THE CALIFORNIA DEPARTMENT OF CORRECTIONS AND REHABILITATION

EXPERT PANEL STUDY OF THE
INMATE CLASSIFICATION SCORE SYSTEM

State of California
Department of Corrections and Rehabilitation
___________________________________________________
Office of Research
Research and Evaluation Branch

December 2011

Distributed by the:
CDCR Research and Evaluation Branch
1515 S Street., Suite 208-S, Sacramento, CA 95811
(916) 322-2919
http://www.cdcr.ca.gov/Reports_Research/index.html

December 19, 2011
Matt Cate
Secretary, California Department of Corrections and Rehabilitation
1515 S Street
Sacramento, CA 95814
Dear Mr. Cate:
Over the past 15 months, our group has served as the Expert Panel to the CDCR effort to
evaluate the inmate classification system. Expert Panel members were drawn from academia and
represent individuals with research, methodological and statistical skills necessary to assist and
review the efforts of CDCR staff and their consultants.
Throughout the course of the project, the Expert Panel has assisted the Department in the
development and review of the study design, evaluated analyses and results, and provided
comments on draft and final reports. Our goal has been to bring the best available methods and
data to bear to key questions related to potential classification changes. The findings and
recommendations presented in the report are supported by all members of the Expert Panel.
We thank the many members of the CDCR who assisted the Expert Panel in our work by
providing extensive documentation on the classification process, gathering and abstracting data,
and answering our many questions.
We hope the findings of the report will prove helpful to the Department classification process.
Sincerely,

_________________________
David Farabee, Ph.D.
University of California, Los Angeles

_________________________
Ryken Grattet, Ph.D.
University of California, Davis

Matt Cate

_________________________
Richard McCleary, Ph.D.
University of California, Irvine

_________________________
Steven Raphael, Ph.D.
University of California, Berkeley

________________________
Susan Turner, Ph.D.
University of California, Irvine

Page 2

December 19, 2011

Table of Contents
Members of the Expert Panel ................................................................................................. i
Department of Corrections and Rehabilitation Representatives ............................................ ii
Glossary of Acronyms .......................................................................................................... iii
Executive Summary ..............................................................................................................1
THE INMATE CLASSIFICATION SCORE SYSTEM STUDY ................................................ 3
Introduction ...........................................................................................................................3
History of the CDCR Inmate Classification Score System (ICSS).......................................... 3
Current ICSS Design.............................................................................................................4
Major Project Goals and Research Questions of Study ......................................................... 4
Project Methodology .............................................................................................................5
Classification Study Team ............................................................................................. 5
Prison Housing Design Tours ........................................................................................ 5
Classification Study Team Meetings .............................................................................. 6
Expert Panel Data Analysis Plan ...........................................................................................6
Data Sources ........................................................................................................................8
Study Findings ......................................................................................................................9
Expert Panel Recommendations ......................................................................................... 13
Supplementary Materials: Misclassification and Other States’ Current Strategies for
Reducing Inmate Population ............................................................................................... 13
Suggestions for Future Research ........................................................................................ 13
Appendix A Overview of the CDCR Inmate Classification Process...................................... 15
Appendix B Overview of the CDCR Inmate Classification Procedures ................................ 19
Appendix C CDC 839 Classification Score Sheet ................................................................ 22
Appendix D CDC 840 Reclassification Score Sheet ............................................................ 23
Appendix E Expert Panel Data Analysis Plan March 2011 .................................................. 24
Appendix F Successful CDCR Escapes: 1999 to 2010 ...................................................... 32
Appendix G Variable List for the Classification Score and Close Custody Designation
Datasets ...........................................................................................................33
Appendix H Data Analysis and Reporting Project Plan........................................................ 35
Appendix I Summary of Findings from the Regression-Discontinuity Analysis of Inmate
Behavioral Outcomes ....................................................................................... 39
Appendix J Statistical Methods and Summarized Results for Gap, Matching and Longitudinal
Analyses .......................................................................................................... 74
Appendix K CDCR Inmate Classification Score System Study Crosswalk......................... 115
Appendix L Cohort Descriptives ........................................................................................ 120

Inmate Classification Score System Study

Appendix M Literature Review on Escape Risk Factors ................................................... 122
Appendix N Literature Review on Misclassification .......................................................... 135

Inmate Classification Score System Study

Members of the Expert Panel
__________________________________________________________________________________________________________________

Ryken Grattet is Professor of Sociology at the University of California, Davis. He writes about
punishment and criminal law, focusing on the development and enforcement of hate crime law and the
California prison and parole system. His research on the policing of hate crime in California led directly
to several provisions of Senate Bill 1234, the California Omnibus Hate Crime Act of 2004. His research
on the California parole system contributed to changes in the way parole violations are handled in the
state. In 2005-6, he took a leave from his teaching and research duties at UC Davis to serve as the
first Assistant Secretary of Research in the new California Department of Corrections and
Rehabilitation. His research has been awarded the Law & Society Association 2001 Article Prize; the
Society for the Study of Social Problems, Crime and Delinquency 2002 Award for Outstanding
Scholarship; and the American Sociological Association, Sociology of Law Section 2007 Distinguished
Article Award. In recognition of his public service contributions, Professor Grattet was awarded the
Pacific Sociological Association’s 2006 Distinguished Practice Award and the UC Davis 2010
Distinguished Scholarly Public Service Award.
David Farabee is Professor of Psychiatry and Biobehavioral Sciences at the University of California,
Los Angeles and a principal investigator at the Integrated Substance Abuse Programs (ISAP). He has
published in the areas of substance abuse, adult and juvenile crime, HIV/AIDS, and offender
treatment, was co-editor of the books Treatment of Drug Offenders (2002; New York: Springer) and
Treating Addicted Offenders: A Continuum of Effective Practices, Volumes I and II (2004, 2007; New
York: Civic Research Institute), author of Rethinking Rehabilitation: Why Can’t We Reform Our
Criminals? (2005; Washington, D.C.: AEI Press), and is co-editor of the Offender Programs Report.
Richard McCleary is Professor of Social Ecology at the University of California, Irvine where he
teaches courses in criminology and statistics. Dr. McCleary's research interests include the evaluation
of corrections and supervision programs. Along with Jan Chaiken and Michael Maltz, Dr. McCleary
recently completed an evaluation of CDCR population projection models. Dr. McCleary has been a
consultant to the U.S. Bureau of Prisons and to more than a dozen state corrections departments.
Susan Turner is a Professor in the Department of Criminology, Law and Society at the University of
California's Irvine campus. She also serves as Director of the Center for Evidence-Based Corrections,
and is a board member of the newly created California Rehabilitation Oversight Board (C-ROB). She
received her Ph.D. in Social Psychology from the University of North Carolina at Chapel Hill. She has
lead a variety of research projects, including studies on racial disparity, field experiments of private
sector alternatives for serious juvenile offenders, work release, day fines and a 14-site evaluation of
intensive supervision probation. Dr. Turner's areas of expertise include the design and implementation
of randomized field experiments and research collaborations with state and local justice agencies. Dr.
Turner has conducted a number of evaluations of drug courts, including a nationwide implementation
study. Her article, "A Decade of Drug Treatment Court Research" (2002) appeared in Substance Use
and Misuse, summarizing over 10 years of drug court research conducted while she was at RAND
Corporation. Dr. Turner is a member of the American Society of Criminology, the American Probation
and Parole Association, and is a Fellow of the Academy of Experimental Criminology.
Steve Raphael is a Professor of Public Policy at the University of California, Berkeley. Dr. Raphael
received his Ph.D. in economics from UC Berkeley in 1996. His primary fields of concentration are
labor and urban economics. Dr. Raphael has authored several research projects investigating the
relationship between racial segregation in housing markets and the relative employment prospects of
African-Americans. Dr. Raphael has also written theoretical and empirical papers on the economics of
discrimination, the role of access to transportation in determining employment outcomes, the
relationship between unemployment and crime, the role of peer influences on youth behavior, the
effect of trade unions on wage structures, and homelessness.

i

Inmate Classification Score System Study

Department of Corrections and Rehabilitation
Representatives
__________________________________________________________________________________________________________________

Executive Office

Classification Services Unit

Matthew L. Cate
Secretary

Tanya Rothchild
Chief

Martin Hoshino
Undersecretary

Elizabeth DeSilva
Correctional Counselor III

Terri McDonald
Undersecretary (A)

Jeff Lynch
Associate Warden

Lee E. Seale
Director

Brian Moak
Classification and Parole Representative

Office of Research

Jim Short
Correctional Counselor III

Jay Atkinson
Deputy Director (A)
Brenda Grealish
Research Manager III
Kevin Grassel
Research Program Specialist II
Jean Leonard
Research Program Specialist II

ii

Inmate Classification Score System Study

Glossary of Acronyms
CC
CCF
CCR
CDCR
CQAU
CSR
CSU
DAI
DRB
DDPS
FTB
FY
ICC
ICSS
ISRS
LWOP
MEPD
MSF
OBIS
OISB
OR
PG
PV
RC
RD
REB
RTC
RVR
SHU
UCB
UCC
UCD
UCI
UCLA
WG

Correctional Counselor
Community Correctional Facility
California Code of Regulations
California Department of Corrections and Rehabilitation
Classification Quality Assurance Unit
Classification Staff Representative
Classification Services Unit
Division of Adult Institutions
Departmental Review Board
Distributed Data Processing System
Franchise Tax Board
Fiscal Year
Institution Classification Committee
Inmate Classification Score System
Institution Staff Recommendation Summary
Life Without Parole
Minimum Earliest Parole Date
Minimum Support Facility
Offender Based Information System
Offender Information Services Branch
Office of Research
Privilege Group
Parole Violator
Reception Center
Regression Discontinuity Design
Research and Evaluation Branch
Return to Custody
Rules Violation Report, also known as a CDC 115
Security Housing Unit
University of California, Berkeley
Unit Classification Committee
University of California, Davis
University of California, Irvine
University of California, Los Angeles
Work Group

iii

Inmate Classification Score System Study

Executive Summary
The California Department of Corrections and Rehabilitation (CDCR) uses an inmate
classification system to ensure that inmates are properly housed and supervised. The proper
housing and supervision of inmates serves goals that are paramount to the correctional
setting: protecting staff and inmates from in-prison misconduct, protecting the public from
inmate escapes, and safeguarding opportunities for inmates to benefit from rehabilitative
programming. All three goals serve public safety by promoting institutional order and inmate
rehabilitation.
California’s prison system presents a multitude of housing and supervision options to achieve
these goals. Housing types range from camps to open dormitories to cells. Some housing is
protected by a low-security perimeter, some secured by an electrified fence. Some areas
have armed coverage, others do not. Within those different types of housing, inmate
supervision levels may vary as well, with some inmates more closely monitored than others.
A successful inmate classification system utilizes this spectrum of choices to ensure an
appropriate balance between liberty and security.
Currently, CDCR uses a classification process that is based on two overlapping systems: the
inmate’s placement score and the inmate’s Custody Designation. The placement score is
determined by the Inmate Classification Score System (ICSS), which is further broken down
into two parts – the preliminary score and any applicable Mandatory Minimums. The
preliminary score predicts risk for institutional misconduct using several variables related to
an inmate’s background and prior incarceration behavior. Additional Mandatory Minimum
scores are then applied to inmates incarcerated for certain violent or sex crimes, crimes of
public notoriety, or crimes carrying life sentences. Mandatory minimums restrict the housing
level to which these inmates can be assigned. The final classification placement score is the
maximum of either the preliminary or the Mandatory Minimum score. Final classification
scores determine the institution or housing level in which an inmate will be placed by
producing four levels of scores that correspond to four institutional housing security levels.
Custody Designations attempt to mitigate an inmate’s risk for escape and threat to the
community if escaped. They determine the level of in-prison supervision that inmates receive
and also present a further opportunity for restricting program access and the housing levels
to which certain inmates can be assigned.
The purpose of this study was to evaluate CDCR’s classification system. The study aims to
assist CDCR in best identifying factors that justify restrictions on liberty while avoiding factors
that could lead to unwarranted impingements on inmate rehabilitation. Analyses focused on
male offenders since the research design relied on the delineation between particular
housing levels that are not applicable to female offenders.
CDCR determined that the best strategy for carrying out the study was to work with outside
correctional experts and statisticians. An “Expert Panel” was created, comprised of scholars
with experience in studying correctional issues. To assist in conducting its analyses, the
department contracted with the University of California (UC), Davis and
UC Berkeley for statistical services.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 1
December 19, 2011

Inmate Classification Score System Study

The major findings from this study are as follows:
1) There are no natural “breaks” in preliminary classification scores that indicate sharp
changes in inmate behavior across housing levels, though the likelihood of behavioral
infractions increases with preliminary score.
2) Mandatory Minimum scores appear to “trap” many well-behaving inmates into higher
housing levels. Inmates crowded above the classification score cut points due to the
Mandatory Minimum scores are relatively well behaved. This better behavior is
explained entirely by age and the lower average preliminary scores of these inmates.
In other words, age and the preliminary classification score provide a better predictor
of behavior for those “trapped” at a specific placement classification score than does
the actual placement classification score determined by binding mandatory
minimums.
3) There is little evidence that housing inmates with preliminary scores slightly above the
Classification Score Level I/II, Level II/III, and Level III/IV thresholds suppresses
institutional misconduct. Furthermore, there is evidence of a criminogenic effect for
inmates who have classification scores just above the Classification Score Level III/IV
threshold who are placed in Level IV housing.
4) There are few escapes, particularly in institutions with electric fences. The risk of
inmate escapes from facilities with electrified fences is nearly zero.
Based on these findings, the Expert Panel developed the following recommendations:
1) Decisions to move inmates into lower housing levels should be guided by the safety
risks those inmates pose to other inmates, staff, and the public. Estimates of risk
should be grounded in the preliminary classification score and should not be
overridden by CDCR Mandatory Minimum factors. Older inmates could also be given
priority in downward housing placements.
2) Inmates with preliminary and placement scores at the threshold (or classification score
cut points) of each housing level can be moved to lower levels with the expectation
that it will not lead to increases in individual or overall rates of serious misconduct
within levels.
3) CDCR should not use Custody Designation as a proxy for the risk of inmate
misconduct. The custody classification system was not designed for this purpose and
does not capture meaningful dimensions of an inmate’s likelihood of bad behavior.
Downward movements in custody should be based upon preliminary classification
score.
4) Moreover, Custody Designations may no longer be justified as a mechanism to reduce
the likelihood of escape. CDCR should consider removing the use of Custody
Designation as markers for escape risks.
Changes to current policy need to be monitored. The Expert Panel advocates the use of
random assignment as the best way to determine the impact. If that is not possible, quasiexperimental methods could provide some evidence of impact, although not as conclusive.
Monitoring will require more extensive data collection on specifics of timing, location and the
nature of violations than is currently collected in automated systems.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 2
December 19, 2011

Inmate Classification System Study

THE INMATE CLASSIFICATION SCORE SYSTEM STUDY
Introduction
The California Department of Corrections and Rehabilitation (CDCR) uses an Inmate
Classification Score System (ICSS) to ensure that inmates are properly housed and
supervised. The ICSS is based on several variables, including inmate social factors and
history of incarceration behavior. This system produces four levels of scores that correspond
to four institutional housing security levels. A preliminary classification score predicts risk for
institutional misconduct using several variables related to an inmate’s background and prior
incarceration behavior. Additional Mandatory Minimum scores are applied to inmates who
have case factors that require they be housed no lower than a designated housing security
level (e.g., convictions for certain violent or sex crimes, public notoriety or life sentences).
The final classification placement score is either the preliminary or the Mandatory Minimum
score, whichever is higher. While classification scores determine the institution and housing
level in which an inmate will be placed, Custody Designations further determine the amount
of supervision an inmate will receive once he is placed in an institution.
The purpose of this study is to evaluate the ICSS to assist CDCR in best identifying factors
that justify restrictions on liberty while avoiding factors that could lead to unwarranted
impingements on inmate rehabilitation.

History of the CDCR Inmate Classification Score System (ICSS)
California was the first state in the country to develop a standardized prison inmate
classification system 1. The initial classification score system implemented by CDCR in the
early 1980s relied on a consensus of opinion rather than on empirical evidence. Since that
time, the CDCR Inmate Classification Score System (ICSS) has evolved based on periodic
validation studies designed to improve the association between classification scores and
institutional misconduct.
The CDCR ICSS was first validated in the mid-1980s. The most recent validation study was
conducted in 1997 by Dr. Richard Berk, University of California, Los Angeles (UCLA). As a
result of the 1997 UCLA study, CDCR implemented a pilot study to evaluate the variables
used to calculate initial classification scores. Based on the pilot study findings, several
changes were made to the scoring practices used with inmates at their classifications. Two
of these changes involved the development and application of Mandatory Minimum Scores
to prevent institutional misconduct and refinement of Close Custody Designations to restrict
particular inmates from having opportunities and from escaping.
In 2000, CDCR implemented the more stringent Close Custody regulations, which dictate
the degree of personal supervision inmates require. In 2002, CDCR implemented an
updated ICSS, consistent with the 1997 UCLA research, which added Mandatory Minimum
Placement Scores to systematize the administrative overrides that were prevalent in the
prior ICSS. These Mandatory Minimum scores, which were not assessed in terms of their

1

Austin, J. & Hardyman, P. Objective Prison Classification: a Guide for Correctional Agencies.
(2004).

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 3
December 19, 2011

Inmate Classification System Study

ability to predict institutional misconduct, applied to specific classes of inmates [e.g., Life
Without the Possibility of Parole (LWOP)], ensuring they would never be housed lower than
a designated level. The net result of the changes in the Close Custody regulations and the
introduction of Mandatory Minimum scores led to an increased need for celled housing.

Current ICSS Design
The current CDCR ICSS still uses calculated scores, referred to as preliminary classification
scores, and Mandatory Minimum scores. The preliminary classification score, validated by
Berk (1997), predicts risk for institutional misconduct and is based on several variables
related to an inmate’s background and prior incarceration behavior. Mandatory Minimum
scores are then applied to restrict the housing levels of particular inmates who are
considered to be threats to staff and other inmates. Final classification scores are either
based on the preliminary or the Mandatory Minimum score, whichever is higher. This final
score is referred to as the placement classification score. Close Custody Designations are
then added, where applicable, to restrict inmate movement throughout the facilities to
prevent escape and to limit the threat to communities if an escape occurs.
An inmate’s classification score will fall into one of four ranges corresponding to the four
institution housing and security levels; greater scores equating to greater security. A score
of 0-18 is Level I, 19-27 is Level II, 28-51 is Level III, and 52+ is Level IV. Each institution is
assigned a housing level based on physical construction. There are six Custody
Designations used in general population housing settings: Close A, Close B, Medium A,
Medium B, Minimum A, and Minimum B. Custody is assigned to denote the level of
supervision the inmate requires within the institution with greater supervision at the higher
custody levels. Close custody inmates require direct and constant staff supervision while
minimum custody inmates may work in the community with little staff supervision.
The final component to the ICSS is the Administrative Determinants, which are conditions
that allow for alternative placement based upon specific case factors that may result in
overrides that are inconsistent with calculated classification scores. An inmate with one or
more Administrative Determinants may be housed in an institution with a housing and
custody level that is inconsistent with his placement score.
More detailed information about the inmate classification process is located in Appendix A.
An overview of the CDCR ICSS procedures is located in Appendix B. The associated
classification scoring forms that support these ICSS processes and procedures are in
Appendices C and D.

Major Project Goals and Research Questions of Study
The major goals of this study are to investigate the possibility of adjusting upward the cut
points between the four classification score levels, as well as potentially modifying the
Mandatory Minimum scores and Close Custody Designations to ensure that the policies for
determining security needs are based on empirical support.
While Administrative
Determinants are used as variables in some of the analyses, they are not in and of
themselves a focus of the study.
Research questions regarding classification scores include whether there are any “tipping
points,” or particular scores above or below the cut points, at which there are noticeable
differences in institutional misconduct. These tipping points might enable classification
scores to be adjusted upward to allow more inmates to be housed in less secure housing, if
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 4
December 19, 2011

Inmate Classification System Study

appropriate. Mandatory Minimum scores are also studied to determine if they contribute any
value in predicting institutional misconduct.
The Close Custody Designation research questions address potential reductions in time an
inmate must be closely supervised. If this time could be reduced, inmates could more
quickly move from celled to dormitory housing. Close Custody Designation criteria used to
prevent escapes are also studied.
The research questions, along with the answers to each question, are presented in
Appendix K, entitled “CDCR Inmate Classification Score System Study Crosswalk.”

Project Methodology
Classification Study Team
Expert Panel
The CDCR Office of Research (OR) determined that the best strategy for addressing the
research questions was to have CDCR staff work with outside correctional experts and
statisticians. Accordingly, an “Expert Panel” was created, comprised of academicians with
experience in studying correctional issues. The panel members included:
David Farabee, Ph.D., University of California, Los Angeles
Ryken Grattet, Ph.D., University of California, Davis
Richard McCleary, Ph.D., University of California, Irvine
Steven Raphael, Ph.D., University of California, Berkeley
Susan Turner, Ph.D., University of California, Irvine
Statisticians
To assist in conducting the statistical analyses the department contracted with the University
of California, Davis (UC Davis) for the services of the UC Davis Statistics Department. The
OR also executed a no-cost Memorandum of Understanding (MOU) with the University of
California, Berkeley (UC Berkeley), for the services of a graduate student who worked
directly under the close supervision of one of the Expert Panel members, Dr. Raphael.
CDCR Classification Services Unit and Office of Research Representatives
The CDCR Classification Services Unit (CSU) and OR worked collaboratively on all aspects
of this study, including the development of the research questions for this study, and
coordination of the introduction and background materials, prison tours and all meetings
associated with the project.
Prison Housing Design Tours
In order for Classification Study Team members to understand the different types of prison
housing that underlie this study, the first meeting was held on September 14, 2010, at
California State Prison, Solano. Team members toured 270 design housing units and
observed institutional classification committees at work (see Appendix B for a description of
the institutional classification committees).
A second meeting was held at California State Prison, Sacramento, on October 6, 2010,
during which team members toured 180 design housing units.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 5
December 19, 2011

Inmate Classification System Study

Classification Study Team Meetings
The prison housing design tours were followed by a series of conference calls in November
and December 2010, during which the study design was discussed.
On January 14, 2011, the Classification Study Team convened in person at CDCR
Headquarters in Sacramento. Decisions were made on the study design. CDCR Secretary
Matthew Cate attended the meeting and discussed the importance of this study with the
Expert Panel members.
From April through November 2011, the Classification Study Team held weekly conference
calls to identify data sources/elements, troubleshoot issues/concerns, review preliminary
findings and monitor the overall progress of the project.

Expert Panel Data Analysis Plan
The Expert Panel drafted a Data Analysis Plan that identified and guided the analyses
performed to address the research questions. Although the preference of the Expert Panel
was to address the research questions using a randomized experiment, this method is not
always feasible. Not only are there time and resource constraints due to the court-ordered
mandates and departmental reduction efforts, but the safety-compromising implications
associated with the potential for error in correctional settings often does not allow for
experimental research. With this in mind, the Expert Panel opted for a quasi-experimental
study design to evaluate the current ICSS for potential changes that could be made using
existing CDCR administrative datasets, acknowledging that a future research
recommendation might be to address data gaps with experimental methodologies. Female
offenders were not examined for this study since the research design relied on the
delineation between the housing Levels I-IV, which are only applicable to male offenders.
Once drafted, input regarding the Data Analysis Plan was sought from a correctional
classification system subject matter expert from the Association of State Correctional
Administrators (ASCA). This feedback was incorporated into the final Data Analysis Plan,
where possible. Completed in March 2011, the final Data Analysis Plan is found in
Appendix E.
The analytical strategies for the classification score and Close Custody Designation
research questions are as follows:
Classification Scores
The Expert Panel sought to address two broad questions regarding the classification
scores. One question is, “does the preliminary score predict the behavior of inmates
whose placement scores are constrained by the Mandatory Minimum scores?” Another
is, “do inmates with large differences between their preliminary and placement scores
behave better than individuals with small differences between their preliminary and
placement scores?” If so, then individuals with large differences could be considered
better candidates for moving to lower levels than individuals with small differences
between preliminary and placement scores. If there is no difference, then moving
individuals with Mandatory Minimums based upon their preliminary classification score
would not be advisable since such inmates would not be expected to behave differently
than inmates with the same placement classification score but a higher preliminary
classification score. The Expert Panel referred to this study design as a “Gap Analysis.”

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 6
December 19, 2011

Inmate Classification System Study

Two Gap Analyses are used to explore the preliminary scores of inmates who have
similar placement scores due to the Mandatory Minimums in order to determine whether
or not actual differences exist in their propensity to engage in institutional misconduct
(also referred to as Rules Violation Reports, or “RVRs”). One uses the difference
between the preliminary and placement classification scores to predict misconduct. A
follow-up analysis examines each particular Mandatory Minimum criterion (e.g., LWOP,
Lifer) in order to see if particular criteria indicate elevated or lowered risk of violation.
Furthermore, a Matching Analysis compares inmates who have similar backgrounds,
with the exception that some have a Mandatory Minimum, to see if differences exist in
their propensity to engage in institutional misconduct (i.e., for a group of similar inmates,
do those with Mandatory Minimum scores behave better or worse than those who do
not).
Both the Gap and Matching Analyses model the likelihood an inmate has an RVR (using
logistic regression models), as well as an inmate’s count of RVRs during the review
period (using Poisson regression models). Separate models are run to isolate results for
A-F violations, A-D violations and A violations (violation types range from A through F,
with A being the most serious violation type and F being the least serious). Additional
statistical methods are employed to account for other relevant factors, such as the fact
that inmates could have multiple reviews over time and that there are varying lengths of
time between classification score reviews.
The second question posed by the Expert Panel is, “would increasing the cut points at
the thresholds separating the levels increase misconduct?” This is addressed through
the application of a Regression Discontinuity (RD) design analysis. Specifically, the
Expert Panel hypothesized that two factors may lead to an impact of security on the
likelihood of a behavioral infraction: a suppression effect (tighter restrictions may
suppress behavioral problems) and a peer effect (associating with poorly behaving
inmates in higher security levels may result in increases in institutional misconduct for
inmates who might otherwise behave better if placed in a lower security level). The RD
analysis serves as an empirical approach to observe the net of these two effects.
Analyses are performed to examine discontinuities at each of the classification score cut
points in relation to RVRs. Multiple models examine both preliminary and placement
scores in relation to: 1) inmate housing level crossing into the review period and 2)
inmate housing level after the reclassification hearing at the end of the review period
(statistically adjusting the model based on an estimate of the amount of time inmates
spent in the housing level endorsed by the classification review). Further analyses
isolate results for any violations, A violations, B/C/D violations and E/F violations.
Collectively, the results from these analyses serve to triangulate around the classification
score research questions.
Close Custody Designations
Since the Close Custody Designation research questions address both the impact the
designations have on escape risk and risk for institutional misconduct, different
methodological approaches were used.
These methodologies include:
CDCR
Successful Escape Reports; manual file reviews; literature reviews; regression, Matching
and Gap Analyses.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 7
December 19, 2011

Inmate Classification System Study

Data Sources
Internal meetings were held between CDCR Classification Services Unit and CDCR Office
of Research staff to identify appropriate and relevant data sources to support the study.
After careful consideration, the following databases were used to compile the datasets for
the statisticians.
Offender Based Information System (OBIS)
OBIS maintains offender demographic, sentencing and parole revocation data, location
where the offender is serving his sentence, length of sentence and time served,
information on holds, warrants or detainers, and DNA collection information. OBIS data
are used in the study to identify inmates’ locations and length of sentences.
Inmate Classification Score System (ICSS)
Information from paper-format classification sheets are entered into the ICSS. This
system provides statistical reports and is used for quality control of the manual calculation
process. ICSS data are used in this study to identify components of classification scores,
including mandatory minimums, points added due to misconduct, and points subtracted
due to good behavior.
Distributed Data Processing System (DDPS)
DDPS contains information on inmates’ housing location within an institution,
classification level; job assignment; trust account activity; restitution collection; visiting
activity; medical, mental health, and developmental disability identification; and canteen
activity. These data are used to manage inmates locally, but are also sent to CDCR
headquarters for system-wide management.
DDPS data are used in this study to
identify prison and housing locations, and medical, mental health or developmental
disability status.
Data Quality
The Offender Information Services Branch (OISB) is CDCR’s primary provider of summary
statistical information about inmates and parolees. It is also responsible for coordinating the
timely, accurate, and consistent coding and entry of data, including classification score data.
OISB staff perform data integrity and quality control functions for OBIS. Within OISB, the
Classification Quality Assurance Unit (CQAU) is responsible for managing the paper-based
process that captures data from the classification scoring sheets of the ICSS, the CDC
Forms 839, 840 and 841. 2
Working Datasets
Five datasets were produced by CDCR OR staff to support this study.
One is a
classification score dataset that includes male felons (non-death row) who were admitted to
CDCR prior to June 30, 2009, had a reclassification review with a review period beginning
date that occurred on or after July 1, 2008, and who had served at least one consecutive

2

Documentation regarding the limitations of the CDCR ICSS data are reported in the Office of
Research Data Evaluation and Recommendation (ORDER) Project, Final Version 1.2. (2008).
Estrada Consulting Inc.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 8
December 19, 2011

Inmate Classification System Study

year in prison (records with inaccurate or missing dates were omitted). The remaining four
datasets were created to address the Close Custody Designation questions pertaining to
four sentence types: those sentenced to 15 to 50 years, 50+ years, life and multiple life.
Each of these longitudinal data files includes classification data collected since November
1998 for all inmates in Level I-IV housing on June 30, 2011.
The primary variables compiled in these datasets are the number and type of RVRs,
preliminary classification scores, Mandatory Minimum classification scores, respective
housing levels, and demographic characteristics (e.g., criminal offense, age). The complete
table of variables is in Appendix G.

Study Findings
The Expert Panel’s Data Analysis Plan lays out the strategies for conducting the statistical
analyses. CDCR OR staff performed the basic descriptive analyses. The UC Davis and UC
Berkeley statisticians performed the more complex RD, Gap, Matching and longitudinal
analyses. Specific responsibilities for the Classification Study Team members are located in
the “Data Analysis and Reporting Project Plan” (Appendix H). Below is a brief description of
the study cohort, followed by the results from the statistical analyses for the classification
score and Close Custody Designation research questions. Detailed documentation of the
RD analyses, entitled “Summary of Findings from the Regression-Discontinuity Analysis of
Inmate Behavioral Outcomes,” may be found in Appendix I. Detailed documentation of the
Gap and Matching Analyses may be found in Appendix J, entitled “Statistical Methods and
Summarized Results for the Gap, Matching, and Longitudinal Analyses.” A crosswalk
between the research questions, the strategies to address each of the research questions
as outlined in the Data Analysis Plan, and the findings for each of the research questions is
presented in Appendix K, entitled “CDCR Inmate Classification Score System Study
Crosswalk.”
Description of Sample
Appendix L provides the demographic characteristics of the classification study cohort,
which represent 98,355 reclassification score reviews. As mentioned above, this study
focused solely on males as there are no housing data (Level I-IV) available for females.
Nearly forty percent of those were for inmates in Level III housing. Just over 25 percent
were in Level IV housing and just fewer than 25 percent were in Level II housing. Only
6.1 percent were in Level I housing. The oldest inmates (average age of 43.5 years) were
housed in Level II and the youngest (average age of 36.0 years) were in Level IV housing.
Race/ethnicity categories were fairly evenly distributed between the four housing levels.
Hispanic inmates made up the largest percentage of inmates in Levels II – IV, while
Black/African American inmates made up the largest percentage of inmates in Level I. Over
75 percent of the inmates in Levels II – IV had a crime against a person as their commitment
offense. Nearly 30 percent of Level I inmates were committed for property, drug, and
crimes against a person. The highest number of those required to register as a sex offender
were in Level III (8,977); however, Level II had the highest proportion (26.5 percent). Most
inmates designated as CCCMS or EOP were housed in Levels III and IV. In Level III, nearly
27 percent of inmates had one of these mental health designations, while just over
32 percent were designated so in Level IV. In Level IV, nearly 99 percent of the inmates had
ever been committed for a serious and/or violent offense while just under 76 percent of the
inmates housed in Level I had ever been committed for serious and/or violent offenses.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 9
December 19, 2011

Inmate Classification System Study

Level II inmates had the lowest proportion of inmate reclassification score reviews where a
serious disciplinary violation occurred (16.1 percent). Levels I and IV had the highest
proportion of inmates with serious disciplinary violations (30.5 percent and 30.0 percent
respectively). However, the majority of the Level IV violations were of the more serious
variety (A1, A2, B, C, D), while the vast majority of the Level I violations were of the least
serious variety (E, F). Looking at the relationship between an inmate’s preliminary score and
placement score (after Mandatory Minimums are factored in) we see that nearly 66 percent
of the inmates with a preliminary score that fall within Level I were elevated to Level II due to
their Mandatory Minimum. Another 7 percent were elevated to Level III and 3.4 percent
were elevated to Level IV. For inmates with a preliminary score that falls within level II,
3.7 percent had a Mandatory Minimum that placed them in Level III and nearly 2 percent
had one that placed them in Level IV. For inmates with a preliminary score in Level III,
2.6 percent had a Mandatory Minimum that placed them in Level IV. With respect to
administrative overrides, 30.1 percent of inmates with a Level I placement score were
actually housed in Level II. Nearly 42 percent of the inmates with a Level II placement score
were housed in Level III and only 12.7 percent of the inmates with a Level III placement
score were housed in Level IV.
Classification Scores
The strongest findings from this study resulted from the examination of the classification
scores. The results from the analyses performed to address the classification score
research questions provide additional support for the work performed by Berk (1997) on the
preliminary classification score in that, of the routinely captured administrative data within
the CDCR, it is the best predictor of institutional misconduct.
The primary classification score research questions asked if there are any natural “tipping
points” or particular scores that indicate distinct increases in RVRs along the continuum of
classification scores, and further questioned if cut points could be adjusted upward to allow
inmates to move to lower housing levels without compromising institutional and public
safety. Statistical analyses reveal that there are no “tipping points” and that there is a
positive correlation between preliminary classification scores and RVRs, i.e., as preliminary
scores increase so do RVRs (see Appendix I, pp. 47-48). Based on this information,
classification score ranges could modestly be adjusted upward to include more inmates in
the lower levels without serious compromise to institutional safety or public safety.
There is little evidence of a suppression effect at each of the current classification score cut
points (i.e., 19, 28, and 52). Housing inmates who are just above the cut points (or
threshold) in the respective higher housing level does not suppress their institutional
misconduct. There is evidence of a criminogenic effect for inmates who have classification
scores just above the Classification Score Level III/IV threshold who are placed in Level IV
housing. These inmates have worse behavioral outcomes than those who have scores just
below
the
threshold
and
are
placed
in
Level
III
housing
(see
Appendix I, pp. 48-51). Thus, it appears that modest adjustments may be made to the
current classification score cut points without jeopardizing institutional or public safety. In
fact, adjusting the Level III/IV cut point could be considered a proactive measure toward
preventing institutional misconduct.
Furthermore, the placement score (determined by the Mandatory Minimum factors) houses
well-behaved inmates in higher levels of security than is necessary relative to in-custody
behavior. Inmates placed in higher housing levels due to Mandatory Minimum scores pose
less risk of engaging in institutional misconduct than individuals who are placed in a higher
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 10
December 19, 2011

Inmate Classification System Study

level determined by their preliminary score. Therefore, it is optimal to determine the
appropriate housing level for inmates based on their preliminary classification score (see
Appendix I, pp. 52-54).
Additional classification score research questions asked if housing level restrictions may be
lowered for two particular Mandatory Minimum scores: Mandatory Minimum score “B” (Life
Without Parole) and Mandatory Minimum score “C” [CCR 3375.2 (a)(7), Life Inmate
(multiple/execution style murders; escapes)].
The research question pertaining to
Mandatory Minimum Score “B” also calls for investigation into the type of housing design
that may be allowed (180 versus 270 housing design). Since data for 270 housing design
are not captured electronically, the analyses performed do not directly address the impact of
housing design on institutional misconduct.
Analyses of the specific reasons for the Mandatory Minimum show that inmates with “B”
(LWOP) and “C” (CCR 3375.2, Lifer) designations engage in less institutional misconduct
than other inmates who do not have a Mandatory Minimum score (see Appendix J, pp. 9395). Again, among the variables considered in these analyses, the preliminary score was
found to be a better predictor of institutional misconduct, particularly when compared to the
placement score. However, these analyses should be viewed with some caution as the
sample is restricted to inmates who have placement scores near housing level thresholds
and may not capture all important relationships and variables which may affect RVRs.
Close Custody Designations
Since the original purpose for the Close Custody Designations is to restrict inmate access
around institutions to prevent escape, it is logical that any statistical analyses for this study
be performed on escape data. However, since few inmates escape from Level II through
Level IV institutions with electric fences, it is not possible to model the impact that changes
in Close Custody Designations may have on inmate escapes.
Because of this issue, the Expert Panel relied on existing CDCR escape reports to address
research questions about escapes. In particular, the Expert Panel examined frequency
counts of successful escapes that occurred from 1999 to 2010 for each CDCR institution
within the various housing levels, focusing on Levels II-IV since these levels are protected
by an electric fence at most CDCR institutions (Appendix F). To supplement this
information, CSU staff performed manual case reviews of successful escapes by inmates
assigned to Level II through Level IV housing. They found that virtually all escapes were
from settings outside of Level II through Level IV institutions 3. Such settings include
reception centers, being out to court, and being out on medical leave, all of which are
irrelevant to this study.
Collectively, this information demonstrates that there is a near-zero escape probability for
inmates who are housed behind electric fences, thus resulting in the finding that the Close
Custody Designations are unnecessary for preventing escapes.

