Predicting High-Risk Behaviors in Veterans With Posttraumatic Stress Disorder

Article (PDF Available)inJournal of Nervous & Mental Disease 193(7):464-72 · August 2005with85 Reads
DOI: 10.1097/01.nmd.0000168238.13252.b3 · Source: PubMed
Abstract
The present study sought to identify posttraumatic stress disorder (PTSD) patients at high risk for negative behavioral outcomes (violence, suicide attempts, and substance use). The Mississippi Scale for Combat-Related PTSD, the Beck Depression Inventory, and demographic and behavioral data from 409 male combat veterans who completed a VA residential rehabilitation program for PTSD were analyzed using signal detection methods (receiver operating characteristics). A validation sample (N = 221) was then used to test interactions identified in the signal detection analyses. The best predictors of behaviors at follow-up were those same behaviors shortly before intake, followed by depressive and PTSD symptoms. However, for each of the models other than that for hard drug use, cutoffs determined at the symptom level did not lend themselves to replication. Recent high-risk behaviors, rather than patients' history, appear to be more predictive of high-risk behaviors postdischarge.

Figures

Full-text (PDF)

Available from: Kent D Drescher, Jun 25, 2014
ORIGINAL ARTICLES
Predicting High-Risk Behaviors in Veterans
With Posttraumatic Stress Disorder
Tamara L. Hartl, PhD,* Craig Rosen, PhD,†‡ Kent Drescher, PhD,† Tina T. Lee, MD,†
and Fred Gusman, MSW†
Abstract: The present study sought to identify posttraumatic stress
disorder (PTSD) patients at high risk for negative behavioral out-
comes (violence, suicide attempts, and substance use). The Missis-
sippi Scale for Combat-Related PTSD, the Beck Depression Inven-
tory, and demographic and behavioral data from 409 male combat
veterans who completed a VA residential rehabilitation program for
PTSD were analyzed using signal detection methods (receiver op-
erating characteristics). A validation sample (N 221) was then
used to test interactions identified in the signal detection analyses.
The best predictors of behaviors at follow-up were those same
behaviors shortly before intake, followed by depressive and PTSD
symptoms. However, for each of the models other than that for hard
drug use, cutoffs determined at the symptom level did not lend
themselves to replication. Recent high-risk behaviors, rather than
patients’ history, appear to be more predictive of high-risk behaviors
postdischarge.
Key Words: PTSD, veterans, violence, suicide, substance use.
(J Nerv Ment Dis 2005;193: 464 472)
P
osttraumatic stress disorder (PTSD) is a life-altering and
complex psychological condition estimated to affect 30%
of Vietnam theater veterans over their lifetime (Kulka et al.,
1990). Despite recent increases in the number of treatment
programs aimed at addressing the myriad of problems asso-
ciated with the disorder (Chief Medical Director’s Special
Committee on PTSD, 1991), high-risk behaviors such as
suicide attempts, violence, and substance use persist among
PTSD veterans even after treatment (Rosenheck and Fontana,
2001).
For example, several studies indicate that veterans with
PTSD endorse greater rates of interpersonal violence com-
pared with veterans without PTSD (Beckham et al., 1997;
Kulka et al., 1990). Greater risk of violence in PTSD veterans
has been shown to be associated with more severe hyper-
arousal symptoms and drinking quantity (Savarese et al.,
2001), lower socioeconomic status and PTSD severity (Beck-
ham et al., 1997), and participation in war zone violence
(Hiley-Young et al., 1995).
In addition to the greater risk of violence associated
with PTSD in veterans, there is substantial evidence that a
diagnosis of PTSD in Vietnam era veterans amplifies the
overall risk of attempting or committing suicide (Bonin et al.,
2000; Fontana and Rosenheck, 1995).
Substance abuse is also common among PTSD veter-
ans, even after completion of PTSD treatment programs.
Perconte and Griger (1991) assessed combat veterans with
PTSD who completed a partial hospitalization program for
PTSD. Compared with patients who relapsed, veterans who
improved in treatment and maintained positive gains at 18-
month follow-up endorsed a lower weekly alcohol intake at
the time of intake and discharge.
On average, PTSD symptoms do not seem to be ame-
liorated by inpatient or outpatient treatment programs (Fon-
tana and Rosenheck, 1996; Hammarbarg and Silver, 1994;
Johnson et al., 1996). Therefore, relapse in this population
might better be defined in terms of high-risk behaviors rather
than symptom reoccurrence or exacerbation. Because of the
chronicity of PTSD and its association with high-risk behav-
iors, there is a great need for predicting which patients are at
greater risk for relapse and serious harm. Better prediction of
outcomes could lead to specialized interventions that result in
a decrease in rehospitalization rates and a substantial en-
hancement to veterans’ quality of life.
The epidemiological and clinical studies discussed here
provide ample evidence that veterans with chronic PTSD can
*Health Services Research and Development, VA Palo Alto Health Care
System, Palo Alto, California; †National Center for PTSD, VA Palo Alto
Health Care System, Palo Alto, California; and ‡Stanford University
School of Medicine, Stanford, California.
Supported in part by the VA National Center for Posttraumatic Stress
Disorder, VA Health Services Research and Development Service, VA
Office of Academic Affiliations, VA Mental Health Strategic Health
Group, and VA Palo Alto Health Care System.
The views expressed in this article are those of the authors and do not
necessarily represent the views of the Department of Veterans Affairs.
Send reprint requests to Tamara L. Hartl, PhD, 116 B, VA Palo Alto Health
Care System, 3801 Miranda Ave., Palo Alto, CA 94304.
Copyright © 2005 by Lippincott Williams & Wilkins
ISSN: 0022-3018/05/19307-0464
DOI: 10.1097/01.nmd.0000168238.13252.b3
The Journal of Nervous and Mental Disease Volume 193, Number 7, July 2005464
be at risk for relapse with respect to a number of high-risk
behaviors, including attempted or completed suicide, vio-
lence, and misuse of alcohol and drugs.
