Felony and Violent Recidivism Among Supermax Prison Inmates in
Washington State: A Pilot Study

David Lovell, Research Associate Professor
Clark Johnson, Research Associate Professor
Department of Psychosocial & Community Health
University of Washington

This research was funded by a grant from the Intramural Research Funding Program in
the School of Nursing, University of Washington. The authors wish to acknowledge the
assistance of the Dean and staff in the Office of Research, School of Nursing, for their
support and helpful suggestions in the development of this project. We are also grateful
to Polly Phipps and colleagues at the Washington State Institute of Public Policy, and to
Peggy Smith at the Office of Planning and Research, Department of Corrections, for
providing data on study subjects. Further helpful suggestions were provided by Elaine
Thompson, by Gregg Gagliardi of the Washington Institute for Mental Illness Research
and Training, and by Steve Aos, Associate Director of WSIPP.
Table of Contents

Lovell & Johnson

Recidivism of Supermax Inmates

Executive Summary ........................................................................................................... iii
Introduction..........................................................................................................................1
Methods
Data Sources ..................................................................................................................2
Data Compilation ...........................................................................................................2
Identification of Subjects ...............................................................................................3
Control Procedures
Post-hoc Controls.....................................................................................................3
Controlling for Mental Health Status.......................................................................5
Dependent and Independent Variables ..........................................................................7
Results
Profile of Subjects, Controls, and All Men Released from Prison ...............................7
IMU Group Membership
Rates of Felony Recidivism .....................................................................................8
Seriousness of New Crimes .....................................................................................9
Confounding Variables ..........................................................................................10
Timing of New Offenses........................................................................................11
Amount of Time in IMU..............................................................................................11
Timing of Release from IMU ......................................................................................13
Discussion
Summary of Findings...................................................................................................15
Two Problems of Interpretation .............................................................................16
Methods of Further Analysis .......................................................................................16
Factors for Further Investigation .................................................................................18
Significance of Preliminary Findings
Timing of Release ..................................................................................................19
IMU Group Membership .......................................................................................19
References..........................................................................................................................20
Appendix A

ii

Lovell & Johnson

Recidivism of Supermax Inmates

Executive Summary
Purpose
Since the early 1980’s, most prison system have built specially designed facilities that
keep selected offenders in lockdown status for lengthy periods of time, sometimes years,
on the grounds that they pose a danger to the prison community. There have been several
successful court challenges to some aspects of these practices, for example confinement
of psychologically vulnerable inmates in such facilities, but there is scarcely any
systematic research on who gets assigned to supermax, how it affects them while they’re
there, whether such facilities actually reduce violence within prison systems, and whether
it has any bearing on their later behavior.
This study compares recidivism in the community by released offenders who were and
who were not subjected to substantial periods of supermax confinement while in prison.
Specifically, we ask:
1. Whether supermax assignment is associated with the probability, seriousness, or
timing of new offenses.
2. Whether the probability, timing, and seriousness of new offenses is associated with
(a) the amount of time offenders spend in supermax environments or (b) the length of
the interval between transfer out of supermax and release to the community.
Methods
The subject pool comprised all offenders released from DOC facilities during the index
years 1997 and 1998.
•

The Department of Corrections provided data on demographics and correctional
behavior, including mental illness indicators, infractions, and assignment to
supermax facilities: Intensive Management Units (IMUs).

•

The Washington State Institute for Public Policy provided criminal history and
post-release offense data.

Files were compiled into a format suitable for analysis. In the 8,000-member male
offender pool, there were 242 IMU subjects, who had spent at least three continuous
months in IMU (most of them much longer); they were individually matched with nonIMU controls on eight felony recidivism predictors, recoded where necessary into
categorical or ordinal variables:
•

Past felonies

•

Past misdemeanors

•

Felony versatility

•

Age of first offense

•

First-time sex offender (Y/N)

•

Past drug felonies

•

Past violent felonies

•

White or Minority Group (2 values)

iii

Lovell & Johnson

Recidivism of Supermax Inmates

Though only slightly correlated with recidivism, ethnicity was included as a control
variable because of the salience of racial identification in prisons and particularly in gang
membership issues that affect assignment to IMU. A separate matching procedure was
conducted for subjects with mental illness, matching them with mentally ill controls on
the first 5 control variables.
•

There were 21% of IMU subjects with strong indicators of serious mental illness
(SMI), vs. 5% of non-IMU pool members.

•

In each group (IMU subjects and controls), there were 52 SMI members, 190 nonSMI.

Felony recidivism was defined as the commission of a new felony within three years of
release, for which the offender was adjudicated guilty. Person offense recidivism was
defined as commission, followed by conviction, of a misdemeanor assault or a felony
offense against persons (robbery, assault, sex offenses, homicide).
Results
IMU Group Membership. IMU group members, non-IMU group members, and the
entire offender pool are compared below on felony offense and new person offenses:

•

Outcome

IMU

Controls

All Men

New Felony

47%

40%

38%

New Person Offense**

36%

24%

21%

These results are not significant for felony recidivism (p=.1); though significant
for new person offenses (p=.004), IMU subjects and controls were not matched on
several significant predictors of new person offenses.

When subjects were partitioned into SMI and non-SMI subgroups, IMU assignment was
significantly associated with felony recidivism among non-SMI offenders:
•

Non-SMI felony recidivism: IMU=47%, controls=38%; χ2=3.12, p=.048 1-tailed

Among felony recidivists, IMU subjects committed more serious offenses than controls:
•

Violent felonies:

Among IMU recidivists, 54%
Among non-IMU recidivists, 39%
χ2=5.98 df=2, p=.05, 1-tailed

Cox regression results indicated that group membership was not significantly associated
with the hazard of committing a new felony, i.e., with how long offenders lasted in the
community before reoffending; but it was significantly associated with the hazard of
committing a new person offense: χ2=7.72, p=.005.
Length of Time in IMU. Among IMU group members, both felony recidivists and
felony non-recidivists averaged approximately 1 year of time in IMU during the index

iv

Lovell & Johnson

Recidivism of Supermax Inmates

incarceration. Length of time in IMU also had no bearing on seriousness of new offenses
or timing of new offenses.
Interval Between IMU Release and Prison Release. This study was stimulated by a
concern about whether releasing offenders directly from IMU into the community,
without any intervening period in normal prison settings to readjust to the presence of
others, was associated with a higher risk of recidivism. Of the IMU group, there were 59
members released to the community immediately after release from IMU.
•

New felonies:

•

This prediction effect remained significant when statistical controls were
introduced for other variables associated with immediate prison release: age at
first offense, number of past misdemeanors.