3

Between 1999 and 2010 there were 18 reported escapes of level II-IV inmates. Over that span of
time the annual Average Daily Population of level II-IV inmates ranged from a low of 127,600 to a
high of 138,670. However, only two of the 18 inmates (both Level II) actually escaped from behind
the walls of a CDCR institution.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 11
December 19, 2011

Inmate Classification System Study

Over time, the Close Custody Designations have evolved to serve as yet another tool to
supervise inmates who engage in institutional misconduct. As such, several of the Close
Custody Designation research questions focus on their relationship to institutional
misconduct rather than escape. Although the Close Custody data available for statistical
modeling is limited by the lack of comparison groups (current CDCR policies dictate that all
inmates who meet certain Close Custody Designation requirements must be supervised at a
designated custody level), the Expert Panel attempted to address these research questions
through a series of exploratory analyses on the Close Custody Designations as they relate
to RVRs.
Analyses of the classification score data, including the longitudinal data, were used to
produce and test multiple statistical models. The results from these various models are
mostly ambiguous and uncertain (Appendices K and L provide documentation of these
inconclusive results). Some analyses reveal that decreasing custody levels appear to
increase RVRs; however, these findings hinge on inmates who are already engaging in
institutional misconduct.
As demonstrated by the classification score analyses, the preliminary score is the best
known predictor for institutional misconduct. Therefore, the benefit of the Close Custody
Designations is proportional to inmate preliminary scores. The higher the preliminary score,
the greater the need for Close Custody. Conversely, the lower the preliminary score, the
less the need for Close Custody. Based on the data used for these analyses, a majority of
the inmates meeting the Close Custody Designation criteria who are not engaging in
institutional misconduct are being closely supervised despite the fact that their preliminary
classification score provides no indication that such supervision is warranted.
After having an opportunity to examine the relationship between Close Custody
Designations and institutional misconduct, the Expert Panel concluded that the existing
administrative data are inadequate and inappropriate for addressing the research questions;
a randomized experiment is necessary. Furthermore, the Expert Panel questioned the
added value of repurposing the Close Custody Designations to address institutional
misconduct since there are already other tools in place to do so (e.g., preliminary
classification score, Administrative Determinants).
Conclusions
The main findings from this study may be summed up as follows:
1) There are no natural “breaks” in preliminary classification scores that indicate sharp
changes in inmate behavior across housing levels, though the likelihood of behavioral
infractions increases with preliminary score.
2) Mandatory Minimum scores appear to “trap” many well-behaving inmates into higher
housing levels. Inmates crowded above the classification score cut points due to the
Mandatory Minimum scores are relatively well behaved. This better behavior is
explained entirely by age and the lower average preliminary scores of these inmates.
In other words, age and the preliminary classification score provide a better predictor
of behavior for those “trapped” at a specific placement classification score than does
the actual placement classification score determined by binding mandatory
minimums.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 12
December 19, 2011

Inmate Classification System Study

3) There is little evidence that housing inmates with preliminary scores slightly above the
Classification Score Level I/II, Level II/III, and Level III/IV thresholds suppresses
institutional misconduct. Furthermore, there is evidence of a criminogenic effect for
inmates who have classification scores just above the Classification Score Level III/IV
threshold who are placed in Level IV housing.
4) There are few escapes, particularly in institutions with electric fences. The risk of
inmate escapes from facilities with electrified fences is nearly zero.

Expert Panel Recommendations
The Expert Panel developed four recommendations based on the study findings:
1) Decisions to move inmates into lower housing levels should be guided by the safety
risks those inmates pose to other inmates, staff, and the public. Estimates of risk
should be grounded in the preliminary classification score and should not be
overridden by CDCR Mandatory Minimum factors. Older inmates could also be given
priority in downward housing placements.
2) Inmates with preliminary and placement scores at the threshold (or classification
score cut points) of each housing level can be moved to lower levels with the
expectation that it will not lead to increases in individual or overall rates of serious
misconduct within levels.
3) CDCR should not use Custody Designation as a proxy for the risk of inmate
misconduct. The custody classification system was not designed for this purpose and
does not capture meaningful dimensions of an inmate’s likelihood of bad behavior.
Downward movements in custody should be based upon preliminary classification
score.
4) Moreover, Custody Designations may no longer be justified as a mechanism to
reduce the likelihood of escape. CDCR should consider removing the use of Custody
Designation as markers for escape risks.

Supplementary Materials:
Misclassification and Other States’ Current
Strategies for Reducing Inmate Population
As a supplement to this study, CDCR Executive staff requested a literature review of factors
related to misclassification, as well as a review of measures other states are taking to
reduce their inmate population to address overcrowding (Appendix N). Despite the fact that
the research is somewhat limited, it is clear that misclassification is a serious problem.
Although underclassification can be a problem, it appears to be relatively insignificant
compared to the repercussions of overclassification. In particular, some studies have
demonstrated that inmates who are overclassified may learn new criminal behaviors through
interaction with more experienced criminals. As mentioned above, this finding is also
evident in this classification study as inmates housed in Level IV institutions were found to
have worse behavioral outcomes than inmates who are in Level III housing.

Suggestions for Future Research
It is the contention of the Expert Panel that changes to current policy need to be monitored
and members collectively advocate the use of an experiment as the best way to determine
the impact. If that is not possible, members assert that quasi-experimental methods could
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 13
December 19, 2011

Inmate Classification System Study

provide some evidence of impact, although not as conclusive. Monitoring will require more
extensive data collection than is currently collected in automated systems (e.g., timing,
location, and nature of RVR).
In addition to monitoring system changes that result from the findings of this study, the
Expert Panel offered the following suggestions for future study:
Preliminary Classification Score
Two suggestions are related to the preliminary score. One is to further examine the
importance of age in the calculation of the preliminary classification scores since this
study resulted in a hypothesis that age may be underweighted in the final score.
Another is a revalidation study, which would investigate whether a new classification
system might perform better, especially given the expected nature of the inmate
population changes due to realignment.
Close Custody Designations
Further research could be pursued to assess the true value of the Close Custody
Designations in predicting institutional misconduct (i.e., do the Close Custody
Designations provide additional value that is over and above the predictive power of the
preliminary classification scores).
Administrative Determinants
Additionally, the use and predictive power of Administrative Determinants could be
examined. Descriptive statistics show that administrative overrides place a substantial
percentage of inmates in housing levels higher than placement scores dictate. The
question to be addressed is similar to the previous question: Do Administrative
Determinants provide additional value that is over and above the predictive power of the
preliminary score?
Cost-Benefit Study
Some form of cost-benefit analyses could be conducted for randomized experiments to
determine not only the impact of RVRs, but also to identify the costs associated with
proposed changes.
Randomized designs could be used to answer questions related to moving certain groups to
lower housing levels. To answer the question about whether inmates with placement scores
just above the thresholds could be safely moved down, inmates with placement scores
several points above the threshold could be randomly assigned to either remain in the
higher housing level or moved one level lower. To determine whether celled or dorm
housing has an independent effect on RVRs, a study could focus on inmates near the II/III
threshold and randomly assign into cell or dorm housing, within level II or level III. A third
randomized design might consider different age groups and randomly assign age cohorts to
different housing levels.
Key to all randomized designs is assuring comparable inmates are in the different study
groups, as well as carefully measuring behaviors using methods that may be more sensitive
than current databases containing RVRs, and following all inmates for at least a period of
one year in order to gauge the impact of the different treatments.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 14
December 19, 2011

Inmate Classification System Study

Appendix A
Overview of the CDCR Inmate Classification Process
The current CDCR inmate classification process has three components: the Inmate
Classification Score System (ICSS), Custody Designations and Administrative
Determinants. Although these factors are used to classify both male and female inmates,
female inmates are housed in institutions irrespective of their classification score and
Custody Designations.
Classification Scores 4
When inmates are first received by CDCR, their classification process begins by establishing
a Preliminary Classification Score using a point system based on the following factors:
-

Age at first arrest
Age at reception by CDCR
Current term of incarceration
Street gang/disruptive group
Prior incarceration(s) (including juvenile, federal, other states)
Prior incarceration behavior
o Favorable prior behavior
o Unfavorable prior behavior

In addition, inmates with certain case factors will receive a Mandatory Minimum Score that
requires he be housed no lower than a specific security level. The Mandatory Minimum
Score was instituted to minimize unnecessary security level overrides. Mandatory Minimum
Scores are given for the reasons shown in Table 1, below.
Table 1. Reasons for Mandatory Minimum Score
Reason
Condemned
Life without possibility of parole
CCR 3375.2 (a)(7) Life inmate (multiple/execution style murders; escapes)
History of escape
Warrants “R” Suffix (sex crimes)
Violence exclusion
Public interest case
Other life sentence

Score
52
52
28
19
19
19
19
19

The final classification placement score is either the preliminary or the mandatory minimum
score, whichever is highest. Currently there are four ranges of classification scores and four
levels of institutions, which are reflected in Table 2, below.: 5

4

CCR Title 15. Crime Prevention and Corrections, Section 3375.3. CDC Classification Score Sheet,
CDC Form 839, Calculation.
5
CCR Title 15. Crime Prevention and Corrections, Section 3375.1. Inmate Placement and Section
3377. Facility Security Levels.
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 15
December 19, 2011

Inmate Classification System Study

Table 2. Institutional Security Levels
Level
Points

Institutional Design

I

0 – 18

Open dormitory facilities and camps and low
security perimeter.

II

19 – 27

Open dormitories with a secure perimeter, which
may include armed coverage.

III

28 - 51

Housing units with cells adjacent to exterior walls
and secure
Perimeter with armed coverage.

IV

52+

Housing units or cell block housing with cells nonadjacent to exterior walls and secure perimeter
with internal and external armed coverage.

The inmate’s preliminary score continues to be computed to the absolute of zero (0), below
which there is no computation. It is one source used to determine in what level prison a
male inmate will serve his sentence. An inmate’s placement score is the primary
determining factor for the institution in which he will serve his sentence. Changes in a
placement score may cause an inmate to be transferred to a different level institution.
Custody Designations 6
In addition to a classification score, an inmate receives a Custody Designation which
determines the level of supervision he will receive once he is in the institution. It may also
impact the jobs or programs to which he may be assigned. The purpose of the Custody
Designation is to determine supervision control levels based upon problematic behavior or
an individual’s potential for escape and threat to the community if an escape occurs. The
designation is primarily based on the following factors, although other reasons may be
considered:
-

6

The inmate’s total term, sentence, or remaining time-to serve
The inmate’s escape history
Receipt of an active law enforcement felony hold
An inmate who is considered to be High Notoriety or is designated as a Public
Interest Case
Identification of a management concern
A finding of guilt for a serious, felony level, Rules Violation Report (RVR)

CCR Title 15. Crime Prevention and Corrections, Section 3377.1. Inmate Custody Designations.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 16
December 19, 2011

Inmate Classification System Study

Table 3. Custody Designations
Custody
Designation

Close A

Housing
Type
Cells within Level III and
Level IV facilities in
housing units located
within an established
facility security perimeter.

Close B

Housing is in cells within
designated institutions in
housing units located
within an established
facility security perimeter.

Medium A

Housing is in cells or
dormitories within the
facility security perimeter.

Medium B

Housing is in cells or
dormitories within the
facility security perimeter.

Minimum A

Housing is in cells or
dormitories within the
facility security perimeter.

Level of Supervision
Required

Custody staff supervision is
required to be direct and
constant.

Custody staff provide direct
and constant supervision at
all times. The work
supervisor is required to
provide direct and constant
supervision during the
inmates’ assigned work
hours.

Custody and/or work
supervisor supervision is
frequent and direct.
Custody staff provide
frequent and direct
supervision inside the facility
security perimeter and direct
and constant supervision
outside the facility security
perimeter.
Staff supervision consists of
at least hourly observation if
assigned outside the facility
security perimeter. Sufficient
staff supervision of the
inmate shall be provided to
ensure the inmate is present
if assigned inside the facility
security perimeter.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Assignment and Activity
Conditions
Program assignments and
activities are only scheduled within
the hours of 6:00 a.m. to 6:00 p.m.
unless hours are extended by the
Warden to no later than 8:00 p.m.
when it is determined that visibility
is not compromised within the
facility security perimeter.
Assignments and activities are
only scheduled within the hours of
6:00 a.m. to 8:00 p.m. in areas
located within the facility security
perimeter, including beyond the
work change area in a designated
Level II, Level III or Level IV
institution. Inmates may
participate in designated work
program assignments until 10:00
p.m. when the work program is in
an assigned housing unit located
within the facility security
perimeter. Inmates may participate
in limited evening activities after
8:00 p.m. until the general evening
lockup and count when the limited
activity is in a designated housing
unit located within the facility
security perimeter.
Assignments and activities are
within the facility security
perimeter.
Assignments and activities are
within the facility security
perimeter. Inmates may be given
daytime assignments outside the
facility security perimeter but must
remain on facility grounds.

Assignments and activities may be
inside or outside the facility
security perimeter.

Page 17
December 19, 2011

Inmate Classification System Study

Minimum B

Housing may be in cells or
dormitories on facility
grounds,
in a camp, in a Minimum
Support Facility (MSF) or
in a
community based facility
such as a Community
Correctional
Facility (CCF).

Sufficient staff supervision of
the inmate shall be provided
to ensure the inmate is
present.

Assignments and activities may
be inside or outside the facility
security perimeter.

Administrative Determinants 7
The third component is Administrative Determinants. These are conditions that allow for
alternative placement based upon specific case factors that may identify overriding factors
inconsistent with the normal scoring process. For example, an inmate who is eligible for an
identified security level who has medical or mental health needs that can be better provided
at an institution with a different security level, or an inmate who is an active gang member
would not be housed in an institution where he would be in danger from members of an
opposing gang, although his classification score qualifies him to be sent there. An inmate
with one or more Administrative Determinants may be housed in a facility with a security
level which is not consistent with the inmate’s placement score. This component of the
classification system was originally outside of the scope of this study.
Because
Administrative Determinants were found to have an impact on Close Custody and housing
placement, a few major determinants were included as variables in the data analyses.

7

CCR Title 15. Crime Prevention and Corrections, Section 3375.2. Administrative Determinants.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 18
December 19, 2011

Inmate Classification System Study

Appendix B
Overview of the CDCR Inmate Classification Procedures
Initial CDCR Inmate Classification
When an individual is sentenced to prison, he will first go to a Reception Center (RC)
located in a prison. 8 There are currently nine CDCR men’s institutions that have a RC.
Upon arrival, a new inmate receives brief physical and mental health evaluations and an
evaluation for any safety concerns, such as gang activity. In addition, the inmate’s legal
documents are reviewed by a case records analyst and then sent to the Inmate Case
Records Office (Records) for processing. Records staff assemble a Central File (C-File) for
each inmate. New commitments receive new CDCR numbers. Parole Violators (PV) keep
their existing CDCR number and their C-File is ordered from the appropriate Parole Records
office. It can take three to six weeks for the complete assembly and processing of C-File,
depending upon the inmate and the Records staff workload. During his time in the RC, the
inmate receives further in-depth evaluations for physical, mental, dental and education
needs. Documentation of these evaluations is added to the C-File.
After a C-File is created and the evaluations are completed, a Correctional Counselor I
(CC I) will review all documentation and complete a CDC 839, 9 Classification Score Sheet
(Appendix C) for inmates who are new to prison. A CDC 841, Readmission Score Sheet, is
completed if the inmate is a PV who returned to prison with a new term or a PV-Returned to
Custody (RTC). The CCI then completes an Institution Staff Recommendation Summary
(ISRS) or a CDC-816 (Readmission Summary), which suggests appropriate housing options
for the inmate. The ISRS is reviewed by a Supervising Correctional Counselor II (CC II) and
provided to a Classification Staff Representative who will endorse the inmate to the prison
where he will serve his sentence. Ideally the total process time is 30 days from RC arrival to
transfer, but complex case factors, staff shortages and overcrowding in living units can
prolong the process time. In addition to the placement score a number of other factors are
considered when making the final placement decision. These include the inmate’s
preferences, the county of last legal residence, mental health or physical disability needs,
program needs and available bed space.
Endorsement and transfer from the RC usually results in arrival at a General Population
(GP) institution for long term housing. An inmate is assigned to housing based on his
endorsement and bed space availability. He is then assigned to a CCI, who will review the
C-File and prepare for the Initial-Unit Classification Committee (UCC). The UCC is supposed
to occur within 14 days of the inmate’s arrival. At this classification hearing, the inmate’s
custody level, Work Group (which determines his credit-earning status) and Privilege Group
(WG/PG), assignment to a waiting list if indicated, and any visiting restrictions are
established.

8

There is one exception: male inmates who are sentenced to death are sent directly to the
Condemned housing at San Quentin State Prison.
9
Forms continued to be referred to as CDC forms, although the department is now CDCR.
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 19
December 19, 2011

Inmate Classification System Study

Subsequent CDCR Inmate Classification Procedures
Inmates are reclassified at least once annually. They may be reclassified after six months,
whenever case factors indicate a change in endorsement, or when program or treatment
needs require it. The form used is the CDC 840 Reclassification Score Sheet (Appendix D).
There are different levels of classification committees. The first and second level committees
are held within an institution. The highest level committee is held in CDCR headquarters.
The committees’ responsibilities are as follows:
•

The Unit Classification Committee (UCC) conducts the majority of hearings dealing
with routine matters of program or housing assignments. It is chaired by a Facility
Captain or his or her designee.

•

The Institution Classification Committee (ICC) considers cases referred by a UCC
when an inmate must be placed in Administrative Segregation or security housing.
This committee is chaired by the Warden or his or her designee.

•

The Departmental Review Board (DRB) considers cases referred by ICCs to resolve
an ICC difference of opinion, transfer of an inmate to a Federal or other state system,
Meritorious Time Reduction cases, or any other case that is unusually complex. The
DRB represents the CDCR Secretary and is comprised of various executive staff in
CDCR headquarters.

Housing
An inmate’s classification determines the type of housing in which he will be placed.
Level I or II inmates may be housed in open dormitory settings. Level III and IV inmates are
placed in 180 degree or 270 celled housing units. The number of degrees refers to view
from a central elevated control booth. The “180-degree” design is a configuration of the
cellblocks (housing units). The cellblocks are partitioned into three separate, self-contained
sections, forming a half circle (180 degrees). The partitioning of sections, blocks, and
facilities ensures maximum control of movement and quick isolation of disruptive incidents,
thereby ensuring effective overall management of inmates.
In addition to open dormitories and cell units there are the following special housing units:
•

Security Housing Unit (SHU): the most secure area within a Level IV prison designed
to provide maximum coverage. These are designed to house inmates that cannot be
housed with the general population of inmates. This includes inmates that are
validated prison gang members or gang leaders. SHU terms can vary in length.

•

Administrative Segregation (ASU): similar in design to a SHU, ASU houses inmates
for up to 30 days, or longer with approval from a Classification Staff Representative
(CSR). Inmates are placed into ASU to resolve issues that concern the safety of the
inmate, the safety of others, or jeopardize the security of the institution. ASU may
also house inmates as Disciplinary Detention for up to ten days as a disposition
resulting from a guilty finding on a serious RVR.

•

Reception Center (RC): provides short term housing to process, classify and
evaluate incoming inmates.

•

Condemned (Cond): Holds inmates with death sentences.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 20
December 19, 2011

Inmate Classification System Study

Lethal Electrified Fences (LEF)
Beginning in 1993, CDCR added LEFs to the perimeters of its prisons. All new prisons with
a security classification level of II and above require the installation of an LEF. Today 27
institutions are surrounded by LEFs, which serve as lethal barriers to assist in the prevention
of escapes.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 21
December 19, 2011

Inmate Classification System Study

Appendix C
CDC 839 Classification Score Sheet

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 22
December 19, 2011

Inmate Classification System Study

Appendix D
CDC 840 Reclassification Score Sheet

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 23
December 19, 2011

Inmate Classification System Study

Appendix E
Expert Panel Data Analysis Plan
March 2011
As outlined in the project Blueprint, the major goal of this study is to determine whether the
CDCR classification system may be modified without jeopardizing institutional security or
public safety. The Expert Panel members developed this data analysis plan to guide the
Office of Adult Research and the statistical consultant in performing analyses that will inform
CDCR management in the revision of its classification practices. The proposed research
design developed by the Expert Panel addresses four major questions:
•

What is the nature of the problem?

•

Does an inmate’s preliminary score add to our ability to predict incidents over and
above placement score?

•

What is the estimated effect on violations of moving up cut-points?

•

Do close Custody Designations serve to reduce escape and violation risk?

We note that the strongest method for answering many of these questions would involve an
experimental research design in which inmates are randomly assigned to different housing
or close custody levels and we would record information on subsequent violations. We do
not have such a design. Our analyses are based on analysis of existing data, controlling for
factors that may cloud our ability to determine the actual impacts of housing and custody
levels. We caution that our conclusions must be interpreted in this light.
For each question, we pose the main and secondary questions and outline proposed
analysis plans for each question. We note that when we discuss violations, our analyses
will consider multiple definitions if possible: 1) any violation; 2) any “serious” violation (i.e., AF); 3) any “violent” violations; 4) the number of violations and 5) the number of serious
violations, and 6) the number of “violent” violations. We note that our analyses of the
Blueprint questions related to custody are dependent upon the ability to tie violations to the
level of custody (close, minimum, medium) the inmate was in when the violation was
recorded.

What is the nature of the problem?
This first question addresses the current state of housing, placement and violations.
How is the current population distributed across housing levels in the system and
how does that distribution compare with design capacity? Which housing levels are
over- and which are under-capacity?
Analysis Plan
Analysis is descriptive. Tables/graphs should be constructed to:
•

Present changes over the last 15 years in the distribution of inmates by level, and
custody levels.

•

Illustrate the “impactedness” of all housing levels

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 24
December 19, 2011

Inmate Classification System Study

How does the existing classification process drive housing level? What is the
relationship between preliminary score, placement score, and housing level? How are
inmates sorted into housing levels in ways that conform to and depart from
preliminary and placement scores?
Analysis Plan
Analysis is descriptive. Construct tables that show:
•

Preliminary ScorePlacement Score Housing Level for each range of Preliminary
Scores (L1, L2, L3, L4).

•

Upward moves (mandatory minimums) for Preliminary ScorePlacement Score

•

Show

upward

and

downward

(administrative

overrides)

for

Placement

ScoreHousing Level
How are incidents distributed by Preliminary Score, Placement Score, and Housing
Level?
Analysis Plan
•

Construct Spline graphs for Preliminary and Placement Scores broken down by
Housing Level and by seriousness of the RVR. Use AIC to make this choice.

Does an inmate’s preliminary score add to our ability to predict incidents over
and above placement score?
Do preliminary scores have predictive power in explaining the behavior of those
whose placement is constrained by mandatory minimum scores?
According to data provided by CDCR, the distribution of placement scores reveals
bunching right above the cut-points for levels II, III, and IV. Specifically, of the 56,614
inmates in the analysis sample, 28 percent have placement scores of 19, 4.5 percent
have placement scores of 28, while roughly 3.1 percent have placement scores of 52.
These masses at the cutoffs exceed the mass of inmates with scores just below and just
above. In conjunction with the fact that we don’t observe such bunching in the
preliminary scores, these patterns suggest that the policy of assigning mandatory
minimums is shifting the distribution of security designations to the right. The first
approach is to focus on the behavior of inmates with placements scores of 19, 28, and
52. There are sufficient observations at each of these levels to do a simple analysis of
the predictors of behavioral problems among these groups.
Analysis Plan
A. Matching Analysis
The first analysis will match individuals with similar backgrounds but who differ in
terms of whether they have a mandatory minimum. Matches will be based upon
preliminary score, race/ethnicity, mental status, etc. using observed categories in the
data. Matches are seldom perfect and will require as much individual level data
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 25
December 19, 2011

Inmate Classification System Study

on each inmate as we can obtain, but it will help us in our attempt to use multiple
approaches to address the questions. We will determine whether “similar” inmates,
except for the mandatory minimums are different on our outcomes of interest.
As supplementary analyses, we will examine CCR3375, 2(a) (7) and LWOP inmates.
B. “Gap-Analysis” Using Regression
The Gap Analysis will use a simple set of regressions to determine whether, among
individuals with similar placement scores due to imposed mandatory minimums,
those with lower preliminary scores have a lower incidence of violation relative to
those with higher preliminary scores. There are two ways we recommend exploring
this. One is to estimate a model where the dependent variable is having committed a
violation and the key explanatory variable is the difference between the preliminary
and placement score. The difference effect (or lack thereof) would provide a global
test of whether some of the inmates given a mandatory minimum are less risky than
others.
A second approach would be to create dummy variables for each of the specific
reasons for mandatory minimums. CDCR uses eight official reasons: condemned,
LWOP, Life, History of Escape, "R" suffix (sex crimes), violence exclusion, public
interest case, and other life sentence. Controlling for preliminary score the effect of
each dummy variable would tell us whether an individual given mandatory minimum
has a higher incidence of 115s relative to individuals with comparable preliminary
scores but who does not have a mandatory minimum. These regressions can be
seen as validation tests for the mandatory minimum exceptions.
The dependent variable in these analyses is a count, which skewed and nonnegative, and which suggests a Poisson or negative binomial model. Such models
are relatively easy to interpret for a lay audience and variable effects can be
expressed in terms of effects on the predicted probability of incidents; however, we
might also consider dichotomizing the dependent variable to contrast those who
have no incidents and those who have one or more, and use model a logit, probit, or
linear probability model. The analyses should be undertaken for all incidents and
then subdivided by violent incidents and the seriousness scale (A/B, C/D, E/F). We
will need to determine what makes the best sense for the outcomes – whether to use
categories such as “D or worse” or narrower categories based on specific types of
violations. Decisions will be informed by the data. Analyses will be done within
housing level.
As supplementary analyses, we will examine CCR3375, 2(a) (7) and LWOP inmates.
The main caveat of these analyses is that they will be flawed by our inability to
assess the amplification or suppression of inmate behavior that may occur in the
different housing levels and types. Nonetheless, put together with the other evidence
we assemble below it may contribute to a fuller understanding of the likelihood of
moving inmates who are subject to mandatory minimums into lower housing levels.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 26
December 19, 2011

Inmate Classification System Study

What is the estimated effect on violations of moving up cut-points?
The preliminary score is intended to provide a measure of the risk that an inmate poses to
an institution in terms of their behavior while incarcerated. Presumably, higher preliminary
scores should predict worse behavior. If we are willing to assume that whatever the
relationship between the risk of bad behavior and the score (linear, quadratic, or governed
by a higher-order polynomial), this underlying relationship is continuous, then we can
estimate the impact of being in different security levels by testing for discontinuities in the
violation-score relationship at key break points.
In our group discussions, we have identified two factors that may lead to an impact of
security level on the likelihood of a behavioral infraction. First, the tighter restrictions and
more circumscribed liberties in higher security levels may suppress behavioral problems.
For the sake of argument, let’s label this a suppression effect. On the other hand, one’s
peers in higher security levels may be relatively bad influences, may create social situations
conducive to behavioral problems etc. We can label this a peer effect. What we can
observe empirically is the net of these two effects. Since they are of opposite sign (or I
would guess that is the case), any significant impact of security level designation on the
likelihood of committing a behavioral infraction would reflect one of these effects dominating
the other.
Analysis Plan
The assignment process of the CDCR in conjunction with a regression-discontinuity
empirical model can be used to estimate these net effects. We can illustrate the
method with an example. Suppose we restrict the sample to those inmates housed
in levels I or II facilities. Define the variable P as an inmate’s preliminary score and
the variable P19 as a dummy variable indicating a preliminary score of 19 or higher
(which should place the inmate in a level II facility). Furthermore, define the variable
LII as being housed in a level II facility. The regression discontinuity model basically
estimates the following two equations:

S115 i = α 0 + α 1 Pi + α 2 Pi + βP19 i + δ 1 Pi P19 i + δ 2 Pi P19 i + ε i
2

2

LII i = λ0 + λ1 Pi + λ 2 Pi + κP19 i + γ 1 Pi P19 i + γ 2 Pi P19 i + η i
2

2

The coefficient β is an estimate of how someone with a preliminary score of 18 would
behave if they were Arbitrarily assigned a score of 19 (and everything that entails in
terms of the probability of being housed at a level II institution). The coefficientĸ
measures the effect of going from level 18 to level 19 on the likelihood of being

assigned to a level II facility. The ratio of one to the other ( β / κ ) provides an
estimate of what would happen to behavior if we moved a person on the margin
from level I to level II or vice versa. We will then redo this analysis for inmates
housed in levels II and III and inmates house in levels III and IV.
Note, in order for this to work there needs to be some discrete increase in
assignment to higher levels as the preliminary score crosses these thresholds.
Looking over some of the tabulations performed by CDCR, there does seem to be
such an impact, especially between levels II and III. Of course, the breaks aren’t as
sharp as they would be if it were not for the mandatory minimums ratcheting up the
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 27
December 19, 2011

Inmate Classification System Study

placement scores. We need to determine which of the six outcomes would be used
in this analysis –this determination will be made based on distributions of the
different outcomes in the data.
Using a quadratic trend on either side of the break does indeed assume that the
relationship between behavior and the preliminary score is smooth. We could also
try higher-order polynomials; use more non-parametric techniques (with various
levels of smoothing) on either side of the discontinuity and also look specifically
within the neighborhood of the break. The latter might be problematic if the sample
size is very small, but we will try it and see what happens.
Regarding interpretation of the coefficient, the trend terms are involved in calculating
the change as we pass through the discontinuity. We can and will calculate effect
sizes in this manner. The way it is written above effectively identifies the impact if
one were to extrapolate the trend on either side of the discontinuity beyond the
break. Both ways are commonly used in applied social science research and we will
try both.
Regarding measuring the effect of different security levels, we actually are not
conditioning the RDD models on housing level. We are simply conditioning on nonlinear functions of the preliminary score with a permit structural break. Ultimately,
these estimates will be equivalent to two-stage least squares estimates of security
level on behavior, where the discontinuous increase in moving levels at the threshold
is used as an instrument for level of housing.

Do close Custody Designations serve to reduce escape risk and violations?
Custody Level and Escape Risk
Analyses of the Custody Designations on escapes may be limited. This is because we
expect to see few escapes, particularly with the advent of electric fences which appear
to have dramatically reduced the occurrence of escapes. Nevertheless, as part of a
comprehensive review of classification procedures, data should be collected to specify
this trend and to describe the primary “drivers” that determine how Custody Designations
are made. From a policy standpoint, both of these goals can be achieved largely through
the use of descriptive statistics.
Analysis Plan
A. Describe Escapes
Documenting escape trends would involve reporting the number of escape events by
year for the past 10 years (or earlier, if possible). To control for the growth in the
CDCR inmate population over time, a chart should be prepared plotting both the raw
numbers of escapes by year, with a separate trend line based on the number of
escapes per 1,000 inmates (i.e., rates of escape). These data should be limited to
inmates escaping from secure facilities, not walk-aways from minimum support
facilities (although a separate plot of walk-aways might also be of interest). If
possible, escapes should be differentiated by facility level, as well as type of custody,
housing type, and whether the facility has electrified fences. Preliminary data show
such a small number of escapes that we will not be able to perform analyses which
attempt to describe those inmates who are at better or worse risk for escape
(however, we propose on analysis based on sentence length below).