However, existing research provides little guidance for
determining which individual PTSD patients are at greatest
risk. There are at least three reasons for this. First, prediction
of individual behavior is notoriously difficult, particularly
when the behavior has a relatively low base rate (McNeil and
Binder, 1997). For an infrequent behavior such as suicide,
even a very good model will generate a high proportion of
false positives to true positives. Second, patient subpopula-
tions that are at the highest risk may be relatively homoge-
nous in terms of risk factors. For example, it is not difficult to
predict that psychiatric inpatients are at greater risk for
violence than are psychiatric outpatients. However, it is more
difficult to predict which inpatients are at highest risk, be-
cause several risk factors (such as past history of violence,
history of substance misuse, and poor social support) may be
nearly universal in this patient population. The final barrier is
statistical. Conventional correlational analyses and logistic
regression models reported in most research are primarily
designed to identify which variables are associated with risk
for an outcome, rather than which individuals are at greatest
risk (Kiernan et al., 2001). Moreover, such analyses do not
provide clinically useful cut-points or decision rules that can
be used to categorize patients.
CURRENT STUDY
The aim of the present study is to assess patient char-
acteristics that predict relapse to high-risk behaviors follow-
ing discharge from a residential rehabilitation program for
PTSD. To evaluate patient characteristics that predict greater
risk of relapse, we used signal detection methods that are
commonly used in medical decision making to evaluate
performance on diagnostic tests (Kraemer, 1992). Receiver
operating characteristics (ROC) is a type of signal detection
analysis that has specific advantage over logistic regression
for the purposes of identifying which patients are at risk.
Specifically, ROC is a person-centered approach that identi-
fies individuals at high or low risk based on their scores on
predictor variables. In contrast, logistic regression is variable-
centered and identifies variables that are related to outcomes
across individuals. The advantage of using ROC analysis for
this investigation is that it yields an empirically derived tool
for identifying high-risk patients and thus can inform clini-
cians which patients are most in need of follow-up support
services. Other advantages of signal detection methods in-
clude the ability to designate a criterion for favoring detection
of false positives over false negatives, reservation of power
amid multicollinear predictors, and the designation of empir-
ically constructed higher-order interactions (Kiernan et al.,
2001). The primary outcome variables of interest for this
study included whether patients (a) attempt suicide, (b) en-
gage in violent behaviors, (c) misuse alcohol, or (d) misuse
cocaine, amphetamines, or opiates within 4 months of
discharging from the controlled environment of residential
treatment.
METHODS
Sample
The sample consisted of 630 male veterans with a
primary PTSD diagnosis who consecutively entered the res-
idential rehabilitation program for PTSD at the VA Palo Alto
Health Care System between July 1994 and December 2000.
The residential rehabilitation program is a highly structured
treatment in which every patient participates in various skills
classes (e.g., conflict resolution, anger management, effective
communication), medication management, relapse preven-
tion, and recreation. Patients are referred to the program by
medical and mental health staff from health care facilities
largely concentrated in the Pacific Northwest.
For patients with multiple admissions during this time
period (N 85), the earliest admission was retained. Length
of stay for the entire sample ranged from 2 to 286 days, with
an average of 71.6 days (SD 35.29). Although there is a
large disparity in length of stay, the program’s standard
length of treatment was 90 days (before 1996) and then 60
days (1996 and on). The mean age of the sample was 50.56
(SD 4.55; range 25.79 –76.26). Ethnicity was reported as
60% Caucasian, 13% Hispanic/Latin American, 13% African
American, 6% mixed ethnicity, 5% Native American, 1%
Asian/Pacific Islander, and 1% other. One percent of the
responses was missing for ethnicity. Ninety percent of the
men were Vietnam era veterans, 86% of whom had served in
combat.
Measures
The data set was constructed to reflect information that
is typically collected at the outset of most treatment pro-
grams. Thus, any decision-making tools that might result
from this investigation might be testable in, and perhaps
generalizable to, other programs.
Demographic and Military Service Variables
Variables from demographic and background history
questionnaires administered at intake included age, ethnicity,
education, marital status, and history of incarceration. Nearly
72% of the sample had a history of incarceration (Table 1),
which speaks to the severity of impairment in this patient
population. Variables related to patients’ war zone trauma
exposure included whether patients served in a war zone,
witnessed incoming fire, witnessed or participated in atroci-
ties, or had been held a prisoner of war. Treatment-related
variables included number of prior admissions to the same
rehabilitation program and whether patients dropped out of
The Journal of Nervous and Mental Disease Volume 193, Number 7, July 2005 High-Risk Behaviors in PTSD
© 2005 Lippincott Williams & Wilkins 465
the program early, defined as staying fewer than 30 days
when the program’s typical length of stay was 60 days and
staying fewer than 45 days when the program’s average
length of stay was 90 days.
Behaviors
Violent behavior and suicide attempts were assessed
with items from the Northeast Program Evaluation Center
survey used in ongoing program evaluation of VA PTSD
patients (Fontana and Rosenheck, 1997; McFall et al., 1999).
Three violence items inquired whether in the last 4 months
patients had threatened someone with physical violence, had
a physical fight with someone, or threatened someone with a
weapon. Two additional items asked whether patients had
attempted suicide in their lifetimes and/or in the last 4
months. Half of the sample had attempted suicide at some
point (Table 1). Substance use was assessed with several
Northeast Program Evaluation Center items drawn from the
Addiction Severity Index (McLellan et al., 1992). Questions
about drug and alcohol use specified that respondents should
report their use in the 30 days prior to a possible qualifying
period (typically 30 days) of no substance use for the pro-
gram. Questions included the number of days the patient
drank alcohol, drank alcohol to the point of feeling drunk or
intoxicated, and used cannabis, cocaine, amphetamines, or
opiates. The last three items, reflecting the most commonly
used hard drugs in this population, were summed into a single
measure of hard drug use. For example, if a participant used
cocaine on 20 days, amphetamines on 15 days, and opiates on
0 days, his total hard drug use score would equal 35.