Immediate prison release group: 64%
Other IMU group members:
41%
χ2=9.8, df=1, p=.001, odds ratio=2.6

Implications
The findings reported above pose two related but distinct problems of interpretation:
•

The problem of statistical prediction. Where we found that IMU assignment or
timing of release predicted recidivism, are these robust findings? Would they
survive if different methods were used, or further statistical controls were
introduced?

•

The problem of causal interpretation. If there are robust statistical relationships
between recidivism and IMU assignment or timing of IMU release, to what extent
can these relationships be attributed to the effect of the IMU experience, rather
than to some unmeasured and uncontrolled disposition in the offender—for
example, psychopathy—that both provokes IMU assignment and leads to further
criminal aggression.

These questions define the terrain for future research, involving both further analysis of
the data collected for this pilot study and collection of additional information. With
additional data, more sophisticated statistical techniques may allow us to assess whether
our findings bear up when appropriate controls are introduced.
Infraction rates and SMI status are important factors for further investigation. When did
infractions occur (before, during, or after IMU assignment) and how serious were they?
If SMI offenders signed to IMU posed no greater risk of recidivism than non-IMU SMI
offenders, what were the differences that led prison staff to assign them to IMU?
In the absence of random assignment, which for obvious reasons must be ruled out in
IMU studies, threats to validity will always remain. Attempts to get at causal explanation
through collection of further data and longitudinal analysis of cases will narrow the range
of alternative explanations that may be considered true to life. Our findings about the
apparent effects of immediate prison release, and the role of mental illness, support
policy concerns that are more than academic.
v

INTRODUCTION
Since the early 1980’s, most prison systems have built specially designed facilities—
either stand-alone or inside larger prisons—to keep selected inmates in lockdown status.
This is a typical lockdown regime: inmates are confined to single cells around the clock,
leaving three times a week for showers and five times a week for solitary exercise; at
these times, they are shackled and escorted by a pair of officers; commissary and
property privileges are restricted; surveillance is continual; and on the very rare occasions
when inmates are in the same room with another person—for example, when meeting
with a review committee—they are caged or bolted down.
Critics of supermax confinement have collected evidence that a disproportionate number
of super-maximum custody prisoners have problems coping with prison due to mental
illness, brain damage, or other factors; that needed treatment is not provided; and that
vulnerable inmates are further damaged by sensory deprivation and other disorienting
features of the environment. Some studies of inmates in isolation indicate that even
those who start out healthy can become withdrawn, incapable of initiating or governing
behavior, suicidal, or paranoid (Grassian & Friedman, 1986; Haney, 1993, 2003). For
these reasons, the use of super-maximum confinement has led to successful litigation in
several jurisdictions (Jones’El v. Berge, 2001; Madrid v. Gomez, 1995).
Defenders of supermax continue to claim that evidence of its damaging effects is partial,
anecdotal, or limited to a few states that have successfully been sued. This defense is
possible because whether a person’s rights are being violated by conditions of
confinement, and whether those conditions produce unwanted outcomes, are distinct
though related questions. The first concerns what is happening now, to individuals, and
reports accepted by courts are entirely appropriate for this purpose. The second question
concerns how classes of persons will fare over time, and requires methods that
standardize observations and compare factors across individuals. The relationship
between the questions is twofold. That someone is suffering now provides reason to fear
about his future welfare and behavior, though it doesn’t settle the matter. More
important, that a practice causes suffering now—assuming it can be justified at all—
provides reason to require that it lead to outcomes we value, such as safer prisons or
communities, and that it be limited to prisoners for whom it is clearly necessary.
It is therefore remarkable how little systematic research has been conducted on who gets
assigned to supermax, how it affects them while they’re there, whether it has any bearing
on their later behavior, and whether such facilities actually reduce violence within prison
systems (Kurki & Morris, 2001; two exceptions are Lovell, Cloyes, Allen & Rhodes,
2000, and Briggs & Sundt, submitted). Nor has any count of supermax inmates
nationwide been conducted since King (1999), whose estimate of 20,000 inmates in 36
states as of 1996 must seriously understate the current scope of the practice. In this
report, we begin to address one research gap by comparing recidivism in the community

Lovell & Johnson

Recidivism of Supermax Inmates

by released offenders who were and who were not subjected to substantial periods of
supermax confinement while in prison. Specifically, we ask:
1. Whether supermax assignment is associated with the probability, seriousness, or
timing of new offenses.
2. Whether the probability, timing, and seriousness of new offenses is associated with
(a) the amount of time offenders spend in supermax environments or (b) the length of
the interval between transfer out of supermax and release to the community.

METHODS
Data Sources
The subject pool comprised all offenders released from DOC facilities during the index
years: calendar years 1997 and 1998. The Office of Planning and Research, Department
of Corrections (DOC), provided three data files:
1. Principal offender file: identifiers, age, ethnicity, sex, current offense, dates of
incarceration and release, mental health data including diagnosis and status as
seriously mentally ill (where available), level of care codes, and days of residence in
various prison residential mental health treatment units. This file contained one
record per incarceration per offender, i.e., offenders released several times during the
index years had multiple records.
2. Infractions file: offender ID, dates, infractions types, and sanctions, one record per
event.
3. Movement file: offender ID, dates, location codes, movement type codes, and
movement reason codes. There was one record per event, with sending and receiving
treated as separate events; furthermore, whether a field corresponded to the sending
or receiving location depended on the movement type codes. These records covered
the entire history of the offender, including terms that ended before or began after the
index years.
The Washington State Institute for Public Policy (WSIPP) with the permission of the
Office of the Administrator for the Courts, provided Washington criminal history data on
all offenders in the subject pool: offense types, dates, jurisdiction (juvenile, district or
superior courts), and disposition.
Data Compilation
Organization of DOC records into files capable of analysis proved the most arduous task
of the study. Omitting details (Appendix A), we used offender movement data to
complete or correct admission and release dates in the principal offender file. From this
file we also retrieved the total amount of time spent in intensive management units