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 28
December 19, 2011

Inmate Classification System Study

To determine whether there is a relationship between sentence length and escape,
we can calculate the mean sentence length (as well as other descriptive statistics)
for those who escape. We can also determine whether the escapes are due to any
particular sentence type (e.g., life, 3-striker). Due to the low number of escapes, we
cannot determine whether there is a suitable range (time remaining to serve) that
can be identified to allow for adjustment and identify risk.
B. Document Custody Rules and Regulations
Assessing the principal drivers of Custody Designations will require a review of the
policies and procedures governing this practice. According to CCR Title 15. Crime
Prevention and Corrections, Section 3377.1, inmates are assigned designations
based (primarily) on the following factors:
•
•
•
•
•
•

The inmate’s total term, sentence, or remaining time-to serve
The inmate’s escape history
Identification of a management concern
Receipt of an active law enforcement felony hold
A finding of guilt for a serious Rules Violation Report (RVR)
An inmate who is considered to be High Notoriety or is designated as a
Public Interest Case

This descriptive analysis will require a compilation of the reasons using Title 15 as a
source.
C. Which Close Custody Designations are Used Most Frequently?
This would start with a descriptive analysis in which a table would be created that
shows a breakdown of close Custody Designations by the factors leading to that
designation. Assuming that these reasons have been relatively stable over time, a
historical plot would not be necessary. Rather, a simple bar chart could be prepared
based on aggregate data from the past 5 years.
This simple, descriptive analysis would show which of the factors listed above
account for the largest share of close-Custody Designations.
Custody Level and Violations
This analysis will examine the relationship between custody level and violations using
crosstabs and regression analyses. These analyses will primarily try and address the
“suppression” effect of being in close custody, as opposed to minimum and medium
security. The variables we may be able to examine in custody analyses include
demographic characteristics (age, ethnicity), sentence length, preliminary score,
placement score, housing level, and custody level.
Analysis Plan
Multiple analyses will be conducted. Perform crosstabs/means, by housing level, that
show violations (using outcome measures for violations) for subgroups defined by
preliminary score, placement score and custody level.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 29
December 19, 2011

Inmate Classification System Study

Regress violations as a function of placement score + preliminary score + close
custody [plus covariates such as age, ethnicity, and sentence length] for the violation
outcomes. Regressions should be done separately by housing level. We will
examine multicollinearity among variables, particularly placement and preliminary
score. We will determine whether differences in outcomes are due to individual
characteristics or close custody. We will try to determine if there are certain types of
offender characteristics associated with better outcomes from the regression
analyses. Matching can also be done here; matching on inmate characteristics
except for close custody status.
Supplemental regression analyses would be conducted (in the same spirit as for the
“gap” analysis) in order to assess which of the reasons used for Custody Designation
are associated with greater or fewer violations. However, in this analysis, it is not
clear a priori which reasons affect which Custody Designations. We will need to
determine which reasons affect which designations in order to conduct this analysis.
If we can, we will create dummies for the reasons and use them in regressions in
place of the custody status.
To determine whether changes in close custody time can be reduced, we propose a
three-step strategy: 1) first determine how many are in each group (in Blueprint
custody question #1); 2) if there are enough inmates, determine if there is a
relationship between time to MERP and violation using regression with individual
level covariates; 3) if there is a relationship, identify inmates at the lowest risk.
To determine whether Minimum A and Minimum B custody can be combined into
one custody level, the first step would be to gather current information on the
numbers of inmates in the two different custody levels. If there were enough in
Minimum A (Kevin Grassel’s data show several hundred), we can predict RVRs as a
function of background inmate variables and Min A vs. Min B status.

How many close custody inmates can be safely removed to lower custody
levels?
In the Blueprint, several questions are asked about how many current inmates can be
moved as a result of analyses undertaken. This applies to Blueprint Custody questions 4,
5). In general, answers to these questions would require additional data on current inmate
populations – not the sample we are working with. In addition, it would require a
determination made in conjunction with CDCR staff about what constitutes “safely” moved.
For example, if one were willing to tolerate 5% increase in violations, the number that could
be “safely” moved would be less than if one were willing to tolerate 20%. We could also do
different scenarios of risk tolerance, but would require guidance from CDCR on what are
reasonable parameters.

Miscellaneous Blueprint questions related to custody
The Blueprint asks several questions that may be best answered with literature reviews. For
example, in order to determine whether the current regulatory Close Custody accurately
identify escape risk based on evidence based practices or to identify criteria that could be
used to change an inmate’s custody classification, a literature review would be appropriate.
This might be conducted by AOR staff.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 30
December 19, 2011

Inmate Classification System Study

In order to determine whether electrified fences have reduced the number of escapes, we
would require data on escapes and whether institutions had electrified fences at the time of
the escape (Kevin has sent data on this). This is a descriptive analysis since the numbers
of escapes are so low.
To determine whether data are available that identifies age and physical impairment as
factors that would allow for reduced custody, we would need to ask OAR if these data are
captured in automated files that can be used. We can address age as a risk factor in
analyses described above that identify individual level factors associated with violations.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 31
December 19, 2011

Inmate Classification System Study

Appendix F
Successful CDCR Escapes: 1999 to 2010
1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010

Yearly Avg

Camps
CA Corr Center
Sierra Cons Cente
CA Men's Colony
Total

7
8
0
15

4
7
0
11

3
1
1
5

3
3
0
6

0
5
1
6

3
6
0
9

2
11
0
13

6
5
0
11

3
1
0
4

6
4
0
10

3
12
0
15

4
5
0
9

9.50

Level I
CA Corr Center
Wasco
Chuckawalla
Pelican Bay
RJ Donovan
Sierra Cons Cente
CA Medical Facility
Centinela
High Desert
North Kern
Salinas
CA Men's Colony
Folsom
CA Corr Institution
Calipatria
CA Inst for Men
Pleasant Valley
San Quentin
Mule Creek
Deuel
Corr Training Facil
Ironwood
Corcoran
CSP-LA
CSP-SAC
Total

0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
1
1
0
0
0
1
0
0
4

0
1
0
0
2
1
0
0
0
0
0
0
3
0
1
0
0
1
0
0
1
1
0
0
0
11

0
0
0
1
1
0
0
1
0
1
0
0
1
0
0
0
1
1
0
1
0
0
0
0
0
8

0
0
0
0
1
0
0
0
0
0
0
0
2
0
0
0
0
0
2
1
0
0
0
2
1
9

0
2
2
0
0
1
0
0
1
0
1
0
0
0
1
1
0
0
0
0
0
0
0
0
0
9

0
2
0
0
0
0
2
5
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
9

0
0
1
1
2
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
4

1
0
0
0
0
0
1
2
1
1
2
0
0
0
0
0
0
0
0
0
0
0
0
0
0
8

1
0
0
0
1
0
0
0
0
0
0
1
1
0
0
0
0
0
0
0
0
0
0
0
0
4

0
0
2
0
1
1
2
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
1
8

1
0
0
0
2
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
4

0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
1
0
1
3

6.75

Level II
Pitchess
San Quentin
Avenal
Folsom
CA Rehab Center
Total

2
0
0
0
0
2

0
1
0
0
1
2

0
0
0
0
0
0

0
0
0
0
0
0

0
0
1
0
0
1

0
0
0
0
1
1

0
1
0
0
0
1

0
0
1
0
0
1

0
0
0
0
0
0

0
0
0
0
0
0

0
0
1
0
2
3

0
0
0
1
0
1

1.00

Level III
Ironwood
CA Inst for Men
Wasco
Deuel
RJ Donovan
Total

1
0
0
0
0
1

0
0
0
0
0
0

0
0
0
0
0
0

0
0
0
0
0
0

0
0
0
0
0
0

0
1
0
0
0
1

0
0
1
0
0
1

0
0
0
0
0
0

0
0
0
0
0
0

0
0
0
1
1
2

0
0
0
0
0
0

0
0
1
0
0
1

0.50

Level IV

0

0

0

0

0

0

0

0

0

0

0

0

0.00

Other
Deuel I/II
Salinas I/II
Kern Valley I/II
Pleasant Valley I/II
SATF-CORC
Total

0
0
0
0
0
0

0
0
0
0
0
0

0
0
0
0
0
0

0
0
0
0
0
0

0
0
0
0
0
0

0
0
0
0
0
0

0
0
0
0
0
0

1
0
0
0
0
1

0
2
0
0
0
2

0
0
2
0
1
3

0
0
0
0
0
0

0
0
0
1
0
1

0.58

Grand Total 22

24

13

15

16

20

19

21

10

23

22

15

18.33

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 32
December 19, 2011

Inmate Classification System Study

Appendix G
Variable List for the Classification Score and Close
Custody Designation Datasets
Variable
A1A2_Viol
adjust840
age
BCD_Viol
C1
C2
C3
C4
C5
C6
C7
C8
C9
CLVL
contmin840
count_RVR
csrcep840
csrina840
csrinb840
DDPS
det1
det1_840
det2
det2_840
det3
det3_840
det4
det4_840
det5
det5_840
DLVL
DPP
dppdate
EF_Viol
EOPDate
Ethnicity
FACILITY
Firstarrest
HCAT
HCLV
HPAS
ID
IN_DATE
L4_180_840
L4_180R840
lastd840
Length_Stay

Type
Char
Num
Num
Char
Num
Num
Num
Num
Num
Num
Num
Num
Num
Char
Num
Num
Char
Char
Char
Char
Char
Char
Char
Char
Char
Char
Char
Char
Char
Char
Char
Char
Num
Char
Num
Char
Char
Char
Char
Char
Char
Num
Num
Char
Char
Num
Num

Description
Has an A-1/A-2 Violation (Y/N)
Additional adjustments (+/-) to score
Age at REV BEG date
Has an B,C, or D Violation (Y/N)
Number of A-1/A-2 Violations
Number of B,C, or D Violations
Number of E or F Violations
Number of 'battery on non-prisoner' enhancements
Number of 'battery on prisoner' enhancements
Number of 'drug distribution' enhancements
Number of 'deadly weapon' enhancements
Number of 'inciting a disturbance' enhancements
Number of 'cause serious injury' enhancements
custody level
points subtracted for Continuous Minimum Custody
Total number of RVRs
Reason for administrative or irregular placement
Endorsed Institution (location they should move to)
Endorsed level/program (level/program they should move to)
DDPS data available on REV BEG date (Y/N)
if = *, then corresponding Administrative Determinant Code is removed
Administrative Determinant Code
if = *, then corresponding Administrative Determinant Code is removed
Administrative Determinant Code
if = *, then corresponding Administrative Determinant Code is removed
Administrative Determinant Code
if = *, then corresponding Administrative Determinant Code is removed
Administrative Determinant Code
if = *, then corresponding Administrative Determinant Code is removed
Administrative Determinant Code
DDP level (developmental disability) (via DDPS)
Disability Placement Program Code
Date of DPP Code Designation
Has an E or F Violation (Y/N)
Date Designated EOP
Race/Ethnic Group
Facility name
Age at time of first arrest ever (Categorical)
Housing category (Cell/Dorm)
Housing level (Housing Level: I, II, III, IV)
Housing program assignment
Pseudo Identification Number
TERM GROUP START DATE
Level IV 180 Degree Design Status
Reason Code for Level IV 180 Degree Design
Date of previous 840 review
Number of days served as of REV BEG date

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 33
December 19, 2011

Inmate Classification System Study

Variable List for the Classification Score and Close
Custody Designation Datasets (continued)
Variable
MHcode
netchange

Type
Char
Num

NEW_840_CALC
nodisp840
now1_840
now2_840
OFFSGRP
old_840
Old_Placement
OUT_DATE
PRELDATE
PRELDATE_TYPE
rdate840
revbeg
revend
reversed
Risk_Level
RVR
SCORE840_CALC
scoref840
scorem840
SCOREM840_CALC
SEN_TYPE
sent_Length
Sex
SexReg_Date
SexReg_Flag
Streetgang
Stays
Sum_RVRs

Num
Num
Char
Char
Num
Num
Num
Num
Num
Char
Num
Num
Num
Char
Num
Char
Num
Char
Num
Num
Char
Num
Char
Num
Char
Char
Num
Char

SV
voc840

Char
Num

Description
Mental Health Status
Net change in preliminary score during review period from old
preliminary score
New Preliminary Score (Computer Calculated)
points subtracted for No Serious Disciplinary incidents
Institution at time of review
Program/Level at time of review
Commitment Offense Group
Old Preliminary Score
Old Placement Score
TERM GROUP END DATE
Projected Release Date
Source of Projected Release Date
Date 840 review form was completed
Start of review period
End of review period
Revend and revbeg dates were switched (Y/N)
CSRA Risk Score
Has an RVR (Y/N)
New Placement Score (Computer Calculated)
Reason code for Mandatory Minimum (A through H)
MANDATORY MINIMUM SCORE (Paper)
New Mandatory Minimum (Computer Calculated)
SENTENCE TYPE (D,2,3,L,R,W) - via OBIS
Length of Sentence in Months
Gender
Date Sex Reg Flag was Assigned
Required to Register as a Sex Offender (Y/N)
Street gang member at time of incarceration (YES/NO)
Number of incarcerations at CDCR ever
RVRs Received (None, One, Multiple)
Has been incarcerated at CDCR for a serious and/or violent offense
EVER
points subtracted for average (or above) performance in programming

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 34
December 19, 2011

Inmate Classification System Study

Appendix H
Data Analysis and Reporting Project Plan
July 2011
Background
An overview of this project and its scope are well-summarized in the “Inmate Classification
Study, Project Data Analysis” and the “crosswalk” that were prepared by the Expert Panel,
which lists questions to be addressed and the proposed analyses for each question. One of
the original strategies to approach this project was to train California Department of
Corrections and Rehabilitation (CDCR), Office of Research, Adult Research Branch (REB),
staff to perform these statistical analyses under the direction of the project’s University of
California, Davis (UCD), Statistical consultant. However, since that plan was envisioned, the
project has encountered delays (related both to the contracting process and to the dataset
development) that have resulted in a need to change this strategy. Rather than training
REB staff as the analyses are performed, the Statistical consultant is first going to work
independently to perform these analyses and, once the analyses are finalized, will then train
REB staff on the steps that were taken to perform the analyses. Thus, the analyses will not
be a product of the training sessions, but rather will be used to guide a subsequent training
portion of the project. Furthermore, some analyses that would have initially been delegated
to the UCD Statistical consultant will now be directed to the University of California, Berkeley
(UCB), graduate student who will work under the close supervision of Expert Panel member,
Dr. Raphael.
The “Inmate Classification Study, Project Data Analysis” plan and the “crosswalk” are
subject to modification because data that would enable the planned analyses are not always
available in the anticipated form, or because some additional data have been acquired that
enable questions to be addressed in more sophisticated ways. While the intent of this
document is to reflect current thinking about the questions and the approaches to be used, it
should be recognized that even this project plan may change as the study progresses.
Project Steps
Step One: Creation of Data Sets for Subsequent Analysis
REB staff will create the working data sets to be used in the analysis phase. They will also
maintain any associated code used to manipulate the data in standardized ways. Two
primary data sets will be created and used for these analyses. The first is a Classification
Score System data set that includes data collected on and reported from the CDCR 840
Reclassification Score Sheet for annual classification reviews that began in Fiscal Year
08/09 and ended during FY 09/10 (no later than June 30, 2010). Note: This data set
contains data on rules violation reports (RVRs), which are of most interest when examining
institutional misbehavior (or lack thereof). Most of the observations reflect an annual review
on individual inmates, but there will be some inmates who receive more than one review
during a 12-month period. To avoid biasing the results toward the types of inmates who
receive a single annual review, the inmates with multiple reviews need to be included in this
data set. This data set has been created and provided to the Statistical consultants at UCD
and UCB.
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 35
December 19, 2011

Inmate Classification System Study

The second is a longitudinal data set intended to include currently incarcerated inmates who
were sentenced for crimes requiring Close Custody. This data set is comprised of four
different groups of inmates: lifers, life without possibility of parole, 15 to 49 year sentences,
and 50 year plus sentences. These groups vary in the amount of time they are in Close A
Custody. The goal is to create a longitudinal data set going back ten years. This data set
is under construction as of June 2011.
Step Two: Quality Checks
An analysis that is based on erroneous data will be intrinsically flawed. Therefore, an
essential part of the data analysis is to review the data for possible errors and investigate
the validity of any questionable observations. The UCD Statistician and UCB graduate
student will perform quality checks on all data provided to them. An REB staff person will be
the point of contact for all communication on correcting data and will be responsible for
managing any code associated with correcting identified errors. It is incumbent upon UCD
and UCB staff to keep REB apprised of any data issues.
Step Three: Projected Analyses
RVR Analyses
The UC Davis Statistician will perform the following analyses:
1.

A matching analysis will be performed to compare the RVR records (likelihood
and frequency for various severity levels) of subjects with similar preliminary
scores and other demographic information, comparing inmates who have
mandatory minimum sentences against those without such sentences. The
exact form of the matching algorithm will be determined based on preliminary
analysis of the data, though initial attempts will be based on a callipered
Mahalanobis distance algorithm. Similar analyses will be run for other categories
of inmates whose preliminary and placement scores are not equal for other
reasons.

2. The gap analyses will examine whether subjects with similar placement scores,
but different preliminary scores, have a similar likelihood or number of RVRs.
The variable of central interest added to this model would be the difference
between the preliminary and placement scores. Separate analyses would be run
for subjects in the vicinity of each of the mandatory minimum cutoff points (19, 28
and 52). These analyses will be performed using logistic regression methods to
determine whether a given inmate has one or more RVRs or not. When
analyzing the number of RVRs, mixed Poisson models will be used. Because
the response with respect to the placement/preliminary score difference is of
unknown form, the linearity of the fit will be assessed using a spline approach (a
generalized additive model, or GAM)
A second gap analyses would fit the likelihood or frequency of RVRs using the
preliminary score by including dummy variables for each of the administrative
reasons that a subject might have for having a higher placement score. It should be
noted that this analysis will partly address concerns raised by Dr. Patricia Hardyman,
Ph.D., Association of State Correctional Administrators, about whether the existing
preliminary (or placement) score is in need of calibration (redefinition).
3. A crude analysis relating preliminary score, placement score and housing level to the
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 36
December 19, 2011

Inmate Classification System Study

likelihood and frequency of RVRs was done using GAMs while the Expert Panel was
still devising the analysis plan. This analysis needs to be revisited based on the most
current data and with a more deliberate choice of the complexity of the spline
functions that are fit to these relationships. Because this analysis merely involves a
reworking of existing results, it is a lower priority relative to other (new) analyses.
The UCB graduate student will perform the following analyses under the supervision of
Dr. Raphael:
1. The regression discontinuity analyses will use local polynomial fitting to estimate the
likelihood and frequency of RVRs “in the neighborhood of” one of the current housing
level cutoffs in order to estimate the changes in violation levels that would occur if the
cutoffs were changed.
Close Custody Analyses
The UCD Statistician will perform the following analyses:
1. Logistic and Poisson analyses based on longitudinal data. For example, for subjects
who are within their first 10 years of Close A Custody these analyses will help to
determine whether (and how) violation levels decrease as a function of the number of
years in Close A, with the presumption that once those levels have decreased, there
is less rationale for retaining the subjects within Close A. The original intent of these
analyses had been to have longitudinal data going back 10 years for all subjects, but
not all of the variables of interest are available in an electronic format for that length of
time. For that reason, the longitudinal data set will include four groups of inmates
(lifers, life without possibility of parole, 15 to 49 year sentences, and 50 year plus
sentences) who have been in Close A Custody for up to 10 years. In that way, an
analysis involving variables that are available for the full 10 years can be based on a
static set of subjects, while an analysis involving variables that do not go back that far
will be based on within-subject changes in RVR levels over the range of years for
which the data are available.
To supplement these analyses, Classification Services Unit (CSU) staff will examine twelve
years of escape data to identify escapes that are relevant to this study for subsequent
qualitative review. In addition, REB staff will perform a literature review on escapes.
Step Four: Findings Report
The report on findings from the study will have several components. The UCD Statistician
and UCB graduate student will provide to REB written technical explanations of their
analyses, and will support REB staff in “translating” these technical reports into information
that is presentable to a lay person audience. The report outline is as follows:
•
•
•
•
•
•
•

Introduction
Current CDCR Inmate Classification System
Initial CDCR Inmate Classification Procedures
Subsequent CDCR Inmate Classification Procedures
Major Project Goals and Research Questions of This Study
Classification Study Team
Study Methodology

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 37
December 19, 2011

Inmate Classification System Study

•
•
•
•
•
•
•
•
•
•
•
•
•
•
•

Results
Discussion (including limitations)
Expert Panel Recommendations
Implications for Future Research
Blueprint (Appendix 1)
CDC 839 Classification Score Sheet (Appendix 2)
CDC 840 Reclassification Score Sheet (Appendix 3)
180 Housing Design (Appendix 4)
270 Housing Design (Appendix 5)
Expert Panel Data Analysis Plan (Appendix 6)
UCD Statistical Consultant Data Analysis Plan (Appendix 7)
Crosswalk (Appendix 8)
Literature Review on Misclassification (Appendix 9)
Literature Review on Escape Risk Factors (Appendix 10)
Statistical Methodology/Results Technical Report (Appendix 11)

Step Five: Presentation to Executive Staff
The Expert Panel, UCD Statistician, UCB graduate student, and CDCR REB and CSU staff
will present the study findings and recommendations to CDCR Executive staff.
Step Six:

Training REB Staff

Dr. Willits will guide REB staff through the steps of the statistical analyses performed for the
study, addressing the following:
•

Documentation procedures

•

Quality control checks performed on the data

•

Writing SAS code

•

Preparing charts and/or graphs to illustrate analyses

•

Preparing presentations for a non-technical audience

While the focus of the training will be on the current study, it should also provide guidance
on conducting statistical analyses, in general.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 38
December 19, 2011

Inmate Classification System Study

Appendix I
Summary of Findings from the Regression-Discontinuity
Analysis of Inmate Behavioral Outcomes
Steven Raphael
Goldman School of Public Policy
UC Berkeley
Sarah Tahamont
Goldman School of Public Policy
UC Berkeley
October 19, 2011
This memo summarizes the results from our empirical analysis of the security classification
system currently used by the California Department of Corrections and Rehabilitation
(CDCR) and its impact on inmate behavioral infractions. Our primary goal in undertaking
this analysis was to assess the net effect of being placed in a higher security level on the
likelihood that inmates receive serious rules violation reports (RVRs) of various kinds; a
secondary objective is to provide an analysis of the system of mandatory minimum scores
used to determine housing level placement for a sizable minority of inmates. To conduct the
analysis, we analyze administrative records provided to us by CDCR research staff
documenting the behavior of roughly 80,000 inmates over one complete review period
during the calendar year 2008. We employ a regression-discontinuity (RD) design
framework to isolate the effects of higher security placement on behavioral problems. Our
main findings are the following.
•

•
•

•

Finding 1: We find little evidence that placement in level II institutions suppresses or
exacerbates behavioral problems relative to placement in level I institutions for
inmates whose preliminary scores are fairly close to the points threshold between
levels I and II.
Finding 2: We find little evidence that placement in level III institutions suppresses
or exacerbates behavioral problems relative to placement in level II institutions for
inmates with preliminary scores near the level II/III threshold.
Finding 3: We find some evidence that inmates in level IV have worse behavioral
outcomes than inmates in level III. In particular, there is some evidence that inmates
placed in level IV are more likely to acquire division B, C, or D RVRs over the
course of the review period.
Again, this finding applies only to those with
preliminary scores that are fairly close to the level III/IV threshold.
Finding 4: Inmates with binding mandatory minimum placement scores (those with
preliminary scores below the mandatory minimum) are quite well-behaved and
acquire RVRs at rates that are notably lower than inmates with placement scores that
are slightly above or slightly below the mandatory minimum levels across all types of
RVRs.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 39
December 19, 2011

Inmate Classification System Study

•

Finding 5: The relatively better behavior of inmates with binding mandatory
minimums can be explained entirely by their lower average preliminary scores and by
the fact that they are older.

Our interpretation of these findings is as follows. The lack of an impact of higher security
level placement on behavior (and the evidence suggesting possible criminogenic effects of
level IV placement relative to level III) suggests that the best prediction for how inmates with
scores just above the security level threshold will behave if moved to a lower security level is
their behavior at the higher security level. Alternatively stated, these results imply that small
changes in the point classification system (such as moving the level II/level III cutoff from
28 to 30) would probably not result in a system-wide increase in behavioral problems.
With this interpretation in mind, the relatively good behavior of those with binding
mandatory minimum scores is particularly intriguing. The statistical analysis we present
below shows that these inmates are relatively older, and have relatively low average
preliminary scores that are ratcheted up for the purposes of determining security placement.
Given that inmate characteristics readily observable to the CDCR explain their relatively
good behavior, we believe that these inmates in particular provide perhaps the best prospects
for targeted reforms intended to transfer portions of the inmate population to lower security
levels.
Regarding policy implications, we believe that the findings support the following.
• Small increase in the security level cutoff thresholds will probably not result in large
increases in behavioral problems. If CDCR is facing capacity constraints in higher
security institutions, small adjustments in these scores should provide relief while not
compromising safety.
• Inmates facing binding mandatory minimums who have low preliminary scores are
particularly good prospects for moving to lower security levels. In fact, the system of
mandatory minimums should be rethought and perhaps abandoned in exchange for a
system that allows for more case-by-case discretion in establishing minimum security
levels rather than blanket administrative determinations with broad applicability.
1. Summary of security classification process and the data analyzed in this memo
Security Classification Process - The process determining which security level an inmate will
be housed in can be summarized as follows. For new prison admissions, the CDCR collects
information on a variety of factors for each new inmate including but not limited to sentence
length, age, gang affiliation, past behavioral problems during prior incarceration. Inmates are
assigned a preliminary classification score based on background characteristics and prior
behavior while incarcerated. The classification tool assigns weights to each of the predictive
factors; factors that predict higher criminality are awarded more points (for example, longer
sentence length, gang membership, being under 25 years of age). For most inmates, this
preliminary score is the final classification score. However, some inmates qualify for a
mandatory minimum point allocation. Mandatory minimum points are triggered by certain
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 40
December 19, 2011

Inmate Classification System Study

characteristics of the sentence (life without the possibility of parole (LWOP), other life
sentences) or of the offense (offenses with violent or sexual component). This initial, or
“preliminary score” is then compared to applicable mandatory minimum points and the
inmates are assigned a final classification score, or “placement score,” which is the
maximum of preliminary score or the mandatory minimum points. For example, an inmate
with a preliminary score of 32 who is sentenced to life without the possibility of parole
(LWOP) would be assigned a placement score of 52, because the mandatory minimum score
(52) is higher than the preliminary score. Whereas, an inmate with a preliminary score of 32
who has an Immigration and Customs Enforcement (ICE) hold would be assigned a
placement score of 32, because the mandatory minimum score (19) is lower than the
preliminary score.
During their incarceration, inmates go through a reclassification process at least once a year.
The classification periods consist of 6 month intervals, for the most part each annual review
encompasses two review periods. 10 At each reclassification hearing, behavior (good and
bad) since the last hearing is reviewed, behavioral problems lead to upward point adjustments
and good behavior is rewarded with downward point adjustments to the preliminary score
assigned at the previous hearing. Placement scores are still set as the maximum of the new
preliminary score and any applicable mandatory minimum.
Placement scores generally determine an inmate’s housing security level, although there are
many inmates who are housed outside of their security levels due to discretionary
administrative placements, pending transfers or housing Custody Designation that effectively
preempts the security classification. These administrative placements aside, inmates are
assigned to security levels based on the following schema;
•
•
•
•

Inmates with placement scores of 18 or lower are assigned to level I
Inmates with placement scores from 19 to 27 are assigned to level II
Inmates with placement scores from 28 to 51 are assigned to level III
Inmates with placement scores of 52 or higher are assigned to level IV.

The Data - The data that we analyze for this report includes all males housed in a CDCR
institution for all of FY08/09 that are not on death row and for whom we can observe a
complete review period between reclassification hearings. For each inmate, the data set
includes information on all serious RVR’s acquired during the review period, demographic
information about each inmate, information regarding sentencing and controlling offense,
information on housing and security level, and information on several other personal and
institutional characteristics.

10

Occasionally, inmates are reviewed after 1 period (6 months instead of 1 year).

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 41
December 19, 2011

Inmate Classification System Study

Key variables in our empirical analysis are the preliminary and placement scores for each
inmate. We were provided with three sets of scores. The first are the scores for each inmate
one day prior to the beginning of the observed review period. Hence, if a hypothetical
inmate’s review period runs from January 1, 2008 through December 31, 2008, the initial set
of score apply to this inmate on December 31, 2007 and should determine housing at the
beginning of the review period. We also observe new preliminary and placement scores at
the beginning of the review period (after the initial January 1, 2008 classification hearing for
our hypothetical inmate) as well as the scores at the end of the review period following the
terminal point reclassification hearing.
Figures 1 and 2 show the empirical relative frequency distributions of the inmates in our
analysis sample across preliminary and placement scores. Figure 1 shows the empirical
distribution of inmates according to their preliminary score one day prior to the beginning of
the review period while figure 2 shows the distribution of corresponding placement score. In
both figures inmates with 100 or more points are lumped together as one group. For
reference, each figure also shows where the point cutoff levels are between the security
housing levels.
Figure 1 shows a fairly even distribution of inmates across the preliminary score values.
There are large masses of inmates with preliminary scores of zero (nearly 15 percent) and
with scores of 100 or more (nearly 10 percent). Figure 2 shows the effects of the system of
mandatory minimums on placement scores. There are notable masses of inmates at 19, 28
and 52 points (the points just above the security threshold cutoffs). Nearly 28 percent of the
inmates in our sample are constrained at these minimum levels and are unable to move to
lower security levels as a result. The lion’s share of the inmates with mandatory minimum
scores is constrained at placement scores of 19, effectively keeping them out of level I
institutions.
Figure 3 shows the distribution of the change in preliminary scores occurring over the review
period that we observe. Roughly 80 percent of inmates experience a drop in score between
review periods (suggesting that most are fairly well behaved over the course of the year).
Tables 1 and 2 present some descriptive statistics pertaining to inmates that are housed in
levels I through IV. 11 The offense distributions reveal that inmates with violent controlling
offenses inmates are more heavily represented among those in levels III and IV, while
inmates with non-violent controlling offenses are more likely to be in housed in lower
security levels. This is to be expected, because violent offenses usually carry longer sentence
lengths which result in higher preliminary scores. Inmates in higher security levels also tend
to be younger, more likely to be serving their first term at a CDCR institution, are more
likely to be mentally ill (have an EOP or CCCMS designation) and are more likely to be in a
gang.

11

The analysis is limited to inmates housed in levels I-IV. As a consequence, inmates housed in reception
centers and other types of housing placements are excluded.
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 42
December 19, 2011

Inmate Classification System Study

2. Methodology
Being placed in higher security level housing may impact the incidence of behavioral
infractions through several channels. First, the tighter restriction on movement, personal
liberties, and time outside of cells that accompanies higher security levels may effectively
suppress rules violations among inmates who would otherwise commit such infractions. We
refer to this as the “suppression effect” of higher security levels. Alternatively, the average
inmate in higher security level institutions is younger, more likely to be convicted for a
violent offense, and through the system of classification and reclassification, more likely to
have acquired serious RVRs in the past. Being housed with such inmates may increase the
likelihood of getting into trouble through peer influences, a higher likelihood of conflict with
another inmate, or possibly through an enhancement of anti-social attitudes associated with
being housed with a more hardened population. We refer collectively to these potentially
criminogenic effects of higher security level placement as “peer effects.” The net effect of
higher security level placement will be the sum of the suppression and peer effects and can
be either positive or negative. It is this net effect that we seek to measure.
A major methodological challenge that we face in measuring such net effects concerns the
fact that inmate assignment to alternative security levels is not random; in fact the assignment
process assigns those inmates with a high likelihood of poor behavior to higher security
institutions. An ideal research design would randomly assign inmates to security levels and
then observe their behavior over an evaluation period. Random assignment would ensure
that inmate characteristics (both observed and unobserved) are not systematically related to
housing security assignment and that any observable differences in behavior between inmates
in different security levels could be attributed to difference in housing assignment. Such
experimental analysis, of course, would require randomization in the assignment process, a
condition that certainly does not describe the process determining security levels that we
described above.
In the absence of random assignment, we must employ non-experimental empirical methods
to estimate the effects of higher security level placement on behavior. Such methods usually
involve statistically controlling for observable characteristics in an attempt to isolate the
effects of security level, seeking out and exploiting exogenous shocks to security level that
generate variation in housing level that is “as good as random,” or some combination of these
two approaches.
In this project, we exploit the discontinuity in housing level assignment created by the point
thresholds to identify exogenous variation in security level assignment. Specifically, as
assignment is determined in part by variation in one’s preliminary score, we would expect
that inmates who are just above and just below a given points thresholds will experience
discretely different treatments in terms of their assigned housing levels. Since inmates with
such similar scores are likely to be quite similar to one another, and as we are able to model
the general relationship between bad behavior and preliminary score, any discontinuous
change in behavior occurring around the point thresholds can be attributable to the
corresponding change in the housing security level.
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 43
December 19, 2011

Inmate Classification System Study

This research design depends crucially on there being discontinuous treatment at the points
cutoffs, but also requires that all other variables that may determine behavioral infractions
(age, offense history, risk scores etc) vary continuously through the cutoffs. These two
requirements relate to the study design in the following manner. Regarding the first
requirement, the larger the proportion of inmates experiencing a change in housing
assignment as we move through a cutoff, the more precise our estimates of the effect of
housing security level on behavior will be. For example, a relatively large difference in
treatment (e.g. assignment to level II) between inmates with scores of 18 relative to inmates
with scores of 19 makes it more likely that we would detect an effect of housing level on
behavior if it exists and would also improve the precision (the margin of error) of our
estimate of this net behavioral effect.
The second requirement (that all other variables vary continuously through the cutoffs)
basically ensures that there are not large differences in observable or unobservable
characteristics between inmates just above and just below points thresholds and that any
discrete change in behavior can be attributable to the change in housing security level.
In conducting our analysis, a key design choice that we face is whether to use preliminary
score or placement score as the key running variable for our regression-discontinuity design.
Placement score has the advantage that it is more predictive of housing security level than
preliminary score, since many inmates have placement scores that differ from their
preliminary scores due to a binding mandatory minimum. In other words, the predictive
power of the point thresholds in determining security level is greater with placement than
preliminary score. A key weakness of using placement score, however, is that through the
mandatory minimum system those inmates just above a threshold are notable and discretely
different from those inmates just below a threshold. In particular, inmates above the
thresholds via placement scores have key differences in sentences and controlling offense
(more likely to be LWOP, a convicted sex offender etc) and, as we will show later, have
much lower average preliminary scores than those with placement scores just below.
For these reasons, our analysis will be based on variation in housing security level
assignment associated with preliminary score before mandatory minimums are factored in.
While the predictive power of the point thresholds in determining security level assignment
is lower, there are no identifiable selection mechanisms that create large discontinuities in
inmate characteristics around point thresholds when points are measured with preliminary
score.
Figure 4 displays the proportion of inmates housed in each security level at the beginning of
the review period by their preliminary score measured just prior to the beginning of the
review period. The proportion of inmates with preliminary scores that would place them in
level I (less than 19) that are actually housed in level I is quite small (under 0.2). This is not
too surprising since figure 2 shows a very large proportion of inmates with binding
mandatory minimum scores of 19. Nonetheless, as we pass through the level I/level II point
threshold, the proportion in level I drops by roughly 0.1, while the proportion in level II