Symptoms
Posttraumatic stress disorder symptoms were assessed
with the Mississippi Scale for Combat-Related PTSD (Keane
et al., 1988). The Mississippi Scale is a reliable and well-
validated 35-item measure of PTSD symptoms in combat
veterans (Keane et al., 1987; Kulka et al., 1990). Total scores
from the Beck Depression Inventory-II (BDI-II; Beck et al.,
1996) were used to assess depressive symptoms.
Due to multicollinearity between patient symptom
scores from intake and discharge assessments, we included
only intake scores as predictors of postdischarge outcome.
Intake scores were selected because these data were more
complete, and if predictive, could be most helpful in inform-
ing treatment planning based on patients’ profiles at intake.
Additional post hoc analyses (detailed below) tested whether
using discharge rather than intake symptom scores improved
our prediction of postdischarge outcomes.
Diagnoses
Structured Clinical Interview for DSM-IV Disorders
(First et al., 1995) diagnoses were included for alcohol,
cannabis, and hard drugs. The last was a composite variable
reflecting a diagnosis of cocaine, amphetamine, or opioid
dependence. These three variables were coded in order of
increasing severity from “no diagnosis” to “dependence in
early or partial remission.” Number of hard drug diagnoses
(from 0 –5: cocaine, amphetamines, opioids, sedatives, and
hallucinogens) was included as another potential predictor.
Major depression, bipolar disorder, and anxiety disorder di-
agnoses were not included because preliminary bivariate
TABLE 1. Sample Characteristics at Intake in a Sample of
630 Male Veterans
Variable Mean SD Range
Education (y) 13.30 1.89 6–20
Mississippi Sale scores 138.02 15.74 88–175
BDI scores 32.26 10.56 0–59
Number of admissions 1.17 .52 1–8
Number of types of
violence
1.14 1.09 0–3
Number of hard drug
SCID Dx’s
.65 .93 0–5
Number of days drank
alcohol prior to
intake
3.2 8.08 0–30
Number of days
intoxicated prior to
intake
2.22 7.01 0–30
Number of days used
cannabis prior to
intake
1.67 6.27 0–30
Number of days used
hard drugs prior to
intake (aggregate)
1.68 7.69 0–90
Variable Percentages
Marital status 36.5% married
Served in a war zone 98.4%
Witnessed incoming
fire
97.6%
Prisoner of war 1.6%
Witnessed or
participated in
atrocities
60.8%
Suicide attempt 4 mo
prior to intake
8.1%
Suicide attempt ever 50.2%
Had been incarcerated 71.6%
Left treatment early 10.3%
SCID Dx of alcohol
abuse or dependence
72%
SCID Dx of cannabis
abuse or dependence
27%
SCID Dx of hard drug
abuse or dependence
39%
Hartl et al. The Journal of Nervous and Mental Disease Volume 193, Number 7, July 2005
© 2005 Lippincott Williams & Wilkins466
analyses showed no relationship between these variables and
the outcomes of interest.
Outcomes
Four dependent variables (DVs) were selected to reflect
relapse to high-risk behaviors. All DVs were binary to enable
ROC analyses to be applied to the data set. Violence was
defined as engaging in violent behavior (physical fight,
threatening someone, or threatening with a weapon) in the 4
months after discharge. Suicide was defined as attempting
suicide within the 4 months following discharge. Alcohol
misuse was defined as “drinking to the point where you felt
drunk or intoxicated; that is, had three or more drinks in one
sitting” on 2 or more days in the month prior to follow-up.
Intoxication two or more times in the past 30 days, as
opposed to 1 or more days, was chosen to discriminate from
drinking to intoxication on only 1 day (the most frequently
endorsed response). Hard drug use was defined as any use of
cocaine, amphetamines, or opioids in the month prior to
follow-up. Whereas the questions about violence and suicide
attempts assessed these behaviors in the past 4 months,
questions about substance abuse applied only to the 30 days
prior to completing the follow-up questionnaire.
Data Analytic Plan
Software specifically designed to compute ROCs
(Kraemer, 1992) was used in an effort to identify interactions
between predictors. ROC is a signal detection method that
entails the use of a cutoff point for each predictor to procure
the best differentiation on the dependent variable in question
based on the assigned sensitivity/specificity weight. For this
investigation, the weight was set at .70 to favor sensitivity
and avoid false negatives. In this type of analysis, the sample
is divided into two subsamples at the optimal cutpoint on
each variable that predicts the specified clinical criterion (the
DV). The process is then repeated in an iterative manner
across all remaining predictors in the two subsamples to
detect the next best predictor of the criterion. The process
continues until too few individuals for analysis remain in the
subgroup (e.g., n 10).
Receiver operating characteristics analyses were first
conducted on a random selection of two thirds (409 patients)
of the sample to establish prediction models for each out-
come. The remaining one third of the patients (N 221) were
used as a replication sample to test the fit of the models
developed in the exploratory analyses.
RESULTS
Descriptive Statistics
In the 30 days prior to a prequalifying period of
abstinence, 11.6% of the sample reported having drunk to the
point of intoxication on 2 or more days, and 8.1% had used
drugs. In the 4 months prior to intake, 61.9% of the sample
reported having committed one or more violent acts, and
8.1% reported having attempted suicide. Reports of these
behaviors at the 4-month follow-up assessment indicate a
decrease in both violence (49% of the sample committed one
or more violent acts) and suicide attempts (4.9% of the
sample) over the preceding 4 months. However, in the 30
days preceding follow-up, drinking to intoxication and drug
use both increased slightly relative to preintake reports
(14.3% and 9.5%, respectively). This is likely a reflection of
individuals returning to drinking after a required period of
abstinence for program admission.