2

Lovell & Johnson

Recidivism of Supermax Inmates

(IMU)—Washington’s name for its supermax facilities—during the offender’s entire
DOC history and during the index incarceration, the number of IMU stays, the length of
each stay, and the date of the last release from IMU before prison release.
To compile criminal history and recidivism data, we classified events as offenses only if
they resulted in a conviction, counting one offense (the most serious) per offense date.
We retrieved numbers of misdemeanors and felonies, dates of the first historical offense,
dates and types of the first post-release offense, the most serious post-release offense, and
numbers of felonies according to four principal types: drug felonies, property felonies
(burglary and theft), sex offenses, and violent offenses (robbery, assault, homicide).
Recidivism included all supervision violations, misdemeanors, and felonies occurring
three years or less after the index release.
Identification of Subjects
From the 10,520 offenders released during the index years, we removed those who had
not served continuous prison terms of six months or more. Since the intensive
management program officially applies only to male inmates, we also eliminated all
female inmates, leaving a potential subject pool of 8,307 offenders.
Not every prisoner who spends time in IMU qualifies as an IMU subject. Depending on
the prison and the circumstances, these facilities may be used to store an offender safely
who presents an acute problem, to hold offenders serving short terms of disciplinary
segregation (as opposed to long-term preventive detention, the stated purpose of intensive
management), or to isolate an offender while an incident or report is being investigated
(which can last up to 12 weeks before he must be returned to general population or
assigned intensive management status). Once assigned intensive management status,
offenders normally stay 180 days in IMU before being considered for release to general
population, but are occasionally released earlier. To maximize statistical power, we
classified as IMU subjects all offenders who had spent at least one continuous period of
more than 12 weeks in IMU or who had shorter stays that added up to 40% of more of
their prison term. This procedure yielded 242 IMU subjects among men released from
prison.
Control Procedures
Post-Hoc Controls. Our previous study (Lovell et al., 2000) showed that offenders
assigned to IMU differ from other offenders in several respects that also predict higher
recidivism, such as young age and extensive criminal history. To limit the confounding
influence of these variables, we matched IMU subjects with non-IMU controls on the
major variables that predicted recidivism. Multivariate regression techniques are also
available that facilitate statistical controlling for possible confounding influences on
recidivism. In a pool of 8,307 cases, however, we found that IMU group membership,
applying to only 242 subjects, made no contribution to a logistic regression recidivism

3

Lovell & Johnson

Recidivism of Supermax Inmates

equation, despite the fact IMU group members had substantially higher rates of
recidivism than others in the pool. We therefore decided to stay with the matching
strategy for the first stage of analysis. Potential controls were male offenders who had no
more than 30 days of IMU time during their index incarceration (30 days is the limit for
disciplinary segregation).
We recognize, of course, that unmeasured factors may predispose some prisoners both to
IMU assignment and to later recidivism; for this reason, significant results using a posthoc control group do not necessarily warrant any particular causal interpretation. Given
the dearth of knowledge about supermax, however, we believed it important to determine
whether IMU subjects had higher rates of recidivism than one would expect, based on
standard age and criminal history predictors.
From other studies (Barnoski & Aos, 1999; Beck, 1997; Gendreau, Little & Goggin,
1996) and our past research on mentally ill prisoners, we identified a set of eight
variables that were significantly correlated with felony recidivism in our pool of 8,307
released male offenders, and which made independent contributions to logistic regression
felony recidivism prediction equations (Table 1). Continuing analysis of predictor sets
from our previous studies of recidivism among mentally ill offenders (Lovell, Gagliardi
& Peterson, 2002) showed that very little predictive power is lost when continuous
variables are recoded into ordinal or categorical variables, for both mathematical and
theoretical reasons: equation coefficients are better behaved if variable ranges are roughly
equivalent; and the difference in proclivity for crime between a man of 35 and a man of
25 is far more significant than the difference between 35 and 50 or even 60. Limiting the
range of potential values for the variables allows matching on a greater number of
variables. So we recoded the continuous variables as ordinal variables with 2-4 values,
selecting cut points to provide clear differences in average rates for each ordinal value
and significant numbers of IMU offenders in each category. Breakdowns are presented
in Appendix A.
Using these strategies, we were able to achieve a 1:1 match for almost every combination
of eight predictor values; as explained below, a separate matching process on 5 variables
was conducted for IMU subjects with mental illness. In 5 cases for non-mentally ill
subjects, and 6 cases for mentally ill subjects, there were no exact matches and controls
from the next closest combination of predictor scores were selected. Where multiple
matching controls per subject were available, controls were selected at random. Table 1
presents average scores (for continuously distributed variables) and rates (for categorical
variables) of IMU subjects, controls, and the entire released male offender pool. There
was wide variability in most of the continuous variables, and a small number of offenders
with values at the extreme high end tended to raise average values above the median and
cause high standard deviations. In some cases, therefore, average values differ between
IMU subjects and controls despite equal numbers of cases in each ordinal category of the
variable.

