
8
4
4
8
7
8
8
-6
-5
-8
-6
-2
-2
-4
-4
-3
-3
-9
-8
-5
-4
-2
-2
-4
-4
-4
-4
217
5
3
7
6
2
2
4
4
3
3
7
7
4
4
2
2
4
3
4
4
-.31
-.31
.04
.08
.01
.01
-.08
-.06
-.06
.01
.01
.02
-.15
-.12
-.02
.05
-.02
-.02
-.17
-.17
.856
.804
2.456
2.520
.562
.572
.761
.833
1.019
1.023
2.159
2.064
1.023
.963
.668
.733
.759
.773
1.287
1.395
Table 3.100 Multivariate Domains Raw Change for Probation
Race
Age
Gender
Time at Risk T2
Risk Category T1
Raw Change Criminal History
Raw Change Education/Employment
Raw Change Financial
Raw Change Family/Marital
Raw Change Accommodation
Raw Change Alcohol/Drug Problem
Raw Change Emotional/Personal
Raw Change Leisure/Recreation
Raw Change Companion
Raw Change Attitudes/Orientation
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
Df
Sig.
Exp(B)
.205
-.008
.120
.001
.433
-.224
-.015
-.147
-.043
-.006
-.185
.018
-.095
-.049
-.077
-2.770
.146
.007
.142
.000
.062
.060
.024
.100
.072
.055
.085
.071
.027
.053
.047
.382
1.950
1.050
.709
49.961
48.409
13.828
.391
2.133
.355
.013
4.720
.064
12.599
.881
2.698
52.655
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
.163
.305
.400
.000
.000
.000
.532
.144
.551
.909
.030
.800
.000
.348
.100
.000
1.227
.992
1.127
1.001
1.542
.799
.985
.864
.958
.994
.831
1.018
.909
.952
.926
.063
142.539 (15)
2193.285
.070
.100
218
95.0% C.I.for EXP(B)
Lower
Upper
.921
.978
.853
1.001
1.365
.710
.939
.709
.831
.893
.703
.885
.862
.858
.845
1.635
1.007
1.491
1.001
1.743
.899
1.033
1.051
1.104
1.106
.982
1.171
.958
1.055
1.015
Table 3.101 Multivariate Domains Raw Change for Parole
Race
Age
Gender
Time at Risk T2
Risk Category T1
Raw Change Criminal History
Raw Change Education/Employment
Raw Change Financial
Raw Change Family/Marital
Raw Change Accommodation
Raw Change Alcohol/Drug Problem
Raw Change Emotional/Personal
Raw Change Leisure/Recreation
Raw Change Companion
Raw Change Attitudes/Orientation
Constant
Model Chi-Square (df)
-2 Log Likelihood
Cox and Snell R2
Nagelkerke R2
B
S.E.
Wald
Df
Sig.
Exp(B)
-.207
-.006
.227
.001
.524
-.387
-.048
.019
-.155
.134
-.129
-.122
-.034
-.021
-.132
-2.786
.216
.011
.249
.000
.092
.095
.037
.150
.098
.084
.124
.109
.042
.086
.065
.570
.918
.330
.829
15.692
32.390
16.657
1.711
.016
2.492
2.555
1.075
1.257
.661
.058
4.091
23.855
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
.338
.566
.363
.000
.000
.000
.191
.901
.114
.110
.300
.262
.416
.809
.043
.000
.813
.994
1.254
1.001
1.688
.679
.953
1.019
.857
1.143
.879
.885
.966
.979
.876
.062
76.905 (15)
966.888
.084
.121
219
95.0% C.I.for EXP(B)
Lower
Upper
.533
.972
.770
1.000
1.410
.564
.888
.759
.707
.970
.689
.715
.890
.828
.770
1.241
1.016
2.043
1.001
2.022
.818
1.024
1.368
1.038
1.347
1.122
1.096
1.049
1.159
.996
CHAPTER 4
CONCLUSION: THE FUTURE OF THE LSI-R
The Level of Supervision Inventory Revised (LSI-R) is a third-generation risk and needs
assessment instrument selected for use in a number of correctional settings (Andrews & Bonta,
1995) and with various offender populations (Bonta, 1989; Shields, 1993; Coulson, 1996;
Gendreau et al, 1996; Flores et al, 2006; Hollin & Palmer, 2006; Bechtel et al., 2007). Given the
versatility of the instrument and the ease of its implementation, administration, and scoring, it is
perhaps not surprising that the LSI-R is one of the most widely-used assessment instruments.
The extensive use of this assessment instrument by correctional agencies in the United States and
internationally has not gone unnoticed by researchers. To date, the LSI-R and related
instruments (e.g., LSI, YLS, YLS-CMI) have been the topic of research in more than 40
published studies.
This dissertation attempts to contribute to the growing body of literature on the LSI-R in
three ways. First, the predictive validity of the LSI-R is tested to see if total LSI-R score is a
significant predictor of recidivism. The majority of research on the LSI tests the predictive
validity of the LSI-R at a single point in time. The current project is unique because the
predictive validity of the LSI-R is tested at two distinct points in time (time 1 and time).