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 44
December 19, 2011

Inmate Classification System Study

increases. Interestingly, a small proportion of inmates with level II points are housed in level
I institutions (roughly 0.1 to 0.15).
The biggest discontinuities in security level assignment are observed at the level II/III
threshold and the level III/IV threshold. As preliminary score crosses the level II/III
threshold (from 27 to 28 points) the proportion housed in level II institutions drops by nearly
0.4, while the proportion housed in level III increases by a similar amount. Similarly, as
preliminary score crosses the level III/IV point threshold, the proportion housed in level III
drops by roughly 0.6 while the proportion housed in level IV increases.
We estimate the effect of being assigned to higher security levels on the RVR incidence by
statistically testing for breaks in the relationship between the proportion of inmates with
RVRs and preliminary score as at the point thresholds between levels. For example, in
Figure 4 we observe a large increase in the proportion of inmates in level III when the
preliminary score passes from 27 to 28 points. If level III suppresses behavior relative to
level II we should see a discrete downward shift in the relationship between preliminary
score and the proportion with RVRs at 28 points. Conversely, if being placed in higher
security levels is criminogenic, we should see a discrete upward shift in this relationship at
28 points.
To formally illustrate this test, suppose we wish to estimate the net effect of placement in
level IV relative to level III. Define the variable Pi as the preliminary score for inmate i
defined relative to the level III/level IV cutoff, where Pi =0 when preliminary score equals
52, Pi =1 when preliminary score =53, Pi =-1 when preliminary score equals 51 and so on.
Define the variable GE52i as an indicator variable equal to one for inmates with preliminary
scores greater or equal to 52. Finally, define RVRi as an indicator variable equal to one if the
inmate acquires an RVR over the review period. Restricting the sample to inmates with
preliminary scores that would place them in level III or level IV (i.e., preliminary scores
greater than 27), one would test for a structural break in the RVR incidence at 52 points by
estimating the equation
(1)

RVRi = α + βPi + δPi + γGE 52i + λPi ∗ GE 52i + κPi ∗ GE 52i + ε i
2

2

where we allow for a quadratic relationship between the RVR incidence and preliminary
score that is permitted to differ above and below the point threshold via the inclusion of two
interaction terms. The parameters to be estimated are α, β, δ, γ, λ, and
ĸ, and ε i is an error
term. The parameter γ provides an estimate of the discrete change in the RVR incidence
occurring at 52 points. Our principal empirical test essentially involved testing whether this
parameter is statistically and significantly different from zero. A finding of γ > 0 implies a
net criminogenic effect of being in level III while a finding of γ < 0 implies a net suppression
effect. Below we estimate various specifications of equation (1) using several alternative
measure of rules violations and omitting and including a large set of observable covariates.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 45
December 19, 2011

Inmate Classification System Study

Equation (1) provides a reduced-form estimate of the effect of higher security placement on
RVR incidence since the coefficient γ is basically the product of the effect of having 52
points or more on the likelihood of being placed in level IV and the effect of level IV
placement on the RVR incidence. If the discontinuity in assignment at 52 points were sharp,
meaning that all 52s are assigned to level IV while all 51s are assigned to level III, then γ
would be interpretable as the effect of level IV assignment on behavior problems. However,
since the discontinuity in assignment at 52 points is less than complete (in the parlance of
non-experimental analysis, the RD is fuzzy), estimating the effect of being placed in level IV
on the RVR incidence requires that we scale the estimate γ by the effect of being over 52
points on the likelihood of being placed in level IV. To do so, we estimate the following
two-equation system using two-stage least squares

LevelIVi = φ0 + φ1 Pi + φ2 Pi + νGE 52i + θ 0 Pi ∗ GE 52i + θ1 Pi ∗ GE 52i + η i
2

(2)

2

RVRi = α + βPi + δPi + γLevelIVi + λPi ∗ GE 52i + κPi ∗ GE 52i + ε i
2

2

where the variable LevelIVi is an indicator variable for being housed in a level IV facility
and all other variables are as defined above. The two equation system in (2) employs the
dummy variable GE52i as an instrumental variable for being housed in level IV. In other
words, the estimated effect of level IV housing on RVR incidence measured by the parameter
γ makes use of the variation in the proportion assigned to level IV caused by crossing the
point threshold at 52.
We estimate various specifications of equation (1) and the two-equation system in (2) for
sub-samples of the data for the three thresholds between level assignments. For all models,
we restrict the sample to inmates housed in state who are in level I through level IV housing
(i.e., we drop inmates in reception centers or in other types of housing). To estimate the
effect of level II placement relative to level I, we restrict the analysis to inmates with
preliminary scores below 28 points. To estimate the effect of level III placement relative to
level II placement, we restrict the analysis to inmates with preliminary scores greater than 18
but less than 52 points. Finally, to estimate the effect of placement in level IV relative to
level III we restrict the analysis sample to inmates with preliminary scores that are greater
than 27 but less than or equal to 70.
Before proceeding to the estimation results, we must note an important caveat. The estimates
presented below should be interpreted as the effects of higher level placement on RVR
incidence for inmates with preliminary scores that put them “in the neighborhood” of the
point threshold between security levels. That is to say, our estimate of the effect of level IV
placement on RVR incidence applies to inmates with, say, 50 to 53 or 54 points on their
preliminary score. The research design basically employs those inmates below the threshold
as a control group for those slightly above, and uses the overall relationship between RVR
incidence and preliminary score to project what RVRs would be if those above the threshold
were assigned to a lower security level. The result does not apply to an inmate with a
preliminary score of, for example, 100 points. Such an inmate is sufficiently far from the

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 46
December 19, 2011

Inmate Classification System Study

discontinuity threshold that extrapolations of an effect estimates for an inmate with 52 points
is likely to be erroneous.
Throughout the analysis to follow, we estimate the effect of security placement on the
incidence of any serious RVR (defined as an A through F violation), acquiring an A1 or A2
violation, acquiring a B, C, or D violation, or acquiring and E or F violation. In separate
results not reported here, we also tested for impacts of security level assignment on the
number of RVR’s acquired over the course of the review period. The results are qualitatively
and quantitatively similar to what we find for the more simple RVR incidence outcomes.
3. Results of RD analysis using preliminary score
As our description of the data indicated, we have several alternative measures of preliminary
scores that differ in terms of the date of measurement. Specifically, we observe the
preliminary score coming into the review period (i.e., the score prior to the commencement
of the review period). We also observe the preliminary score after the reclassification
hearing at the beginning of the review period as well as the preliminary score after the
reclassification hearing at the end of the review period. The last preliminary score is clearly
inappropriate as it will reflect behavior over the observation period and may or may not
coincide with one’s housing security level over the course of the review. The first observed
preliminary score (measured prior to the beginning of the review period) provides the best
predictor of where inmates are housed at the beginning of the review period. However, as
many inmates improve their scores from one period to the next, many will spend much of the
review period housed at a level that is different from where they are at the beginning of the
review period. For example, an inmate with 55 points coming into the review period who
has 51 points after the reclassification hearing at the beginning of our data will likely move
down to level III shortly thereafter, and may spend the majority of his time during the
observed review period at the lower level. Hence, for many inmates the preliminary score
after the beginning of the review period provides a better predictor of where they do their
time over the course of the review. In this section, we provide estimates using both
preliminary scores. Using the latter score requires that we adjust our measurement of where
the inmate is housed in the two-stage least squares analysis to account for those who move
levels during the course of the review period. We discuss the details of this adjustment with
the discussion of the results below.
Figure 5 graphically displays the reduced-form RD analysis using the preliminary score
coming into the review period (that is to say, prior to the reclassification hearing at the
beginning of our data). Each figure plots the proportion of inmates that acquire an RVR by
their preliminary score value. The figures also fit separate quadratic relationships allowing
for breaks at the point thresholds between the security levels. We present results for any
RVR, an A1/A2 violation, a B/C/D violation, or an E/F violation. The data generally show a
positive relationship between preliminary score and RVR incidence, especially for A1/A2
violations and B/C/D violations. There are no visible breaks in the RVR incidence levels
around the point thresholds, suggesting little evidence of an effect of higher security
placement on RVRs for those near the cutoffs. This is particular notable for the level II/level
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 47
December 19, 2011

Inmate Classification System Study

III and level III/level IV thresholds given the large change in housing assignment at these
thresholds documented in Figure 4.
Table 3 presents the results from formal tests for discontinuity in RVR incidence at the point
thresholds via estimation of equation (1) above. For each of the four RVR outcome
variables, we estimate the break associated with being above the threshold omitting
covariates from the analysis as well as including controls for offense, type of sentence,
number of prior incarcerations served with CDCR, age, race/ethnicity, gang affiliation and
length of the review period. The table only reports the coefficient on the dummy indicating
being above the threshold (corresponding to γ in equation (1) above). Note: a positive
significant estimate of γ would be evidence of a net criminogenic effect while a negative
significant estimate of γ would be evidence of a net suppression effect). The table presents
separate estimate for the three point thresholds.
There is no evidence of significant changes in RVR incidence across the level I/II threshold
(results in Panel A) or the level II/III threshold (results in Panel B). There is weak evidence
of a positive effect of being above the level III/IV threshold for having acquired any RVR.
In the model with no covariates, the positive coefficient is statistically significant at the ten
percent level of confidence. This effect, however, becomes insignificant once covariates are
added to the specification. All of the other coefficients in Panel C are near zero and
statistically insignificant. Since none of the reduced form estimates based on estimation of
equation (1) are significant, we do not present further analysis using the two-stage least
squares model in equation (2) using this earliest measured preliminary score.
As we noted in the introduction to this section, the preliminary score prior to the beginning of
the review period provides the best predictor of housing security level at the beginning of the
review period. However, it does not provide the best predictor of where inmates do their
time during the review period for the many inmates slated to move after the reclassification
hearing that initiated the review period. To see this, Figure 6 graphs the proportion of
inmates that change security levels during the review period. 12 Roughly twelve percent of
inmates change security levels, though these twelve percent are certainly not randomly
distributed across the preliminary score distribution. The first figure in figure 6 shows the
proportion changing security levels by the preliminary score prior to the beginning of the
review period. The proportion moving is particular high above the level II/III threshold and
above the level III/IV threshold. When movers are disaggregated into those who move up in
security level and those who move down, we see bunching of those who move up below the
point thresholds and bunching of those who move down above the point threshold.

12

We were provided with administrative records on all facility moves for all inmates in our sample during the
review period. The data includes the date of the move and the security level of the new institutions.
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 48
December 19, 2011

Inmate Classification System Study

Tables 4 and 5 provide additional information on inmates that change security levels over the
course of the review period. Tables 4 shows key percentiles from the empirical distribution
of a variable measuring the number of days between the initiating reclassification hearing
and the move date for those inmates that move. The median mover changes security levels
within three months of their hearing, with those moving up taking slightly longer than those
moving down. Most moves occur in within six months of the reclassification hearings.
Table 5 shows that most moves occur between consecutive levels, although we do see some
inmates skip levels in both downward and upward movements.
These movers complicate the analysis and raise questions about the results using the earliest
measured preliminary scores displayed in Table 3. To be specific, many of those who are
predicted to be in higher security levels based on the earliest preliminary score actually spent
much of their time in lower security levels due to point deductions at the beginning of the
review period. In essence, the initial analysis may be misclassifying the housing conditions
for inmates near the threshold and drawing incorrect inference as a result. Moreover, given
the large proportions that move near the threshold, this may be a particularly severe problem.
To address this issue, we re-estimate the RD models using the preliminary score measured
after the beginning of the review period. To situate these estimates within the two-stageleast-squared framework we also need to alter our characterization of their housing
assignment during the review period. We do so in the following manner. For inmates that
move during the review period, we measure the proportion of time that they spent in each
level. Hence, an inmate that spends one month in level IV and eleven months in level III
during a twelve-month review period is measured as serving 1/12 of his time in level IV and
11/12 of his time in level III. By contrast, non-movers that spend the whole review period in
level IV would be assigned a value of one for proportion of time spent in level IV and values
of zero for proportion of time spent in other security levels.
Figure 7 displays the relationship between the proportion of time spent in each level and
preliminary score measured after the beginning of the review period. The discontinuity at the
level I/II threshold becomes even murkier than that observed in Figure 4. However, we still
see notable and sizable discontinuities in the proportion of time spent in level III at the level
II/III threshold and the proportion of time spent in level IV at the level III/IV threshold.
Hence, the test for discontinuities in behavior will be sharpest for these higher thresholds
when we use the later preliminary score as the running variable for the RD analysis.
Figure 8 graphs the incidence of RVRs against preliminary score measured after the
beginning of the review period. Each data point provides the proportion of inmates with the
particular preliminary score that acquire an RVR over the review period. Again, we fit
quadratic functions to this relationship that allow for breaks (both in terms of intercept and
slope) at the security level thresholds. The figures reveal little evidence of discrete increases
or decreases in RVR incidence at the level I/II and level II/III thresholds. However, we do
observe an increase in the incidence of any RVR and in B/C/D/ RVRs at the level III/IV
threshold.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 49
December 19, 2011

Inmate Classification System Study

Table 6 present formal tests for discontinuous breaks in RVR incidence at the threshold from
estimates of the model laid out in equation (1). Within each panel, we first present estimates
from models with no covariates. We then present estimates from models that add observable
covariates on sentencing, offense, age and other personal characteristics. The final
specification adds dummy variables for whether the inmate is an up-mover or down-mover.
Again, we only present estimation results for the coefficient γ from equation (1).
Beginning with the results for the level I/II cutoff, we see a slight but significant increase in
RVRs for those above the cutoff for the any RVR outcome and for the B/C/D outcome. We
will see, however, that the first-stage relationship between being above the threshold and
assignment to level II is quite weak, and hence we do not place great confidence in this
particular finding. The results for the level II/III cutoff in panel B show significant increases
in RVR incidence for those above the threshold for any RVR and for E/F violations when we
do not control for whether one is an up or down-mover. Adding controls however eliminates
these effects, yielding an estimate of zero consistent with the results from the prior analysis
using the earliest measured preliminary score. We should note that the first-stage
relationship between being above the level II/III threshold and proportion of review period
spent in level III is quite strong, and hence, we have more confidence in these estimates than
those presented in panel A.
Panel C presents the estimation results for the level III/IV threshold. Here we see consistent
evidence of a significant increase in the incidence of any RVR and B/C/D RVRs associated
with being slightly over the point threshold. The most complete specification suggests that
those just over the thresholds are 3.5 percentage points more likely to acquire an RVR during
the review period and 3.2 percentage points more likely to acquire a B/C/D RVR.
Table 7 presents two-stage-least-squares estimates of the effect of serving time in a higher
security level on RVR incidence using the two-equation model in (2) above. Here we present
results only for the most complete mode specification from Table 7 (inclusive of all
covariates and controls for having moved up or down in security levels). For each set of
models, we report the F-statistic testing the significance of the instrumental variable in the
first-stage regression model. As is evident, the instrument is quite weak in predicting the
time spent in level II for inmates with preliminary scores surrounding the level I/II threshold.
For this reason, we do not discuss these results further. However, the first-stage relationships
in predicting time spent in level III for the level II/III models and for time spent in level IV
for the level III/IV models are quite strong.
Consistent with the reduced-form estimates in Table 7, we find no evidence of an impact,
positive or negative, of assignment to level III as opposed to level II on RVR incidence. On
the other hand, the results in panel C suggest that assignment to level IV increases the
likelihood of acquiring any RVR by 9 percentage points, with this effect being driven
entirely by an increase in the likelihood of a B, C, or D rules violation.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 50
December 19, 2011

Inmate Classification System Study

In additional analysis that we do not report here, we re-estimated all models dropping movers
from the analysis and using the preliminary score after the initiating reclassification hearing
as the running variable. The results from this model yield no evidence of either peer or
suppression effects for at the level I/II cutoffs and level II/III cutoffs, and some mixed
evidence suggestive of a criminogenic effect of being placed in level IV.
Overall, the analysis finds little evidence that the behavior near the security level thresholds
differs depending on which side of the threshold one falls, although some of the results are
indicative of a positive causal effect of level IV placement on RVR incidence. These results
lead to us to the general conclusion that for those within the neighborhood of the security
level thresholds, the average behavior of those with placement scores slightly above the
cutoff is the best predictor of how they would behave should they be moved to the next
lowest security level.
4. Mandatory minimums
In our analysis of the effect of housing security level on behavioral outcomes, we made the
deliberate research design choice to use preliminary scores as the running variable in the RD
analysis rather than placement score. The reasoning behind this choice is that the
administrative rules determining placement score are likely to create discontinuous breaks at
the security level thresholds in observable characteristics, such as controlling offense and
mean preliminary score, and perhaps unobservable characteristics, such as unobserved
propensity towards violence or willingness or ability to follow orders. An important
unobservable or perhaps unquantifiable factor that may also change for those bound by a
mandatory minimum concerns the incentives to behave well. To the extent that moving to
lower security levels is desirable, inmates bound by a mandatory minimum face less of an
incentive than inmates not bound by mandatory minimums to comply with institutional rules
and to engage in positive programming.
In this section, we explore the relative behavioral outcomes of inmates bound by mandatory
minimum point assignments. In particular we reanalyze the RD models using placement
score as the running variable to highlight the relatively better behavioral outcomes for
inmates with binding mandatory minimum constraints and assess whether this better
behavior can be explained by observable characteristics. 13

13

Inmates with mandatory minimum point assignments fall into one of the following categories: Condemned –
minimum score of 52, Life without Parole (LWOP) – minimum score of 52, CCR3375.2(a)(7) Life Inmate with
special circumstances (multiple life sentences, torture, execution style) – minimum score of 28, History of
Escape – minimum score of 19, Warrants “R” Suffix (Sex Offender) – minimum score of 19, Violence
Exclusion – minimum score of 19, Public Interest Case – minimum score of 19, and other Life Sentence –
minimum score of 19.
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 51
December 19, 2011

Inmate Classification System Study

By construction, placement and preliminary score will be the same for inmates not facing
mandatory minimum and for inmates with mandatory minimums whose preliminary scores
exceed their placement scores. However, for inmates who over time have improved their
scores through good behavior, mandatory minimums will eventually become binding. Figure
9 displays this fact. The figure presents a plot of the average preliminary score of all inmates
in our sample by single placement score values (using preliminary and placement score
measured prior to the initiating review period). 14 For inmates with placement scores not
equal to 19, 28, or 52, average preliminary score equals placement score. For inmates above
at the point thresholds however, average preliminary scores are discretely and substantially
lower relative inmates with slightly higher and slightly lower placement scores. Note: Figure
2 revealed that a substantial minority of inmates has placement scores at these mandatory
minimum levels (especially at 19). In essence, the scores of these inmates continue to
improve with time, yet it is not reflected in the one score that actually determines security
level placement.
To be sure, this would not be an issue if placement score (inclusive of the mandatory
minimums) was more predictive of behavioral problems than preliminary score. Of course,
this is a question that we can explore. Figure 10 presents a scatter plot of the proportion of
inmates that acquire an RVR over the review period by placement scores measured prior to
the beginning of the review period. For reference, we have placed circles around the data
points for inmates with placement scores of 19, 28, or 52. For all of the RVR outcomes
displayed in the figures, inmates with placement scores at the mandatory minimums behave
discretely better than inmates with slightly higher placements as well as inmates with slightly
lower placement scores. Notably, the inmates bound at the higher minimums (28 and 52) are
among the best behaved relative to all other placement score levels. Figure 11 reproduces
these figures using placement score measured after the reclassification hearing at the
beginning of the review period. The patterns are nearly identical.
These patterns beg the question of why inmates with binding mandatory minimums behave
better. Certainly, the incentives associated with moving to a lower security level cannot
explain this difference, as improvements in preliminary score cannot alter their housing
arrangements. It is possible that persistent good behavior is driven by an incentive stay at the
lowest possible Custody Designation, which determines many of the conditions of
confinement within security level. It is also possible that some aspect of being placed under a
mandatory minimum acts to suppress behavior in a manner difficult to observe in
administrative data. An additional possibility is that differences in observable characteristics
such as preliminary score and age explain the better behavior of inmates trapped at these
minimum placement scores. To the extent that this is the case, a clear implication would be
that the assignment process is ignoring useful information that is predictive of inmate
behavioral outcomes in following a rigid set of administrative placement rules.

14

This figure looks similar regardless of the scores used.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 52
December 19, 2011

Inmate Classification System Study

To explore this question, we first apply the RD equation (1) to the data in Figures 10 and 11
using placement score as the running variable. The discontinuity here provides a formal
empirical estimate of the size and statistical significance of the difference in behavior of
those with mandatory minimum scores relative to other inmates. Given the patterns in the
figures we expect the structural break coefficients to be negative, sizable, and statistically
significant. Next, we assess whether controlling for observable characteristics explains these
differentials. Specifically, we re-estimate the models using several alternative sets of
covariates and assess whether doing so eliminates the significant breaks at the point
thresholds. The first specification adds controls for preliminary score and age (both entered
as quadratic control functions) to the specification of equation (1). The second specification
includes the more complete set of covariates employed in our regression analysis above. The
one departure from equation (1) is that we control for the running variable as a cubic
polynomial interacted with being above the point threshold. We make this modification due
to the sharper discontinuities in behavior that we observe in the data and due to specification
tests suggesting that a cubic polynomial provides a better overall fit.
Table 8 presents the results of this analysis using placement score measured prior to the
beginning of the review period. Beginning with the results for the level I/II cutoff in panel
A, the first row of estimates provides the structural break estimates for each RVR dependent
variable when no controls are included in the specification other than the cubic polynomial in
the running variable, the dummy for being above the threshold, and the interaction term
between the threshold dummy and the placements score polynomial terms. 15 The second row
provides the estimates of the structural break in behavior controlling for variation in
preliminary score and age. The final row adds all covariates to the model (the set of
additional explanatory variables are described in the notes to the table). For the level I/II
cutoff, the results with no covariates show large decreases in RVR incidence for those above
the threshold. Controlling for preliminary score and age explains nearly half of this break (as
is evident by the sharp decline in the absolute value of the coefficients). Adding all
explanatory variables yields estimates of the structural break that are statistically
insignificant.
The results for the level II/III threshold are even more pronounced. The results for the model
without control variables find that RVR incidence declines by 14 percentage points for those
just above the thresholds (with this estimate statistically significant at the one percent level of
confidence). Controlling for preliminary score and age (the results in the next row) explains
all of this decline, as can be seen in the large decline in the absolute value of the coefficient
estimate that also becomes statistically insignificant. Similar results are observed when we
look at specific types of RVRS.

15

Similar to our previous analysis, we measure placement score relative to the point threshold. This permits
interpreting the coefficient on the structural break as the magnitude of the break in the dependent variable as we
move though the threshold.
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 53
December 19, 2011

Inmate Classification System Study

Finally, we find similar patterns at the level III/IV threshold. Being above the threshold is
predicted to result in a 13 percentage point decline in the proportion of inmates with an RVR.
Controlling for preliminary score and age completely explains away this relationship. Again,
similar patterns are observed when we look at specific types of RVRs. To be complete,
Table 9 reproduces these results using placement score measured after the beginning of the
review period. The findings are qualitatively identical and quantitatively comparable.

0

Proportion of inmates
.05
.1

.15

Figure 1: Empirical Distribution of Preliminary Scores One-Day Prior to the
Beginning of the Review Period

0

20

40
60
Preliminary Score

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

80

100

Page 54
December 19, 2011

Inmate Classification System Study

0

.05

Proportion of inmates
.1
.15

.2

Figure 2: : Empirical Distribution of Placement Scores One-Day Prior to the Beginning
of the Review Period

0

20

40
60
Placement Score

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

80

100

Page 55
December 19, 2011

Inmate Classification System Study

0

.1

Proportin of inmates
.2
.3

.4

Figure 3: Distribution of Inmates by Change in Preliminary Score over the Review
Period Covered by the Data

-10

10
20
0
Change in preliminary score over the review period

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

30

Page 56
December 19, 2011

Inmate Classification System Study

Figure 4: Proportion of Inmates Housed in Each Security Level at the Beginning of the
Review Period by Placement Score Measured One-Day Prior to the Beginning of the
Review Period

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 57
December 19, 2011

Inmate Classification System Study

60
20
40
Preliminary Score

80

0

60
20
40
Preliminary Score

80

Proportion with an A1A2
0 .01 .02 .03 .04

0

0

40
60
20
Preliminary Score

80

0

40
60
20
Preliminary Score

80

Proportion with any EF
.05 .1 .15 .2

Proportion with any BCD
.05 .1 .15 .2

Proportion with any RVR
.1 .15 .2 .25 .3 .35

Figure 5: Incidence of Serious Rules Violation Reports by Preliminary Score Measured
One Day Prior to the Beginning of the Review Period

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 58
December 19, 2011

Inmate Classification System Study

Figure 6: Proportion of Inmates Changing Housing Security Levels Over the Review
Period by Preliminary Score Measured One Day Prior to the Beginning of the Review
Period

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 59
December 19, 2011

Inmate Classification System Study

Figure 7: Proportion of Review Period Spent in Each Housing Level by Preliminary
Score Measured After the Start of the Review Period

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 60
December 19, 2011

Inmate Classification System Study

20
40
60
Preliminary Score

80

0

20
40
60
Preliminary Score

80

Proportion with an A1A2
0 .01.02.03.04.05

0

0

40
60
20
Preliminary Score

80

0

20
40
60
Preliminary Score

80

Proportion with any EF
.1 .15 .2 .25

Proportion with any BCD
.05 .1 .15 .2 .25

Proportion with any RVR
.1 .15 .2 .25 .3 .35

Figure 8: Incidence of Serious Rules Violation Reports by Preliminary Score Measured
After the Beginning of the Review Period

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 61
December 19, 2011

Inmate Classification System Study

0

Average Preliminary Score
20
40
60

80

Figure 9: Average Preliminary Score by Placement Score Measured Prior to the
Beginning of the Review Period

0

20

40
Placement Score

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

60

80

Page 62
December 19, 2011

Inmate Classification System Study

20
40
60
Placement Score

80

0

60
20
40
Placement Score

80

Proportion with an A1A2
0 .01 .02 .03 .04 .05

0

0

20
40
60
Placement Score

80

0

40
20
60
Placement Score

80

Proportion with any EF
.1 .15 .2 .25

Proportion with any BCD
0 .05 .1 .15 .2

Proportion with any RVR
.1 .15 .2 .25 .3 .35

Figure 10: RVR Incidence by Placement Score Measured One Day Prior to the
Beginning of the Review Period

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 63
December 19, 2011

Inmate Classification System Study

20
40
60
Placement Score

80

0

20
40
60
Placement Score

80

Proportion with an A1A2
0 .01.02.03.04.05

0

0

60
20
40
Placement Score

80

0

20
40
60
Placement Score

80

Proportion with any EF
.1 .15 .2 .25 .3

Proportion with any BCD
0 .05 .1 .15 .2 .25

Proportion with any RVR
.1 .2 .3 .4

Figure 11: RVR Incidence by Placement Score Measured After the Beginning of the
Review Period

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 64
December 19, 2011

Inmate Classification System Study

Table 1: Controlling Offense Distribution of Inmates in our Analysis Sample by
Housing Security Level
Level I

Level II

Level III

Level IV

Murder

0.00

0.26

0.20

0.39

Manslaughter

0.01

0.02

0.04

0.04

Robbery

0.21

0.09

0.18

0.20

Assault

0.10

0.13

0.15

0.14

Sex Offense

0.00

0.20

0.15

0.07

Kidnapping

0.00

0.04

0.02

0.03

Burglary

0.13

0.06

0.07

0.05

Larceny

0.07

0.02

0.02

0.01

Vehicle theft

0.05

0.01

0.01

0.01

Forgery/Fraud/Other Property

0.02

0.01

0.01

0.00

Drugs

0.32

0.11

0.09

0.04

DUI

0.04

0.01

0.00

0.00

Other

0.05

0.03

0.04

0.02

Proportion of regularly-housed
inmates

0.07

0.25

0.39

0.29

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 65
December 19, 2011

Inmate Classification System Study

Table 2
Average Characteristics by Housing Security level
Level I

Level II

Level III

Level IV

Age

38.15

43.27

39.07

35.81

First stay

0.313

0.594

0.546

0.575

EOP/CCCMS

0.056

0.168

0.277

0.303

Gang

0.086

0.045

0.172

0.209

48
72
108

72
96
148

84
144
192

144
204
288

Black

0.333

0.292

0.312

0.365

White

0.301

0.267

0.243

0.173

Hispanic

0.324

0.365

0.373

0.400

Sentence
P25
P50
P75

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 66
December 19, 2011

Inmate Classification System Study

Table 3
RD Estimates of the Effects of Being Above the Security Classification Threshold on the
Likelihood of Serious Rules Violations During the Review Period Using Preliminary
Score One Day Prior to Review Period
Panel A: Level I/Level II Threshold
Any RVR
A1/A2
B/C/D
E/F
No Covariates
0.010
-0.002
0.008
0.009
(0.015)
(0.003)
(0.009)
(0.013)
Covariates
0.005
-0.002
0.007
0.005
(0.014)
(0.003)
(0.009)
(0.013)
Panel B: Level II/Level III
Any RVR
A1/A2
B/C/D
E/F
No Covariates
-0.015
0.006
-0.006
-0.022
(0.021)
(0.005)
(0.015)
(0.018)
Covariates
-0.018
0.005
-0.010
-0.023
(0.020)
(0.005)
(0.015)
(0.017)
Panel C: Level III/Level IV
Any RVR
A1/A2
B/C/D
E/F
c
No Covariates
0.033
0.002
0.004
0.016
(0.018)
(0.005)
(0.013)
(0.015)
Covariates
0.028
0.001
0.002
0.015
(0.017)
(0.005)
(0.013)
(0.015)
Standard errors are in parentheses. The figures in the table provide the coefficient from a
dummy variable indicating a preliminary score (measured one day before the beginning of
the review period) above the security classification threshold for the comparison indicated.
All models include a quadratic function in preliminary score interacted with the threshold
cutoff. The models with covariates also include thirty two offense dummies, dummy
variables indicating the time of sentence (second-strike, third-strike, LWOP, determinate),
dummy variables indicating first, second, third, or fourth or higher stay with CDCR, a
quadratic in age, seven race/ethnicity dummies, a dummy indicating documented gang
affiliation, and a linear measure of the length of the review period.
a. Statistically significant at the one percent level of confidence.
b. Statistically significant at the five percent level of confidence.
c. Statistically significant at the ten percent level of confidence.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 67
December 19, 2011

Inmate Classification System Study

Table 4
Distribution of the Days Between the Beginning of the Review Period and the Move
Date for Inmates that Changes Housing Custody Levels During the Study Period
Percentiles
All Movers
Those who move up Those who move
down
10th
23
24
23
th
25
46
59
42
Median
88
115
76
75th
154
174
135
th
90
227
262
196
Tabulations based on the 9,747 identified inmates that move custody levels and that have
complete information on the custody level at the new institution.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 68
December 19, 2011

Inmate Classification System Study

Table 5
Distribution of Inmates Who Change Security Levels During the Review Period by
Origin and Destination Security Level
Destination Security Level
Origin Security
I
II
III
IV
Level
I
0
957
319
36
II
799
0
1,509
78
III
530
2,345
0
967
IV
67
40
2,100
0
Tabulations based on the 9,747 identified inmates that move custody levels and that have
complete information on the custody level at the new institution

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 69
December 19, 2011

Inmate Classification System Study

Table 6
RD Estimates of the Effects of Being Above the Security Classification Threshold on the
Likelihood of Serious Rules Violations During the Review Period Using Preliminary
Score After the Beginning of the Review Period
Panel A: Level I/Level II Threshold
Any RVR
A1/A2
B/C/D
E/F
a
b
No Covariates
0.038
0.001
0.022
0.009
(0.015)
(0.004)
(0.009)
(0.013)
b
b
Base Covariates 0.037
0.001
0.021
0.007
(0.015)
(0.004)
(0.009)
(0.013)
Plus indicators
0.037a
0.007
0.021b
0.008
for up and down (0.015)
(0.004)
(0.009)
(0.013)
moves
Panel B: Level II/Level III
Any RVR
A1/A2
B/C/D
E/F
a
No Covariates
0.063
0.004
0.028
0.039b
(0.023)
(0.006)
(0.017)
(0.019)
Base Covariates 0.052
0.005
0.022
0.031
(0.022)b
(0.006)
(0.017)
(0.019)c
Plus indicators
0.026
0.005
0.010
0.015
for up and down (0.022)
(0.006)
(0.017)
(0.019)
moves
Panel C: Level III/Level IV
Any RVR
A1/A2
B/C/D
E/F
a
a
No Covariates
0.077
0.002
0.047
0.033b
(0.019)
(0.005)
(0.015)
(0.015)
Base Covariates 0.064a
0.005
0.043a
0.020
(0.018)
(0.005)
(0.014)
(0.015)
Plus indicators
0.035c
0.003
0.032b
-0.000
for up and down (0.019)
(0.006)
(0.015)
(0.016)
moves
Standard errors are in parentheses. The figures in the table provide the coefficient from a
dummy variable indicating a preliminary score (measured one day before the beginning of
the review period) above the security classification threshold for the comparison indicated.
All models include a quadratic function in preliminary score interacted with the threshold
cutoff. The models with base covariates also include thirty two offense dummies, dummy
variables indicating the time of sentence (second-strike, third-strike, LWOP, determinate),
dummy variables indicating first, second, third, or fourth or higher stay with CDCR, a
quadratic in age, seven race/ethnicity dummies, a dummy indicating documented gang
affiliation, and a linear measure of the length of the review period. The final models add
indicator variables for inmates that move up in security levels and inmates that move down in
security levels.
a. Statistically significant at the one percent level of confidence.
b. Statistically significant at the five percent level of confidence.
c. Statistically significant at the ten percent level of confidence.
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 70
December 19, 2011

Inmate Classification System Study

Table 7
IV Estimates of the Effects of Being at a Higher Security Level Using Points Above the
Threshold as an Instrument for Housing Level
Panel A: Level I/Level II Threshold (First Stage F-statistic = 4.66)
Any RVR
A1/A2
B/C/D
E/F
Coefficient on
-1.478
-0.028
-0.840
-0.308
proportion of
(1.155)
(0.141)
(0.679)
(0.545)
period in level II
Panel B: Level II/Level III (First Stage F-statistic = 180.63)
Any RVR
A1/A2
B/C/D
Coefficient on
0.126
0.023
0.047
proportion of
(0.108)
(0.030)
(0.081)
period in level
III
Panel C: Level III/Level IV (First Stage F-statistics = 1,285.94)
Any RVR
A1/A2
B/C/D
c
Coefficient on
0.091
0.006
0.084b
proportion of
(0.050)
(0.015)
(0.039)
period in level
IV

E/F
0.070
(0.091)

E/F
-0.000
(0.041)

Standard errors are in parentheses. The figures in the table provide the second-stage
coefficient on the noted housing security level variable. The second stage specification
include proportion of time in the upper security level of the range analyzed, a quadratic
function in preliminary score (measured after the beginning of the review period) interacted
with the threshold cutoff, thirty two offense dummies, dummy variables indicating the time
of sentence (second-strike, third-strike, LWOP, determinate), dummy variables indicating
first, second, third, or fourth or higher stay with CDCR, a quadratic in age, seven
race/ethnicity dummies, a dummy indicating documented gang affiliation, a linear measure
of the length of the review period, and indicator variable for those who move up security
levels and an indicator variable for those who moved down. The first stage employs a
dummy for being above the threshold as an instrument for the proportion of time in the
indicated housing level.
a. Statistically significant at the one percent level of confidence.
b. Statistically significant at the five percent level of confidence.
c. Statistically significant at the ten percent level of confidence.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 71
December 19, 2011