4

Lovell & Johnson

Recidivism of Supermax Inmates

Table 1
Comparisons of Eight Felony Recidivism Predictors Among IMU Subjects (N=242),
Controls (N=242), and All Male Offenders Released in 1997 and 1998 (N=8,307)*
IMU

Controls

All
Males

p IMU
vs. Cntrl

+ Past felonies (avg)

4.69

4.53

3.90

.607

+ Past misdemeanors (avg)

4.26

4.30

4.45

.945

+ Felony Versatility (avg, ordinal 1-4)

1.79

1.82

1.68

.732a

Variable

–

Age of First Offense (avg)

19.6

20.3

23.58

.308b

–

First-Time Sex Offender (pct)

10%

9%

9%

.876

+ Past Drug Felonies (avg)

.55

.60

0.98

.593

+ Past Violent Felonies (avg)

1.15

1.12

0.66

.753

+ African American or Other Minority (pct)c

34%

31%

29%

.438

*Variables are listed in the order of their univariate correlations with felony recidivism. The plus or minus
signs indicate whether higher values (or positive values, in the case of yes/no variables) were associated
with increased or decreased rates of recidivism.
a. p value based on chi-square comparison of distributions across 4 ordinal categories (1-4), reflecting the
number of different major felony types (drug, property, sex, violent) in the offender’s history.
b. p value for comparison of distributions across 4 ordinal categories (<19 yrs old, 19 thru 25, 36 thru 35,
and >35) = .943.
c. Ethnicity has a relatively low correlation with felony recidivism (.114) but was included as a control
variable because of the salience of racial identification in prisons and particularly in the issues (e.g.,
gang membership) that affect assignment to IMU.

Controlling for Mental Health Status. Mental health status is not integral to
correctional operations, and has been incorporated into DOC’s electronic classification
system only since 1997. Federal court precedents requiring medically necessary
treatment for vulnerable inmates have led corrections systems to define a class of inmates
labeled seriously mentally ill (SMI), i.e., those who are so functionally impaired by a
recognized mental disorder that the constitutional requirement applies. Because no
completely reliable method is available for inspecting electronic databases and
identifying offenders with mental illness, we used a conservative combination of
indicators to identify inmates in our sample pool with probable serious mental illness:
1. recorded status as SMI; or
2. two of the following:
a. qualifying diagnosis;
b. 30 days residential mental health unit residency;
c. level of care codes indicating need for regular psychotropic medication

5

Lovell & Johnson

Recidivism of Supermax Inmates

Improvements in documentation since 1997 mean that identification of offenders with
mental illness in our release cohort is more likely for men who have returned to prison.
This may explain why, in contrast to our previous study which used case-by-case
archived chart reviews to identify SMI subjects (Lovell et al., 2002), we found that in the
overall sample pool, offenders with mental illness were much more likely than others to
have committed new felonies (48% vs. 37%). Regardless of the general relationship
between mental illness and recidivism, mental health status is a major issue in the use of
supermax assignment. Table 2 shows that rates of probable mental illness are
substantially higher for IMU subjects than for others in our sample pool (21% vs. 5%),
and that mentally ill prisoners were over 5 times as likely as other prisoners to have been
IMU subjects.
Table 2
Mental Illness and IMU Status Among Men Released from Washington Prisons in
1997 and 1998 (N=8,307)*
Probable Mental Illness

IMU Status

Yes

No

Total

Yes

52

190

242

No

390

7675

8065

Total

442

7865

8307

*χ2=129.3, df=1, p=.000, odds ratio=5.39
Knowing that IMU offenders had a disproportionate rate of serious mental illness,
reviewers of our proposal warned us that mental illness might interact with IMU
assignment and affect our results. Given the relatively low prevalence of inmates with
mental illness in our pool, we judged that adding mental illness to other variables in a
single matching process would unduly restrict our ability to match on other variables.
We decided instead to conduct a separate matching procedure for offenders with mental
illness, using only the first 5 variables in Table 1 to accommodate the lower number of
available matches. As our recidivism results will show, controlling for mental illness
proved a wise precaution.
Dependent and Independent Variables
Our principal outcome measure was commission of a new felony within three years after
release from prison, as determined by a disposition of guilty. For this phase of the study,
only Washington state data were available to us. Our previous study of mentally ill
offenders (Lovell et al., 2002) indicated that consideration of out-of-state data would
increase the felony rate by approximately 2%; whether IMU subjects and controls would
differ in rates of out-of-state offense is unknown. We used chi-square techniques to test

6

Lovell & Johnson

Recidivism of Supermax Inmates

whether IMU group membership was associated with higher felony recidivism. We also
wanted to test whether IMU group membership was associated with more serious new
crimes and shorter survival times in the community before committing new offenses.
In addition to the fundamental question about differences in recidivism rates according to
group membership, we were concerned with two other possible independent variables:
the amount of time the offender had spent inside IMU, and the length of the interval
between release from IMU and release from prison. In particular, observers concerned
about supermax have wondered whether inmates released directly from such facilities to
the community would be more likely to commit new offenses. Whether total IMU time
and time between IMU and prison release were associated with recidivism outcomes
(yes/no, seriousness, and timing) were tested within the 242-member IMU group.

RESULTS
Profile of Subjects, Controls, and All Men Released from Prison
Table 3
Index Offenses of IMU Subjects (N=242), Controls (N=242), and
All Men Released from Washington Prisons in 1997 and 1998 (N=8,307)
Index Felony

IMU Subjects

Controls

All Males

Unclassified

16

7%

20

8%

516

6%

Drug

39

16%

44

18%

2871

35%

Property

51

21%

54

22%

1912

23%

Robbery & Assault

85

35%

85

35%

1846

22%

Sex Offenses

36

15%

29

12%

911

11%

Homicide

15

6%

10

4%

219

3%

Table 3 shows that IMU subjects and controls resembled each other in index offenses, but
that both groups had substantially lower rates of index drug offenses and higher rates of
index violent offenses—especially robbery and assault—than the entire male release
cohort. As Table 1 indicated, study subjects were younger and had more extensive
criminal histories than other male prisoners in the release cohort. IMU subjects and
controls were similar to each other and to the overall release cohort in ethnic breakdown:
68% white, 24% African-American, and 9% Native American or Pacific Islander. There
were 66 subjects (13%) with Hispanic origin.
In a previous study of recidivism among mentally ill offenders (Lovell et al., 2002), we
found that new violent felonies were extremely rare (10%); among subjects in this study,
they were more common (100 cases, 21%), but still too few to provide sufficient power
for equations predicting violence. As in the previous study, therefore, we added
7