Second, multiple assessments provide the information necessary to calculate change
scores. That is, the difference in total score at time 1 compared to the total score at time 2.
Change scores are tested to determine whether a change in score impacts the ability of the LSI-R
to predict recidivism. The review of research revealed that few studies have previously
considered the effect of change on the ability to predict recidivism (O’Keefe, Klebe, & Hromas,
220
1998; Hollin, Palmer, & Clark, 2003; Miles & Raynor, 2004; Raynor, 2007). To that end, this
dissertation will add to existing research on change scores and recidivism.
Third, this project examines the individual domains of the LSI-R to consider if one
domain is a better predictor than the others at time 1 and time 2. Again, the multiple assessment
points allow for the calculation of change scores. In turn, the change in domain total score from
time 1 to time 2 is tested to see if change in a particular domain of the LSI is more or less
important than change in the other domains. The review of result did not reveal any past
research that specifically analyzed change in individual domain scores. As such, this dissertation
will begin to bridge the gap of research in this area.
This chapter will begin by providing a summary of the dissertation’s empirical findings
regarding the predictive validity of the LSI-R. The next two sections will discuss the findings’
theoretical and policy implications. The chapter will conclude by examining the future of the
LSR-R.
SUMMARY OF RESULTS
The bivariate results for the sample and all subgroups indicate that the LSI-R is a valid
predictor of recidivism at time 1 and time 2. There is a positive and significant relationship
between offender total LSI-R score and recidivism. This finding suggests that the higher the
total LSI-R score, the higher the likelihood of recidivism. Conversely, the lower the total LSI-R
score, the lower the likelihood of recidivism. As previously mentioned, the LSI-R is a valid
predictor of recidivism at time 1 and time 2. Comparisons of confidence intervals at time 1 and
221
time 2 suggest that there is no statistically significant difference in the instrument’s ability to
predict recidivism at time 1 or time 2.
Multivariate logistic regression models are estimated for the sample for all subgroups at
time 1 and time 2. Consistent with the findings from the bivariate analysis, the multivariate
analyses indicate that LSI-R total score is a valid predictor of recidivism at time 1 and time 2 for
the sample and all subgroups. The direction of the total LSI-R coefficients in each of the models
suggests a positive relationship between LSI-R total score and recidivism. That is, the higher the
total LSI-R score, the greater the likelihood of recidivism. A comparison of the regression
models at time 1 and time 2 revealed that there is no significant difference between the time 1
and time 2 models.
The change models for the sample and subgroups provide answers to two important
questions. First, does change matter? That is, are changes in LSI-R scores associated with
changed in the likelihood of recidivism? Second, does change in total LSI-R score impact rate of
recidivism the same across categories of risk? That is, does change in LSI-R score impact rate of
recidivism the same for low-risk offenders as it does high-risk offenders?
Percent change is a significant predictor of recidivism for the sample. Percent change is
also a significant predictor of recidivism for males, whites, probationers, and parolees. Percent
change is not a significant predictor for blacks or females. The direction of the coefficients
suggests a negative relationship between change and recidivism. As such, more change results in
a lower likelihood of recidivism. The interaction between risk level and percent change is
significant for the entire sample and for males but fails to predictor for the remaining subgroups.
This finding indicates that percent change is more meaningful for high-risk offenders than it is
for low risk offenders.
222
Raw change is a significant predictor for the entire sample. Raw change is also a
significant predictor for all subgroup samples. This finding suggests that an offender whose total
LSI-R score increases from time 1 to time 2 has an increased likelihood of recidivism.
Conversely, an offender whose total score decreases from time 1 to time 2 has a decreased
likelihood of recidivism. The interaction between risk category and raw change is a significant
predictor for the sample but fails to predict for any of the subgroups. This finding indicates that
change in raw score is more meaningful for high risk offenders than it is for low risk offenders.
From the information provided in the multivariate analysis, it is possible to predict the
impact a 10% change in total LSI-R score has on rate of recidivism. The findings suggest that a
10% change in total LSI-R score for high-risk offenders results in a 6% reduction in recidivism
while a 10% change in total LSI-R score for low risk offenders results in a 1% reduction. To that
end, change is not the same across categories of risk. This finding has practical implications that
will be discussed later in the chapter.
The results from the bivariate and multivariate domain analyses suggests that no one
domain is a significantly predictor of recidivism than any other domain at time 1 or time 2. This
finding holds true for the sample and all subgroups. Percent change and raw change in the
domains were analyzed for the sample and then subsamples of males, females, blacks, whites,
probationers, and parolees.
Percent change in domain total score is calculated for each of the ten domains. Percent
change in a specific domain total score fails to predict recidivism for the sample and for
subsamples of females, blacks, whites, probationers, and parolees. The one exception is in the
sample of whites where percent change in attitudes/orientation is a significant predictor of
223
recidivism. The direction of the coefficient suggests that a reduction in total score in
attitudes/orientation corresponds with a reduced likelihood of recidivism for white offenders.
The results from the raw change analysis suggest that change in criminal history is
significant for the sample and for subsamples of males, females, whites, probation, and parole.