Signal Detection: Exploratory Sample
Figures 1 to 4 depict the ROC models for the four
behavioral DVs assessed at follow-up. Each figure depicts a
decision tree type of graph to delineate partitioning off of
subgroups of patients based on cutoff scores for the signifi-
cant predictors in each model. The base rate of the behavior
being predicted in the exploratory samples was used as the
cut-point to discriminate high-risk versus low-risk groups for
each model, with groups scoring at or above the base rate
considered high-risk. Sensitivity indices were calculated by
summing the high-risk patients in each of the subgroups at the
end of a tree branch and then dividing the number of actual
relapses (true positives) by the total number of patients that
relapsed. The specificity index equaled the number of true
negatives divided by the total number of patients who did not
score positively on the outcome variable.
Replication
The predictive validity of the four ROC models was
evaluated by using the cut-point determined in the explor-
FIGURE 1. Prediction model for suicide. Figures are for the
exploratory sample (N 408); figures in parentheses are for
the replication sample (N 221).
The Journal of Nervous and Mental Disease Volume 193, Number 7, July 2005 High-Risk Behaviors in PTSD
© 2005 Lippincott Williams & Wilkins 467
FIGURE 4. Prediction model for hard
drug use in the past month. Figures
are for the exploratory sample (N
408); figures in parentheses are for
the replication sample (N 221).
FIGURE 2. Prediction model for vio-
lence. Figures are for the exploratory
sample (N 408); figures in paren-
theses are for the replication sample
(N 221).
FIGURE 3. Prediction model for al-
cohol intoxication (three or more
times in the past month). Figures are
for the exploratory sample (N
408); figures in parentheses are for
the replication sample (N 221).
Hartl et al. The Journal of Nervous and Mental Disease Volume 193, Number 7, July 2005
© 2005 Lippincott Williams & Wilkins468
atory sample to categorize subjects in the replication sample.
The percentages of positive scores on the DVs within the
replication sample at each branch or cutpoint are shown in
parentheses in Figures 1 to 4. Overall sensitivity and speci-
ficity for each model were also calculated in the replication
sample.
Suicide
Exploratory Sample
Signal detection methods identified predictors that dis-
tinguished patients more and less likely to attempt suicide
after discharge. In Figure 1, two high-risk subgroups were
identified. The single best predictor of a suicide attempt after
discharge was having attempted suicide in the 4 months prior
to intake (
2
1,296兴⫽15.03; p 0.001). Among those who
had not attempted suicide shortly before intake (left side of
Fig. 1), the next optimal predictor was patients’ BDI scores:
15% of patients whose BDI scores were greater than or equal
to 46 had attempted suicide at follow-up, as opposed to 2%
whose BDI scores were less than 46 (
2
1,252兴⫽10.54; p
0.001). Sensitivity for this model was calculated at .63, with
a specificity of .80.
Replication
For the suicide model, the first cutoff determined by
prior attempt within the previous 4 months did not replicate
well. Compared with 19% of the original sample who both
had a recent attempt and attempted in the 4 months postdis-
charge, 0% of the replication sample who had a recent prior
attempt attempted postdischarge. Sensitivity and specificity
for the model in the replication sample were .11 and .84,
respectively.
Violence
Exploratory Sample
For violence (Fig. 2), the single best predictor of
committing a violent act at follow-up was having committed
one or more violent acts prior to intake (
2
1,405兴⫽45.96;
p 0.001). Among those who had committed one or more
violent acts, the next best predictor was BDI scores
(
2
1,247兴⫽9.40; p 0.001), with patients scoring 34 or
greater comprising a higher risk group (68% vs. 48%). There
were no further distinctions for patients scoring less than 34
on the BDI. The number of past violent acts reemerged as the
best predictor (
2
1,149兴⫽7.09; p 0.001) among the
subgroup of patients with higher BDI scores, such that pa-
tients who committed two or more violent acts represented
the highest risk group in the model (77%).
Among patients who reported no violence prior to
intake, the best predictor of postdischarge violence was
Mississippi PTSD scores (
2
1,128兴⫽9.56; p 0.001), with
patients scoring greater than or equal to 121 comprising
another high-risk group (34%). In comparison, patients scor-
ing less than 121 on the Mississippi constituted a low-risk
group (6% likelihood of being positive for follow-up vio-
lence). Overall sensitivity of the model was .81, while spec-
ificity was .49.
Replication
The first cutoff in the violence model replicated well,
with similar percentages of patients reporting violence post-
discharge among the first split’s two subgroups, defined by
past number of types of violent episodes. However, at the
second split on the left side of the model, determined by a
Mississippi cutoff of 121, replication was poor in that 53% of
individuals with Mississippi scores lower than 121 reported
violence at follow-up, compared with only 6% of the original
sample. Similarly, the subgroup that had BDI scores greater
than or equal to 34 and fewer than two past violent episodes
comprised a high-risk group in the exploratory analysis but a
low-risk group upon replication. Sensitivity for the replica-
tion model was .72, while specificity was .60.
Intoxication
Exploratory Sample
The optimal predictor for intoxication at follow-up was
having been intoxicated 2 or more days during the 30-day
period prior to the month of required abstinence before intake
(
2
1,406兴⫽14.58; p 0.001; Fig. 3). Among those who
had been intoxicated more than two times prior to intake, the
next distinction was the commission of violent acts prior to
intake, with patients who committed one or more violent act
at greater risk (
2
1,83兴⫽9.30; p 0.001). An additional
split for the subgroup of patients who had not committed
violence was age, but the marginal means (cell sizes) were
too small for this to be reliable, and thus it was excluded from
the overall model. On the left side of the figure, among
patients who had been intoxicated fewer than two times in the
month prior to intake, the best predictor of violence was
Mississippi scores (
2
1,284兴⫽9.81; p 0.001), with
patients scoring greater than or equal to 159 representing a
high-risk group (30% vs. 9%). In the exploratory sample,
sensitivity for this model was .49, with a specificity of .84.