Inmate Classification System Study

Table 8
RD Estimates of the Effects of Being Above the Security Classification Threshold on the
Likelihood of Serious Rules Violations During the Review Period Using PLACEMENT
SCORE Measured One Day Prior to the Beginning of the Review Period
Panel A: Level I/Level II Threshold
Any RVR
A1/A2
B/C/D
E/F
No Covariates
-0.168a
-0.008
-0.058a
-0.138a
(0.026)
(0.006)
(0.017)
(0.023)
Controlling for
-0.087a
-0.011
-0.025
-0.076a
preliminary
(0.027)
(0.006)
(0.017)
(0.023)
score and age
All Covariates
-0.041
-0.008
-0.004
-0.044c
(0.027)
(0.007)
(0.018)
(0.023)
Panel B: Level II/Level III
Any RVR
A1/A2
B/C/D
E/F
a
b
No Covariates
-0.141
-0.006
-0.050
-0.121a
(0.029)
(0.007)
(0.020)
(0.025)
Controlling for
-0.031
-0.002
-0.008
-0.044c
preliminary
(0.030)
(0.008)
(0.021)
(0.026)
score and age
All Covariates
-0.032
-0.003
-0.009
-0.045c
(0.030)
(0.008)
(0.021)
(0.026)
Panel C: Level III/Level IV
Any RVR
A1/A2
B/C/D
E/F
a
b
a
No Covariates
-0.133
-0.011
-0.079
-0.069a
(0.019)
(0.005)
(0.015)
(0.016)
Controlling for
-0.036
-0.005
-0.027
-0.015
preliminary
(0.022)
(0.006)
(0.017)
(0.018)
score and age
All Covariates
-0.025
-0.006
-0.021
-0.006
(0.022)
(0.007)
(0.017)
(0.018)
Standard errors are in parentheses. The figures in the table provide the coefficient from a
dummy variable indicating a placement score (measured one day before the beginning of the
review period) above the security classification threshold for the comparison indicated. All
models include a cubic function in placement score interacted with the threshold cutoff. The
models with all covariates include a quadratic in preliminary score, thirty two offense
dummies, dummy variables indicating the time of sentence (second-strike, third-strike,
LWOP, determinate), dummy variables indicating first, second, third, or fourth or higher stay
with CDCR, a quadratic in age, seven race/ethnicity dummies, a dummy indicating
documented gang affiliation, and a linear measure of the length of the review period.
a. Statistically significant at the one percent level of confidence.
b. Statistically significant at the five percent level of confidence.
c. Statistically significant at the ten percent level of confidence.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 72
December 19, 2011

Inmate Classification System Study

Table 9
RD Estimates of the Effects of Being Above the Security Classification Threshold on the
Likelihood of Serious Rules Violations During the Review Period Using PLACEMENT
SCORE Measured After the Beginning of the Review Period
Panel A: Level I/Level II Threshold
Any RVR
A1/A2
B/C/D
E/F
a
a
No Covariates
-0.109
0.004
-0.063
-0.087a
(0.028)
(0.007)
(0.018)
(0.025)
Controlling for
0.007
0.008
-0.015
-0.002
preliminary
(0.029)
(0.007)
(0.019)
(0.025)
score and age
All Covariates
0.051c
0.013c
0.008
0.021
(0.029)
(0.007)
(0.019)
(0.025)
Panel B: Level II/Level III
Any RVR
A1/A2
B/C/D
E/F
a
a
No Covariates
-0.104
-0.004
-0.065
-0.054b
(0.033)
(0.008)
(0.023)
(0.028)
Controlling for
0.049
0.006
-0.000
0.042
preliminary
(0.033)
(0.009)
(0.000)
(0.028)
score and age
All Covariates
0.047
0.007
-0.001
0.039
(0.033)
(0.009)
(0.024)
(0.028)
Panel C: Level III/Level IV
Any RVR
A1/A2
B/C/D
E/F
a
a
No Covariates
-0.109
-0.004
-0.059
-0.076
(0.021)
(0.006)
(0.016)
(0.017)
Controlling for
0.026
0.009
0.015
-0.009
preliminary
(0.023)
(0.007)
(0.018)
(0.019)
score and age
All Covariates
0.032
0.010
0.021
-0.007
(0.023)
(0.007)
(0.018)
(0.019)
Standard errors are in parentheses. The figures in the table provide the coefficient from a
dummy variable indicating a placement score (measured after beginning of the review
period) above the security classification threshold for the comparison indicated. All models
include a cubic function in placement score interacted with the threshold cutoff. The models
with all covariates include a quadratic in preliminary score, thirty two offense dummies,
dummy variables indicating the time of sentence (second-strike, third-strike, LWOP,
determinate), dummy variables indicating first, second, third, or fourth or higher stay with
CDCR, a quadratic in age, seven race/ethnicity dummies, a dummy indicating documented
gang affiliation, and a linear measure of the length of the review period.
a. Statistically significant at the one percent level of confidence.
b. Statistically significant at the five percent level of confidence.
c. Statistically significant at the ten percent level of confidence.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 73
December 19, 2011

Inmate Classification System Study

Appendix J
Statistical Methods and Summarized Results
for Gap, Matching and Longitudinal Analyses
Neil Willits, Ph.D.
Statistical Laboratory, Department of Statistics
UC Davis
November 17, 2011
The analyses described and summarized in this Appendix have been divided into three types.
The gap analyses were used to address Classification questions 1 through 4, as well as some
additional questions that were raised as an offshoot from those analyses. A variant of the gap
analyses was used to address custody Question 2 as well. Most of the other close custody
questions were addressed based on the longitudinal analyses. The matching analyses were
used as an alternate way of addressing the second classification question.
The Data:
The data for these analyses were of two general forms. The first data set represented a yearlong “snapshot” of prison behavior (CDCR 840) reviews for review periods that began in the
2008/9 fiscal year, restricted to prisoners who were imprisoned for the entire fiscal year.
This comprised 121,374 reviews on a total of 72,322 distinct inmates. Many inmates
received semiannual (or even more frequent) reviews during this period, accounting for the
larger number of reviews than inmates. This data was used as the basis of the gap analyses
and its offshoots. The matching analyses were based on exact matches between inmates with
and without constraining mandatory minimum scores, which were drawn from this data set.
Various quality control checks were run on the data, to eliminate obvious errors. For
example, there were numerous reviews for which the listed beginning of review period was
later than the listed end of review period, there were some obvious errors in various dates
(e.g., review periods beginning prior to 1950), misspelling in various categorical variables.
These errors were corrected when the appropriate correction was unambiguous and treated as
missing values otherwise. The data were quite high in overall quality, with misspelling
occurring in fewer than 1% of observations.
The other data sets were longitudinal in nature, following groups of inmates through up to
approximately 10 years of 840 reviews. These data sets focused on particular subgroups of
inmates, including (1) inmates with sentences of at least 15 years and below 50 years who
were in the start of their prison term, (2) inmates with sentences of 50+ years at the start of
their terms, (3) inmates with single life sentences, and (4) inmates with multiple life
sentences. The longitudinal analyses were based on these data sets.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 74
December 19, 2011

Inmate Classification System Study

Dependent Variables: Occurrence of an RVR and RVR Counts:
The responses (dependent variables) used in the statistical analyses were similar for all of
these data sets. One form looked simply at whether an inmate had committed one or more
violations of a given type. The other form of outcome looked at the number of violations of
a given type that occurred during the review period. Since the review periods varied in
length (even though most of them covered a period of either six or twelve months), the
analyses needed to adjust for this variability. This adjustment is necessary since you
wouldn’t expect an inmate to run up as many violations in a six month period as in a twelve
month period. A more detailed explanation of this adjustment will appear in conjunction
with the description of the analyses.
To examine the form of the relationship between a continuous predictor and one of the
outcome variables, generalized additive models (or GAMs) were fit to the data, which use
spline methods (in this case) to fit more general curves to a relationship like this. The
standard algorithms for fitting GAMs don’t allow for the inclusion of random effects (such as
an inmate effect) and so the GAM models were used to estimate the form of the relationships
between continuous predictors and an outcome measure, after which a mixed model GLIM
was used to insert the necessary random effects into the model. Logistic (binomial) GLIMs
were used for modeling the occurrence of one or more violations, and Poisson GLIMs were
used for modeling RVR counts.
Gap analyses:
These analyses were based on the data from the 2008/09 fiscal year, restricted to include only
the inmates who were listed as being housed in Level I through Level IV. Each analysis
focused on inmates whose placement scores were close to one of the threshold levels used to
classify inmates into housing levels. Thus each analysis was split into three sub-analyses,
one looking at placement scores in the range 19±7, a second at the range 28±7 and a third at
the range 52±7. There were also parallel analyses run on the dichotomous and count
outcomes, and for all serious (A through F) RVRs, on A through D violations, and restricted
to A violations only. This meant that each analysis was divided into 18 sub-analyses (three
thresholds levels times 2 forms of outcome times three levels of RVRs).
Matching analyses:
These analyses contrasted inmates who have a constraining mandatory minimum (one that’s
greater than their preliminary score) with inmates who have the same preliminary score but
no constraining mandatory minimum. These pairs of inmates were also matched on their
ethnicity, mental health code, the administrative determinants for gang affiliation,
psychological condition and sex offender status. A more restrictive set of matches were run
which added offense category to the criteria for an exact match. Offense category was
defined based on the primary offense group that was defined for each inmate, though
offenses were aggregated more broadly:

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 75
December 19, 2011

Inmate Classification System Study

Violent
Sex-related
Property
Drug-related
Other

offense groups 1 through 8 and 14,
offense groups 9 through 13,
offense groups 15 through 22,
offense groups 23 through 31,
offense groups 31 through 37.

Matches were made without regard to the housing level at which an inmate was being held,
since when one inmate has a constraining mandatory minimum, he would typically be housed
in a more secure housing level than the matched inmate who lacked the mandatory minimum.
The primary focus of the matching analyses was the difference in offense counts between the
matched inmates with- and without- constraining mandatory minimums. Additional analyses
were run to examine how these differences differ as a function of the preliminary score level.
Longitudinal Analyses:
The longitudinal analyses were similar to the gap analyses, except that there was more than
one year of data on most inmates and each analysis focused on a single additional factor of
interest, outlined in the first Custody question:
-

The impact of time served on behavior among inmates with sentences of 15 to 50
years, who are in the first half dozen years of their sentence,
The impact of time served on behavior among inmates with sentences of 50 or more
years, who are in the first ten years of their sentence,
The impact of years to the minimum earliest parole date (MEPD) on behavior among
inmates with single life sentences, who are within ten years of MEPD, and
The impact of years to MEPD on behavior among inmates with multiple life
sentences, who are within ten years of MEPD.

To get a clear picture of the importance of the factor of interest, these analyses all included
the preliminary score and inmate age, which have been seen to be the most important factors
in predicting inmate behavior.
Summary of Results (gap analyses):
Since the classification score questions were phrased in terms of placement scores rather than
preliminary scores, the first of these analyses looked at the impact of placement score,
housing level in the middle of the review period (hclv), and the difference between the
placement and preliminary score (called ppdiff in the tables). For inmates without a
constraining mandatory minimum, this predictor will be equal to zero, whereas for inmates
with a constraining minimum, this predictor will indicate how much of an “upgrade” was
caused by their mandatory minimum. The mid-review housing level was used, since at the
start of a review period, many inmates will be scheduled to move to a different level, based
on the results of the previous review. Thus the mid-review housing level is more indicative
of where an inmate spent most of his time during the subsequent review period. Although
each of these analyses were restricted to inmates whose placement score was within ±7 of a
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 76
December 19, 2011

Inmate Classification System Study

placement threshold, the sample sizes were still extremely large. For the analysis on the
inmates with placement scores of 19±7, that subset of the data base included data from 36849
reviews. For the analysis run at the 28±7 threshold, the subset included data from 20742
reviews. For the analysis run at the 52±7 threshold, the subset included data from 14446
reviews. When sample sizes are this large, even relatively trivial effects can be highly
significant (i.e., having small p-values), and so the results will be summarized primarily in
terms of estimates of relative effect size (e.g., F-statistics) to facilitate comparisons between
analyses and between predictors.
The following table contains the F-statistics for these three predictors from an analysis that
assumes that the relationship between ppdiff and the likelihood of one or more violation or
the number of violations is linear in the corresponding generalized linear model. The table
contains parameter estimates (slopes) for each of the factors, along with the F-statistic that
was used to test for statistical significance. The F-statistic can be thought of as a signal to
noise ratio, indicating how much larger a given effect is relative to the noise level in the data.
These relationships are all highly significant (p < .0001) except where noted.
Outcome
Any RVR
RVR count
Any RVR
RVR count
Any RVR
RVR count
Any ABCD
ABCD count
Any ABCD
ABCD count
Any ABCD
ABCD count
Any A
A count
Any A
A count
Any A
A count

cutoff
19
19
28
28
52
52
19
19
28
28
52
52
19
19
28
28
52
52

b(placement)
.0204
.0118
.0412
.0357
.0236
.0185
.0260
.0199
.0432
.0442
−.0286
.0218
.0330
.0288
.0246
.0203
-.0010
.0009

F(placement)
14.43 *
9.95 *
69.97
105.73
12.75 *
16.21
20.20
17.13
65.30
107.81
14.42 *
16.34
20.54
19.72
13.63 *
11.03 *
0.01 (ns)
0.01 (ns)

b(ppdiff)
−.0557
−.0546
−.0577
−.0521
−.0300
−.0271
−.0561
−.0531
−.0502
−.0479
−.0334
−.0317
−.0347
−.0420
−.0708
−.0719
−.0252
−.0229

F(ppdiff)
870.58
1378.81
358.13
427.82
209.82
261.76
690.16
870.08
214.30
271.01
170.53
219.47
177.11
298.90
139.84
160.34
73.29
81.14

F(hclv)
67.55
331.98
19.65
48.46
4.78 (ns)
14.31 *
35.26
33.72
35.26
23.34
6.16 (ns)
12.68 *
169.32
233.44
45.92
49.21
10.57 *
11.64 *

In all instances, the F-statistic for the placement/preliminary difference is greater than the
corresponding F-statistic for the placement score itself, often by an order of magnitude or
more. In all of these analyses, the coefficient for the placement score is small and positive,
whereas the coefficient for ppdiff is larger and negative, meaning that for inmates without
mandatory minimums, there’s a slight increase in the likelihood/number of violations as the
preliminary score (equal to the placement score) increases, but for inmates with mandatory
minimums, the greater the discrepancy between the preliminary and placement scores, the
lower the likelihood/number of violations is. For example, for RVR counts near the 19
threshold, an increase of one in the placement score (i.e., for an inmate without a
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 77
December 19, 2011

Inmate Classification System Study

constraining mandatory minimum) would result in an approximately 1% increase in the
expected numbers of RVRs, given that the corresponding slope is around .01. However for
inmates with a constraining mandatory minimum, for each point their preliminary score falls
below 19, there would be an approximately 5% decrease in the expected numbers of RVRs,
given that the slope with respect to ppdiff is around −.05.
The significance of the housing level as a predictor of inmate misbehavior generally takes the
form that the rate of misbehavior increases as the housing level becomes less restrictive. The
question about the form of housing level effects was first raised, some crude analyses were
done based on summary statistics, rather than the full data set. Two of these graphs appear
next, one for the most broadly-defined RVR response (A through F) and the other for A
through D violations. In the first graph (for all serious RVRs), the rate of violations in Level
I (as a function of the preliminary score) is considerably higher than for comparable inmates
housed in higher housing levels:

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 78
December 19, 2011

Inmate Classification System Study

However, when the analysis is restricted to A through D violations, the difference is much
less pronounced, suggesting that the suppression effect in Level II is predominately due to
the least serious (E and F) violations.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 79
December 19, 2011

Inmate Classification System Study

The initial gap analyses all assumed that the relationships with the preliminary or placement
score were linear.
Subsequent analyses fitted smooth, nonlinear curves to those
relationships, to gain insight into the form of the relationships. In most cases, the nonlinear
curves fit significantly better than the linear ones, though this result needs to be viewed in
light of the large sample sizes on which the analyses were based. For example, when looking
at the likelihood of any serious RVRs as a function of the preliminary and placement scores,
along with the housing level, the estimated linear relationships were as follows:

In these graphs, the value old_placement represents an inmate’s placement score at the
beginning of the review period in question. Equivalently, the prior preliminary score is the
preliminary score at the beginning of the review period. These relationships both have strong
linear components to their trends, representing the overall linear trend; for a given
preliminary score, the likelihood of one or more violation decreases with the placement
score, whereas for a given placement score, the likelihood of one or more violations increases
with the preliminary score. At the same time, there are nonlinear aspects of these curves,
most notably that the placement score plot is relatively flat for placement scores below 15,
and that there’s a flattening of the preliminary score plot for scores between 5 and 15. In this
analysis, the significance of the linear and nonlinear components of these trends were as
follows:
Component
Placement (linear)
Placement (nonlinear)
Preliminary (linear)
Preliminary (nonlinear)
Housing level

F-statistic
92.98
10.64
892.00
62.37
163.52

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

p-value
<.0001
.0011
<.0001
<.0001
<.0001
Page 80
December 19, 2011

Inmate Classification System Study

This result is typical of those for other types of RVRs, as well as for the “RVR count”
responses. It’s noteworthy that the F-statistic for the (linear) preliminary score is nearly ten
times as large as for the corresponding F-statistic for the placement score. This can be taken
as evidence that the preliminary score is a more useful (or essential) variable than the
placement score when it’s being used to predict inmate behavior. It’s also noteworthy that
there aren’t any sharp changes in these functions throughout the range of values that are
plotted. This is in part due to the fact that the method is based on splines which are chosen to
be smooth, but it also speaks to the fact that there aren’t any sharp tipping points (scores
where sharp changes in inmate behavior occurs). These observations are typical of these
analyses and supports the conclusion that the preliminary is a better predictor of inmate
behavior than the placement score.
To focus more closely on the impact of preliminary score on behavior, and to simplify these
models, a series of analyses were run on the inmates that have constraining mandatory
minimums and who are housed at the expected level for their placement score. Thus there
are three such analyses: one for inmates with a placement score of exactly 19, housed at
Level II, one for inmates with a placement score of exactly 28, housed at level III, and one
for inmates with a placement score of exactly 52, housed at level IV. The factor that was
examined in these analyses was the preliminary score, which could range from 0 up to the
threshold level in a given analysis. The graphs will be presented for the dichotomous
analyses only, since the analyses of RVR counts are qualitatively very similar.
For A-F RVRs, the 19 cutoff and inmates in housing level II, the estimated curve was:

Here we see the lowest violation rates for the very lowest preliminary scores, increasing to a
plateau around a score of 2 through 15 and then a second increase after that.
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 81
December 19, 2011

Inmate Classification System Study

For A-D RVRs, the 19 cutoff and inmates in housing level II, the estimated violation rates
were:

In this graph, we again see an increase for the inmates with the very lowest preliminary
scores, followed by a long plateau and an increase beyond a preliminary score of around 15.
The corresponding curve for the most serious (A) violations is noisier, due primarily to the
fact that A violations are comparatively rare and so a given sample will contain less
information about those violations.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 82
December 19, 2011

Inmate Classification System Study

The recurring aspect of these curves is that the inmates with very low preliminary scores (say
two or less) are less likely to have violations than inmates with higher scores.
When these analyses are run on the inmates with a placement score of 28, housed at level III,
the curves are quite similar to the first three graphs:
For A-F RVRs, the 28 cutoff and inmates in housing level III,

For A-D RVRs, the 28 cutoff and inmates in housing level III,

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 83
December 19, 2011

Inmate Classification System Study

For A RVRs, the 28 cutoff and inmates in housing level III,

When the analysis is run on inmates with a placement score of 52 in housing level IV, the
graphs are again quite similar:

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 84
December 19, 2011

Inmate Classification System Study

For A-D RVRs, the 52 cutoff and inmates in housing level IV,

Finally, for A RVRs, the 52 cutoff and inmates in housing level IV,

The next extension of these analyses asked whether there were any additional factors, beyond
the preliminary and placement scores and the housing level, that would help predict an
inmate’s likelihood or propensity for accruing violations. The main focus of this analysis
was the administrative determinants that CDCR records for each inmate. The models
included both preliminary and placement scores, the housing level, and one of the
administrative determinants at a time. The following table summarizes those results for each
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 85
December 19, 2011

Inmate Classification System Study

threshold and dependent variable (type of RVR). When a determinant is significantly
significant, the direction of the effect is indicated in a column and colored to help identify
patterns in the data. Cells colored orange through red indicate that an inmate with this
administrative determinant is more likely to offend than one who lacks the determinant. The
opposite is true for cells that are colored green. One set of red cells has been singled out,
corresponding to the administrative determinant for the disciplinary flag (DIS). Since the
determinants were defined at the 840 review at the end of a review period, having this flag
set could indicate misbehavior within that review period, and so this likely isn’t a predictive
relationship. The fact that this flag is highly significant signifies little; just that inmates who
have this administrative determinant are apt to have misbehaved recently.
The same isn’t true for the other determinants, which aren’t likely to have changed in the
current review period due to events that transpired during that period, and so these are more
likely to be interpretable as predictors of behavior. Among those determinants, several that
stand out are that the enemies flag is typically predictive of improved behavior, as are the
flags for life sentences, sex offender status and psychological conditions. For the other flags,
the administrative determinants are either insignificant or else there isn’t a clear pattern of
being a positive (or negative) risk factor for rules violations. The table includes the results
for the determinants for discipline problems, life sentences, sex offender status and
psychological conditions, along with the flag for age, which is included as a representative
determinant that isn’t strongly associated with behavioral changes.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 86
December 19, 2011

Inmate Classification System Study

change due to
determinant

threshold

determinant

type

(values for DIS)

19

AGE

RVR

0.245467

19

AGE

RVR

0.375682

28

AGE

RVR

1.414904

0.234244

28

AGE

RVR

0.555453

1.974283

0.159994

52

AGE

RVR

0.517954

2.945305

0.086127

52

AGE

RVR

1

1.403091

8.108672

0.004405

-75.42

19

AGE

ABCD

0

1

1.441947

10.2663

0.001355

-76.35

19

AGE

ABCD

ad_AGE

0

1

0.210367

0.395886

0.529222

28

AGE

ABCD

count

ad_AGE

0

1

0.176071

0.433865

0.510098

28

AGE

ABCD

anyRVR

ad_AGE

0

1

1.045863

3.696161

0.054538

52

AGE

ABCD

count

ad_AGE

0

1

1.093033

5.272633

0.021663

52

AGE

ABCD

anyRVR

ad_AGE

0

0

20.2857

19

AGE

A

anyRVR

ad_AGE

0

1

-0.25667

0.553284

0.456979

28

AGE

A

count

ad_AGE

0

1

-0.20361

0.39316

0.530643

28

AGE

A

anyRVR

ad_AGE

0

1

-0.11978

0.077978

0.780056

52

AGE

A

count

ad_AGE

0

1

-0.1137

0.088164

0.766524

52

AGE

A

anyRVR

ad_DIS

0

1

-0.92437

202.3492

6.41E-46

152.03

19

DIS

RVR

count

ad_DIS

0

1

-0.77535

335.7295

5.43E-75

117.14

19

DIS

RVR

anyRVR

ad_DIS

0

1

-0.71844

95.15115

1.76E-22

105.12

28

DIS

RVR

dependent

Parameter

Level1

DF

Estimate

ChiSq

ProbChiSq

anyRVR

ad_AGE

0

1

0.186849

0.534113

0.464883

count

ad_AGE

0

1

0.241441

1.348921

anyRVR

ad_AGE

0

1

0.27132

0.784785

count

ad_AGE

0

1

0.272303

anyRVR

ad_AGE

0

1

count

ad_AGE

0

1

anyRVR

ad_AGE

0

count

ad_AGE

anyRVR

change due to
determinant

-66.48

count

ad_DIS

0

1

-0.569

159.1013

1.78E-36

76.65

28

DIS

RVR

anyRVR

ad_DIS

0

1

-0.27289

21.451

3.63E-06

31.38

52

DIS

RVR

count

ad_DIS

0

1

-0.21145

29.43909

5.77E-08

23.55

52

DIS

RVR

anyRVR

ad_DIS

0

1

-1.45314

694.3488

5.1E-153

327.65

19

DIS

ABCD

count

ad_DIS

0

1

-1.33063

1066.323

6.9E-234

278.34

19

DIS

ABCD

anyRVR

ad_DIS

0

1

-1.06

253.6683

4.12E-57

188.64

28

DIS

ABCD

count

ad_DIS

0

1

-0.97912

448.4227

1.6E-99

166.21

28

DIS

ABCD

anyRVR

ad_DIS

0

1

-0.57021

99.19824

2.28E-23

76.86

52

DIS

ABCD

count

ad_DIS

0

1

-0.4891

136.5437

1.52E-31

63.08

52

DIS

ABCD

anyRVR

ad_DIS

0

1

-3.22514

7693.384

0

2415.70

19

DIS

A

count

ad_DIS

0

1

-3.26715

10748.81

0

2523.66

19

DIS

A

anyRVR

ad_DIS

0

1

-2.89395

3809.144

0

1706.46

28

DIS

A

count

ad_DIS

0

1

-2.90039

5211.892

0

1718.12

28

DIS

A

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 87
December 19, 2011

Inmate Classification System Study

anyRVR

ad_DIS

0

1

-2.11796

1606.988

0

731.42

52

DIS

A

count

ad_DIS

0

1

-2.02903

1947.586

0

660.67

52

DIS

A

anyRVR

ad_ENE

0

1

0.421813

172.4046

2.21E-39

-34.41

19

ENE

RVR

count

ad_ENE

0

1

0.355491

205.1835

1.54E-46

-29.92

19

ENE

RVR

anyRVR

ad_ENE

0

1

0.445228

123.3197

1.19E-28

-35.93

28

ENE

RVR

count

ad_ENE

0

1

0.336955

136.8361

1.31E-31

-28.61

28

ENE

RVR

anyRVR

ad_ENE

0

1

0.188371

18.51348

1.69E-05

-17.17

52

ENE

RVR

count

ad_ENE

0

1

0.151701

25.18232

5.22E-07

-14.08

52

ENE

RVR

anyRVR

ad_ENE

0

1

0.328474

87.61986

7.93E-21

-28.00

19

ENE

ABCD

count

ad_ENE

0

1

0.322397

118.5441

1.32E-27

-27.56

19

ENE

ABCD

anyRVR

ad_ENE

0

1

0.385402

76.88188

1.81E-18

-31.98

28

ENE

ABCD

count

ad_ENE

0

1

0.330677

88.77737

4.42E-21

-28.16

28

ENE

ABCD

anyRVR

ad_ENE

0

1

0.14839

10.24223

0.001373

-13.79

52

ENE

ABCD

count

ad_ENE

0

1

0.098557

7.758101

0.005347

-9.39

52

ENE

ABCD

anyRVR

ad_ENE

0

1

0.637895

179.0877

7.67E-41

-47.16

19

ENE

A

count

ad_ENE

0

1

0.684234

232.1492

2.03E-52

-49.55

19

ENE

A

anyRVR

ad_ENE

0

1

0.807503

156.9522

5.24E-36

-55.40

28

ENE

A

count

ad_ENE

0

1

0.809659

187.8981

9.14E-43

-55.50

28

ENE

A

anyRVR

ad_ENE

0

1

1.145488

236.9446

1.82E-53

-68.19

52

ENE

A

count

ad_ENE

0

1

1.133506

289.9282

5.15E-65

-67.81

52

ENE

A

anyRVR

ad_LIF

0

1

0.235131

49.82789

1.68E-12

-20.95

19

LIF

RVR

count

ad_LIF

0

1

0.213429

66.044

4.41E-16

-19.22

19

LIF

RVR

anyRVR

ad_LIF

0

1

0.552334

133.4498

7.21E-31

-42.44

28

LIF

RVR

count

ad_LIF

0

1

0.43919

159.1896

1.7E-36

-35.54

28

LIF

RVR

anyRVR

ad_LIF

0

1

0.397849

75.58947

3.49E-18

-32.82

52

LIF

RVR

count

ad_LIF

0

1

0.341688

116.2925

4.1E-27

-28.94

52

LIF

RVR

anyRVR

ad_LIF

0

1

0.153155

17.38394

3.05E-05

-14.20

19

LIF

ABCD

count

ad_LIF

0

1

0.079238

6.632196

0.010015

-7.62

19

LIF

ABCD

anyRVR

ad_LIF

0

1

0.378319

54.72549

1.39E-13

-31.50

28

LIF

ABCD
ABCD

count

ad_LIF

0

1

0.354136

74.28377

6.77E-18

-29.82

28

LIF

anyRVR

ad_LIF

0

1

0.399183

67.28713

2.35E-16

-32.91

52

LIF

ABCD

count

ad_LIF

0

1

0.360212

92.29452

7.47E-22

-30.25

52

LIF

ABCD

anyRVR

ad_LIF

0

1

-0.00924

0.042946

0.835827

19

LIF

A

count

ad_LIF

0

1

0.01969

0.220783

0.638444

19

LIF

A

anyRVR

ad_LIF

0

1

0.158662

6.466906

0.01099

28

LIF

A

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

-14.67
Page 88
December 19, 2011

Inmate Classification System Study

count

ad_LIF

0

1

0.218946

14.03088

0.00018

anyRVR

ad_LIF

0

1

0.088974

2.292347

0.130013

-19.66

28

LIF

A

52

LIF

A
A

count

ad_LIF

0

1

0.072017

1.897612

0.168346

52

LIF

anyRVR

ad_PRE

0

1

-0.49587

7.120704

0.00762

64.19

19

PRE

RVR

count

ad_PRE

0

1

-0.37527

7.339193

0.006747

45.54

19

PRE

RVR

anyRVR

ad_PRE

0

1

-0.20314

0.688578

0.406648

28

PRE

RVR

count

ad_PRE

0

1

-0.30215

3.655997

0.055868

28

PRE

RVR

anyRVR

ad_PRE

0

1

-0.13267

0.139963

0.708318

52

PRE

RVR

count

ad_PRE

0

1

-0.07816

0.103379

0.747812

52

PRE

RVR

anyRVR

ad_PRE

0

1

-0.43706

4.967351

0.02583

54.81

19

PRE

ABCD

count

ad_PRE

0

1

-0.35376

4.596965

0.032029

42.44

19

PRE

ABCD

anyRVR

ad_PRE

0

1

0.260641

0.702176

0.402054

28

PRE

ABCD

count

ad_PRE

0

1

0.231985

0.847027

0.357395

28

PRE

ABCD

anyRVR

ad_PRE

0

1

0.086645

0.046362

0.829519

52

PRE

ABCD

count

ad_PRE

0

1

-0.06367

0.047522

0.827433

52

PRE

ABCD

anyRVR

ad_PRE

0

1

-0.71388

12.13194

0.000496

104.19

19

PRE

A

count

ad_PRE

0

1

-0.6432

11.06682

0.000879

90.26

19

PRE

A

anyRVR

ad_PRE

0

0

19.68242

28

PRE

A

anyRVR

ad_PRE

0

1

-1.01588

10.84638

0.00099

176.18

52

PRE

A

count

ad_PRE

0

1

-0.97076

12.77746

0.000351

163.99

52

PRE

A

anyRVR

ad_PSY

0

1

0.02243

0.198069

0.656284

19

PSY

RVR

count

ad_PSY

0

1

0.020077

0.263175

0.607947

19

PSY

RVR

anyRVR

ad_PSY

0

1

0.120102

3.285906

0.069877

28

PSY

RVR

count

ad_PSY

0

1

0.078467

2.755422

0.096925

28

PSY

RVR

anyRVR

ad_PSY

0

1

0.047442

0.420455

0.51671

52

PSY

RVR

count

ad_PSY

0

1

0.028007

0.308828

0.5784

52

PSY

RVR

anyRVR

ad_PSY

0

1

0.057539

1.060723

0.30305

19

PSY

ABCD

count

ad_PSY

0

1

0.023801

0.261571

0.609043

19

PSY

ABCD

anyRVR

ad_PSY

0

1

0.139085

3.659569

0.055748

28

PSY

ABCD
ABCD

count

ad_PSY

0

1

0.127667

4.787736

0.028663

28

PSY

anyRVR

ad_PSY

0

1

-0.01132

0.021699

0.882891

52

PSY

ABCD

count

ad_PSY

0

1

-0.03807

0.429168

0.512398

52

PSY

ABCD

anyRVR

ad_PSY

0

1

0.750281

65.59549

5.54E-16

-52.78

19

PSY

A

count

ad_PSY

0

1

0.767845

76.16955

2.6E-18

-53.60

19

PSY

A

anyRVR

ad_PSY

0

1

0.659025

31.66434

1.83E-08

-48.26

28

PSY

A

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

-11.99

Page 89
December 19, 2011

Inmate Classification System Study

count

ad_PSY

0

1

0.701776

40.3039

2.17E-10

-50.43

28

PSY

A

anyRVR

ad_PSY

0

1

0.358236

10.52989

0.001175

-30.11

52

PSY

A

count

ad_PSY

0

1

0.365937

13.69708

0.000215

-30.65

52

PSY

A

anyRVR

ad_SEX

0

1

0.484565

219.9772

9.15E-50

-38.40

19

SEX

RVR

count

ad_SEX

0

1

0.465668

320.6919

1.02E-71

-37.23

19

SEX

RVR

anyRVR

ad_SEX

0

1

0.640571

181.1008

2.79E-41

-47.30

28

SEX

RVR

count

ad_SEX

0

1

0.51587

215.3073

9.55E-49

-40.30

28

SEX

RVR

anyRVR

ad_SEX

0

1

0.417235

53.50522

2.58E-13

-34.11

52

SEX

RVR

count

ad_SEX

0

1

0.34879

72.70572

1.5E-17

-29.45

52

SEX

RVR

anyRVR

ad_SEX

0

1

0.617968

267.1995

4.63E-60

-46.10

19

SEX

ABCD

count

ad_SEX

0

1

0.637544

386.8886

3.94E-86

-47.14

19

SEX

ABCD

anyRVR

ad_SEX

0

1

0.704491

163.3875

2.06E-37

-50.56

28

SEX

ABCD

count

ad_SEX

0

1

0.624111

195.7221

1.79E-44

-46.43

28

SEX

ABCD

anyRVR

ad_SEX

0

1

0.391222

39.93288

2.63E-10

-32.38

52

SEX

ABCD

count

ad_SEX

0

1

0.333664

47.98699

4.29E-12

-28.37

52

SEX

ABCD

anyRVR

ad_SEX

0

1

0.877655

294.4237

5.4E-66

-58.42

19

SEX

A

count

ad_SEX

0

1

0.897038

344.6077

6.33E-77

-59.22

19

SEX

A

anyRVR

ad_SEX

0

1

1.120464

167.3657

2.78E-38

-67.39

28

SEX

A

count

ad_SEX

0

1

1.124837

200.1309

1.96E-45

-67.53

28

SEX

A

anyRVR

ad_SEX

0

1

1.15244

117.4707

2.26E-27

-68.41

52

SEX

A

count

ad_SEX

0

1

1.14384

143.7162

4.1E-33

-68.14

52

SEX

A

anyRVR

ad_VIO

0

1

0.166736

25.6272

4.14E-07

-15.36

19

VIO

RVR

count

ad_VIO

0

1

0.150357

33.90585

5.78E-09

-13.96

19

VIO

RVR

anyRVR

ad_VIO

0

1

0.197647

18.29368

1.89E-05

-17.93

28

VIO

RVR

count

ad_VIO

0

1

0.150351

20.68565

5.41E-06

-13.96

28

VIO

RVR

anyRVR

ad_VIO

0

1

0.09686

3.685351

0.054892

52

VIO

RVR

count

ad_VIO

0

1

0.122933

12.17693

0.000484

-11.57

52

VIO

RVR

anyRVR

ad_VIO

0

1

0.240521

41.75152

1.04E-10

-21.38

19

VIO

ABCD

count

ad_VIO

0

1

0.179189

33.0422

9.02E-09

-16.41

19

VIO

ABCD

anyRVR

ad_VIO

0

1

0.230273

20.45743

6.1E-06

-20.57

28

VIO

ABCD

count

ad_VIO

0

1

0.201814

24.55987

7.2E-07

-18.28

28

VIO

ABCD

anyRVR

ad_VIO

0

1

0.141901

6.864548

0.008792

-13.23

52

VIO

ABCD

count

ad_VIO

0

1

0.133503

10.20363

0.001402

-12.50

52

VIO

ABCD

anyRVR

ad_VIO

0

1

0.321159

47.4671

5.59E-12

-27.47

19

VIO

A

count

ad_VIO

0

1

0.357401

66.08237

4.32E-16

-30.05

19

VIO

A

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 90
December 19, 2011

Inmate Classification System Study

anyRVR

ad_VIO

0

1

0.234213

12.501

0.000407

-20.88

28

VIO

A

count

ad_VIO

0

1

0.303689

24.0054

9.61E-07

-26.19

28

VIO

A

anyRVR

ad_VIO

0

1

0.155767

5.23188

0.022177

-14.42

52

VIO

A

count

ad_VIO

0

1

0.176064

8.330335

0.003899

-16.14

52

VIO

A

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 91
December 19, 2011

Inmate Classification System Study

The next analyses addressed custody question 4 by considering a range of additional
predictors of RVRs in an attempt to build a more a more complete model of the likelihood
and numbers of violations, to see which predictors combine to make a model that represents
an improvement on the gap models that have been presented so far. A “short list” of
predictors was developed which have been found in other settings to be useful predictors of
in-prison behavior. These included:
Housing category (primarily dormitory versus cells)
Inmate age, fit using spline methods to account for possible nonlinearity,
Offense category (violent, property, sex-related, drug-related and other),
Ethnicity,
Mental health code,
Risk level,
History of serious/violent violations,
Sentence type (determinant-driven, 2nd strike, 3rd strike, life, death and life w/o
parole),
Sex offender status,
Street gang affiliation,
Age at first arrest (categorical),
Years to earliest release, fit using spline methods to account for possible nonlinearity,
and Sentence length, fit using spline methods to account for possible nonlinearity,
Each of these predictors were used in conjunction with the preliminary and placement scores
and the mid-review housing level in models to predict likelihood/numbers of RVRs in three
categories (A, A through D and A through F) and at each of the three cutoffs (19, 28 and 52).
A “consensus” stepwise selection procedure was used to build models to predict these
outcomes. The other factor used to guide model selection was whether a given predictor
could be used in practice for classifying inmates, due to legal concerns or concerns about
public safety. The only way in which this circumvented the normal stepwise selection
procedure was that sex offender status was found to be a highly significant predictor of
inmate behavior at an early stage, with sex offenders significantly less likely to accrue
violations. Since the assignment of sex offenders is also governed by public safety concerns,
lest these inmates be given an opportunity to escape, this predictor was passed over in the
selection process. In order, the predictors selected were as follows, along with a general
description of their impact on the predictive models:
Age was the first predictor added to the model. After adjusting for preliminary and
placement scores and housing level, the likelihood/number of violations decreases with age,
and in a fashion that’s very nearly linear, at roughly 2 to 4% for each additional year of age,
depending on the group of inmates and the severity of violation being analyzed. Since age is
one of the factors considered in the preliminary score, this would indicate that for this
purpose, that score under-adjusts for age.
The offense category was the predictor added to the model at the second step. The strongest
difference here was that inmates convicted on sex-related offenses have lower violation rates,
in some cases up to 50% lower, after adjusting for the other factors in the model. As
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 92
December 19, 2011