Raw change in criminal history is not a significant predictor of recidivism for blacks. Raw
change in leisure/recreation is significant for the sample and for subsamples of males, whites,
and probationers. Raw change in attitudes/orientation is significant for whites but not for the
sample or for any other subgroup. Raw change in the remaining domains fails to significantly
predict recidivism for the sample or respective subgroups.
In sum, five major findings emerge from the current project. First, the LSI-R is a valid
predictor of recidivism irrespective of gender, race, or supervisory status. Second, change
matters. Reduction in total score results in lower rates of recidivism. Conversely, increases in
total score result in higher rates of recidivism. Third, change matters more for some than it does
for others. Specifically, reducing the total LSI-R score of high-risk offenders corresponds with
more dramatic reductions in recidivism than does reducing the total LSI-R score of low-risk
offenders. Fourth, no single domain the LSI-R is a significantly better predictor of recidivism
than the others. Rather, it is the cumulative effect of the domains that predicts recidivism. Fifth
and finally, change in a particular domain does not appear to be more important that change in
the other domains. Again, it appears that the cumulative change across domains is more
important than change in a single domain.
224
IMPLICATIONS FOR THE THEORY OF EFFECTIVE
CORRECTIONAL INTERVENTION
The LSI-R is an assessment instrument based on the theory of effective intervention.
Although this dissertation is not a direct test of this theory, the findings do have theoretical
implications. To reiterate, the three major components outlined in the theory of effective
intervention are the principles of risk, need, and responsivity. The risk principle maintains that
offenders should receive treatment and supervision commensurate with their level of risk. That
is, high-risk offenders should receive more treatment and supervision than low-risk offenders.
The needs principle asserts individuals have certain characteristics (e.g., antisocial attitudes,
criminal peers, antisocial personality) that may increase their likelihood of recidivism. These
factors or “needs” can be targeted for treatment. Finally, the responsivity principle suggests that
offenders should be placed in treatment programs based on their risk, need, and learning style.
The purpose of matching offender to treatment is to maximizes the offender’s potential for
positive change and reduce likelihood of recidivism (Andrews et al.,1990; Andrews & Bonta,
1998).
The findings from the current study relate back to the theory of effective intervention in
two ways. First, the risk principle dictates that high-risk offenders should be the focus of
treatment and supervision efforts. The findings from the change analyses suggest that the impact
that change has on rate of recidivism varies across categories of risk. Specifically, change for
high-risk offenders has a greater impact on rate of recidivism than change for low-risk offenders.
Recall, a 10% reduction in total LSI-R score for high-risk offenders resulted in a 6% reduction in
rate of recidivism. A 10% reduction in LSI-R score resulted in a 1% reduction in recidivism for
low-risk offenders. Although any reduction in recidivism is meaningful, the difference in crime
225
savings between high and low-risk offenders suggests that concentrated efforts with high-risk
offenders may provide the biggest return (reduction in recidivism) on correctional investment
(treatment and supervision services).
Second, the theory of effective intervention implies that individual behavior can be
changed if appropriate criminogenic needs are targeted for change through correctional
treatment. To be clear, this dissertation did not consider or evaluate any type of treatment
services that the offenders in the sample may have received prior to or during the study period.
As such, it is not appropriate to speculate why an offender’s total LSI-R score at time 2 is higher,
lower, or the same as it was at time 1. However, the fluctuation in total LSI-R scores leaves
open the possibility that offenders may have risks/needs that can be identified through proper
assessment and targeted for change through appropriate correctional programming. Again,
speculating why change in total LSI-R score occurred is beyond the scope of the current project,
but this line of research is recommended for future studies involving change scores and the LSIR.
POLICY IMPLICATIONS
As of 2005, there were approximately 5 million U.S. citizens on probation or parole
(Glaze and Bonczar, 2006). The dramatic increase in the offender population over the last thirty
years has forced correctional agencies to make difficult decisions about how to balance the need
for public safety against the cost of treating and supervising the offender population. To that
end, correctional agencies must resort to managing groups of offenders rather than each
individual offender (Feeley & Simon, 1992). Offender classification instruments are commonly
226
used by correctional agencies to divide offenders into groups based on offender risk level.
Although there are a number of different classification instruments available for use today, the
LSI-R has emerged as one of the most popular. It is within this context, that the policy
implications for use of the LSI-R with the offender population are discussed.
The findings from the majority of studies on the LSI-R conclude that the instrument is a
valid predictor of recidivism (Gendreau et al., 1997; Barnoski & Aos, 2003; Simourd, 2004;
Mills et al., 2005; Holsinger et al., 2006). The findings from this dissertation are consistent in
support of the LSI-R as a valid predictor of recidivism. For that reason, the LSI-R it is
appropriate for correctional agencies to adopt this risk/needs assessment for use in offender
classification (Andrews and Bonta, 1995).
As previously mentioned, the LSI-R is a valid predictor of recidivism. Some suggest that
the LSI-R may not be a valid predictor of recidivism with select offender groups. In particular,
researchers have questioned the use of the LSI-R with female offenders (Reisig et al 2006;
Holtfreter & Culp, 2007). Contrary to those findings, the current study indicates that the LSI-R
is a valid predictor of recidivism across categories of gender, race, and supervisory status. The
notion that the LSI-R is appropriate for general use – that is, for a variety offender populations –
as opposed to specific use – only appropriate for use with a select offender population – will
likely add to the already broad appeal of the LSI-R with correctional agencies in the United
States and internationally.