Lovell & Johnson

Recidivism of Supermax Inmates

misdemeanor assaults to commonly recognized felony offenses against persons (1st
degree arson, 1st degree burglary, robbery, assault, kidnapping, sex offenses, and
homicide) to create the dependent variable, new offense against persons (person offense).
Principal recidivism outcomes are displayed in Table 4.
Table 4
Rates of Felony and Person Offense Recidivism for IMU Subjects (N=242), Controls
(N=242), and All Men Released from Washington Prisons in 1997 and 1998 (N=8,307)
Outcome

IMU

Controls

All Men

New Felony

47%

40%

38%

New Person Offense**

36%

24%

21%

Note: Felony recidivism, IMU vs. controls: χ2=1.89, df=1, p=.1, 1-tailed
**
New person offense, IMU vs. controls, χ2=7.7, df=1, p=.004, 1-tailed
IMU Group Membership
Rates of Felony Recidivism. Although IMU Group members had a higher rate of felony
recidivism than controls, this difference fell short of statistical significance (Table 4).
While exploring our data, however, we noticed an unexpected interaction between
serious mental illness, IMU group membership, and felony recidivism. These
relationships are displayed in Table 5, which partitions the cross-tabulation of felony
recidivism by IMU group into SMI and non-SMI subjects.
There is a significant association between IMU status and recidivism for 380 non-SMI
subjects, and no relationship for SMI subjects. While the association between mental
health status and felony recidivism was almost significant for IMU subjects, it was highly
insignificant for controls. In short, there is a suggested association between SMI status
and recidivism, but only for non-IMU offenders; and our hypothesis that IMU status is
associated with recidivism is confirmed, but only for non-SMI offenders.

8

Lovell & Johnson

Recidivism of Supermax Inmates

Table 5
Relationships Among SMI Status, IMU Status, and Felony Recidivism (N=242)*
IMU Status
Yes

SMI Status

24
46%

26
50%

50
48%

No

28
54%

26
50%

54
52%

52

52

104

Yes

89
47%

72
38%

161
42%

No

101
53%

118
62%

219
58%

Totals

190

190

380

Totals
New Felony
No

Totals

Yes
New Felony
Yes

No

*Percentages are rates of new felonies (yes or no) for each IMU and SMI status.
IMU status and recidivism: for probable SMI χ2=.154, df=1, p=.422 1-tailed
for non-SMI, χ2=3.12, df=1, p=.048 1-tailed1
SMI status and recidivism: for IMU group, χ2=.008, df=1, p=.528 1-tailed
for controls, χ2=2.45, df=1, p=.079 1-tailed
Seriousness of New Crimes. As demonstrated in Table 4, IMU subjects were far more
likely than controls to commit new person offenses. Looking only at felonies, the
distribution of the most serious new felonies committed by recidivists in the IMU subject
and control groups is displayed in Table 6. We combined the few “other” felonies with
property felonies, and 3 new sex offenses and 1 new homicide (committed by an IMU
group member) were added to the far more frequent robberies and assaults in the violent
felony category. Clearly, IMU recidivists committed more serious—i.e., violent—new
felonies than control recidivists did.

1

A similar pattern applies to the other outcome of interest, new person offenses (distributions not shown):
for SMI offenders, new person offenses by IMU status, χ2=.370, df=1, p=.343 1-tailed; for non-SMI, new
person offenses by IMU status, χ2=8.2, p=.003 1-tailed.

9

Lovell & Johnson

Recidivism of Supermax Inmates

Table 6
Types of Most Serious New Crimes Committed by Felony Recidivists
in the IMU Group (N=113) and Control Group (N=98)*
IMU Status

New Felony
Type

Yes

No

Total

Drug

17
15%

24
24%

41
19%

Property

34
30%

36
37%

70
33%

Violent

62
54%

38
39%

100
47%

Total

113

98

211

* Percentages represent the share of each crime type among the most serious new felonies
committed by group members; χ2=5.98 df=2, p=.05, 1-tailed
Confounding Variables. Although IMU subjects and controls, by design, resembled
each other closely with respect to eight felony recidivism predictors, IMU subjects
differed from controls and from the entire male release cohort with respect to several
variables that predict new person offenses. Also, despite roughly similar breakdowns of
index crime types, the particular offenses of IMU subjects must have been more serious.
They entered prison on the index conviction at an earlier age than controls, served longer
terms, and were released at approximately the same age. (In the entire release cohort, age
of admission and age of release were highly correlated at .975, but the latter worked
better in logistic recidivism prediction equations.) Table 7 presents means of several
uncontrolled variables of interest for both IMU subjects and controls, along with
univariate correlations of these variables with felony and person offense recidivism in the
entire male release cohort.
The three major statistical differences between IMU members and controls are index
violent offense (which governs the length of the index prison term), annual infraction
rate, and age of admission.
•

Having a violent index crime is negatively correlated with felony recidivism in all
groups (release cohort, SMI, not SMI, IMU, not IMU), and not correlated with
future violence.

•

Annual infraction rates ranged from 0 to 73 with a mean of 1.5, a median of .64,
and a standard deviation of 2.99. Given this skewed distribution and wide
variation, outliers may drive up the average rate for IMU subjects.

10

Lovell & Johnson
•

Recidivism of Supermax Inmates

Looking at age of admission, 60% of IMU subjects were 25 or younger when
admitted to prison on the index term, vs. 46% of controls and 36% of all men in
the release cohort. In this case, a few older inmates in the IMU group may have
dampened the differences in group means.