Inmate Classification System Study

mentioned, this may not be seen as a reason to house these inmates in less secure housing,
but it’s important to include this predictor in a model in order to get a clearer picture of the
importance of other potential predictors.
The mental health code was the predictor added to the model at the third step. Inmates with
CCCMS and EOP designations can be roughly 30-40% more likely to accrue violations, with
the exception of when the analysis is restricted to A violations. As is often the case, A
violations are comparatively quite rare and so the analyses restricted to A violations only
don’t allow as powerful inferences as analyses that are based on more broadly-defined
violations.
Following these three factors, the analyses became less consistent, with significance often not
extending across different threshold levels or for different groups of violations. There may
be additional factors that have a genuine but less pronounced impact on violation rates, but
these effects are relative minor compared to the factors added at the first three steps in the
model-building process (age, offense category and mental health code).
Finally, the second custody questions asked specifically about the impact of housing
category, most notably cell versus dorm housing, on violation rates. A set of analyses were
run looking at the impact of the preliminary and placement scores, housing level, custody
level and housing category on the usual range of outcome variables. The results here were
mixed. At the 19 cutoff, there were no significant differences in violation rates according to
housing category. At the 28 cutoff, the rate of A through F violations in celled housing was
estimated to be roughly 15% lower than in dormitory housing after adjusting for the other
effects in the model, and roughly 25% lower in celled housing when focusing on A through
D violations. At the 52 cutoff, the rate of A through F violations in celled housing was
estimated to be roughly 27% lower in celled housing and for A through D violations, it was
roughly 32% lower.
An additional question that was raised was whether the reason that for an inmate’s mandatory
minimum was predictive of in-prison behavior. The following table gives the frequency of
each reason for the inmates used in the gap analysis for each of the three placement score
cutoffs.
Group min.

Description

A
B
C
D
E
F
G
H

condemned
LWOP
CCR 3375.2
escape risk
“R” suffix
violence exclusion
public interest case
other life sentence
No mandatory score

52
52
28
19
19
19
19
19
0

cutoff = 19
n
%

cutoff = 28
n
%

517
12194
16074
11
2489
5564

3345
238
3946
6749
5
608
5851

1.40
33.09
43.62
0.03
6.75
15.10

16.13
1.15
19.02
32.54
0.02
2.93
28.21

cutoff = 52
n
%
11
0.08
2469
17.09
326
2.26
179
1.24
2251
15.58
5722
39.61
5
0.03
519
3.59
2964
20.52

These models are similar to the other gap analyses. They all include preliminary score and
housing level as factors in the model, so the comparisons are adjusted as if inmates with and
without a given reason were in the same housing level. Thus the normal consequence of
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 93
December 19, 2011

Inmate Classification System Study

having one of these determinants would be a change in housing level, but these analyses look
at whether they were worse (or better) behaved apart from this consequence.
The following table contains an identifier for each analysis, the F-statistic and p-value for
testing whether inmates are equally likely to accrue violations, regardless of the reason for
their mandatory minimum (or whether they have one), along with pairwise comparisons
between each of the listed reasons for a mandatory minimum and the group of inmates who
have no mandatory minimum. It should be noted that since each mandatory minimum has an
associated placement score, not all of these minimum are what we’ve called constraining
mandatory minimum. For example, the categories D through H (outlined above) each have a
mandatory score of 19 and at the 28 or 52 cutoff, these would not constrain an inmate’s
placement.
Dependent/cutoff
Any RVR/19

RVR count/19

Any RVR/28

RVR count/28

Any RVR/52

overall F/p
level
Δ(level,no mandatory)
54.36 / <.0001 D (escape)
+21.5%
E (R suffix)
-47.6%
F (violence)
-21.2%
G (pub. interest)
+105%
H (other life)
-18.0%
90.85 / <.0001 D
+16.2%
E
-48.0%
F
-22.7%
G
+74.1%
H
-20.0%
40.37 / <.0001 C (Life inmate)
-50.5%
D
-19.3%
E
-53.7%
F
-20.8%
G
+64.0%
H
-35.5%
47.69 / <.0001 C
-41.8%
D
-16.0%
E
-44.6%
F
-16.5%
G
+152%
H
-25.8%
12.67 / <.0001 A (condemn.)
+126%
B (LWOP)
-48.6%
C
-51.4%
D
-22.3%
E
-43.1%
F
-26.4%
G
+12.4%
H
-24.3%

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

p-value
.0647
<.0001
<.0001
.2690
.0037
.0400
<.0001
<.0001
.2325
<.0001
<.0001
.1642
<.0001
<.0001
.6097
<.0001
<.0001
.1017
<.0001
<.0001
.0371
<.0001
.2205
<.0001
<.0001
.1735
<.0001
<.0001
.9050
.0128

Page 94
December 19, 2011

Inmate Classification System Study

RVR count/52

Any ABCD/19

Any ABCD/28

Any ABCD/52

Any A/19

Any A/28

Any A/52

18.72 / <.0001 A
B
C
D
E
F
G
H
61.27 / <.0001 D
E
F
G
H
27.05 / <.0001 C
D
E
F
G
H
9.93 / <.0001 A
B
C
D
E
F
G
H
59.03 / <.0001 D
E
F
G
H
54.37 / <.0001 C
D
E
F
G
H
29.58 / <.0001 A
B
C
D

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

+102%
-42.9%
-34.8%
-6.3%
-37.7%
-23.1%
-25.9%
-22.6%
+19.6%
-49.2%
-16.1%
+253%
+5.5%
-38.0%
-10.8%
-50.3%
-15.1%
+67.1%
-2.3%
+17.0%
-47.1%
-38.9%
-20.0%
-42.6%
-26.4%
-11.6%
-27.5%
+11.4%
-47.5%
+15.3%
-100%
+23.9%
-7.7%
-65.8%
-68.5%
+29.4%
-100%
-9.8%
+993%
+84.9%
-10.2%
+7.0%

.0519
<.0001
<.0001
.5767
<.0001
<.0001
.6712
.0007
.1060
<.0001
.0002
.0219
.4471
<.0001
.4828
<.0001
.0003
.5869
.8149
.8314
<.0001
.0014
.2484
<.0001
<.0001
.9038
.0065
.4782
<.0001
.0231
.9977
.0170
.4956
.0017
<.0001
<.0001
.9987
.4519
<.0001
<.0001
.5909
.7881

Page 95
December 19, 2011

Inmate Classification System Study

E
F
G
H

-64.2%
+42.4%
-100%
+0.6%

<.0001
<.0001
.9989
.9680

Summary of Results (matching analyses):
When inmates were matched on their preliminary score, ethnicity, mental health code, the
administrative determinants for gang affiliation, psychological condition and sex offender
status, there were a total of 6568 unique matched pairs that had complete agreement on these
criteria. Each matched pair contained one inmate with a constraining mandatory minimum
and one without one. The first part of this analysis examined the difference between RVR
characteristics (and age) between the two groups. The results were as follows:
Dependent
Any RVR
RVR count
Any A-D
A-D count
Any A
A count
Age

mean difference
−.073
−.143
−.009
−.012
.004
.003
3.876

t-statistic
−9.73
−10.949
−1.707
−1.885
1.749
1.330
30.31

p-valueComments:
<.0001 higher in non-mandatory group
<.0001 higher in non-mandatory group
.0879
.0595
.0803
.1834
<.0001 higher in mandatory min. group

As can be seen, the likelihood and count of broadly-defined RVRs were higher within the
non-mandatory group, and the comparison for A through D RVRs showed higher numbers in
the non-mandatory group as well, though these results were marginally insignificant. For A
violations, the numbers were insignificantly higher in the mandatory minimum group. The
mandatory minimum group was also significantly older than the no-minimum group (by
about 4 years), which may tie into these differences. These differences are hard to interpret,
since due to the standard way of assigning inmates to housing, most of these pairs would
contain a no-mandatory inmate at one housing level and a “mandatory minimum” inmate
housed at a more secure housing level. Thus the mandatory/no-mandatory differences are
confounded with possible differences due to housing level, as well as with age.
Analyses were also run that compared the likelihood or count of RVRs of a given type in
matched inmates, as a function of the matched preliminary score. These analyses were run
using generalized additive models within each of the groups. The following graphs
summarize those results for this group of matched inmates, focusing on the “Any RVR”
outcomes, rather than the “RVR count” outcomes. The set of graphs for the count outcomes
were extremely similar to the ones presented here.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 96
December 19, 2011

Inmate Classification System Study

Any RVR’s:

Any A-D violation:

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 97
December 19, 2011

Inmate Classification System Study

Any A violation:

In the first two of these graphs, the likelihood of RVRs is higher for inmates without
mandatory minimums throughout most of the range of preliminary scores. By contrast, the
two lines are quite similar for most of the range when focusing entirely on A violations.
A second set of analyses was run in which the offense category was added as another
matching criterion. When this criterion was added, the number of exact matched reduced to
4700.
The results in most ways were quite similar to the first set of matching analyses:
Dependent
Any RVR
RVR count
Any A-D
A-D count
Any A
group
A count
Age

mean difference
−.080
−.150
−.013
−.015
.006

t-statistic
−8.82
−9.23
−1.943
−1.882
2.496

p-valueComments:
<.0001 higher in non-mandatory group
<.0001 higher in non-mandatory group
.0521
.0599
.0126 slightly higher in mandatory

.005
5.993

1.901
32.77

.0573
<.0001 higher in mandatory min. group

The only qualitative difference here is that the slight increase in A violations in the
mandatory minimum group is now marginally significant rather than marginally
insignificant.
Similar patterns are also seen in the graphs that plot the likelihood of various types of
violation in each group, as a function of the preliminary score:

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 98
December 19, 2011

Inmate Classification System Study

Any RVRs:

Any A-D RVRs:

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 99
December 19, 2011

Inmate Classification System Study

Any A violation:

For both A-F and A-D RVRs, there’s a higher likelihood of an RVR in the no-mandatory
group than for the mandatory minimum group. Unlike the first set of graphs, the lines cross
for very low preliminary scores. This is an artifact of the matched sample, since for a
preliminary score of zero, there were only 27 matches, and housing assignments were quite
irregular among those inmates, with most of the mandatory minimum inmates being housed
in Level I, while most of the no-mandatory inmates were housed in Level II. Other than this
artifact, the most notable difference relative to the first set of graphs is that the difference due
to having a mandatory minimum (lower likelihood of violations) in the A-D graph doesn’t
become apparent until a preliminary score of around 20. Below that level, the two groups are
quite comparable.
As has been mentioned previously, the interpretation of the matching analyses is
problematical, since these comparisons often involve inmates who have identical preliminary
scores but who are housed at different housing levels.
Summary of Results (longitudinal analyses):
The longitudinal analyses were based on subsets of data that were developed, focusing on
four subgroups of inmates:
1.
2.
3.
4.

Inmates with sentences of 15 to 50 years (11252 reviews on 6168 inmates),
Inmates with sentences of 50+ years (391 reviews on 217 inmates),
Inmates with single life sentences (40292 reviews on 20573 inmates), and
Inmates with multiple life sentences (14194 reviews on 6729 inmates),

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 100
December 19, 2011

Inmate Classification System Study

who were followed over as many past years as possible, given that most types of data aren’t
available before a point in time where they started being compiled electronically. The key
variables in these analyses included two of the ones used in previous analyses (preliminary
score, housing level), along with custody level and at least one additional variable, either the
number of years than an inmate had served to date, or else the number of years until the
inmate’s minimum (earliest possible) parole date (MEPD). These analyses used generalized
additive models to adjust for all of the factors except the one additional variable, since the
question of interest was how behavior changed over time with respect to that additional
factor. These analyses included the individual inmate as a random factor in the analyses,
since over half of the inmates were represented by more than one 840 review in these
analyses.
As with the other analyses that fit nonlinear functions to responses, the results can be
summarized in terms of the significance of the various effects, but given the sample size, the
form of an effect is perhaps more to the point than the significance. Thus while a table of pvalues will be presented, the accompanying graphs are more germane in getting at the
questions that had been posed. The following table contains all of those p-values. In the
table, the significance of each model effect is summarized in terms of an F-statistic and the
corresponding p-value.
Significance of effects in longitudinal analyses
Data set public15
Dependent
anyRVR
RVRcount
anyABCD
ABCDcount
anyA
Acount

hclv
2.83
.0592
5.47
.0042
0.57
.5628
0.71
.4931
0.00
.9952
0.02
.9795

clvl
9.45
<.0001
15.88
<.0001
11.02
<.0001
16.47
<.0001
13.37
.1977
14.80
.0028

prelim(linear)
96.85
<.0001
100.54
<.0001
86.80
<.0001
90.83
<.0001
44.59
<.0001
40.64
<.0001

prelim(spline)
21.07
<.0001
25.59
<.0001
11.02
.0009
14.01
.0002
6.11
.0135
6.22
.0126

years(linear)
6.94
.0084
5.27
.0217
3.51
.0609
2.19
.1391
0.02
.8844
0.01
.9154

years(spline)
5.05
.0247
6.50
.0108
4.09
.0431
3.35
.0671
13.13
.0003
12.66
.0004

clvl
0.06
.8022
2.71
.1008
0.00
.9779
2.40

prelim(linear)
9.66
.0020
12.76
.0004
4.39
.0369
7.19

prelim(spline)
9.07
.0028
13.14
.0003
8.91
.0030
10.10

years(linear)
0.09
.7699
0.69
.4081
0.55
.4574
0.00

years(spline)
3.41
.0657
3.76
.0534
4.09
.0440
3.86

Data set public50
Dependent
anyRVR
RVRcount
anyABCD
ABCDcount

hclv
0.09
.9174
0.02
.9912
0.97
.3789
0.87

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 101
December 19, 2011

Inmate Classification System Study

anyA
Acount

.4206
1.26
.2836
0.01
.9852

.1222
3.80
.0521
0.05
.8257

.0077
6.98
.0086
0.02
.8989

.0016
116.59
<.0001
0.01
.9365

.9945
2.48
.1165
0.02
.8925

.0501
85.74
<.0001
0.01
.9026

prelim(spline)
117.04
<.0001
125.93
<.0001
33.95
<.0001
43.56
<.0001
23.48
<.0001
21.35
<.0001

years(linear)
12.70
.0004
9.08
.0026
1.79
.1811
1.32
.2501
0.81
.3668
0.61
.4335

years(spline)
42.12
<.0001
37.76
<.0001
14.29
.0002
21.76
<.0001
7.62
.0058
8.13
.0044

prelim(spline)
50.98
<.0001
42.19
<.0001
18.26
<.0001
19.00
<.0001
12.61
.0004
10.98
.0009

MEPD(linear)
3.86
.0496
0.65
.4218
0.21
.6459
0.06
.8045
3.66
.0557
3.67
.0553

MEPD(spline)
9.37
.0002
4.23
.0396
6.37
.0116
5.10
.0239
5.42
.0199
4.67
.0306

prelim(spline)
29.19
<.0001
25.61
<.0001
16.46
<.0001
13.05

years(linear)
9.99
.0016
4.11
.0425
5.88
.0153
2.41

years(spline)
17.23
<.0001
9.69
.0019
12.55
.0004
7.04

Data set public single (life sentences)
Dependent
anyRVR
RVRcount
anyABCD
ABCDcount
anyA
Acount

hclv
7.25
.0007
15.64
<.0001
3.93
.0197
6.55
.0014
3.75
.0235
4.00
.0182

clvl
8.32
.0002
3.87
.0208
4.30
.0136
2.00
.1350
3.13
.0438
3.00
.0497

prelim(linear)
279.91
<.0001
304.93
<.0001
171.72
<.0001
202.52
<.0001
36.28
<.0001
34.98
<.0001

Data set public single (life sentences)
Dependent
anyRVR
RVRcount
anyABCD
ABCDcount
anyA
Acount

hclv
3.60
.0273
6.55
.0014
2.35
.0955
2.53
.0798
1.37
.2543
1.53
.2157

clvl
11.28
<.0001
6.42
.0016
5.15
.0058
3.61
.0270
2.32
.0978
2.11
.1216

prelim(linear)
150.12
<.0001
141.33
<.0001
82.22
<.0001
93.54
<.0001
15.89
<.0001
14.49
.0001

Data set public multiple (life sentences)
Dependent
anyRVR
RVRcount
anyABCD
ABCDcount

hclv
4.22
.0147
4.91
.0074
2.94
.0530
3.15

clvl
2.51
.0814
1.50
.2232
2.11
.1216
1.45

prelim(linear)
98.40
<.0001
93.77
<.0001
89.59
<.0001
81.89

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 102
December 19, 2011

Inmate Classification System Study

anyA
Acount

.0427
0.42
.6570
0.44
.6447

.2353
2.33
.0976
1.99
.1369

<.0001
26.76
<.0001
26.02
<.0001

.0003
8.34
.0039
7.54
.0060

.1204
0.15
.6948
0.04
.8355

.0080
6.86
.0088
7.00
.0081

prelim(spline)
7.02
.0081
10.77
.0010
3.21
.0733
3.51
.0612
5.93
.0149
7.78
.0053

MEPD(linear)
0.31
.5780
0.42
.5147
0.40
.5292
0.38
.5355
1.02
.3127
0.89
.3467

MEPD(spline)
2.91
.0882
2.44
.1180
2.21
.1373
1.99
.1586
7.43
.0065
7.69
.0056

Data set public multiple (life sentences)
Dependent
anyRVR
RVRcount
anyABCD
ABCDcount
anyA
Acount

hclv
1.47
.2289
1.45
.2343
0.71
.4915
0.78
.4600
1.88
.1522
1.93
.1451

clvl
1.48
.2286
1.57
.2091
0.53
.5911
0.49
.6149
1.38
.2521
1.37
.2538

prelim(linear)
17.60
<.0001
21.07
<.0001
8.68
.0032
9.56
.0020
2.85
.0914
2.91
.0879

From this table, it’s clear that the single best predictor of inmate behavior is still the
preliminary score and so the impact of the other predictors need to be interpreted in the
context provided by an inmate’s preliminary score. Thus an inmate with a low preliminary
score will pose a minimal risk of misbehavior even if some of the other factors indicate a
slight elevation in the risk of a violation. Moreover, the preliminary scores for most
prisoners decrease over time, and because these models include preliminary score as a
predictor, the effect of years in prison are viewed conditional on the preliminary score,
meaning that it’s as if the preliminary score isn’t changing. The next graph looks at the
relationship between preliminary score and years served in the group serving 15-50 year
sentences, demonstrating this decreasing pattern in preliminary scores:

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 103
December 19, 2011

Inmate Classification System Study

Until the sample sizes get particularly small (for inmates more than five years into their
sentence), the relationship is nearly linear, with the score decreasing on the average by about
two points per year. The patterns in the other three subgroups are quite similar to this one.
The following graphs are being presented to convey the form that the relationship between
violations and either the years served in a sentence or the years to earliest parole (MEPD).
Since these are complicated models, they have to assume something about what the values of
the other predictors are. Because these are fairly arbitrary assumptions (and not germane
when looking at the form of a predictor’s impact), these are labeled “comparative plots”. To
keep the number of graphs manageable, only the graphs for the likelihood of one or more
RVR of a given type will be presented. As an aside, the graphs for an RVR count are very
similar to the corresponding graph for one or more RVR of the same type.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 104
December 19, 2011

Inmate Classification System Study

The following graphs look at the impact of the number of years served in a sentence (labeled
“stay_yrs” in the graphs). Within the group of inmates with 15 to 50 year sentences (the
“temp15” data set), the graph for the model of all A through F violations is:

In this graph, as in most of these graphs, the likelihood of one or more RVRs is fairly stable
through the first four years of a sentence, after which the likelihood begins to increase.
The corresponding graph for A-D violations is:

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 105
December 19, 2011

Inmate Classification System Study

The corresponding graph for one or more A violation is:

In the subset of inmates with sentences of 50+ years (temp50), the graphs show a somewhat
different pattern, with the likelihood of violations increasing at first and then decreasing
toward the end of the 10 year period. It should be noted that this subset was considerably
smaller than for the 15-50 group, so these graphs tend to be more erratic. The graph for one
or more A through F violations is:

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 106
December 19, 2011

Inmate Classification System Study

The graph for one or more A through D violations is:

The graph for one or more A violations is:

This pattern is somewhat different, though it should be emphasizes that not only is the
number of inmates small, but the likelihood of A violations within this group is also small.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 107
December 19, 2011

Inmate Classification System Study

Within the single life sentence group (tempsingle), the likelihood of one or more A through F
violations increases over time, assuming that the other model predictors are unchanged. This
is a fairly fanciful assumption, since we’ve seen that it’s typical for an inmate’s preliminary
to decrease over time.

For A through D violations, the corresponding graph is

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 108
December 19, 2011

Inmate Classification System Study

and for A violations, the graph is

Thus in this group, an inmate whose preliminary score was decreasing steadily would be
increasingly a good behavior risk, while one whose preliminary score wasn’t decreasing
would be less of a good risk as his sentence progressed.
The other question posed for this group was whether an inmate’s behavior would tend to
improve as he approached his MEPD. The following three graphs portray these
relationships. It should be noticed that as time passes, the MEPD decreases, so an inmate
would move “right to left” in these graphs as time passed.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 109
December 19, 2011

Inmate Classification System Study

The graph for one or more A through F violations is:

In this graph, we see that improvement steadily improves when an inmate is within 10 years
of his MEPD, with about half of this improvement having taken place when he’s
approximately 8 years from his MEPD.
The corresponding graph for one or more A through D violations exhibits a similar pattern as
MEPD approaches zero:

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 110
December 19, 2011

Inmate Classification System Study

For one or more A violations, the graph is somewhat different, though it should be noted that
A violations are rare and so these trends are poorly estimated:

Similar patterns were examined in the group of inmates with multiple life sentences. The
relationship with the number of years an inmate had served to date for one or more A through
F violations is:

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 111
December 19, 2011

Inmate Classification System Study

For A through D violations, the estimated relationship is:

For one or more A violations, the graph is:

All of these graphs show a generally increasing worsening in behavior over time, albeit
conditional on the preliminary score.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 112
December 19, 2011

Inmate Classification System Study

When looking at the impact of MEPD on behavior within the subgroup of inmates with
multiple life sentences, the graph for one or more A through F violations is:

The graph for one or more A through D violations is;

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 113
December 19, 2011

Inmate Classification System Study

Finally, the graph for one or more A violations is:

All three of these graphs show a general improvement over time as an inmate gets within 10
years of his MEPD (reading right to left on the graphs). This effect is in addition to the
improvement in behavior that you’d expect to see as the inmate ages and/or his preliminary
score decreases.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 114
December 19, 2011

Inmate Classification System Study

Appendix K
CDCR Inmate Classification Score System Study Crosswalk
CLASSIFICATION SCORE RESEARCH QUESTIONS
Research Question(s)
Findings
1. After reviewing ranges of scores for
the current Classification System (0 –
18, 19 – 27, 28-51 and 52+), can
natural “tipping points” or scores be
There are no natural tipping points.
identified which indicate an increase
in the predicted probability of inprison violence?
2. Does analysis of the in-prison CCR 3375.2 inmates with a Mandatory Minimum
behavior of inmates sentenced under score of 28 pose no greater risk of misconduct
“CCR 3375.2 (a) (7) Life inmate than other inmates with similar scores. Inmates
(multiple/execution style murders; with CCR 3375.2 Mandatory Minimums AND low
escapes)” (who have a mandatory preliminary scores pose less risk for misconduct
minimum classification score of 28) than inmates with no Mandatory Minimum and
indicate the need to restrict their preliminary/placement scores of 28.
housing to Level III institutions?
Inmates at each of the cutoff thresholds, including
the Level III threshold of 28, can be moved down
to a lower housing level without the expectation
that misconduct will increase.
3. Could the Mandatory Minimum Score LWOP inmates with placement scores of 52 due
code “B” for LWOP, currently at 52 to Mandatory Minimums pose no greater risk of
(Level IV) be reduced to allow LWOP misconduct than other inmates with similar scores.
inmates to house in Level III 270, LWOP inmates with placement scores of 52 due
new
design
facilities,
without to Mandatory Minimums AND lower preliminary
compromise to institutional security scores are less at risk for misconduct than
or public safety?
inmates with no Mandatory Minimum and
placement scores of 52.

4. Could the Mandatory Minimum Score
code “C” (CCR 3375.2) be adjusted
to cover Close Custody inmate’s
exclusively, without compromise to
institutional security or public safety?

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Inmates at each of the cutoff thresholds, including
the Level IV threshold of 52, can be moved down
without the expectation that misconduct will
increase. Data were not available to determine
differences between 270 degree and 180 degree
housing.
Custody Designations were not designed to
predict inmate misconduct and should not be used
to do so (see below).
Moreover, Custody
Designations do not predict escape (or
consequent threat to public safety).
Thus,
changing an inmate’s Custody Designation would
not alter the institutional security or public safety.

Page 115
December 19, 2011

Inmate Classification System Study

CUSTODY DESIGNATION RESEARCH QUESTIONS
Research Question(s)
Findings
1. Can the following proposed changes Questions 1a – 1d presume that Custody
be made without jeopardizing safety? Designation relates to misconduct in ways that are
a. Can Close A Custody be reduced unfounded. Custody Designations were created
from five to two years, and Close B to manage inmates who were seen as posing
Custody be increased from within high-risk to public safety if they escaped. Lethal
seven years to within ten years, of Electrified Fences have lowered the possibility of
their Minimum Earliest Parole Date escape to near zero (Appendix F).
(MEPD) for inmates with multiple life
sentences? [CCR-3377.2 (b)(3)(B) Because Custody Designations are unrelated to
& (c)(3)(B)]
the risk of misconduct, changes to inmate custody
b. Can Close B Custody be reduced status is not expected to negatively affect
from ten to five years for inmates behavior or institutional security.
with sentences of 50 or more years?
[CCR-3377.2(c)(3)(A)]
c. Can Close B Custody be increased
from within seven years to within ten
years of their MEPD for inmates
with life sentences? [CCR-3377.2
(b) (3) (C)]
d. Can Close B Custody be reduced
from four to two years for inmates
with sentences of 15 to 50 years?
[CCR-3377.2(c)(3)(D)]
2. Does the need for increased The escape data presented in Appendix F show
supervision for Close B Custody that stated “need” for increased supervision is
equate to the need to be assigned to questionable. Manual case reviews performed on
celled housing rather than dormitory those cases where inmates assigned to Level II
housing within a secure perimeter? through Level IV housing successfully escaped
In other words, can this population of reveals that all but two of the escapes are
inmates be moved from celled irrelevant to this study since they actually occurred
housing to dorm housing without an when the inmates were authorized to be outside of
increase in the predicted probability their secured housing level (e.g., out to court, out
for medical treatment).
of in-prison violence?
Regression analyses are used to examine the
impact of preliminary and placement scores,
housing level, custody level and housing category
(cells vs. dormitories) on RVRs. The results are
mixed.
At the 19 threshold, there are no
significant differences in RVRs with respect to
housing category. At the 28 threshold, the rate of
A-F violations in celled housing is approximately
13 percent lower than in dormitories and about 24
percent lower for A-D violations. At the 52
threshold, the rate of A-F violations in celled
housing is approximately 26.5 percent lower than

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 116
December 19, 2011

Inmate Classification System Study

CUSTODY DESIGNATION RESEARCH QUESTIONS
Research Question(s)
Findings
in dormitories and about 31.5 percent lower for AD violations (see Appendix J, pp. 93).
3. What criteria could be used to A literature review that explores factors related to
change
an
inmate’s
Custody inmate escapes reveals that the research on
Designation?
classification practices related to escapes is
limited, and thus no current practices may be
conceived as being ‘evidence-based.’
The sparse literature on escapes leaves no
definitive criteria that could be used to change an
inmate’s Custody Designation.
Despite this
limitation, CDCR’s Close Custody criteria are, for
the most part, supported by some research. All
criteria except High Notoriety / Public Interest /
Management Concern have been identified as
potential escape risk factors. The full literature
review on escapes may be found in Appendix M.
4. Can the attributes of inmates that
increase or decrease the risk of inprison violence be identified? If so,
should those attributes be used to
identify a subpopulation of inmates
that can be removed from Close
Custody? And, if so, how many
inmates currently assigned Close
Custody could be reduced to Medium
custody?

The attributes of inmates that increase or
decrease the risk of in-prison violence are
identified by the inmate’s preliminary classification
score.
Because Custody Designations do not measure
risk of inmate misconduct ALL Close Custody
inmates can be moved to Medium Custody with
the expectation that it will not affect levels of
misconduct.
However, if CDCR wants to move groups of
inmates downward in terms of supervision or
security level, the inmates that pose the lowest
risk of misconduct can be identified based upon
their preliminary score. Age may also be taken
into consideration.
The number of inmates could easily be estimated
based upon preliminary classification score.

5. How many inmates could safely have
their Custody Designation changed
without an increase in the predicted
probability of in-prison violence?
6. Do the current regulatory criteria for
Close Custody accurately identify
escape risk potential based upon
evidence based practices?

Follow-up analysis to be performed by CDCR,
Office of Research, Offender Information Services
Branch staff.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 117
December 19, 2011

Because escapes in Levels II – IV are near zero,
and because all of the Custody Designations are
unrelated to risk of misconduct, there is little
evidence to support them.

Inmate Classification System Study

CUSTODY DESIGNATION RESEARCH QUESTIONS
Research Question(s)
Findings
Moreover, the literature review (see Appendix M)
shows that the Close Custody criteria currently
used by CDCR have been identified as escape
risk factors. However, not enough research has
been conducted to conclusively state that any of
these criteria are in fact evidence based
7. Is there any specific data that reflects No. There is not enough escape data available to
the tendency for an offender to develop a cohort of escapees and conduct an
attempt to escape based on analysis on length of sentence. The CDCR
sentence length? If so, is there a Successful Escapes Report (Appendix F) shows
suitable range (time remaining to that other than walk-aways, there have been very
serve) that can be identified to allow few escapes since 1999. The literature review
for adjustment and identify risk?
completed for question #6 did not show that
sentence length was a risk factor for escape.
However, it did show that most escapes occur
during the first half of a inmates prison sentence.
8. Have electrified fences reduced the A review of escapes of Level III inmates which
number of escape attempt from those have occurred in the last 12 years shows that
secure environments?
none occurred in institutions with electrified
fences. This suggests that inmates who are
deemed public safety risks could be housed in a
lower level provided an electrified fence was
present.
9. Has it been determined that longer A literature review of misclassification (see
sentences correlate with greater Appendix N) was completed and a number of
instances of misconduct?
factors were found to contribute to the
misclassification of inmates. One of those factors
was length of sentence. Although there is little
research available, multiple studies found that
LWOPS (life without the possibility of parole) were
less likely to engage in violent behavior than
inmates sentenced to 10-14 years, 15-19 years
and more than 30 years.
10. Has it been determined that those
offenders with initial placement
scores of Level III or Level IV, have a
N/A
greater tendency to move up or down
in
points
(excluding
LWOP
offenders)?
11. Does an offender’s security level
along with the application of Close
See previous results.
Custody have an effect on in-custody
behavior?
12. Does the existing Close Custody
Regulatory
Policy
identify
an
N/A
offender’s propensity for misconduct
or is security level a better indicator?
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 118
December 19, 2011

Inmate Classification System Study

CUSTODY DESIGNATION RESEARCH QUESTIONS
Research Question(s)
Findings
13. Is data available that identifies age In the Gap Analysis, age was the first factor
and physical impairment as factors applied to the model.
After adjusting for
that would allow for reduced custody preliminary score, placement score and housing
- regardless of term length and level it’s estimated that the likelihood/number of
security level?
violations decrease between one and four percent
for each additional year of age, depending on the
cutoff point used in the gap analysis and the type
of RVR. This was a linear decrease, so there is
no ‘age cut-point’ where violations drop
dramatically. Furthermore, the literature review on
escapes showed that escapes typically occur
when inmates are in their 20s or mid 30s. These
two findings suggest that as inmates age, the
need for Close Custody decreases. It is also
possible that age is being underutilized in the
calculation of the preliminary score as well as the
utilization of Close Custody.
14. Can it be determined if Minimum A
and Minimum B custody can be
combined into one custody level?
a. Do we have data to support the
need for both Minimum Custody
levels?
b. Which facilities currently house
N/A
offenders
with
Minimum
A
Custody?
c. How many offenders currently
excluded from Camp/MSF/CCF
due to new medical classification
criteria are assigned Minimum A
Custody?