The statistical analysis conducted for the current work provided information regarding
rate of recidivism by risk category on the LSI-R at two distinct points in time. Not surprisingly,
the rate of recidivism increased as risk level increased. Alone, this finding supports the need for
correctional agencies to utilize the LSI-R to identify high-risk offenders for treatment and
227
supervision because high-risk offenders are most likely to recidivate. However, the previous
finding, coupled with the idea that a 10% change in total LSI-R results in greater reductions in
recidivism for high-risk offenders (6%) compared to a smaller reduction in recidivism for lowrisk offenders, should further motivate correctional agencies to focus treatment and supervision
efforts on high-risk offenders. High-risk offenders pose the greatest likelihood of failure but also
have the most potential for positive change.
The finding that change in LSI-R score matters in predicting recidivism and that the
degree to which it matters varies across categories begs the question whether or not it is
necessary to administer the LSI-R to offenders at multiple points in time. Administering the
LSI-R multiple times is recommended for two reasons. First, multiple assessment points
provides an opportunity for the correctional agency to monitor an offender’s rehabilitative
progress. For example, if an offender is assessed at intake and placed in a treatment program
that corresponds with his or her risks, criminogenic needs, and responsivity factors, then a
reduction in total LSI-R score should occur between time 1 and time 2 assessments. If the
offender is reassessed six months after his or her initial assessment, having received treatment in
the months between initial and follow-up assessments, and no change in total LSI-R score has
taken place, then the agency may need to assign the offender to a different treatment program
that may better address the offender’s risks and criminogenic needs.
No change in total LSI-R score may indicate a problem with the program to which the
offender has been assigned. Research suggests that on average, treatment programs reduce
recidivism by 10% (Lipsey, 1992; Losel, 1995). Moreover, some treatment programs work
better than others (Gendreau & Ross, 1979, Cullen & Gendreau, 2000). To that end, in addition
228
to helping agencies monitor the rehabilitative progress of offenders, multiple assessments may
also help agencies monitor the effectiveness of their treatment programs.
The second reason to administer assessment at multiple points in time involves
supervisory placement and release decisions. Given that the probability of recidivating varies
across categories of risk and that change in total scores impacts rates of recidivism differently
across categories of risk, it is possible that offender risk level could change enough to warrant
either a reduction in supervisory level or a complete release from supervision. For example,
offenders in the sample initially assessed as moderate-risk have a 31.3% likelihood of failure.
Moderate-risk offenders whose risk level decreased to low-risk upon reassessment have a 13.3%
failure rate. A dramatic decrease in likelihood of recidivism may inspire an agency with limited
financial and human resources either to reduce the treatment and supervision for the now lowrisk offender or to release the low-risk offender from supervision so that treatment efforts may be
directed towards offenders who have a greater likelihood of failure.
FUTURE RESEARCH
Although there have been over forty studies on the LSI-R to date, very few of them have
considered the impact of change in LSI-R score on likelihood of recidivism (O’Keefe, Klebe, &
Hromas, 1998; Hollin, Palmer, & Clark, 2003; Miles & Raynor, 2004; Raynor, 2007). As such,
the findings from the current dissertation bring to light new research questions to consider in
studies.
First, there is the need to replicate this study with other samples of offenders. While
change matters with this particular sample of probationers and parolees from Iowa, change may
229
or may not matter with samples of offenders from other states or countries. To that end, does
change matter with juvenile offenders? Given that this is only one study and that the
demographic characteristics of offenders Iowa do not represent the entire population of
offenders, it is recommended that researchers carry out similar studies on variety of offender
populations.
Second, the results from the current project indicate that change in total LSI-R score
occurred but cannot offer an explanation as to why the change occurred. Research that considers
the type of treatment the offender receives may help to explain why some offenders’ scores
increase, some decrease, and some stay the same from one assessment point to the next. As
previous research has shown, on average, treatment programs reduce rates of recidivism by 10%
(Lipsey, 1992; Losel, 1995) and that some treatment programs work better than others (Gendreau
& Ross, 1979, Cullen & Gendreau, 2000). To that end, it is important to examine the type of
treatment received and the degree to which type of treatment may or may not impact change in
total LSI-R score.
Third, it is important for future research efforts to question when the LSI-R should be
administered. If an agency is only able to administer the LSI-R once, should that take place at
intake or perhaps six months after the offender has received treatment and supervisory services?
At what point in time is the instrument most accurate in predicting recidivism? Similarly, if an
agency is able to administer the LSI-R at multiple points in time, what is the appropriate number
of days between initial and follow up assessment?
Perhaps the best way to address these questions is through research using an experimental
design. This method would allow for the assignment of offenders to an experimental group or
control group. Moreover, the researcher can control when offenders are assessed and reassessed.
230
Also, this method would allow the researcher to dictate the type of treatment the offender
receives between initial assessment and follow-up assessment. Ultimately, the experimental
design method of research has the potential to overcome many of the limitations of the current
study. Accordingly, it is recommended that experimental design research is used in future
research that attempts to assess the impact of change in LSI-R score on the prediction of
recidivism.