Table 7
Mean Values of Uncontrolled Variables for IMU Subjects (N=242), Controls
(N=242), and All Men Released from Washington Prisons in 1997 and 1998
(N=8,307), and Correlations with Felony and Person Offense Recidivism

Mean Values

Recidivism Correlations

Variable

IMU

Controls

All Men

Felony

Violence

Index Violent Offense (pct)

59%

53%

38%

-.168

.008

Current Prison Term (mos.)**

56

29

25

-.103

-.072

Annual Infraction Rate**

8.2

1.6

1.3

.119

.154

Age of Release

30

30

32

-.112

-.149

Age of Admission**

26

28

30

-.093

-.138

Timing of New Offenses. Among recidivists, the average time to the first new offense
of any type was 8 months for subjects, 7 months for controls (p=.23); for both groups, the
average time to the first new felony was 9 months, and to the first new person offense, 10
months. In short, group membership played no role in timing of new offenses among
recidivists. Cox regression techniques reflect the hazard of re-offending over the entire
post-release period and take account of subjects who survived without re-offending.
Using this more sophisticated test, group membership had no significant association with
the hazard of committing any new offense or of committing a new felony, but was
significantly associated with the hazard of new person offenses (χ2=7.72, df=1, p=.005).
Statistical controls for the influence of other variables, which in the case of matched
samples require special techniques beyond standard logistic regression methods (e.g.,
Kleinbaum, 1994), were not applied at this stage of inquiry.
The felony survival curve (Figure 1) displayed a familiar logarithmic pattern, with most
failures occurring within 9 months and the failure curve beginning to level off around two
years after release.
Amount of Time in IMU
Table 8 displays the proportions of IMU subjects with varying lengths of time in IMU.
Felony recidivists and non-recidivists were similar: about 1 year in IMU during the index
incarceration. This variable also had no bearing on the seriousness or timing of new
11

Lovell & Johnson

Recidivism of Supermax Inmates

offenses. Neither taking previous incarcerations into account, nor partitioning IMU time
into ordinal categories, made any difference to the finding that for the class of IMU
regulars, almost all of whom had spent long periods of time in this environment, the
amount of IMU time bore no relationship to community outcomes.
Figure 1. Community Survival Until First Felony:
All Subjects (N=484)
100

90

Pct of Subjects w/o Felony

80

70

60

50

40
0

6

12

18

24

30

Months to New Felony

Table 8. Amounts of IMU Time for IMU Subjects (N=242)
IMU Time

Number

Percent

6 months or less

65

27%

6 months to 1 year

70

29%

1 to 3 years

92

38%

Over 3 years

15

6%

12

36

Lovell & Johnson

Recidivism of Supermax Inmates

Timing of Release from IMU
A major concern raised by IMU practices, and one motivation for this study, was that
offenders released directly from IMU into the community would be too disoriented,
jumpy, or hostile to cope with the challenges of society. Assuming some deterioration of
social functioning, there is reason to believe that many coping skills can return fairly
quickly after emerging from isolation (Grassian and Friedman, 1986). We found that the
time between subjects’ release from IMU and their release into the community (Time to
Release) was correlated with felony recidivism, new person offenses, and length of
survival in the community before committing new offenses. But if the real issue is
immediate release to the community, using the continuous Time to Release variable may
ignore that marker, while giving unwarranted weight to the 50 long-term prison inmates
with more than three years (ranging up to twelve years) between IMU and prison release.
We therefore defined an Immediate Prison Release variable, classifying as positive the 59
subjects released directly from IMU into the community. When entered into logistic
regression equations, Immediate Prison Release showed more robust associations with
outcomes than Time to Release did. Results are displayed in Tables 9 and 10.
Table 9
Immediate Prison Release and Felony Recidivism Among IMU Subjects (N=242)*
Immediate Prison Release

New Felony

Yes

No

Total

Yes

38
64%

75
41%

113
47%

No

21
36%

108
58%

129
53%

Total

59

183

242

* Percentages are rates of new felonies (yes or no)
χ2=9.8, df=1, p=.001 1-tailed, new felony odds ratio for immediate yes vs. no =2.6

13

Lovell & Johnson

Recidivism of Supermax Inmates

Table 10
Immediate Prison Release and Person Offense Recidivism
Among IMU Subjects (N=242)*
Immediate Prison Release

New Person
Offense

•

Yes

No

Total

Yes

30
51%

57
31%

87
36%

No

29
49%

126
69%

155
64%

Total

59

183

242

Percentages are rates of new person offense (yes or no)
χ2=7.5, df=1, p=.005 1-tailed, person offense odds ratio=2.3

On average, members of the immediate prison release group were younger than the others
at prison admission and release, committed their first offense at a younger age, and had
more previous misdemeanors. Stepwise logistic regression selected the last two for a
significant prediction model. We then controlled for age of first offense and number of
previous misdemeanors, and found that immediate prison release still made a significant
independent contribution to prediction of felony recidivism (χ2=4.382, df=1, p=.036), but
not to new crimes against persons (χ2=3.086, df=1, p=.079).
Membership in the immediate prison release group was also associated with shorter time
to new offenses. But age at first offense, previous misdemeanors, and (for new felonies)
status as a first-time sex offender were also significantly associated with the timing of
new offenses. As Table 11 indicates, controlling for these variables in Cox regression
equations suppressed the effect of immediate prison release on timing of new felonies
and new crimes against persons, i.e., the addition of immediate prison release to the
equation did not significantly increase the likelihood of correctly predicting hazard; but a
significant relationship persisted between immediate prison release and the timing of any
new offense (i.e., any kind of new felony or misdemeanor).

14

Lovell & Johnson

Recidivism of Supermax Inmates

Table 11
Cox Regression Results on Timing of New Offenses with Addition of Immediate Prison
Release to Control Variables*
Results w Control Vbls

Change w Prison Release Vbl

χ

df

p

χ2

df

p

Any New Offense

107.9

2

.000

4.805

1

.028

New Felony

70.9

3

.000

3.262

1

.071

New Person Offense

50.6

2

.000

2.684

1

.101

Outcome

2

* Age at first offense: higher age, decreased hazard;
Previous misdemeanors: higher number, increased hazard;
For new felonies only, status as a first-time sex offender: decreased hazard

DISCUSSION
Summary of Findings
We recap here the research questions stated in the Introduction:
1. Whether supermax assignment is associated with the probability, seriousness, or
timing of new offenses.
2. Whether the probability, timing, and seriousness of new offenses is associated with
(a) the amount of time offenders spend in supermax environments or (b) the length of
the interval between transfer out of supermax and release to the community.
We thus have three independent variables: group membership, time in IMU, and interval
between transfer out of IMU and release to the community; and three dependent
variables: probability, seriousness, and timing of new offenses. Table 12 displays a
matrix of results, with a + or an X indicating significant associations, and 0 no
association. Associations are assigned an X rather than a + if they apply to some groups
and not others (e.g., non-SMI vs. SMI) or if differences between groups pose difficulties
of interpretation.