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 119
December 19, 2011

Appendix L
Cohort Descriptives 16

Table 1.

HOUSING LEVEL
I
Age

n

II
%

n

III
%

n

IV
%

n

TOTAL
%

n

%

18-19

10

0.2%

31

0.1%

304

0.7%

249

0.9%

594

20-24

485

8.0%

733

3.2%

4,691

11.5%

3,179

11.2%

9,088

0.6%
9.2%

25-29

956

15.8%

1,853

8.0%

6,182

15.1%

5,478

19.3%

14,469

14.7%

30-34

902

14.9%

2,556

11.1%

5,759

14.1%

5,273

18.5%

14,490

14.7%

35-39

952

15.7%

3,283

14.3%

5,162

12.6%

4,545

16.0%

13,942

14.2%

40-44

1,013

16.8%

3,962

17.2%

5,265

12.9%

3,751

13.2%

13,991

14.2%

45-49

941

15.6%

4,023

17.5%

5,480

13.4%

2,972

10.5%

13,416

13.6%

50-54

481

8.0%

3,042

13.2%

3,891

9.5%

1,669

5.9%

9,083

9.2%

55-59

204

3.4%

1,794

7.8%

2,091

5.1%

768

2.7%

4,857

4.9%

60+

101

1.7%

1,751

7.6%

2,023

5.0%

550

1.9%

4,425

4.5%

6,045

100.0%

23,028

100.0%

40,848

100.0%

28,434

100.0%

98,355

100.0%

Total
Mean Age

38.2

43.5

38.8

36.0

39.0

Race
White

1,798

29.7%

6,234

27.1%

9,640

23.6%

5,019

17.7%

22,691

23.1%

Hispanic

1,968

32.6%

8,384

36.4%

15,917

39.0%

11,219

39.5%

37,488

38.1%
32.1%

Black

2,010

33.3%

6,744

29.3%

12,342

30.2%

10,482

36.9%

31,578

Asian

28

0.5%

283

1.2%

456

1.1%

233

0.8%

1,000

1.0%

Native American

48

0.8%

225

1.0%

394

1.0%

293

1.0%

960

1.0%
0.2%

Pacific Islander
Other
Total

11

0.2%

63

0.3%

89

0.2%

65

0.2%

228

182

3.0%

1,095

4.8%

2,010

4.9%

1,123

3.9%

4,410

4.5%

6,045

100.0%

23,028

100.0%

40,848

100.0%

28,434

100.0%

98,355

100.0%

Commitment Offense
Category
Person

2,024

33.5%

17,380

75.5%

30,845

75.5%

24,654

86.7%

74,903

76.2%

Property

1,607

26.6%

2,338

10.2%

4,506

11.0%

2,002

7.0%

10,453

10.6%

Drug

1,914

31.7%

2,377

10.3%

3,639

8.9%

1,027

3.6%

8,957

9.1%

498

8.2%

928

4.0%

1,857

4.5%

749

2.6%

4,032

4.1%

2

0.0%

5

0.0%

1

0.0%

2

0.0%

10

0.0%

6,045

100.0%

23,028

100.0%

40,848

100.0%

28,434

100.0%

98,355

100.0%

Other
Missing

Sex Reg Flag
Yes
No
Total

12

0.2%

6,100

26.5%

8,977

22.0%

3,961

13.9%

19,050

19.4%

6,033

99.8%

16,928

73.5%

31,871

78.0%

24,473

86.1%

79,305

80.6%

6,045

100.0%

23,028

100.0%

40,848

100.0%

28,434

100.0%

98,355

100.0%

Mental Health Code
CCCMS/EOP
No Mental Health Code
Total

346

5.7%

4,018

17.4%

10,996

26.9%

9,128

32.1%

24,488

24.9%

5,699

94.3%

19,010

82.6%

29,852

73.1%

19,306

67.9%

73,867

75.1%

6,045

100.0%

23,028

100.0%

40,848

100.0%

28,434

100.0%

98,355

100.0%

94.5%

Serious/Violent
Yes

4,589

75.9%

21,551

93.6%

38,720

94.8%

28,083

98.8%

92,943

No

1,456

24.1%

1,477

6.4%

2,128

5.2%

351

1.2%

5,412

5.5%

6,045

100.0%

23,028

100.0%

40,848

100.0%

28,434

100.0%

98,355

100.0%

Total

16

Only includes reviews where housing level is available.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 120
December 19, 2011

Cohort Descriptives (continued)
Table 1. (cont’d.)
HOUSING LEVEL
I

II

III

IV

TOTAL

Custody
Max

5

0.1%

108

0.5%

503

1.2%

815

2.9%

1,431

Close A

0

0.0%

0

0.0%

2,020

4.9%

5,714

20.1%

7,734

1.5%
7.9%

Close B

0

0.0%

1,766

7.7%

15,333

37.5%

14,195

49.9%

31,294

31.8%

Medium A

90

1.5%

17,619

76.5%

21,913

53.6%

7,660

26.9%

47,282

48.1%

Medium B

414

6.8%

2,200

9.6%

475

1.2%

16

0.1%

3,105

3.2%

Minimum A

531

8.8%

127

0.6%

66

0.2%

0

0.0%

724

0.7%

Minimum B

4,991

82.6%

1,175

5.1%

300

0.7%

6

0.0%

6,472

6.6%

14

0.2%

30

0.1%

156

0.4%

27

0.1%

227

0.2%

0

0.0%

3

0.0%

82

0.2%

1

0.0%

86

0.1%

6,045

100.0%

23,028

100.0%

40,848

100.0%

28,434

100.0%

98,355

100.0%

Unclassified
Unknown
Total
RVR

n

%

n

%

n

%

n

%

Yes

1,841

30.5%

3,703

16.1%

9,383

23.0%

8,533

30.0%

No
Total

n
23,460

%
23.9%

4,204

69.5%

19,325

83.9%

31,465

77.0%

19,901

70.0%

74,895

76.1%

6,045

100.0%

23,028

100.0%

40,848

100.0%

28,434

100.0%

98,355

100.0%

A1/A2 Violations
Yes
No
Total

70

1.2%

171

0.7%

522

1.3%

880

3.1%

1,643

1.7%

5,975

98.8%

22,857

99.3%

40,326

98.7%

27,554

96.9%

96,712

98.3%

6,045

100.0%

23,028

100.0%

40,848

100.0%

28,434

100.0%

98,355

100.0%

B, C, D Violations
Yes
No
Total

502

8.3%

1,318

5.7%

4,284

10.5%

4,867

17.1%

10,971

11.2%

5,543

91.7%

21,710

94.3%

36,564

89.5%

23,567

82.9%

87,384

88.8%

6,045

100.0%

23,028

100.0%

40,848

100.0%

28,434

100.0%

98,355

100.0%

E,F Violations
Yes

1,461

24.2%

2,577

11.2%

5,685

13.9%

4,065

14.3%

13,788

14.0%

No

4,584

75.8%

20,451

88.8%

35,163

86.1%

24,369

85.7%

84,567

86.0%

6,045

100.0%

23,028

100.0%

40,848

100.0%

28,434

100.0%

98,355

100.0%

Total

Preliminary
Score Level
I
II
III
IV

Placement
Score Level
I
II
III
IV

I
n
8,770

Table 2. Placement Score Level
II
III
%
n
%
n
%
23.7% 24,389
65.9%
2,600
7.0%
10,020
94.5%
389
3.7%
24,662
97.4%

I
n
5,158
761
119
7

%
58.8%
2.2%
0.4%
0.0%

Table 3. Housing Level
II
III
n
%
n
%
2,636
30.1%
952
10.9%
19,119
55.6% 14,375
41.8%
1,155
4.2% 22,868
82.7%
118
0.4%
2,653
9.6%

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

IV
n
%
1,257
3.4%
199
1.9%
671
2.6%
25,398 100.0%

IV
n
24
154
3,509
24,747

%
0.3%
0.4%
12.7%
89.9%

TOTAL
n
%
37,016 100.0%
10,608 100.0%
25,333 100.0%
25,398 100.0%

TOTAL
n
%
8,770 100.0%
34,409 100.0%
27,651 100.0%
27,525 100.0%

Page 121
December 19, 2011

Appendix M
Literature Review on Escape Risk Factors
Executive Summary
The California Department of Corrections and Rehabilitation (CDCR) is currently examining both
the external and internal components of its classification system. An external system places an
inmate in a prison, while an internal system places an inmate in housing and programming
within a prison. The purpose of this literature review is to determine whether the current
regulatory Close Custody criteria are based on evidence-based practices that accurately identify
escape risk. An additional goal of this review was to identify other criteria that might be used to
change an inmate’s custody classification. Specifically, this literature review addresses the
following research question: Do the current regulatory criteria for Close Custody accurately
identify escape risk potential based upon evidence-based practices?
The research on escape is first discussed, followed by a comparison made between the
research and CDCR’s Close Custody criteria.
The main findings are:
•

The research on classification practices is so limited that the Close Custody criteria
cannot be confirmed as evidence-based. While there are various studies that appear to
support the criteria, the quality and amount of research is inadequate to conclude that
they are evidence-based.

•

There are important limitations in the escape literature. In particular, multiple counting
methods are employed, definitions of escape vary, and sample sizes are small. In
addition most of the literature examines minimum security inmates because they escape
more frequently than medium or maximum security inmates.

•

Many of the publications are dated. Most research on escapes was conducted before
valid classification systems were in place and before electrified perimeter fences
became standard security features of prisons. Demographic changes may also affect
the relevance of some studies to the current inmate population.

•

Risk factors can be divided into two groups: static (which cannot be changed) and
dynamic (which can be changed).
The most frequently identified static risk factors for escape attempts include:
°

Age: escapees are typically in their 20s or early 30s

°

Adult Criminal History: escapees often have previous convictions and incarcerations

°

Juvenile Criminal History: escapees were found to have a record of juvenile
convictions and incarcerations

°

Property Crimes: escapees tend to have convictions for property-related crimes

°

Length of Time Already Served: escapes typically occur early on in an inmate’s
sentence

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 122
December 19, 2011

°

Previous Escape Attempts: inmates who escape often have previous escape
attempts in their record

°

Race/Ethnicity: in older studies, escapees were more frequently white than black,
although more recent research did not support this finding

The most frequently identified dynamic risk factors for escape attempts include:
°

Holds, Detainers, Denial of Parole: escapees who have legal actions that lengthened
their stay in prison beyond their expected release date are more likely to escape

°

Institutional Misconduct: escapees are more likely to have disciplinary problems
while in prison

°

Relationship Problems: inmates may have be more likely to escape because of
relationship problems, such as a relationship ending, divorce, or death in the family

°

Substance Abuse: escapees are likely to have alcohol or drug addictions, although
the significance of these are unclear as substance abuse is also found in inmates
who do not attempt escapes

Table A compares CDCR’s Close Custody Regulations with the static and dynamic escape risk
factors.
Table A

Escape Risk Factor
Static
Age
Adult Criminal History
Juvenile Criminal History
Property Crimes
Length of Time Already Served
Previous Escape Attempts
Race/Ethnicity
Dynamic
Holds, Detainers, Denial of Parole
Institutional Misconduct
Relationship Problems
Substance Abuse
High Notoriety/Public Interest/Management
Concern

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

CDCR Close
Custody
Criteria

x
x

x
x

Risk Factors Identified
in
the Literature
x
x
x
x
x
x
x
x
x
x
x

x

Page 123
December 19, 2011

CDCR’s Close Custody criteria are, for the most part, supported by some research. All criteria
except High Notoriety/Public Interest/Management Concern have been identified as escape risk
factors. However, not enough research has been conducted to conclusively state that any of
these criteria are evidence-based.
Furthermore, there are escape risk factors found in the literature that are not addressed by the
CDCR Close Custody criteria. Age and relationship problems might be explored as additional
criteria, although the latter could be difficult to objectively define or identify.
In sum, there is not enough empirical research available to confirm whether or not CDCR’s
Close Custody criteria are evidence-based, although there is some research that implicitly
supports them. This does not mean they are wrong, but rather that there is not enough
research to confirm that they are or are not the right criteria to use. As a result, decisions to
maintain or change current CDCR Close Custody practices should not be based upon the
available research alone.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 124
December 19, 2011

Literature Review on Escape Risk Factors
Full Review
Introduction
An inmate classification system is critical for determining where inmates should be placed within
a prison system in order to ensure the safety and security of individuals working and living in
institutions, as well as those in the community. The literature sometimes discusses two types of
systems: external and internal. An external system places an inmate in a prison, while an
internal system places an inmate in housing and programming within a prison.
Within the California Department of Corrections and Rehabilitation (CDCR), the
pointand-level Inmate Classification Score System is considered an external system, while the
CDCR Close Custody Designations are considered to be part of an internal system. The
Custody Designation an inmate receives while in custody at CDCR determines the type of
supervision s/he receives once s/he is in the institution. It may also impact the jobs or programs
to which s/he may be assigned. The purpose of the Custody Designation is to determine
supervision control levels based upon problematic behavior or an individual’s potential for
escape and threat to the community if an escape occurs. The CDCR designation is primarily
based on the following factors (although other reasons may also be considered):
-

The inmate’s total term, sentence, or remaining time-to serve

-

The inmate’s escape history

-

Receipt of an active law enforcement felony hold

-

An inmate who is considered to be High Notoriety or is designated as a Public Interest
Case or a Management Concern 17

-

A finding of guilt for a serious, felony level, Rules Violation Report (RVR)

CDCR is currently examining both the internal and external components of its classification
system. The purpose of this literature review is to compare the current regulatory CDCR Close
Custody criteria with evidence-based practices, to validate current classification practices, and
to potentially identify different/additional criteria that should be used to determine an inmate’s
Custody Designation. This literature review addresses the following research question in the
Inmate Classification System Study: Do the current regulatory criteria for Close Custody
accurately identify escape risk potential based upon evidence-based practices?

17

Per CCR Title 15. Crime Prevention and Corrections, Section 3000. Definitions, a Management
Concern is defined as a behavior observed or documented in the inmate’s criminal history that
demonstrates to a classification committee that the inmate has a propensity towards violence against
self or others; has a history of inciting or pressuring others toward criminal behavior; preys on more
vulnerable members of society; or portrays a level of criminal sophistication and/or access to large
amounts of drugs, money, or power. This may include disruptive groups and prison gang members or
affiliates.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 125
December 19, 2011

Methodology
The following databases were searched for relevant literature on inmate escape attempts and
successful escapes:
• Academic Search Complete
•

Google Scholar

•

JSTOR

•

MEDLINE with Full Text

•

PsychARTICLES

•

PsychINFO

•

Psychology and Behavioral Sciences Collection

•

SocINDEX with Full Text

The search terms varied across the databases due to varying design and use of subject
vocabulary, but the predominant terms used were:
•

Escapes

•

Fugitives from Justice

Free-text searching terms included:
•

Escape Risk Factors

•

Prison Escapes

•

Inmate Escapes

•

Predicting Inmate Escapes

•

Absconders

Criteria for inclusion:
•

Adult male prison inmates

•

Escapes from all security-level institutions

•

Mental health literature, where the focus is on individuals are committed for criminal
behavior

Criteria for exclusion:
•

Escapes from confinement of a military or political nature, such prisoner-of-war camps,
where confinement is not due to behavior as an individual

•

Female inmate escapes, as the Classification Study includes only male inmates

•

Validity of assessment instruments to predict escape

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 126
December 19, 2011

Twenty-six articles were selected and either downloaded or ordered from the State Library. After
this initial review, 16 were found to be useful. Publication dates were between 1959 and 2011.

Limitations
Age of Literature
Most research on escape-risk factors was conducted before the use of validated classification
systems, better-designed secure housing, and electrified fences—all of which may reduce the
need to look at inmates’ personal escape risk factors. Also, because much of the literature is
dated, it is not evident if the identified risk factors apply to today’s inmate population. This is an
important limitation, as technological advances may have rendered much of the prior research
obsolete. For example, there were only six inmate escapes from CDCR’s secure-custody
facilities in 2010. Likewise, the New York Department of Correctional Services (2011) reported
only one inmate escape in 2009 and no escapes in 2008. Therefore, it is likely that many
individual-level factors that have historically been associated with escape risk may have lost
some of their predictive value in modern institutions where such events are increasingly rare.
Multiple Counting Methods and Definitions
There is no one national database that tracks prison escapes. A few national surveys suggest
approximate numbers and rates of escapes, but the difference between the highest and lowest
estimates varies by 176% (Culp, 2005).
In addition, the literature does not always specify the type of facility from which an escape was
made. While escapes are attempted from all security level facilities, the majority are from
minimum security facilities. Culp (2005) estimated that around 89% of escapes are from
minimum security environments. This type of escape is frequently called a “walkaway” or “awol”
(absent without leave). Some studies make a distinction between these types of escapes and
attempts from secured facilities, while other studies do not. Also, “escape” and “escape
attempt” are not clearly differentiated in the research. For the sake of consistency the general
term “escape” will be used in this literature review.
Small Samples
Escape attempts are not routine events in any correctional system. Most inmates do not try to
escape. As a result, the number of escapees in any one study is small. This is especially true
of escapes from maximum security settings.
Study Population
The research focuses on lower security level inmates because they escape more frequently
than inmates in medium or maximum security prisons. It is not known if the risk factors
identified in most of the research would apply to inmates in medium- or maximum-level prisons.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 127
December 19, 2011

Escape Risk Factors
Risk factors can be divided into two groups: static and dynamic. Static risk factors cannot
change (e.g., gender, age at first arrest, age at first incarceration). Dynamic risk factors can
change (e.g., drug use, number of rules violation reports received in a one year period).
Static Escape Risk Factors
Static risk factors are those that cannot change either because they are demographic in nature
(e.g., race) or they are events that occurred in the past (e.g., juvenile escape attempt). Culp
(2005), in his review of static and dynamic risk factors, concluded that static risk factors are
more reliable in predicting escapes. The following are the main static escape risk factors
discussed in the literature:
Age
Young age, ranging from 20 to 31 years, is the most consistently identified risk factor for
escape. Several studies found that the majority of escapees in their studies were under the
age of 30 (Cochrane, 1948 as cited in Loving, Stockwell and Dobbins, 1959; Morrow, 1969).
Morgan (1967) compared a group of South Carolina escapees with a control group of similar
inmates who did not attempt to escape and found that “significantly more” escapees were
younger than 25 years. Muir-Cochrane, Gorta and Sillavan (1991), in their analysis of 812
escapees in New South Wales, Australia, indicated that 58% were under the age of 24 and
Mosel (1996) described the typical absconder as under the age of 26. Sandhu (1996) found
the average age of the inmates he studied to be 27. Culp (2005), in his review of 88
escapes by 135 individuals, was able to identify the average age of 117 individuals as 27
years old. The State of New York Department of Correctional Services (2011) analysis of
escapes between 2006 and 2010 found that, of the individuals who attempted to escape,
80% were less than 31 years old, while only 36% of the inmate population as a whole was
less than 31 years old.
Anson (1983) approached escape attempts from a slightly different perspective, but
obtained similar results. He looked at 11 characteristics of 17 male prisons, such as age
and size of the institution, ratio of staff to inmates, level of supervision, and age of their
populations. He also found that more escapes were attempted where the inmate population
was younger.
Young age is consistently identified as an escape risk factor. Some possible explanations
are that younger inmates have not been incarcerated long enough to adjust to the
environment or that they are more physically capable to attempt an escape than older
inmates.
Adult Criminal History
Inmates who escape tend to have committed more crimes in the past than inmates who do
not escape. In a Kentucky Bureau of Corrections study (1978), only 16% of escapees were
first-time adult offenders, while 30% of
non-escapees were adult first-time
offenders. Sixty-five percent of the walkaways in the Montiuk and Johnson (1992) study had
20 or more previous convictions as an adult.
Juvenile Criminal History
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 128
December 19, 2011

Several studies identify a link between a juvenile criminal history and adult prison escapes.
Dobbins and Stockwell (1960), while not giving specific percentages, did find that escapees
had significantly more juvenile commitments than non-escapees. Seventeen percent of the
escapes in the Kentucky Bureau of Corrections (1978) had been incarcerated as juveniles,
compared to 8% of non-escapees. In studying walkaways from a Canadian minimum
security facility, Johnston and Motiuk (1992) noted that every walkaway had a juvenile
conviction. The convicted murderers in Sandhu’s (1996) study were also described as
having juvenile records, although no specifics were given.
Property Crimes
Property crimes were the most frequently mentioned crimes associated with the risk of
escape (Gorta and Sillivan, 1991; Kentucky Bureau of Corrections, 1978; Johnston and
Motiuk, 1992; Loving, Stockwell, Dobbins, 1959; State of New York Department of
Correctional Services, 2011; Sturrock, Porporino, Johnston, 2008). A couple of reasons
may explain the prevalence of property-related crimes in research about escapees. Because
most escapes are from minimum-security settings, research has focused on minimumsecurity inmates who are less likely to have been convicted of violent crimes that would
place them in medium- or maximum-security housing. Furthermore, escapees are more
likely to be younger offenders who commit lower-level property crimes at an early stage in
their criminal career.
Although property crime is the most frequently identified crime that is associated with
escapes, research has also been conducted on inmates convicted of violent crimes. Sandhu
(1996) compared a group of 31 escapees convicted of first or second degree murder to 88
individuals with similar convictions who did not attempt escape. Escapees were found to be
characterized by other risk factors, such as a more lengthy adult criminal record, a juvenile
record, substance abuse problems, and more disciplinary actions than non-escapees.
These are all risk factors of escapees convicted of property crimes, so the type of crime an
individual was convicted of may not be useful in predicting escapes.
In sum, the frequency of property crimes among escapees could be due to the fact that the
escapees are typically housed in minimum-security settings, where most escapes occur.
Length of Time Already Served
The majority of studies that addressed time served indicated that escape attempts occur
before inmates serve half of their sentence. Dobbins and Stockwell (1960) cited a 1948
study that found that most escapees had served less than 40% of their term. Morgan (1967)
found that “significantly more” escapees served less than half their sentences than nonescapees. The average sentence of escapees studied by Johnston and Motiuk (1992) was
four years, with the average length of time from admission to escape being 371 days, or a
little over one year. Twenty percent of the escapees in Sandhu’s (1996) study escaped in
the first year of sentences, ranging from 10 years to life or condemned to death.
Gorta and Sillivan’s (1991) study of Australian escapees presents a mixed picture. They
made two comparisons. The first compared a group of individuals who attempted to escape
within two weeks of arrival to prison to those who attempted escapes after more than two
weeks. The second compared a group of individuals who attempted to escape within 60
days of their release to those who attempted to escape who had more than 60 days left on
their sentence. In both comparisons, inmates who tried to escape within two weeks of their
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 129
December 19, 2011

arrival and those who had less than 60 days to serve were younger, had shorter sentences,
were more likely to have had been convicted of a property-related crime, and had served
less time on a prior sentence. The inmates who escaped early in their sentence fit the
profile of a young inmate convicted of a property-related crime.
Moore and Hammond (2000) did not find a link between time served and escape among
individuals committed to secure psychiatric facilities for crime-related reasons. There have
also been descriptive accounts of individual inmates spending years preparing for escapes
by gaining the trust of prison staff with their good behavior (Singer, 2006).
While research findings differ on the relationship between time served and escape risk,
there is some support for the idea that inmates will attempt to escape before serving half of
their sentence. One theory is that inmates eventually get “invested” in the time they are in
prison; after serving most of their sentence they do not want to risk postponing their parole
date.
Previous Escape Attempts
A number of studies demonstrated that previous escape attempts are an indication of a
higher risk of future attempts. The Kentucky Bureau of Corrections (1978) escapees had
31% prior escapes compared to 6% of non-escapees. Over 40% of walkaways in one study
had previously escaped at least one time (Johnston and Motiuk, 1992). Sandhu (1996)
found that 20% of escapees made one or more previous attempts.
Race/Ethnicity
Several studies observed that more white inmates escaped than inmates who are members
of racial or ethnic minority groups, both in number and percentage of the total inmate
population. Loving, Stockwell and Dobbins (1959), in their study of 100 escapees from the
Louisiana State Penitentiary, found that between 1949 and 1959, twice as many white
inmates attempted to escape than black inmates, even though black inmates outnumbered
white inmates by three to one. In Sandhu’s (1996) study of escapees convicted of murder
or other violent crimes, 74% were white, followed by black (17%) and Native American (9%).
Moore and Hammond (2000), in their study of criminal, mentally ill absconders from secured
correctional mental health facilities, also found the majority of escapees were white. In both
cases these rates were disproportionate to the racial composition of the population as a
whole.
More recently, though, Culp (2005) examined nationwide surveys of correctional agencies
and found that racial differences in escapees are declining. The State of New York
Department of Correctional Services (2011) reported that 60% of escapes were made by
black inmates, who comprised 51% of the inmate population, 30% were by white inmates,
(21% of the population), and 10% were made by Hispanic inmates, ( 26% of the population).
Although Race/ethnicity was cited as an escape risk in older research, no studies were ever
performed to find an explanation.
Other Static Risk Factors Identified in the Literature
Some other static risk factors mentioned in the literature include age at first arrest (Dobbins
and Stockwell, 1960; Loving, Stockwell, & Dobbins, 1959), childhood abuse or neglect
(Johnston and Motiuk, 1992; Morgan, 1967; Sandhu, 1996), and unstable work history
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 130
December 19, 2011

(Johnston and Motiuk, 1992; Morrow, 1969). However, no research has been conducted to
explain why these factors appear among inmates who escape.
Dynamic Factors Associated with Escape Attempts
Dynamic factors are those that can change, such as behavior while in prison or use of drugs.
Although the research findings on dynamic risk factors are not as clear as those on static
factors, the following are some dynamic factors that have been examined:
Holds, Detainers, Denied Parole
Inmates know their expected release date. The risk of escape increases when they are told
pending legal action could interfere with that date. Loving, Stockwell and Dobbins (1959)
cited a study from 1948 that found escapees were more likely to have detainers, but no
specific numbers were provided. Being denied parole was the reason for attempting to
escape given by 8.5% of the inmates who were interviewed in another study (Duncan and
Ellis, 1973). The Kentucky Bureau of Corrections (1978) found that 34% of escapees, but
only 12% of non-escapees, had time added to their stay in prison.
Institutional Misconduct
Some studies suggest that inmates who escape also get into trouble within the prison more
often inmates who do not escape. The Kentucky Bureau of Corrections (1978) found that
escapees tend to be in trouble with contraband more often than non-escapees. Over 75% of
the walkaways in the Johnston and Motiuk (1992) study had one or more disciplinary
reports. Sturrock, Porporino, Johnston (2008) cite a 1975 dissertation that concluded that
inmates who escape have been placed in solitary confinement (comparable to CDCR
Administrative Segregation or Security Housing Units) escape more than inmates who have
not been placed in such housing.
Relationship Problems
Inmates may escape in order to deal with relationship or family problems, such as
separation, divorce, illness or death in the family, problems with relatives, or economic
difficulties of relatives. In one study these problems were given as the main motivation for
escape by over 15% of inmates who had escaped and were then caught (Duncan and Ellis,
1973). When asked what was on their mind at the time of escape, 42% of walkaways cited
family/marital problems (Johnson and Motiuk, 1992). Sandhu (1996) listed news of a wife’s
illness and problems with a girlfriend among the reasons given by inmates for their escapes.
Relationship problems have been shown to be a risk factor for escape attempts. Inmates
will escape when faced with a personal problem that they believe they cannot address from
a distance.
Substance Abuse
Escapees have been found to have substance abuse problems (Morrow, 1969; Johnston
and Motiuk, 1992; Sandhu, 1996; Muir-Cochrane and Mosel, 1998). However, the
significance of this is unclear. Definitions of substance abuse vary and many inmates who
do not escape also have substance abuse problems. Studies have also produced conflicting
results. Morrow (1969) reported higher rates of alcoholism in escapees (42%) compared to
non-escapees (25%), while Sandhu however found that 29% of escapees were serious
alcoholic abusers as opposed to 48% of non-escapees. Two studies suggested differences
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 131
December 19, 2011

in the use of alcohol and drugs. The Kentucky Bureau of Corrections (1978) asserted that
escapees were more likely to use alcohol and none-escapees more likely to use drugs,
although no data were provided to support this. Sandu (1996) found that 32% of escapees
were serious drug users, compared to 47% on non-escapees.
Other Dynamic Risk Factors Identified in the Literature
Additional dynamic risk factors identified in the literature include problems with prison staff or
dissatisfaction with available programs or services (Duncan and Ellis, 1973; Sandhu, 1996).
Anson (1983) and Johnston and Motiuk (1992) identified too much free time or boredom as
reasons why inmates escaped.
Despite its notable limitations, the research literature on inmate escape risk factors identifies
some factors repeatedly. Age at escape attempt is the most consistently documented static
factor. There is research to support adult and juvenile criminal histories and length of time
served, but the evidence is not as strong.
Although dynamic escape risk factors have not been studied in depth, a few appear consistently
in the literature. Family or relationship problems are the most documented dynamic risk factors.

Comparison Between Escape Research and CDCR Close Custody Criteria
Table A compares CDCR’s Close Custody Regulations with the static and dynamic escape risk
factors. As depicted in this table, CDCR’s Close Custody criteria are, for the most part,
supported by some research. All criteria except High Notoriety/Public Interest/Management
Concern have been researched as possible escape risk factors. However, there is not enough
research to conclusively state that the criteria are evidence-based.
Furthermore, there are escape risk factors found in the literature that are not addressed by the
CDCR Close Custody criteria. Age and relationship problems might be explored as additional
criteria, although the latter could be difficult to objectively define or identify.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 132
December 19, 2011

Table A

Escape Risk Factor
Static
Age
Adult Criminal History
Juvenile Criminal History
Property Crimes
Length of Time Already Served
Previous Escape Attempts
Race/Ethnicity
Dynamic
Holds, Detainers, Denial of Parole
Institutional Misconduct
Relationship Problems
Substance Abuse
High Notoriety/Public Interest/Management
Concern

CDCR Close
Custody
Criteria

x
x

x
x

Risk Factors Identified
in
the Literature
x
x
x
x
x
x
x
x
x
x
x

x

Do the current regulatory criteria for Close Custody accurately identify escape
risk potential based upon evidence-based practices?
In sum, there is not enough empirical research available to confirm whether CDCR’s Close
Custody criteria are evidence-based, although there is some research that implicitly supports
them. This does not mean they are wrong, but rather that there is not enough research to
confirm that they are or are not the right criteria to use. As a result, decisions to maintain or
change current CDCR Close Custody practices should not be based upon the available
research alone.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 133
December 19, 2011

Bibliography
Anson, R. H. (1983). Correlates of Escape: a Prelimiary Assessment of Georgia Prisons.
Criminal Justice Review, 38-42.
Culp, R. F. (2005). Frequency and Characteristics of Prison Escapes in the United States: an
Analysis of National Data . Prison Journal, 85 (3), 270-291.
Dobbins, D., & Stockwell, F. (1960). Individual and Social Correlates of Prison Escapes. Journal
of Consulting Psychology, 24 (1), 95.
Duncan, D., & Ellis, T. (1973, May-June). Sitiuational Variables Associated with Prison Escapes.
American Journal of Corrections, 29-30.
Gorta, A., & Sillavan, T. (1991). Escapes from New South Wales Gaols: Placing the Risk in
Pespective. ANZJ Crim, 24, pp. 204-218.
Johnston, J. C., & Motiuk, L. L. (1992). Factors Related to Unlawful Walkaways from Minimum
Security Institutions. Correctional Service Canada.
Kentucky Bureau of Corrections. (1978). Blackburn Correctional Complex Escape Study 1978.
Loving, W., Stockwell, F., & Dobbins, D. (1959). Factors Associated with Escape Behaviour in
Prison Libraries. Federal Probation, 23 (3), 49-51.
Moore, E., & Hammond, S. (2000). When Statistical Models Fail; Problems in the Prediction of
Escape and Absconding Behaviour from High-Security Hospitals. Journal of Forensic
Psychiatry, 11 (2), 359-371.
Morgan, D. (1967). Individual and Situational Factors Related to Prison Escapes. American
Journal of Correction, 29 (2), 30-31.
Morrow, W. R. (1969). Escapes of Psychiatric Offenders. Journal of Criminal Law, Criminology,
and Police Science, 60 (4), 464-471.
Muir-Cochrane, E., & Mosel, K. (1998). Absconding: a review of the Literature 1996-1998.
International Journal of Mental Health Nursing (17), 370-378.
Sandu, H. (1996). A Profile of Murderer Escapees. Journal of the Oklahoma Criminal Justice
Research Consortium, 3.
Singer, M. (2006, October 9). Escaped. New Yorker, 82 (32).
State of New York Department of Correctional Services. (2011). Inmate Escape Incidents 20062010.
Sturrock, R. C., Porporino, F. J., & Johnston, J. C. (2008). Literature Review on the Factors
Related to Escape from Correctional Institutions. Correctional Service Canada.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 134
December 19, 2011

Appendix N
Literature Review on Misclassification
Executive Summary
To support the current California Department of Corrections and Rehabilitation (CDCR) Inmate
Classification Score System Study, this literature review examines research on inmate
misclassification and presents information on current steps other state correctional systems are
taking to reduce their inmate population.
A total of 26 journal articles, research reports and government reports formed the basis of this
review. Additional information was obtained from conversations with officials in other state
correctional systems.
Misclassification
•

Misclassification research focuses more on overclassification than underclassification.

•

Factors in misclassification can be broadly separated into three groups:
Convictions and Sentencing
Basing classification decisions on the type of crime for which an individual was convicted
and the length of their sentence. Behavior as a criminal, especially with respect to violent
crimes such as murder, may not be indicative of behavior as an inmate, and longer
sentences do not necessarily lead to more misconduct.
Inappropriate Use of Demographic Factors
Unaddressed mental illness has been found to increase misconduct, leading to higher
classification requirements, and classification systems created for male inmates may or
may not be applicable to female inmates.
Staff Use of Classification Systems
Effective use of any classification system is all too often hindered by inadequate training,
limited experience, and the use of subjectivity in making classification decisions.