Although a considerable amount of research has been conducted on the LSI-R, there are
still many questions to be answered. For that matter, there are still many questions to be asked.
The findings from this dissertation help contribute to the existing body of research on the LSI-R
and to extend the literature by examining the impact of change on rates of recidivism.
Additional research efforts assessing offenders at multiple points in time and the impact of
change on rates of recidivism are necessary. The findings from this project and similar projects
in the future may provide correctional agencies with valuable information to better treat,
supervise, and manage the burgeoning offender population.
231
REFERENCES
Andrews, D. A. (1982). The level of supervisory inventory (LSI): The first follow-up. Ontario,
Canada: Ministry of Correctional Services.
Andrews, D. A. (1990). Classification for effective rehabilitation rediscovering psychology.
Criminal Justice and Behavior, 17, 19-52.
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 metaanalysis. Criminology, 28, 369-404.
Andrews, D. A., Bonta, J. (1995). LSI-R the level of service inventory revised user’s manual.
North Tonawanda, NY: Multi-Health Systems Inc.
Andrews, D. A., Bonta, J. (1998). The psychology of criminal conduct (2nd ed.). Cincinnati, OH:
Anderson Publishing Company.
Andrews, D. A., & Bonta, J. (2003). The psychology of criminal conduct (3rd ed.). Cincinnati,
OH: Anderson Publishing Company.
Andrews, D. A., Bonta, J., & Wormith, J. S. (2006). The recent past and near future of risk
and/or need assessment. Crime & Delinquency, 52, 7-27.
Andrews, D. A., Kiessling, J. J., Mickus, S., & Robinson, D. (1986). The construct validity of
interview-based risk assessments in corrections. Canadian Journal of Behavioral Science,
18, 460-471.
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 metaanalysis. Criminology, 28, 369-404.
Austin, J., Coleman, D., Peyton, J., & Dedel Johnson, K. (2003). Reliability and validity study of
the LSI-R risk assessment instrument. Washington, DC: The Institute on Crime, Justice and
Corrections.
Barnoski, R., & Aos, S. (2003). Washington's offender accountability act: An analysis of the
department of corrections' risk assessment. Olympia, WA: Washington State Institute for
Public Policy.
Bechtel, K., Lowenkamp, C. T., & Latessa, E. J. (2007). Assessing the risk of re-offending for
juvenile offenders using the youth level of service/case management inventory. Journal of
Offender Rehabilitation, 45(3/4), 85-108.
Bonta, J. (1989). Native inmates: Institutional response, risk, and needs. Canadian Journal of
Criminology, 31, 49-62.
232
Bonta, J. (2002). Offender risk assessment guidelines for selection and use. Criminal Justice
and Behavior, 29, 355-379.
Bonta, J., & Motiuk, L. L. (1985). Utilization of an interview-based classification instrument: A
study of correctional halfway houses. Criminal Justice and Behavior, 12, 333-352.
Bonta, J. & Motiuk, L. L. (1987). The diversion of incarcerated offenders to correctional
halfway houses. Journal of Research in Crime and Delinquency, 24, 302-323.
Bonta, J. & Motiuk, L. L. (1990). Classification to halfway houses: A quasi experimental
evaluation. Criminology, 28, 497-506.
Bonta, J. & Motiuk, L. L. (1992). Inmate classification. Journal of Criminal Justice, 20, 343353.
Bonta, J., Wallace-Capretta, S. & Rooney, J. (2000). A quasi-experimental evaluation of an
intensive rehabilitation supervision program. Criminal Justice and Behavior, 27, 312-329.
Clear, T. R., & Gallagher, K. W. (1985). Probation and parole supervision: A review of current
classification practices. Crime & Delinquency, 31, 423-443.
Clear, T. R. (1994). Harm in American penology: Offenders, victims and their communities.
Albany, NY: State University of New York Press.
Cohen, R. L. (1995). Probation and parole violators in state prison, 1991. (NCJ149076).
Washington, DC: U.S. Department of Justice.
Coulson, G., Ilacqua, G., Nutbrown, V., & Cudjoe, F. (1996). Predictive utility of the LSI for
incarcerated female offenders. Criminal Justice and Behavior, 23, 427-439.
Cullen, F. T. (2005). The twelve people who saved rehabilitation: How the science of
criminology made a difference. Criminology, 43, 1-42.
Cullen, F. T., & Gendreau, P. (2000). Assessing correctional rehabilitation: policy, practice, and
prospects. In J. Horney (Ed.) Criminal justice 2000: Volume 3 – Policies, processes, and
decisions of the criminal justice system (pp. 109-175). Washington, DC: National Institute of
Justice.
Cullen, F. T., & Gilbert, K. E. (1982). Reaffirming rehabilitation. Cincinnati, OH: Anderson
Publishing Company.
Cullen, F. T., Wright, J. P., & Applegate, B. K. (1996). Control in the community: The limits of
reform. In A. T. Harland (Ed.), Choosing correctional interventions that work (pp. 69-116).
Thousand Oaks, CA: Sage Publications.