15

Lovell & Johnson

Recidivism of Supermax Inmates
Table 12
Matrix of Findings
Dependent Recidivism Variable

Independent
IMU
Variable

a.
b.
c.
d.

New Felony

Seriousness

Timing

IMU Group
Y/N

Xa

Xb

Xc

Time in
IMU

O

O

O

Timing of IMU
Release

+

Xb

Xd

Applies only to Non-SMI offenders.
Differences between groups pose difficulties of interpretation.
Applies to new crimes against persons but not to other recidivism measures.
With statistical controls in place, applies to any new offense regardless of type, but
not to new felonies or new crimes against persons.

Two Problems of Interpretation. The findings reported above pose two related but
distinct problems of interpretation:
•

The problem of statistical prediction. Where we found that IMU assignment or
timing of release predicted recidivism, are these robust findings? Would they
survive if different methods were used, or further statistical controls were
introduced?

•

The problem of causal interpretation. If there are robust statistical relationships
between recidivism and IMU assignment or timing of IMU release, to what extent
can these relationships be attributed to the effect of the IMU experience?

These questions define the terrain for future research, involving both further analysis of
the data collected for this pilot study and collection of additional information.
Methods of Further Analysis
Perhaps the first step in further analysis is to subject the data to systematic diagnostic
tests to ensure that the methods used so far are appropriate. Selection of variables for
multivariate regression equations in this pilot study has been guided by tables of
univariate correlations among the independent and dependent variables of interest.
Further diagnostics should be applied, however, to measure collinearity among groups of
variables, assess interaction effects, and identify outliers.
Using logistic regression on the entire pool of 8,307 men released from prison, we found
that IMU group membership, applying to only 242 subjects, made no contribution to a
prediction equation, despite the fact that IMU group members had substantially higher
16

Lovell & Johnson

Recidivism of Supermax Inmates

rates of recidivism than others in the pool. To prevent the large size of the pool from
swamping the predictive effect of IMU assignment, we could take repeated random
samples of 250 offenders from the non-IMU pool and apply logistic regression to the
combined sample. Assuming that IMU assignment often predicted recidivism in these
mini-studies, meta-analysis could be applied to the results to determine how robust this
finding is. Separate runs could be conducted with SMI and non-SMI offenders to assess
whether the predictive power of IMU assignment depends on mental health status. If
these procedures replicate the significant predictive effects found with the matching
strategy, our confidence would be strengthened in the conclusion that IMU assignment is
independently associated with a higher likelihood of recidivism.
Additional methods would be required to assess the role of measured but uncontrolled
differences between the IMU and non-IMU groups, such as index violent offense, age at
release, length of prison term, age at admission, and infraction rates. Using standard
logistic regression to control for these variables is not appropriate with matched samples,
but logistic methods especially designed for matched samples (Kleinbaum, 1994,
Chapter 8) may be used to test whether the relationship persists if we enter additional
uncontrolled differences between groups into the equations. Whether and how to control
for infraction rates by this technique, however, raises the issue of causal interpretation.
Infractions are potent measures of inmate behavior and of how staff view the inmate, and
consequently are highly associated with IMU assignment. Controlling for infraction rate,
therefore, may render insignificant the independent statistical contribution of IMU
assignment to recidivism outcome. Does this mean that infraction rate, rather than IMU
assignment, actually “explains” recidivism outcomes? Why then do infraction rates show
such a weak relationship with recidivism among the population of men released from
prison? These questions require closer examination of the processes at work in prisons, a
task for which further statistical manipulation of existing data may not suffice.
Propensity matching is an alternative strategy, which comes closer to replicating random
assignment than either multivariate regression or matching on outcome predictors
(Barnoski & Aos, 2003). In propensity matching, logistic regression is used to build
equations that predict membership in the “experimental” group, i.e., IMU assignment,
and the control group is selected by matching subjects on the resulting propensity scores.
Propensity matching is worth a trial with the data developed for this study. Assuming
some degree of rationality in the IMU assignment process, however, it is likely that
infraction rate—for reasons mentioned above—will prove to be an indispensable
propensity variable, and that controls with matching scores on the resulting propensity
equations will be too scarce to support the analysis.
Taking the concept of propensity matching a step further, we could attempt to identify
instrumental variables, i.e., variables that predict assignment to IMU but not recidivism
outcomes (Wooldridge, 2003, ch. 15). Controls could be identified by matching scores
on equations that successfully predict IMU assignment using only such instrumental

17

Lovell & Johnson

Recidivism of Supermax Inmates

variables. We don’t consider it likely that such variables will emerge in the aggregate
IMU sample data. Nevertheless, this strategy is worth exploring; particularly because our
principal finding, that SMI status interacts with IMU assignment in predictions of
recidivism, may be interpreted as revealing that SMI offenders have been assigned to
IMU for reasons that have nothing to do with their risk of recidivism.
Factors for Further Investigation
In the preceding discussion, two variables—mental illness and infraction rates—have
loomed large as sources of complexity in interpretation of findings. Our previous
research (Lovell et al., 2000) revealed a diversity of patterns among inmates assigned to
IMU, and it is likely that the mental illness and infraction rate variables themselves
reflect multiple factors that further investigation may bring to light.
We mentioned in the Methods section that ascription of probable mental illness was
based on a conservative application of available OBTS indicators, which were likely to
have preferentially identified recidivists.
These indicators are also clinically
impoverished, providing little information about differences between men who were and
men who were not assigned to IMU. What were the clinical or behavioral factors that
provoked IMU assignment but evidently had little bearing on behavior after release from
prison? Examination of archived medical files, along with case management narratives
available through OBTS, may allow us to distinguish relevant syndromes among subjects
classified as mentally ill and bring us a step closer to describing causes.
Together with living unit assignments (e.g., IMU or segregation vs. general population),
and loss of good time, infractions are reliable indicators of how successfully an inmate is
coping with prison—at least from the staff’s point of view. But when did the infractions
occur and how serious were they? Inmates can continue to accumulate infractions while
in IMU, but the physical setting is designed to foreclose the possibility of serious assault.
Our data permit a sequential ordering of infractions and IMU residency, and together
with narrative case management records may provide more specific indications of
reasons for placing or retaining subjects in IMU. This information, in turn, may shed
further light on factors underlying the various recidivism patterns of IMU subjects and
help us understand why variables such as age at first offense and number of previous
misdemeanors would be associated with release directly from IMU into the community.
In the absence of random assignment, which for obvious reasons must be ruled out in
IMU studies, threats to validity will always remain. Attempts to get at causal explanation
through collection of further data and longitudinal analysis of cases will enhance
interpretability in two ways. First, the application of further methods to a richer data set
may result in additional findings. Second, additional qualitative detail about longitudinal
patterns among IMU and non-IMU offenders will indicate which explanatory hypotheses
are true to life. Accumulation of statistical findings and qualitative evidence may narrow