Replication
For alcohol intoxication, there were several replication
problems. The first split, determined by number of days
within the last 30 that alcohol was used, yielded results
similar to the original model, with 28% of the sample getting
intoxicated 2 or more days at follow-up if they had used prior
to intake 2 or more days, and 15% of the sample getting
intoxicated at follow-up if they had drank alcohol fewer than
2 days prior to intake. However, there were large percentage
differences on the outcome variable for patients who had used
alcohol 2 or more days in the 30 days prior to intake and who
had no past violent episodes (6% in original model vs. 42%
The Journal of Nervous and Mental Disease Volume 193, Number 7, July 2005 High-Risk Behaviors in PTSD
© 2005 Lippincott Williams & Wilkins 469
in the replication sample). Furthermore, Mississippi scores
among those who had fewer than 2 days of intoxication prior
to intake determined different outcomes between the original
and replication samples. Sensitivity was 100%, while speci-
ficity was 0%.
Hard Drug Use
Exploratory Sample
The best predictor of drug use at follow-up was the
number of days in the 30 days prior to the month of required
abstinence before intake that patients used drugs, with those
using on 3 or more days having significantly higher risk
(
2
1,405兴⫽45.96; p 0.001; Fig. 4). Among those who
had used drugs 3 or more days, an additional significant
predictor was BDI scores, with patients scoring 40 or above
posing the greatest risk for drug use at follow-up (
2
1,28兴⫽
8.50; p 0.001). Sensitivity of this model was .29, with a
specificity of .95.
Replication
The model for hard drug use replicated well, with no
more than 6% point differences among the two splits in the
model. For this model, sensitivity was .17, and specificity
was .96.
Post Hoc Tests of Alternate Models
While the low base rate for attempted suicide at fol-
low-up likely accounted for the replication problems in this
model, the models for intoxication and violence included
cut-off scores on the Mississippi PTSD scale that were near
the tails (extremely high or low), and thus might be unreliable
and unlikely to replicate in a new sample. In an attempt to
build more replicable models for violence and intoxication,
we replaced Mississippi Scale scores with BDI scores, as
scores for the BDI were generally similar to those of the
Mississippi Scale (Table 2). However, when BDI scores were
retained as the only symptom measure of distress and Mis-
sissippi scores were excluded from the model, the problems
with replication remained.
Additional ROC models were run to determine whether
better replication would ensue as a result of using discharge
scores on the Mississippi and BDI instead of intake scores.
However, none of the models, except for the violence model,
were better replicated by this substitution. The violence
TABLE 2. Optimal Cutpoints and
Values at the First Cut for Variables in the ROC Analysis
Variable
Suicide Violence Alcohol Hard Drugs
Best
cutoff
Best
cutoff
Best
cutoff
Best
cutoff
Age 45 .059 51 .135 51 .064 48 .075
Atrocities 2 .111 1 .076 2 .121 2 .055
Education 17 .048 14 .097 15 .069 14 .059
MISS intake 161 .212 134 .230 145 .106 153 .075
BDI intake 46 .203 32 .223 42 .128 40 .148
Number admits 2 .061 2 .046 2 .036 2 .014
Suicide attempt in past 4 mo 1 .246 1 .028 1 .033 1 .040
Suicide attempt ever 1 .056 1 .062 1 .034 1 .001
Incarceration 1 .009 1 .057 1 .031 1 .032
Marital status 1 .003 1 .009 1 .018 1 .023
Ethnicity 1 .014 1 .006 1 .047 1 .046
Dropout 1 .180 1 .018 1 .012 1 .009
Improvement 1 .058 0 .022 0 .074 0 .045
Violence 2 .028 1 .366 1 .086 1 .044
Alcohol Dx 1 .043 3 .052 2 .091 3 .059
Cannabis Dx 1 .018 2 .082 2 .032 1 .032
Hard drug Dx 1 .030 2 .093 2 .041 1 .085
Number of hard drug Dx 1 .043 2 .043 2 .117 2 .150
Alcohol use last 30 1 .045 3 .048 2 .205 2 .148
Drunk last 30 1 .025 28 .031 3 .180 2 .024
Cannabis use last 30 2 .031 30 .021 2 .132 28 .147
Drug use last 30 1 .091 25 .026 17 .115 3 .266
Hartl et al. The Journal of Nervous and Mental Disease Volume 193, Number 7, July 2005
© 2005 Lippincott Williams & Wilkins470
model replicated well primarily because past violence was the
only significant predictor that emerged.
DISCUSSION
Our goal in this study was to identify which patients in
a residential rehabilitation program for PTSD constituted
high-risk subgroups likely to re-engage in problematic be-
haviors after discharge. We had mixed success.
For drug use, we were successful in reliably differen-
tiating groups of patients at very high risk for drug use after
discharge (i.e., those with recent drug use and very severe
depression symptoms at intake) from those at moderate risk
(those with recent drug use and BDI scores below 40) and
relatively low risk (those without recent drug use). The
exploratory model replicated well but did not achieve good
sensitivity (.29 in the exploratory model and .17 in the
replication model). The models were highly specific, however
(exploratory model .95; replication model .96). Al-
though recent drug use and depression are intuitive and
already well-documented risk factors, the model developed
here provided specific cut-points and decision rules that could
be used to target follow-up services to those PTSD patients
who are most at risk.