•

Underclassification appears to sometimes foster better behavior. Inmates placed in less
restrictive housing than their classification score requires tend to behave no worse than the
inmates appropriately placed in that environment and they also tend to have lower
recidivism rates.

•

Overclassification appears to be criminogenic, i.e., it is associated with increased criminal
behavior. Inmates placed in more restrictive housing than their classification score requires
will act like the inmates who were appropriately housed in restrictive housing, and inmates
who were overclassified return to prison more than inmates who were correctly classified.

•

Although research is somewhat limited, it is clear that misclassification is a serious problem.
Although underclassification can be a problem, it appears to be relatively insignificant
compared to the repercussions of overclassification. In-prison behavior and recidivism
appear to be lower when inmates are placed in less secure settings than in more secure
ones.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 135
December 19, 2011

•

A number of factors need to be considered in reducing misclassification: 1) Basing inmate
supervision levels on violent crime convictions does not correctly identify risk of institutional
violence nor recidivism; 2) Insensitivity to the impact of mental illness on behavior can
produce incorrect classification decisions; 3) Female classification can probably be improved
if based on need factors, and not solely behavior.

Research indicates that a classification system should have the following components in order
to minimize the risk of misclassification:
 Implementation of a classification system that has been proven to be both valid and
reliable.
 Full automation and recording of each classification decision.
 A centralized, adequately staffed classification unit that is responsible for monitoring and
preparing policies and procedures.
 A process for the initial and annual reclassification of each inmate.
 Use of over-rides for agency-approved reasons.
Recent Developments in State Correctional Classification Systems
•

By the first decade of the 21st century, many state correctional agencies began efforts at
validating their classification systems, exploring the need for gender-responsive
classification systems, and determining how to best use computer systems in all areas of
classification work. These concerns were prompted by sentencing laws such as ThreeStrikes, which increased the prison population (including a growing female population),
litigation by inmates regarding conditions of confinement, and the increasing use computer
systems.

•

Currently the majority of state correctional agencies are focused on reducing costs, but are
not looking to their classification systems to achieve this reduction. Rather, they are
expanding eligibility for programs that allow inmates to earn sentence reductions and
developing programs for parolees that are intended to reduce the number returning to
custody. Most of new programs were initiated in the last two years. Their actual effect on the
size of the inmate population remains to be seen.

CDCR’s Inmate Classification System Study is characteristic of the research on classification
found in the literature. It uses data analysis rather than experimental designs with randomized
groups. Its Mandatory Minimum Scores, which require inmates to be placed in security levels
based on their convictions for specific crimes, is not evidence-based. It is important to note that
research suggests that overclassification may be criminogenic for some inmates.
Unique among state correctional agencies, CDCR is proposing to amend its classification
system in order to relieve overcrowding. This would come about by moving appropriately
reclassified inmates from more-crowded secure-housing units to the less-crowded facilities for
lower-level offenders.
.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 136
December 19, 2011

Literature Review on Misclassification
Full Review
Introduction
Inmate classification systems are developed and implemented to assign inmates to custodylevel appropriate placements that maximize safety and security for the staff and inmates within
and the community without its walls. The literature sometimes discusses two types of systems:
external and internal. An external system places an inmate in a prison, while an internal system
places an inmate in housing and programming within a prison. Within the California Department
of Corrections and Rehabilitation (CDCR), the external system is characterized by the pointand-level Inmate Classification Score System. The internal system is characterized by the
CDCR Close Custody Designations. The consensus in the literature is that external systems are
fairly well developed while internal systems need further refinement.
Current CDCR projections anticipate great availability of lower security level housing in the near
future due to the transfer of lower custody-level inmates to the counties. In order to maximize
the use of this anticipated housing space, CDCR is examining both the internal and external
components of its classification system. The purpose of this literature review is to inform this
process.
In particular, this literature review addresses two issues regarding inmate
classification systems: 1) misclassification, and 2) what other states are doing with their
classification systems.
The first issue, misclassification, refers to the incorrect placement of an inmate in a prison.
Overclassification refers to the placement of an inmate in a setting that is unnecessarily
restrictive; underclassification refers to a placement in a setting that does not provide adequate
security and supervision. Misclassification may be due to incorrect classification decisions or to
deficiencies in a classification system.
The second issue concerns the changes that have occurred in the past few years to
classification systems in other states.
Methodology
The following databases were searched for relevant literature on inmate classification systems:
•

Academic Search Complete

•

JSTOR

•

Google Scholar

•

MEDLINE with Full Text

•

PsychARTICLES

•

PsychINFO

•

Psychology and Behavioral Sciences Collection

•

SocINDEX with Full Text

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 137
December 19, 2011

The search terms varied across the databases (which varied in design and subject terms), but
the predominant terms used were:
•

Prisoners and Classification

•

Criminal Rehabilitation

•

Female Offenders and Classification

•

Prisoners and Mental Health and Classification

Free-text searching terms included:
•

Misclassification and Prisoners

•

Underclassification and Prisoners

•

Overclassification and Prisoners

Criteria for inclusion in order to focus on misclassification:
•

Adult jail or prison inmates

•

Classification policies that are not evidence-based

•

Factors that negatively affect the quality of classification actions, such as subjectivity

•

Factors that may have been incorrectly or inadequately addressed in classification
systems, such as mental illness or gender

Criteria for exclusion that do not address misclassification:
•

Descriptive-only reports on current systems that did not address misclassification

•

Experimental designs of proposed classification systems

•

Studies of statistical or other methodologies used to validate systems

Fifty-four articles or reports were selected and either downloaded or ordered from the State
Library. After this initial review, 26 were found to be relevant. Publication dates were between
1970 and 2011. Additional information on recent efforts by other state correctional systems to
address overcrowding was obtained through telephone and email conversations with
Classification staff in the New York State and Illinois Departments of Correction.
I. Misclassification
Inmate classification systems basically implement correctional policy and, therefore, tend to
reflect the concerns of the era in which they were developed. Concerns about “super predators”
in the 1980s and 1990s are reflected in more restrictive classification systems. Later research
has found that the actual threat posed by super predators was much less than originally thought
and was based largely on anecdotal evidence that the mass media magnified into mythic
proportions. By the first decade of the 21st century, classification research and development
began focusing on validating classification systems, exploring the need for gender-responsive
classification systems, and automating all areas of the classification process. These concerns
have been prompted by sentencing laws such as Three-Strikes, which increased the prison
population (including a growing female population), litigation by inmates regarding conditions of
confinement, and the increasing use of computers in correctional work. More recent attention to
prison costs and lack of rehabilitative services have prompted correctional researchers and
policy makers to look at ways to reduce inmate restrictions and at the same time maintain – or
even increase -- institutional and community (public) safety.
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 138
December 19, 2011

No single article identified for this literature review focused solely on underclassification. Where
underclassification is discussed, it is primarily in terms of something to avoid. Berk (2006) in his
study of the CDCR classification system described how CDCR officials preferred the possibility
of ten inmates being placed in overly restrictive housing than to have one inmate placed in
housing that was not restrictive enough. Furthermore, there may be some reluctance to
publishing studies on underclassification as it would imply a failure in public safety.
On the other hand, overclassification is discussed from a number of angles, which will be
described below. While overclassification may prompt litigation, exacerbate mental illness and
even increase recidivism, it is not seen as an immediate threat to public safety.
Limitations in the Research
Limited Research
Experimental research on misclassification, in which inmates are randomly assigned to study
and control groups and subjected to various conditions to test their responses, is extremely
limited due to safety and legal concerns. Most studies analyze data on previous behavior of
different groups of inmates. Therefore, study findings for the most part are predictions of
possible future behavior, not assessments of actual recent behavior.
There is limited empirical research (experimental or observational) to support classification
practices and policies that are intended to control inmate behavior. In their meta-analysis, Byrne
and Hummer (2007) only identified seven research studies published between 1984 and 2006
that examined the relationship between classification decisions and inmate behavior. They
concluded that the classification systems they examined do not predict inmate behavior and do
not reduce prison violence.
Sample Sizes
The number of subjects in studies ranged from widely. The percentage of the total inmate
population within a system used for research also varied. As a result, it is difficult to compare
and generalize findings.
Definitions of Behavior Differ
The literature on classification frequently refers to violent misconduct on the part of prison
inmates. However, the specific behavior is usually not stated. It could be murder, assault,
pushing or shoving, taking part in a group disturbance under pressure from other inmates, etc. It
is also important to note that behavior that is tolerated without taking disciplinary measures in
one correctional system can cause disciplinary action in another. Also, within one system,
individual institutions, and even individuals within an institution, can differ in their disciplinary
attitudes.
Inmate Populations May Vary
The characteristics of inmate populations (e.g., race, ethnicity, age) vary based on the
geographic location and age of the literature. Research conducted in southern states typically
refers to white and black inmates, while newer reports on inmates in southwestern or western
states include white, black, and Hispanic inmates.
Correctional System Organization
The majority of the articles in this literature review focus on state prison inmates. However,
states differ in how their prisons are organized and how their offenders are sentenced. For
example, individuals classified as minimum security inmates in California prisons may be placed
in state jails in Texas or in community correctional facilities in Ohio.
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 139
December 19, 2011

Contributing Factors to Misclassification
The available research has identified multiple factors that contribute to misclassification. They
can be broken into three categories: convictions and sentencing, inmate demographic factors,
and staff use of classification systems.
Convictions and Sentencing
Criminal Conviction
Superficially it may appear logical that an individual convicted of a violent crime like murder
requires the strictest level of supervision in prison. However, research does not definitively
support a connection between behavior as a criminal and behavior as a prison inmate.
Research has shown that classification policies based on conviction for violent crimes result
in overclassification. Individuals convicted of violent crimes tend to be better behaved in
prison than individuals convicted of nonviolent crimes (Waldo, 1970; Alexander, 1994; Austin,
2003; Sorenson and Cunningham, 2010). For example, one study found that individuals
convicted of murder were 8.5% less likely to commit rule violations that are defined as having
the potential to result in a violent outcome when compared to inmates with drug, property, or
public order convictions.
Sentence Length
“Nothing left to lose” is a phrase used when discussing inmates with life sentences, or
“LWOPs” (life without possibility of parole). LWOPs are seen as inmates who have no
incentive to behave because they have no possibility of leaving prison. However, the little
research that has been conducted on the relationship between length of sentence and
behavior found that sentence length is not a predictor of in-prison behavior (Austin, 2003;
Cunningham and Sorensen, 2006). Cunningham and Sorenson (2006) compared the
behavior of inmates sentenced to life without the possibility of parole to the behavior of
inmates with various sentences ranging from 10 to 30 or more years. The LWOPS had the
same average age at admission as those in the comparison groups and were similar
demographically in other respects as well. They found that LWOPS were 21% less likely to
engage in violent behavior than inmates serving 10 – 14 years, 15% less likely than inmates
serving sentences of 15 – 19 years, and 7% less like than inmates serving sentences of 30
or more.
Inappropriate Use of Demographic Factors
Mental Illness
Mental illness contributes to preventable overclassification of some inmates. Those with
mental illness are more likely to misbehave than those who do not suffer from mental illness
and, consequently, are more likely to be disciplined more often. Their classification scores
rise, which can lead them to be placed in adminstrative segregation. This restrictive
environment can exacerbate their mental illness, and thus lead to more misbehavior. Kupers
(2009) described this downward spiral of mentally ill inmates in his report on administrative
segregation (Ad Seg) in Mississippi. The litigation that prompted the Mississippi Department
of Corrections (MDOC) to create a new classification system also resulted in mental health
treatment for inmates in Ad Seg who have been diagnosed as being seriously mentally ill.
The Correctional Association of New York (2004) also studied mentally ill inmates in New
York State prisons by reviewing records and conducting interviews with staff and inmates.
Findings were that inmates with known mental health issues comprised approximately 11%
of the total inmate population, but represented 20% to 60% of the population in secure
housing, depending on the institution under examination. Recommendations resulting from
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 140
December 19, 2011

this study included expanding treatment programs, training correctional officers in mental
health issues, creating suicide prevention programs in all secure units, and creating an
independent review board to ensure that reforms are actually made. Due to litigation and
changes in the law that took place in 2008 and 2009, the New York State Department of
Corrections has expanded treatment for mentally ill inmates. For example, the Special
Housing Unit (SHU) Exclusion Law of 2008 provides for alternative placement for all inmates
with serious mental illness and disciplinary confinement requirements.
A contributing factor to the over-respresentation of mentally ill inmates in Ad Seg is the “bad
versus mad” dilemma (Toch, 2007). This occurs when custody staff refer a mentally ill
inmate who is difficult to manage to the mental health staff for treatment. Mental health staff
will then affirm that the inmate is mentally ill, but state that his misbehavior was not due to his
illness, but rather due to malingering. It appears that custody staff and mental health staff
each want to the other to take responsibility for managing these inmates.
Gender
The increase in the female prison population over the past two decades created an interest in
gender-responsive classification of female inmates. All states use some form of classification
system for female, as well as male, inmates. Most use the same system for both genders,
but questions have been raised about the validity of this practice.
In an examination of 10 state classification systems, Hardyman, Austin and Tulloch (2002)
found that 4 of 10 states that were in the process of revalidating their classification systems
wanted to assess the need for either a separate system for female offenders, or potential
modifications to current systems for female classification. Oklahoma concluded that
modifying the cutoff points between classification levels in their current system for women
was appropriate, but a separate system was unnecessary. Oregon, Rhode Island, and
Virginia did not find a need for separate systems or modifications to their current systems.
Montana, while not making gender issues central to their research, found that a newly
developed instrument to assess predatory behavior was valid for male and female inmates.
Van Voorhis et.al., (2001) identified the following concerns with female offenders through a
survey of all 50 state correctional systems:
•

Ten states overrode more than 15% (rates ranged from 18 to 70%) of female inmates’
classification scores, which indicates that their systems were not working well with
women.

•

Respondents from 49 states identified needs and problems that they believed were
unique to female inmates, such as the need for trauma treatment and the impact of their
incarceration on their children, but only 8 states attempted to classify female inmates
differently from male inmates.

While research has demonstrated that female inmates were violent less often than male
inmates, seven risk predictors of assault (severity of prior offense, severity of current offense,
number of prior felony convictions, current age, stability factors, escape history, and history of
institutional violence) were the same for both genders (Van Voorhis, et.al., 2001).
Research on female offender classification differs from research on male offender
classification in that it sometimes includes recommendations to add a component based on
needs such as mental health treatment, substance abuse treatment, or recovery from abuse
(Austin, 1993; Farr, 2000; Hardyman, 2002). No explanation is given as to why these needs
should be addressed for women, but not men.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 141
December 19, 2011

At this time the literature does not provide conclusive evidence that female inmates require
their own classification system; however, research strongly suggests a need to at least
include gender-responsive items for female offender classification.
Staff Use of Classification System
Inadequate Staff Training and Experience
Any classification system is only as good as the people using it. Inadequate training and/or
experience will affect the quality of classification decisions. A number of studies emphasize
the need for strong training and ongoing monitoring to ensure the system is used correctly
(Alexander, 1994; Owens, Will and Camp, 1995; Austin, 2003).
Subjectivity
Kupers (2009) described Mississippi’s experience with the overclassification of inmates
assigned to the Ad Seg unit in the Mississippi State Penitentiary. Prior to 2002, classification
was based on subjective decision-making. In response to litigation, the MDOC, working in
conjunction with Dr. James Austin, created a new, objective classification system. After its
introduction, more than 75% of Ad Seg inmates were reclassified to a lower security level.
After these inmates were moved to less restrictive housing, serious incidents in the institution
fell by 70%. As of 2009, 1% of the inmates in the Mississippi State Penitentiary were in Ad
Seg, compared to 3% of inmates in other states’ prisons.
Bonta and Motiuk (1990) found that when subjective assessments were used, 16% of jail
inmates were found to be eligible for halfway houses. However, when an objective
assessment instrument was used, 51% were considered eligible. Austin and Chan (1993), in
a study on classifying female inmates, revealed that 40% of all classification decisions in the
Indiana Department of Corrections were overrides (classification decisions that conflict with
established criteria), 37.8% of which were upward (primarily from minimum to low-medium
and high-medium to maximum).
Potential Consequences of Misclassification
There are a number of possible consequences of misclassification, ranging from better-thanexpected behavior to murder. In Alexander’s (1994) study, prison staff and inmates in
Pennsylvania and Nevada were interviewed as part of a qualitative assessment of classification.
Instead of using statistical analysis, Alexander looked at a small set of cases in order to discern
factors that influence inmate behavior. He looked at both under- and overclassified individuals.
His findings were that most high-custody inmates are not involved in serious misconduct, but
have higher infraction rates than inmates housed in lower custody settings. Lower custody
inmates who engaged in serious misconduct had lower infraction rates than inmates housed in
high security settings.
Underclassification
The findings of research on underclassification are mixed. There is some research indicating
that there were no significant differences in overall or serious misconduct among inmates who
were housed in environments that were less restrictive than their classification level would
suggest necessary (Bench and Allen, 2003; Camp and Gaes, 2005). Fewer restrictions did not
result in more misbehavior.
However, underclassification can result in obvious problems. One example cited in the literature
is that regarding a private prison in Youngstown, Ohio that housed 1,700 male offenders from
the District of Columbia (Clark, 1998). A new security level, “high medium,” was informally
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 142
December 19, 2011

created because the private prison’s contract prohibited accepting high security; excessive
overrides were used and 274 inmates who had been classified as maximum elsewhere became
high medium. In addition, over 200 inmates who should have been kept separate due to enemy
concerns were mixed in with the population as a whole. After two murders, several riots, and
the escape of 6 inmates at once, inmates were reclassified, resulting in 19 transferring to
maximum security settings. Misclassification, however, was not the only problem at this prison
and it is not clear to what degree it contributed to the behaviors mentioned above.
Overclassification
Overclassification may have an effect on inmates beyond their time in prison. Chen and
Shapiro (2004; 2007) conducted a study on the Federal Bureau of Prisons classification system,
which has five levels: minimum, low, medium, high, and administrative. They found that moving
an inmate from a minimum security (least supervised) to a low security (one level more
supervised than minimum) setting doubles the inmate’s chances of being rearrested within three
years. Gaes and Camp (2009) found similar results in their study on a group of CDCR Level III
inmates, some of whom were randomly assigned to Level III environments and some of whom
were placed in Level I environments. Level III inmates housed in Level III settings had a 31%
higher chance of returning to prison than Level III inmates in Level I settings. These studies
suggest there may be some truth to the cliché about prison being a crime school. Inmates who
are overclassified may learn new criminal behaviors through interaction with more experienced
criminals.
Components of Classification Systems That Can Minimize Misclassification
Research has been conducted to identify the components needed for a classification system to
minimize the risk of misclassification. Austin (2003), under the auspices of the National Institute
of Corrections (NIC), concluded the following components are crucial to have in a classification
system in order to reduce the potential for misclassification:
•

The use of criteria that have been demonstrated through research to use both reliable
and valid factors to assess a prisoner’s custody level;

•

A centralized classification unit that is adequately staffed with well trained
professional personnel who have control over all inter-agency transfers;

•

A centralized classification unit that is responsible for monitoring institutional
classification units and preparing all policies and procedures that pertain to classification;

•

A fully automated classification system such that each classification decision, and the
factors used to make each decision, is recorded and available for analysis;

•

An initial and reclassification process where all prisoners are reviewed at least annually
to update and possibly modify the prisoner’s current classification level; and,

•

The use of over-rides to allow staff to depart from the scored classification level for
reasons approved by the agency.

II. Recent Developments in State Classification Systems
The second part of this literature review looks at recent changes to state classification systems
and how individual state correctional departments are attempting to modify their policies in light
of current economic hardships.
In the first few years of the 21st century, state correctional departments were eager to analyze,
validate, and improve their classification systems. A tremendous growth in inmate population
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 143
December 19, 2011

due to truth-in-sentencing and three strike laws, more female inmates, and the building of
“supermax,” expensive maximum custody housing contributed to this interest. Other reasons
include inmate lawsuits regarding living conditions and the growing use of automated systems.
The National Institute of Corrections (NIC) funded a 15-month studiy in which 10 states
addressed issues regarding their classification systems (Hardyman, Austin and Tulloch, 2002).
States volunteered to take part in the studies and were selected based on their commitment to
analyzing and improving their classification systems and practices. NIC technical specialists
assisted the participants in identifying goals and methologies, and provided training for any new
practices resulting from the studies. The final study report illustrated both the similarity and
diversity of issues state correctional departments faced with respect to their classification
systems. Each state had its own reasons for participating, but the most frequent was the need
to validate their classification system.
Delaware Department of Corrections
This department’s goal was to design and validate an objective classification system in order
to improve the efficiency and effectiveness of its classification practices. Initial and
reclassification custody assessment instruments were developed and refined to create
statistically strong instruments for identifying the safety and security risk an offender poses.
The combination of the factors of current offense, other offense/bail status, escape history,
current age, criminal history, time to serve and institutional program performance was found to
be effective in assessing an inmate’s level of risk. The new classification tools were included
in the department’s automated information system.
Montana Department of Corrections
In response to litigation, the department agreed to modify its classification policy and
instruments to identify predatory and vulnerable inmates. Data analyses demonstrated that
the modified instrument differentiated predatory and vulnerable inmates and suggested that it
was appropriate for both male and female inmates.
Oklahoma Department of Corrections
This department’s goal was to refine its classification risk factors to better assess the risks
posed by female offenders. In particular, the age, current offense, criminal history, and
escape items were revised. The custody scale cut points were also adjusted to create
statistically distinct custody levels. Oklahoma has implemented the revised instruments and
custody scales and addressed the availability of its services and programs to ensure that
women are placed within the least restrictive custody level possible.
Oregon Department of Corrections
Alternative classification scoring criteria were created and simulated in order to assess the
negative impacts of a truth-in-sentencing initiative and other sentencing reforms that had
altered the number and type of offenders and expected lengths of stay. Because the
classification system in place at the time relied heavily upon an inmate’s expected time to
serve, it was anticipated that the system would overclassify its inmate population. As a result
of this validation effort, minor revisions to the instruments and custody matrix were
implemented. The department also studied the need for a separate classification system for
women, which led to plans to pilot test
Rhode Island Department of Corrections
The department wanted to revalidate and adjust its objective classification system to ensure
its appropriateness for both male and female inmates and to design a systematic
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 144
December 19, 2011

administrative review process to document the transfer to minimum custody of inmates with
sentences of less than 6 months.
Analysis of the existing system suggested the need to create a subset of predatory
institutional infractions to identify aggressive inmates, develop an offense severity index that
focused on institutional risk, and modify the custody scale cut points. Pilot testing of the
administrative screening process suggested that the instrument would serve as a simple,
objective mechanism for identifying cases appropriate for minimum custody.
The analysis also suggested the need to revise the department’s disciplinary code; develop a
public safety screening instrument for work release and community housing decisions;
discontinue regular custody assessments for minimum-custody inmates; restrict discretionary
overrides; provide intensive, ongoing training to all classification staff; develop a strong,
centralized classification unit; and upgrade the automated information system.
Tennessee Department of Correction
Before this study, the department discontinued the use of an initial classification instrument
because it had been found to frequently overclassify inmates. For example, over 90% of
inmates who were initially classified as close custody when they were received by the
department dropped to minimum custody after four months (the time of their first
reclassification). While a revised initial classification tool had been developed, the department
wanted to determine if it should use it or continue without an initial classification form.
Findings indicated that the revised instrument was an improvement over the original.
Following some minor adjustments, plans were made to begin using the new initial
classification instrument.
Texas Department of Criminal Justice
Policies, procedures, and operations pertaining to inmates in Ad Seg were reviewed with the
goal being to reduce the number of inmates in this type of housing. Based on statistical data
and a review of departmental policies, the following recommendations were made: continue
single-celling, revise the criteria for placement and retention; and develop ongoing
management reports to monitor and evaluate the Ad Seg policies. In addition, it was
suggested that an anti-gang housing unit program be created.
Virginia Department of Corrections
This department wanted to create an instrument that would tie basic classification information
to housing and work/program assignments. Information on an inmate’s demographic factors
and criminal and incarceration history were used to assess security needs. This information
was then used to make decisions on housing and education or work assignments. The
preliminary instrument was tested and refined. Six levels of risk and appropriate housing
based upon risk were identified. The department adopted use of this new system.
The department also wanted to know if it should create a separate classification for female
inmates. The new system was found to be valid for women as well as men, and it was
determined that a separate system was not needed.
Wisconsin Department of Corrections
The reliability and validity the department’s classification system were assessed, as well as
prevalence of racial bias in the system and the frequency of discretionary overrides. The
analyses indicated that the system was reliable; however, the validity of the risk assessment
was questionable and it appeared that the system overclassified many inmates. For example
approximately 10% of the inmate population were classified as high risk but assigned to
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 145
December 19, 2011

minimum custody. No racial bias was evident in the risk or Custody Designation process
because equal proportions of black and white inmates were assigned to the various risk and
Custody Designations. Key recommendations for improving the reliability, validity, and
automation of the system were provided. Stricter controls on overrides and additional training
for staff were also suggested.
Wyoming Department of Corrections
This department had two goals: assistance with staff training and the development of
organizational and/or procedural changes necessary for the full implementation of their
classification system. Intensive staff training that included reliability testing was provided, as
was a detailed classification manual. A comprehensive classification policy that provided for a
centralized classification unit and independent audits to periodically review a random sample
of the classification instruments for accuracy and completeness was implemented. Audits
completed at the four facilities indicated scoring error rates of less than 10%.
The Current Situation
Large correctional systems are a financial burden to state governments. Recent economic
conditions have prompted state correctional departments to reevaluate their management of
inmates.
The Vera Institute’s national survey, The Continuing Fiscal Crisis in Corrections: Setting a New
Course (2010), in which 44 states participated (California did not), identifed short-term and longterm efforts being made to reduce costs. In the short-term, every state was implementing at
least one type of cost-cutting measure, the most common being staff reductions or hiring
freezes.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 146
December 19, 2011

States are hoping to reduce costs long-term in other ways, as described below:
Reducing Technical Violations:
•

Alabama now limits to 90 days the sentences for eligible nonviolent technical offenders;

•

Kentucky allows its parole officers to sentence violators to county jail for up to 10 days,
with a total of 30 days per year;

•

Colorado allows community punishment for low-level nonviolent offenders;

•

Iowa can either revoke probation or extend it for up to one year.

Alternatives to Incarceration:
•

Vermont has made probation the standard sentence for misdemeanors and nonviolent
felonies;

•

Florida eliminated prison sentences for certain nonviolent felonies;

•

Louisiana and Vermont now allow courts to sentence offenders to home confinement;

•

Florida now diverts nonviolent offenders to recidivism reduction programs;

•

Vermont created a system of community reparative boards to determine alternative
sentences for nonviolent offenders; eligibility for alternative sentencing has also been
expanded to second time offenders;

•

South Carolina crated alternative sentences for some drug offenses;

•

Washington created alternative sentences for nonviolent offenders who have custody of
children under the age of 18.

Relaxing Mandatory Sentences:
•

New York eliminated mandatory minimum sentences and reinstated judicial discretion in
low-level drug cases;

•

New Jersey amended its drug laws to allow judges to apply mandatory minimum
sentences or probation for certain offenses;

•

Minnesota allows judges to deviate from mandatory minimum drug cases if the
prosecutor requests it;

•

Rhode Island eliminated some mandatory minimum sentences;

•

Delaware amended its mandatory minimum sentencing policies by allowing courts to
alter sentences of one year or less for individuals with serious medical problems
requiring continuous treatment.

Expanding Release Opportunities:
•

Oregon and Mississippi expanded eligibility for education-based credits to inmates
convicted of high-level offenses;

•

Louisiana made good-time credits retroactive to 1992, except for inmates convicted of
violent crimes and sex offenders;

•

Colorado increased the number of days nonviolent, program-compliant inmates can earn
to reduce their sentences each month from 10 to 12 days.

Restructuring Conditions of Release:
California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 147
December 19, 2011

•

Indiana now requires inmates who have served 21 years and earned four years of credit
time to be referred to the parole board for possible release;

•

Kansas and New York created early release programs for terminally ill inmates who are
incapacitated and pose no threat to the community;

•

Kentucky allows nonviolent offenders convicted of low-level felonies with 180 days left
on their sentences to be released to home incarceration;

•

South Carolina now requires nonviolent offenders who have been in prison for at least
two years to be released before their release date; this state has also expanded
eligibility for parole or work release within three year of release date to include inmates
convicted of drug offenses;

•

New Hampshire now automatically releases inmates who have not previously paroled
within nine months of their maximum sentence;

•

Louisiana lowered requirements for parole, now requiring only a majority vote by parole
board members instead of unanimity;

•

West Virginia created an accelerated parole program whereby inmates are eligible for
early release if they complete individualized rehabilitative programs; this state has also
expanded annual parole reviews to eligible individuals who are serving life sentences;

•

Vermont now releases all nonviolent offenders to furlough programs if they have served
their minimum sentences and completed rehabilitative program goals; individuals are
now eligible within 180 days before the end of a minimum sentence, instead of the
previous 90 days.

It is important to note that all of these strategies to reduce state prison populations were
enacted either in 2009 or 2010. It is too soon to tell what kind of impact they are having. The
actual number of inmates who are leaving prison early or who do not recidivate due to these
policy changes has not yet been determined.
In addition, states may not be able to fully implement these initiatives due to political pressures.
For example in 2009, Illinois developed a program that allowed some offenders to earn credits
to shorten their sentences. Inmates who were convicted of driving under the influence (DUI)
were included in the category of nonviolent offenders who qualified for the program. Public
anger over letting felons with DUIs out early became so intense that the program was shut
down. It remains to be seen how many of the states’ new efforts to reduce the prison population
will actually achieve their goals.
Conclusion
Although research is somewhat limited, it is clear that misclassification is a serious problem.
Although underclassification can be a problem, it appears to be relatively insignificant compared
to the repercussions of overclassification. In-prison behavior and recidivism appear to be lower
when inmates are placed in less secure settings than in more secure ones.
A number of factors need to be considered in reducing misclassification: 1) Basing inmate
supervision levels on violent crime convictions does not correctly identify risk of institutional
violence nor recidivism; 2) Insensitivity to the impact of mental illness on behavior can produce
incorrect classification decisions; 3) Female classification can probably be improved if based on
need factors, and not solely behavior.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 148
December 19, 2011

The research literature is clear regarding the components required for a classification system
that minimizes misclassification. These components are: 1) The classification system must be
one that has been proven to be both valid and reliable on similar populations; 2) It must be fully
automated and must record each classification decision; 3) It must be operated from an
adequately staffed, centralized unit that is given the responsibility and authority for its monitoring
as well as for the development of its policies and procedures; 4) It must include a process for
the initial and the regular reclassification of each inmate on at least an annual basis and; 5) Its
use of over-rides must be limited to agency-approved reasons.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 149
December 19, 2011

Bibliography
Alexander, J. A. (1994). Understanding Why Inmates are Misclassified; the Use of Qualitative
Methods in Evaluating Objective Prison Classification Systems. National Council on Crime and
Delinquency.
Annin, P. (1996, January 22). 'Superpredators' Arrive. Newsweek, 127 (4), p. 57.
Austin, J. (2003). Findings in Prison Classification and Risk Assessment. Issues in Brief,
National Institute of Corrections.
Austin, J., and Chan, L. E. (1993). Indiana Department of Corrections Women Classification
Study. National Council on Crime and Delinquency.
Bench, L. L., and Allen, T. D. (2003). Investigating the Stigma of Prison Classification. Prison
Journal, 83 (4), 367-382.
Berk, R. A. (2006). Forecasting Dangerous Inmate Misconduct. University of California .
California Policy Research Center.
Bonta, J., and Motiuk, L. L. (1990). Classification to Halfway Houses: a Quasi-Experimental
Evaluation. Criminology, 28 (3), 497-506.
Byrne, J., and Hummer, D. (2007). In Search of the "Tossed Salad Man" and Others Involved in
Prison Violence. Aggression and Violent Behavior .
Camp, S., and Gaes, G. (2005). Criminogenic Effects of the Prison Environment on Inmate
Behavior; Some Experimental Evidence. Crime and Delinquency, 51, 425-442.
Chen, K. M., and Shapiro, J. M. (2004). Does Prison Harden Inmates? A Discontinuity
Approach. Cowles Foundation for Research in Economics. Yale University.
Chen, M. K., and Shapiro, J. M. (2007). Do Harsher Prison Conditions Reduce Recidivism?
Amerian Law and Economics Review, 9 (1), 1-29.
Clark, J. L. (1998). Report to the Attorney General Inspection and Review of the Northeast Ohio
Correctional Center (NEOCC).
Correctional Association of New York. (2004). Mental Health in the House of Corrections.
Cunningham, M. A. (2006). Nothing to Lose? A Comparative Examination of Prison Misconduct
Rates Among Life-Without-Parole and Other Long-Term High-Security Inmates. Criminal Justice
and Behavior, 33 (6), 683-705.
Cunningham, M. D., and Sorenson, J. R. (2007). Predictive Facotrs for Violent Misconduct in
Close Custody. Prison Journal, 87 (2), 241-253.
Cunningham, M. R. (2008). Assertions of "Future Dangerousness" at Federal Capital
Sentencing: Rates and Correlates of Subsequent Prison Misconduct and Violence. Law and
Human Behavior, 32, 46-63.
Farr, K. A. (2000). Classification for Female Inmates: Moving Forward. Crime and Delinquency,
46 (1), 3-17.
Gaes, G. and Camp, S. (2009). Unintended conseqences: experimental evidence for the
criminogenic effect of prison security level placement on post-release recidivism. Journal of
Experimental Criminology, 139-162.

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 150
December 19, 2011

Hardyman, P. L.,
Austin, J. and Tulloch, O.C. (2002). Revalidating External Prison
Classification Systems: the Experience of Ten States and Model for Classification Reform.
George Washington University, Institute on Crime, Justice and Corrections.
Kupers, T. D. (2009). Beyond Supermax Administrative Segregation; Mississippi's Experience
Rethinking Prison Classification and Creating Alternative Mental Health Programs. Criminal
Justice and Behavior OnlineFirst, 1-14.
Owens, C. E., Will, J., and Camp, H. (1995). Inmate Risk and Needs: Developing an Objective
Means for Classifying Florida's Inmate Population. University of North Florida, Northeast Florida
Center for Community Initiatives.
Sorenson, J. and. (2010). Conviction Offense and Prison Violence: a Comparative Study of
Murderers and Other Offenders. Crime and Delinquency, 56 (1), 103-125.
Toch, H. (2007). Violence in Prisons, Revisited. Journal of Offender Rehabilitation , 45 (3/4), 128.
Van Voorhis, P., Peiler, J., Presser, L. Spiropoulis, G. and Sutherland, J. (2001). Classification
of Women Offenders: A National Assessment of Current Practices and the Experience of Three
States. Center for Criminal Justice Research. University of Cincinnati.
Vera Institute of Justice ( 2010). The Continuing Fiscal Crisis in Corrections: Setting a New
Course.
Waldo, G. P. (1970). The "Criminality Level" of Incarcerated Murderers and Non-Murderers .
Journal of Criminal Law, Criminology and Police Science, 61 (1).

California Department of Corrections and Rehabilitation
Office of Research/Research and Evaluation Branch

Page 151
December 19, 2011