233
Dahle, K. P. (2006). Strengths and limitations of actuarial prediction of criminal reoffence in a
German prison sample: A comparative study of LSI-R, HCR-20 and PCL-R. International
Journal of Law & Psychiatry, 29, 431-442.
Dowdy, E. R., Lacy, M. G., & Unnithan, N. P. (2001). Correctional prediction and the level of
supervision inventory. Journal of Criminal Justice, 30, 29-39.
Feeley, M. M., & Simon, J. (1992). The new penology: Notes on the emerging strategy of
corrections and its implications. Criminology, 30, 449-474.
Flores, A. W., Lowenkamp, C. T., Holsinger, A. M., & Latessa, E. J. (2006). Predicting outcome
with the level of service inventory-revised: The importance of implementation integrity.
Journal of Criminal Justice, 34, 523-529.
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), 4449.
Folsom, J., & Atkinson, J. L. (2007). The generalizability of the LSI-R and the CAT to the
prediction of recidivism in female offenders. Criminal Justice and Behavior, 34, 1044-1056.
Fulton, B., Latessa, E. J., Stichman, A., & Travis, L. F. (1997). The State of ISP: Research and
Policy Implications. Federal Probation, 61(4), 65-75.
Gardner, W., Lidz, C. W., Mulvey, E. P., & Shaw, E. C. (1996). Clinical versus actuarial
predictions of violence in patients with mental illness. Journal of Consulting and Clinical
Psychology, 64, 602-609.
Gendreau, P. (1996). Offender rehabilitation what we know and what needs to be done.
Criminal Justice and Behavior, 23, 144-161.
Gendreau, P. (1996). The principles of effective intervention with offenders. In A. T Harland
(Ed.), Choosing correctional options that work: Defining the demand and evaluating the
supply (pp. 117-130). Thousand Oaks, CA: Sage.
Gendreau, P. Goggin, C. E., & Law, M. A. (1997). Predicting prison misconducts. Criminal
Justice and Behavior, 24, 414-431.
Gendreau, P., Goggin, C. E., & Smith, P. (2002). Is the PCL-R really the "unparalled" measure
of offender risk? A lesson in knowledge cumulation. Criminal Justice and Behavior, 29,
397-426.
Gendreau, P., Little, T., & Goggin, C. E. (1996). A meta-analysis of the predictors of adult
offender recidivism: What works! Criminology, 34, 575-607.
234
Gendreau, P., & Ross, R. R. (1979). Effective correctional treatment: Bibliotherapy for cynics.
Crime and Delinquency, 25, 463-489.
Gendreau, P., & Ross, R. R. (1987). Revivification of rehabilitation: Evidence from the 1980s.
Justice Quarterly, 4, 349-407.
Gendreau, P., Smith, P., & French, S. A. (1996). The theory of effective intervention: Empirical
status and future directions. In F. T. Cullen, J. P. Wright, & K. R. Blevins (Eds.), Taking
stock the status of criminological theory (pp. 419-446). New Brunswick, NJ: Transaction
Publishers.
Girard, L., & Wormith, J. S. (2004). The predictive validity of the level of service inventoryOntario revision on general and violent recidivism among various offender groups. Criminal
Justice and Behavior, 31, 150-181.
Glaser, D. (1985). Who gets probation and parole: Case study versus actuarial decision making.
Crime & Delinquency, 31, 367-378.
Glaze, L. E., & Bonczar, T. P. (2006). Probation and parole in the United States, 2005.
(NCJ215091). Washington, DC: U.S. Department of Justice.
Glaze, L. E., & Bonczar, T. P. (2007). Probation and parole in the United States, 2006.
(NCJ220218). Washington, DC: U.S. Department of Justice.
Gnall, K. A. & Zajac, G. (2005). Assessing for success in offender reentry. Corrections Today,
67(2), 94-97.
Gottfredson, S. D., & Moriarty, L. J. (2006). Statistical risk assessment: Old problems and new
applications. Crime & Delinquency, 52, 178-200.
Grove, W. M., & Meehl, P. E. (1996). Comparative efficiency of informal (subjective,
impressionistic) and formal (mechanical, algorithmic) prediction procedures: The clinicalstatistical controversy. Psychology, 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.
Hendricks, B., Werner, T., & Shipway, L. (2006). Recidivism among spousal abusers:
Predictions and program evaluations. Journal of Interpersonal Violence, 21, 703-716.
Hirschi, T. & Gottfredson, M. R. (1983). Age and the explanation of crime. American Journal
of Sociology, 89, 552-584.
Hoffman, P. B. (1983). Screening for risk: A revised salient factor score (SFS 81). Journal of
Criminal Justice, 11, 539-547.
235
Hoffman, P. B. (2003, May). History of the federal parole system. Retrieved August 11, 2007,
from http://www.usdoj.gov/uspc/history.htm
Hoge, R. D. (2002). Standardized instruments for assessing risk and need in youthful offenders.
Criminal Justice and Behavior, 29, 380-396.
Hollin, C. R., & Palmer, E. J. (2006). The level of service inventory-revised profile of English
prisoners: Risk and reconviction analysis. Criminal Justice and Behavior, 33, 347-366.