Our efforts to predict relapse to violence and alcohol
misuse were less successful. On the one hand, recent behavior
(violence or drinking to intoxication) prior to intake consis-
tently predicted risk of relapse in both models. The additive
effects of symptom severity, however, were unclear. More
severe PTSD and/or depression symptoms were associated
with greater risk of violence or alcohol misuse in the explor-
atory sample but failed to predict these outcomes reliably in
the replication sample. Moreover, we were entirely unable to
reliably differentiate patients at higher and lower risk for
attempted suicide after discharge. None of the predictors
identified in the exploratory sample (not even recent suicide
attempts prior to intake) replicated in the second sample. The
difficulty of predicting suicide in this population is likely a
function of both the low base rate (in statistical terms,
although not in epidemiological terms) of suicide attempts
within any month period in contrast with the very high
proportion of patients at potential risk for suicide (roughly
half had a prior history of suicide attempts).
Considering the results from all four of these models
leads us to several conclusions. Our first conclusion is that
statistical replication is essential in testing the predictive
value of any given ROC model. ROC analysis, like stepwise
regression or factor analysis, is an empirically driven descrip-
tive technique in which models are built by maximizing their
fit with a given data set. The robustness of the resultant
models can be tested only through replication. Without rep-
lication, we would have had much more confidence than is
warranted in our ability to use specific cut-points on PTSD or
depression symptom measures to identify patients at higher or
lower risk for violence and alcohol misuse.
Our second conclusion is that recent behavior is the
strongest predictor of future behavior. This is hardly surpris-
ing. However, it was somewhat surprising that lifetime
history of behavior, at least in the case of lifetime history
of suicide attempts and Structured Clinical Interview for
DSM-IV Disorders substance use diagnoses, had little pre-
dictive value and may be much less informative in treatment
planning. Demographic information and severity of military
trauma also were not predictive of relapse in this sample.
These findings highlight the importance of assessing and
explicitly integrating recent problem behaviors into treatment
planning. For example, typical chart notes in a mental health
clinic may indicate that a patient carries a diagnosis of
substance use in remission but may not provide information
on last use and may not flag patients in early recovery as
requiring extra follow-up.
Homogeneity may reduce the predictive effects of
symptoms. Although there was considerable statistical vari-
ability within our sample in level of symptoms, all patients
were strongly symptomatic. For example, the mean Missis-
sippi PTSD Scale score in our sample was over a standard
deviation higher than in a validation sample drawn from an
outpatient VA Vet Center (Keane et al., 1988). The predictive
utility of PTSD symptom scores may be reduced when the
scores are so severe. In essence, the symptom severity in our
sample may have been so severe as to have lost some of its
meaningfulness in terms of discriminating patient outcomes.
In chronic patient populations, symptoms have a large
component that is stable over time. For example, in our sample,
the correlation between PTSD symptoms at intake and at dis-
charge was .57. However, symptoms also fluctuate markedly
over time. It may be the fluctuations in symptoms that are most
predictive of high-risk behaviors, i.e., patients accustomed to a
certain level of chronic distress, may become more likely to
engage in high-risk behaviors at times when their symptoms
become even worse. Thus, predicting risk of relapse may require
more proximal measurement of symptoms than those obtained at
intake, or even at discharge. The possible predictive value of
changes in acute symptoms also suggests the importance of
community-based interventions and frequent postdischarge fol-
low-ups to help patients better manage increases in their distress
levels (McNeil and Binder, 1997).
Moreover, it may not be symptoms per se, but skills and
confidence in being able to manage symptoms that is most
predictive of relapse. The data used in this study (like most
typical clinical data) do not include measures of self-efficacy for
managing distressing symptoms, or measures of commitment to
abstain from high-risk behaviors when under distress. Our anal-
yses were limited by the types of data that are perhaps most
commonly collected in PTSD treatment programs as part of a
standard assessment protocol. Inclusion of measures of self-
The Journal of Nervous and Mental Disease Volume 193, Number 7, July 2005 High-Risk Behaviors in PTSD
© 2005 Lippincott Williams & Wilkins 471
efficacy to maintain sobriety and remain abstinent from high-
risk behaviors at the commencement of treatment might yield
important information about which subsets of patients continue
to place themselves at risk after treatment.
Finally, the results from this study underscore the
importance of evaluating high-risk behaviors in this patient
population, given the high rates of each of the problem
behaviors occurring in a very short period after discharge
from a highly structured and controlled therapeutic environ-
ment. This may speak to the difficulty in changing established
behavior patterns such as violence, suicide attempts, and
substance use. However, it may also be indicative of the
difficulty current treatment programs may have in targeting
these problems, especially in patients who pose substantial
risk. For individuals with chronic treatment-resistant PTSD, it
would seem that reducing high-risk behaviors should indeed
be a major goal of treatment.
There are several limitations to this study. First, it is
unclear to what extent these findings generalize to other
PTSD samples in and outside of the VA system of care. For
example, patients in our sample might present with more
chronic and resistant PTSD syndromes that are associated
with and driven by different behavioral, demographic, and
psychological factors. Therefore, different models may have
emerged if tested in different samples. In addition, no thera-
peutic contextual variables were included in the analysis to
assess the impact of treatment factors or the therapeutic
milieu on outcomes. Patients’ response to treatment, includ-
ing perhaps their perception of support among treatment staff,
might influence behavioral outcomes. Finally, we have not
considered variables in postdischarge treatment that might be
predictive of patient outcomes.
CONCLUSION
High-risk behaviors persist among veterans with PTSD
despite treatment completion. We have presented models that
emerged from empirically driven analyses using signal de-
tection methods in hopes of identifying subgroups of patients
at highest risk for negative behavioral outcomes following
treatment. However, ROC analysis aimed at identifying high-
risk patients did not yield models that were replicable beyond
the level of identifying recent behavior as the most robust
predictor of future high-risk behaviors. This may suggest that
current standard psychometric measures are poor predictors
of future behavior. Future research is needed to determine the
extent to which empirically based prediction models can be
useful in terms of sensitivity and ability to replicate findings,
which could then be used to tailor interventions and fol-
low-up procedures to minimize negative outcomes.