18

Lovell & Johnson

Recidivism of Supermax Inmates

the scope for interpretation and thereby address threats to validity inherent in nonexperimental designs.
Significance of Preliminary Findings
We have described lines of further inquiry suggested by the results of this pilot study. In
this concluding section, we argue that the preliminary results are robust and important
enough to make such inquiry worthwhile, not only for the investigators but for agencies
to which external funding requests will be directed.
Timing of IMU Release. We found, as suspected, that IMU group members released
directly from IMU into the community had significantly higher rates of felony recidivism
than those who had some prerelease period in social prison settings:
IMU Group Membership. For interpreting the contribution of IMU group membership
to prediction of recidivism, the principal threat to validity lies in the possibility that a
special form of behavior (1) elicits the judgment that an inmate is dangerous while in
custody, provoking IMU assignment; and (2) predicts recidivism. Some students of
criminality would suggest an underlying personality structure, psychopathy, to explain
the relationship. We can, rule out the fantastic hypothesis that IMU-provoking behavior
predicts lower recidivism. Our study proposal laid out three possible findings and their
interpretive implications:
1. A finding that IMU assignment predicted lower recidivism would suggest that IMU
confinement is an effective treatment. Our findings do not support this hypothesis.
2. A finding that IMU assignment does not predict recidivism would suggest either
(1) that IMU-provoking behavior and the IMU experience are both neutral; or (2) that
the behavior predicts recidivism as a main effect, but this effect is neutralized by the
IMU experience. Neither of these hypotheses is supported.
3. We found, with qualifications, that IMU assignment predicts higher recidivism. We
may conclude that IMU confinement does not appear to help control recidivism,
which advances knowledge beyond its present null state. But we do not know
whether the predictive effect is due to the IMU experience or to some psychological
process that leads prison staff to see the offender as threatening and which, after
release, leads to further criminal aggression.
Despite difficulties of interpretation, our findings make an important contribution, indeed
the first contribution of systematic research, to assessing the effects on recidivism of our
nation’s experiment with supermax confinement. Furthermore, our unanticipated finding
of an interaction between mental illness and the predictive effect of IMU assignment,
suggests that assignment to IMU does not uniformly respond to a character trait that
predisposes prisoners to future crime. Perhaps, in the case of prisoners with mental
illness, it simply indicates that they are unfit for the rigors of prison existence. This
19

Lovell & Johnson

Recidivism of Supermax Inmates

possibility leads to the question: are prison staff mistaken about the dangerousness of the
mentally ill people they assign to IMU, or is their dangerousness a function of being in
prison? Either way, we may call into question the legitimacy of current procedures, and
recognize the need for further study as more than academic.

REFERENCES
Barnoski, R. & Aos, S. (1999). Presentation to the Standards and Ranges Committee,
Sentencing Guidelines Commission. Olympia, WA: Washington State Institute for
Public Policy.
Barnoski, R. & Aos, S. (2003). Washington State’s Drug Courts for Adult Defendants:
Outcome Evaluation and Cost-Benefit Analysis. Olympia, WA: Washington State
Institute for Public Policy.
Beck, A. (1997). Recidivism of Prisoners Released in 1983. Washington, DC: U.S.
Department of Justice, Bureau of Justice Statistics.
Briggs, C., Sundt, J., Castellano, T. & Vaux, A. Supermaximum security prisons and
institutional violence: an impact assessment. Submitted.
Gendreau, P., Little, T. & Goggin, C. (1996). A meta-analysis of the predictors of adult
offender recidivism: what works! Criminology 34: 575-607.
Grassian, S. & Friedman, N. (1986). Effects of sensory deprivation in psychiatric
seclusion and solitary confinement. International Journal of Law and Psychiatry 8:
49-65.
Haney, C. (1993). 'Infamous punishment’: the psychological consequences of isolation.
National Prison Project Journal Spring 1993, 3-7, 21.
Haney, C. (2003). Mental health issues in long-term solitary and “Supermax”
confinement. Crime and Delinquency 49(1): 124-156.
Jones'El v. Berge, 164 F.Supp2d 1096 (W.D. Wis 2001).
King, R. (1999). The rise and rise of supermax: an American solution in search of a
problem? Punishment and Society 1: 163-186.
Kleinbaum, D. (1994). Logistic Regression: A Self-Learning Text. New York: Springer
Verlag.

20

Lovell & Johnson

Recidivism of Supermax Inmates

Kurki, L. & Morris, N. (2001). The purposes, practices, and problems of supermax
prisons. Crime and Justice 28: 385-424.
Lovell, D., Cloyes, K., Allen, D. & Rhodes, R. (2000). Who lives in super-maximum
custody? A Washington State study. Federal Probation 64(2): 33-38.
Lovell, D., Gagliardi, G. & Peterson, P. (2002). What happens to mentally ill offenders
released from prison? Findings from Washington's Community Transitions Study.
Psychiatric Services 53(10): 1290-1296.
Madrid v. Gomez, 889 F. Supp. 1146 (N.D. Cal 1995).
Wooldridge, J.M. (2003). Econometrics: A Modern Approach, 2nd Ed. Cincinnati:
South-Western College Publishing.

21