Hollin, C. R., Palmer, E. J., & Clark, D. (2003). The level of service inventory-revised profile of
English prisoners a needs analysis. Criminal Justice and Behavior, 30, 422-440.
Holsinger, A. M., Lowenkamp, C. T., & Latessa, E. J. (2006). Exploring the validity of the level
of service inventory-revised with native american offenders. Journal of Criminal Justice, 34,
331-337.
Holtfreter, K., & Culp, R. (2007). Gender and risk assessment the empirical status of the LSI-R
for women. Journal of Contemporary Criminal Justice, 23, 363-382.
Holtfreter, K., Reisig, M. D., & Morash, M. (2004). Poverty, state capital, and recidivism among
women offenders. Criminology & Public Policy, 3, 185-208.
Hubbard, D. J. (2007). Getting the most out of correctional treatment: Testing the responsivity
principle on male and female offenders. Federal Probation, 71(1), 2-8.
Hughes, T. A., Wilson, D. J., & Beck, A. J. (2001). Trends in state parole, 1990-2000. (NCJ
184735). Washington, DC: U.S. Department of Justice.
Ilacqua, G. E., Coulson, G. E., Lombardo, D., & Nutbrown, V. (1999). Predictive validity of the
young offender level of service inventory for criminal recidivism of male and female young
offenders. Psychological Reports, 84, 1214-1218.
Johnson Listwan, S., Cullen, F. T., & Latessa, E. J. (2006). How to prevent prison re-entry
programs from failing: Insights from evidence-based corrections. Federal Probation, 70(2),
19-25.
Kirkpatrick, B. (1998). Field test of the LSI-R a study of offenders under community
supervision. Perspectives, 22(Fall), 24-28.
Kirkpatrick, B. (1999). Exploratory research of female risk prediction and LSI-R. Corrections
Compendium, 24(5), 1-3, 14-17.
Latessa, E. J., & Allen, H. E. (1997). Corrections in the community (3rd ed.). Cincinnati, OH:
Anderson Publishing Company.
236
Lindner, C., & Savarese, M. R. (1984). The evolution of probation early salaries, qualifications,
and hiring practices. Federal Probation, 48(3), 3-10.
Lipsey, M. W. (1992). Juvenile delinquency treatment: A meta-analytic inquiry into the
variability of effects. In T. D. Cook, H. Cooper, D. S. Corday, H. Hartmann, L. V. Hedges,
R. J. Light, A. Lewis, & F. Mosteller (Eds.), Meta-Analysis for explanation: A casebook (pp.
83-127). New York: Russell Sage Foundation.
Lipsey, M. W. (1999). Can intervention rehabilitate serious delinquents? Annals of the
American Academy of Political and Social Science, 564,142-166.
Lipsey, M. W. & Wilson, D. B. (1998). Effective intervention for serious juvenile offenders: A
synthesis of research. In R. Loeber & D. P. Farrington (Eds.), Serious and violent juvenile
offenders: Risk factors and successful interventions (pp. 313-345). Thousand Oaks, CA:
Sage Publications.
Losel, F. (1995). The efficacy of correctional treatment: A review and synthesis of metaevaluations. In J. McGuire (Ed.), What works: Reducing reoffending (pp. 77-111). New
York, NY: John Wiley.
Lowenkamp, C. T. & Bechtel, K. (2007). The predictive validity of the LSI-R on a sample of
offenders drawn from the records of the Iowa department of corrections data management
system. Federal Probation, 71(3), 25-29.
Lowenkamp, C. T., Holsinger, A. M., & Latessa, E. J. (2001). Risk/Need assessment, offender
classification, and the role of childhood abuse. Criminal Justice and Behavior, 28, 543-563.
Lowenkamp, C. T. & Latessa, E. J. (2005). Increasing the effectiveness of correctional
programming through the risk principle: Identifying offenders for residential placement.
Criminology and Public Policy, 4, 263-290.
Lowenkamp, C. T., Smith, P., & Latessa, E. J. (2008). Can 14,737 women be wrong? A metaanalytic review of the research on the relationship between the LSI-R and recidivism among
female offenders. Unpublished Manuscript, University of Cincinnati, Cincinnati, OH.
Mackenzie, D. L. (2000). Evidence-Based corrections: Identifying what works. Crime and
Delinquency, 46, 457-471.
Marczyk, G. R., Heilbrum, K., Lander, T., & DeMatteo, D. (2003). Predicting juvenile
recidivism with the PCL:YV, MAYSI, and YLS/CMI. International Journal of Forensic
Mental Health, 2, 7-18.
Marlowe, D. B., Festinger, D. S., Lee, P. A., Dugosh, K. L., & Benasutti, K. M. (2006).
Matching judicial supervision to clients’ risk status in drug court. Crime & Delinquency, 52,
52-76.
237
Martinson, R. (1974). What works? Questions and answers about prison reform. Public
Interest, 35(Spring), 22-54.
Miles, H., & Raynor, P. (2004). Community sentences in jersey: Risk, need and rehabilitation.
St. Helier, Jersey: Jersey Probation and After Care Service.
Mills, J. F., Kroner, D.G., & Hemmati, T. (2003). Predicting violent behavior through a staticstable variable lens. Journal of Interpersonal Violence, 18, 891-904.