ACKNOWLEDGMENTS
The authors thank Helena Kraemer for her consultation
regarding data interpretation.
REFERENCES
Beck AT, Steer RA, Brown GK (1996) Beck Depression Inventory (2nd ed,
Manual). San Antonio (TX): Psychological Corp.
Beckham JC, Feldman ME, Kirby AC, Hertzberg MA, Moore SD (1997)
Interpersonal violence and its correlates in Vietnam veterans with chronic
posttraumatic stress disorder. J Clin Psychol. 53:859 869.
Bonin MF, Norton GR, Asmundson GJG, Dicurzio S, Pidlubney S (2000)
Drinking away the hurt: The nature and prevalence of PTSD in substance
abuse patients attending a community-based treatment program. J Behav
Ther. 31:55– 66.
Chief Medical Director’s Special Committee on PTSD (1991) Program
Guide to Specialized Inpatient Post-Traumatic Stress Disorder Units
(SIPUs). Washington DC: Department of Veterans Affairs.
First MB, Spitzer RL, Gibbon M, Williams JBW (1995) Structured Clinical
Interview for DSM-IV Axis I Disorders—Patient Edition (SCID-I/D,
Version 2.0). New York: Biometrics Research Department, New York
Psychiatric Institute.
Fontana A, Rosenheck R (1997). Outcome Monitoring of VA Specialized
Intensive PTSD Programs: FY 1996 Report. West Haven (CT): Northeast
Program Evaluation Center.
Fontana A, Rosenheck R (1996) Improving the efficiency of outpatient
treatment for posttraumatic stress disorder. Admin Policy Mental Health.
23:197–210.
Fontana A, Rosenheck R (1995) An etiological model of attempted suicide
among Vietnam theater veterans. J Nerv Ment Dis. 183:377–383.
Hammarberg M, Silver SM (1994) Outcome of treatment for post-traumatic
stress disorder in a primary care unit serving Vietnam veterans. J Trauma
Stress. 7:195–216.
Hiley-Young B, Blake DD, Abueg FR, Rozynko V, Gusman FD (1995)
Warzone violence in Vietnam: An examination of premiliatry, military
and postmilitary factors in PTSD in-patients. J Trauma Stress. 8:125–141.
Johnson DR, Rosenheck R, Fontana A, Lubin H, Southwick S, Charney D
(1996) Outcome of intensive inpatient treatment for combat-related post-
traumatic stress disorder. Am J Psychiatry. 153:771–777.
Keane TM, Caddell JM, Taylor KL (1988) Mississippi Scale for Combat-
Related Posttraumatic Stress Disorder: Three studies in reliability and
validity. J Consult Clin Psychol. 56:85–90.
Keane TM, Wolfe J, Taylor KL (1987) Post-traumatic stress disorder:
Evidence for diagnostic validity and methods of psychological assessment.
J Clin Psychol. 43:32– 43.
Kiernan M, Kraemer HC, Winkleby MA, King AC, Taylor CB (2001) Do
logistic regression and signal detection identify different subgroups at
risk? Implications for the design of tailored interventions. Psychol
Methods. 6:35– 48.
Kraemer HC (1992) Evaluating Medical Tests: Objective and Quantitative
Guidelines. Newbury Park: Sage.
Kulka RA, Schlenger WE, Fairbank JA, Hough RL, Jordan BK, Marmar R,
Weiss D (1990) Trauma and the Vietnam War Generation: Report of
Findings From the National Vietnam Readjustment Study. New York:
Brunner/Mazel.
McFall M, Fontana A, Raskind M, Rosenheck R (1999) Analysis of violent
behavior in Vietnam combat veteran psychiatric inpatients with posttrau-
matic stress disorder. J Trauma Stress. 12:501–517.
McLellan AT, Kushner H, Metzger D, Peters R, Smith I, Grissom G,
Pettinati H, Argeriou M (1992) The 5th edition of the addiction severity
index. J Subst Abuse Treatment. 9:199 –213.
McNeil DE, Binder RL (1997) The impact of hospitalization on clinical
assessments of suicide risk. Psychiatr Serv. 48:204 –208.
Perconte ST, Griger ML (1991) Comparison of successful, unsuccessful and
relapsed Vietnam veterans treated for posttraumatic stress disorder. J Nerv
Ment Dis. 179:558 –562.
Rosenheck R, Fontana A (2001) Impact of efforts to reduce inpatient costs on
clinical effectiveness: Treatment of posttraumatic stress disorder in the
Department of Veterans Affairs. Med Care. 39:168 –180.
Savarese VW, Suvak MK, King LA, King DW (2001) Relationships among
alcohol use, hyperarousal and marital abuse and violence in Vietnam
veterans. J Trauma Stress. 14:717–732.
Hartl et al. The Journal of Nervous and Mental Disease Volume 193, Number 7, July 2005
© 2005 Lippincott Williams & Wilkins472
    • "Research has found that among veterans with a primary diagnosis of PTSD, recent behavior was the strongest predictor of future behavior in this population. Further, a lifetime history of suicide attempts had little predictive value and, perhaps, was much less informative than recent behavior in treatment planning (Hartl et al. 2005). The risk factors in veterans for PTSD can be divided into three groups. "
    [Show abstract] [Hide abstract] ABSTRACT: Individuals with post-traumatic stress disorder PTSD have higher rates of morbidity and mortality as well as an increased risk for death by suicide. Moreover, research has shown that individuals who are diagnosed with (PTSD) are at an increased risk for both suicide ideation and attempts. Among veterans, suicide is far too common and the risk continues to be high for many years following discharge from service. The PTSD-suicidal behavior link has been substantiated by the presence of common risk factors and high rates of comorbidity. War veterans should be adequately assessed for current suicidal behavior on a routine basis.