Mills, J. F., Jones, M. N., & Kroner, D. G. (2005). An examination of the generalizability of the
LSI-R and VRAG probability bins. Criminal Justice and Behavior, 32, 565-585.
Mills, J. F., & Kroner, D. G. (2006). The effect of discordance among violence and general
recidivism risk estimates on predictive accuracy. Criminal Behaviour & Mental Health, 16,
155-166.
Minor, K. I., Wells, J. B., & Sims, C. (2003). Recidivism among federal probationers –
Predicting sentence violations. Federal Probation, 67(1), 31-36.
Motiuk, L. L., Bonta, J., & Andrews, D. A. (1986). Classification in correctional halfway houses
the relative and incremental predictive criterion validities of the Megargee-MMPI and LSI
systems. Criminal Justice and Behavior, 13, 33-46.
Motiuk, M. S., Motiuk, L. L., & Bonta, J. (1992). A comparison between self-report and
interview-based inventories in offender classification. Criminal Justice and Behavior, 19,
143-159.
O’Keefe, M. L., Klebe, K., & Hromas, S. (1998). Validation of the level of supervision inventory
(LSI) for community based offenders in Colorado: Phase II. Colorado Spring, CO: Colorado
Department of Corrections.
Palmer, E. J., & Hollin, C. R. (2007). The level of service inventory-revised with English
women prisoners a needs and reconviction analysis. Criminal Justice and Behavior, 34, 971984.
Palmer, T. (1975). Martinson revisited. Journal of Research in Crime and Delinquency, 12,
133-152.
Petersilia, J. (1997). Probation in the United States. In M. Tonry (ed.) Crime and justice: A
review of research (Vol. 22, 149-200). Chicago, IL: University of Chicago Press.
Petersilia, J. (1999). A decade of experimenting with intermediate sanctions: What have we
learned? Federal Probation, 62(2), 3-9.
238
Petersilia, J. & Turner, S. (1993). Intensive probation and parole. In. M. Tonry (ed.) Crime and
justice: A review of research (Vol. 17, pp. 281-335). Chicago, IL: University of Chicago
Press.
Raynor, P. (2007). Risk and need assessment in British probation: The contribution of LSI-R.
Psychology, Crime & Law, 13, 125-138.
Raynor, P., Knych, J., Roberts, C., & Merrington, S. (2000). Risk and need assessment in
probation services: An evaluation. Home Office Research Study 211. London: Home Office.
Reisig, M. D., Holtfreter, K., & Morash, M. (2006). Assessing recidivism risk across female
pathways to crime. Justice Quarterly, 23, 384-405.
Rothman, D. J. (1980). Conscience and convenience the asylum and its alternatives in
progressive America. New York: Aldine de Gruyter.
Sarbin, T. R. (1943). A contribution to the study of actuarial and individual methods of
prediction. American Journal of Sociology, 48(5), 593-602.
Schlager, M. D., & Simourd, D. J. (2007). Validity of the level of service inventory-revised
(LSI-R) among african american and hispanic male offenders. Criminal Justice and
Behavior, 34, 545-554.
Schmidt, F., Hoge, R. D., & Gomes, L. (2005). Reliability and validity analyses of the youth
level of services/case management inventory. Criminal Justice and Behavior, 32, 329-344.
Schneider, A. L. (1996). Further exploration of the flight from discretion: The role of risk/need
instruments in probation supervision decision. Journal of Criminal Justice, 24, 109-121.
Shields, I. W. (1993). The use of the young offender-level of service inventory (YO-LSI) with
adolescents. IARCA Journal on Community Corrections, 5, 10, 26.
Simourd, D. J. (2004). Use of Dynamic Risk/Needs Assessment Instruments Among Long Term
Incarcerated Offenders. Criminal Justice and Behavior, 31, 306-323.
Smith, P., Gendreau, P., & Goggin, C. (2007). Correctional treatment: Accomplishments and
realities. In P. Van Voorhis, M. Braswell, & D. Lester (Eds.), Correctional counseling &
rehabilitation (6th ed., pp. 301-310). Cincinnati, OH: Anderson Publishing Company.
Smith, P., Gendreau, P., & Goggin, C. (2004). Correctional treatment: Accomplishments and
realities. In P. Van Voorhis, M. Braswell, & D. Lester (Eds.), Correctional counseling &
rehabilitation (5th ed., pp. 285-294). Cincinnati, OH: Anderson Publishing Company.
Van Voorhis, P. (2000). An overview of offender classification systems. In P. Van Voorhis, M.
Braswell, & D. Lester (Eds.), Correctional counseling & rehabilitation (4th ed., pp. 81-108).
Cincinnati, OH: Anderson Publishing Company.
239
Van Voorhis, P., & Lester, D. (2004). Cognitive therapies. In P. Van Voorhis, M. Braswell, &
D. Lester (Eds.), Correctional counseling & rehabilitation (5th ed., pp. 183-207). Cincinnati,
OH: Anderson Publishing Company.
Whitehead, J. T. (1991). The effectiveness of felony probation: Results from an eastern state.
Justice Quarterly, 8, 525-543.
240