    Chapter · Jan 2016 · Temas em Psicologia
    • "Concerns about aggression are common among Iraq and Afghanistan era veterans as well: a study of 1,397 U.S. Iraq era veterans suggested that as many as 67% reported threatening others or engaging in aggressive behavior within the past month (Wright et al., 2012). Further, veterans of these conflicts have high rates of PTSD (Hoge et al., 2004), and PTSD has been found to be robustly associated with anger (Calhoun et al., 2002; Crawford, Calhoun, Braxton, & Beckham, 2007), intimate partner violence (Byrne & Riggs, 1996; Jordan et al., 1992; Orcutt, King, & King, 2003; Taft, Street, Marshall, Dowdall, & Riggs, 2007; Taft, Watkins, Stafford, Street, & Monson, 2011), and general interpersonal violence (Beckham, Feldman, & Kirby, 1998; Begic & Jokic-Begic, 2001; Freeman & Roca, 2001; Hartl, Rosen, Drescher, Lee, & Gusman, 2005; Jakupcak et al., 2007; McFall, Fontana, Raskind, & Rosenheck, 1999; Taft, Vogt, Marshall, Panuzio, & Niles, 2007), particularly in military samples (Taft et al., 2011). Despite consistent associations between PTSD, combat deployment, and aggression in veterans, the majority of veterans with PTSD are not violent. "
    [Show abstract] [Hide abstract] ABSTRACT: Posttraumatic stress disorder (PTSD) is associated with aggressive behavior in veterans, and difficulty controlling aggressive urges has been identified as a primary postdeployment readjustment concern. Yet only a fraction of veterans with PTSD commit violent acts. The goals of this study were to (1) examine the higher-order factor structure of Personality Assessment Inventory (PAI) scales in a sample of U.S. military veterans seeking treatment for PTSD; and (2) to evaluate the incremental validity of higher-order latent factors of the PAI over PTSD symptom severity in modeling aggression. The study sample included male U.S. Vietnam (n = 433) and Iraq/Afghanistan (n = 165) veterans who were seeking treatment for PTSD at an outpatient Veterans Affairs (VA) clinic. Measures included the Clinician Administered PTSD Scale, the PAI, and the Conflict Tactics Scale. The sample was randomly split into two equal subsamples (n's = 299) to allow for cross-validation of statistically derived factors. Parallel analysis, variable clustering analysis, and confirmatory factor analyses were used to evaluate the factor structure, and regression was used to examine the association of factor scores with self-reports of aggression over the past year. Three factors were identified: internalizing, externalizing, and substance abuse. Externalizing explained unique variance in aggression beyond PTSD symptom severity and demographic factors, while internalizing and substance abuse did not. Service era was unrelated to reports of aggression. The constructs of internalizing versus externalizing dimensions of PTSD may have utility in identifying characteristics of combat veterans in the greatest need of treatment to help manage aggressive urges. Aggr. Behav. 9999:XX–XX, 2014. Published 2014. This article is a U.S. Government work and is in the public domain in the USA.
    Full-text · Article · Nov 2014
    • "Num estudo com 221 veteranos com TEPT verifi cou-se que cerca de 83% tinham sobrepeso ou obesidade tendo em conta o IMC encontrado na população de veteranos e na população geral (Vieweg et al., 2006 ) sendo maior a probabilidade de encontrar obesidade em mulheres veteranas com TEPT do que sem TEPT (Dobie et al., 2004). O TEPT é um preditor do envolvimento em comportamentos de risco nos veteranos de guerra e, quanto mais severo o quadro de TEPT, maior a tendência para o envolvimento em comportamentos de risco independentemente do grau de exposição a combate (Hartl, Rosen, Drescher, Lee, & Gusman, 2005; Svetlicky et al., 2010). Contudo, também a história prévia de exposição ao trauma na infância parece ser um fator de risco para o desenvolvimento de TEPT e adoção de comportamentos de risco. "
    [Show abstract] [Hide abstract] ABSTRACT: Resumo A vivência de experiências adversas na infância, a exposição ao trauma e o quadro clínico de TEPT são fatores de risco para o desenvolvimento de comportamentos de risco (consumo de substâncias, comportamentos sexuais de risco e estilo de vida sedentário). Para além disso, a literatura tem vindo a documentar que existe um fenômeno de transmissão intergeracional de comportamentos de risco dos progenitores para os fi lhos realçando a importância da intervenção com as vítimas primárias de um trauma devido não só ao risco de envolvimento em comportamentos que prejudicam a saúde mas também devido ao seu poder como modelo parental na transmissão dos comportamentos de risco. Este trabalho apresenta estudos, na área do stress traumático na população geral e em veteranos de guerra tendo tam-bém em consideração a investigação sobre o mecanismo intergeracional dos comportamentos de risco da família de origem para a vida adulta. Palavras-chave: Comportamentos de risco, experiências adversas, trauma, TEPT. Abstract Adverse childhood experiences, trauma exposure and PTSD are risk factors for the development of health risk behaviors (substance use, sexual risk behaviors and sedentary life style). In addition, literature has shown that there is an intergenerational transmission of parent´s health risk behaviors to the offspring, emphasizing the importance of intervention with primary victims of trauma due to the risk of their involvement in health damaging behaviors and the power of parents, as roles models, in the inter-generational transmission of risk behaviors. This work presents studies in the area of traumatic stress in the general population and in war veterans taking also in consideration the research regarding the intergeneracional transmission of risk behaviors, from the family of origin to adult life.
    Full-text · Article · Jan 2013
Show more