Order #369341 Topic: Assignment-2
tutorthammy
Predicting_Recidivism_in_Sex_Offenders.pdf
Predicting Recidivism With the Personality Assessment Inventory in a Sample of Sex Offenders Screened for Civil Commitment as Sexually
Violent Predators
Marcus T. Boccaccini Sam Houston State University
Daniel C. Murrie University of Virginia
Samuel W. Hawes, Amber Simpler, and Jeremy Johnson Sam Houston State University
We examined the ability of scores from the Personality Assessment Inventory (PAI; Morey, 1991) to predict postrelease (M � 4.90 years follow-up) arrests in a sample of 1,412 sex offenders. We focused on scores from 4 PAI measures conceptually relevant to offending, including the Antisocial Features (ANT), Aggression (AGG), and Dominance (DOM) scales, as well as the Violence Potential Index (VPI). Scores from several PAI measures demonstrated small- to medium-sized effects in predicting violent nonsexual recidivism, nonviolent recidivism, and sex offender registry violations, with the AGG scale being the strongest (d � 0.50 for violent nonsexual recidivism, d � 0.55 for sex offender registry violations) and most consistent predictor of recidivism.
Keywords: risk assessment, sexually violent predator, personality assessment inventory, sex offenders
Sexually violent predator (SVP) laws allow states to civilly commit certain sexual offenders, even after their prison sentence is completed, if courts determine that they pose a high risk for sexual reoffense (see Miller, Amenta, & Conroy, 2005). SVP laws are unusual in that they rely so heavily on the results of psychological assessments. Across the 20 states (and the federal system) that have SVP laws, the SVP selection process typically begins by administering risk measures to the sexual offenders facing release from prison. On the basis of scores on these measures, a subset of offenders is selected for more intensive assessment, usually with additional assessment instruments and clinical interview. As of 2007, more than 4,534 offenders had been committed under SVP laws (Gookin, 2007), but hundreds times this many had been evaluated during some stage of the SVP selection process.
Typically, psychologists perform SVP evaluations, and most use measures specifically designed to predict recidivism (Jackson & Hess, 2007). Usually, these measures are scored by clinicians on the basis of file review (e.g., STATIC-99; Hanson & Thornton, 1999) or file review plus interview (e.g., Psychopathy Checklist– Revised; Hare, 2003). Often, offenders’ correctional files contain results from self-report multiscale inventories used to assess per-
sonality and psychopathology. For example, in Texas, most of- fender files contain results from the Personality Assessment In- ventory (PAI; Morey, 1991, 2007). One review of Texas SVP evaluator reports showed that 45% mentioned the PAI (Amenta, 2005).
Although most self-report inventories were not designed as risk measures, they often include measures that are conceptually rele- vant to risk. Walters (2006) used meta-analysis to show that scores from “content-relevant” (p. 279) self-report measures, such as the PAI, tend to predict violence, misbehavior, and recidivism just as well as scores from clinician-scored measures specifically de- signed to assess violence or recidivism risk. Moreover, the PAI’s developer has argued that “several PAI indexes are particularly useful in assessing risk to others” (Morey, Warner, & Hopwood, 2007, p. 117), including Antisocial Features (ANT), Aggression (AGG), and Dominance (DOM) scales and the Violence Potential Index (VPI). Clinicians apparently use the PAI to address legal questions related to risk: A recent review of case law showed that the PAI has been allowed as evidence in SVP, criminal sentencing, and death penalty trials (Mullen & Edens, 2008).
Although scores from these content-relevant PAI measures should be associated with postrelease recidivism, only in a few small studies, with brief follow-up periods, have researchers ex- amined the relation with postrelease recidivism. Thus, experts who might consider PAI results during risk assessments have little to help them make empirically informed decisions about PAI scores. Thus, the goal of this research is to help practitioners make informed decisions about the appropriate use of PAI scores in risk assessment. We examined the predictive validity of PAI scores in a large-scale (N � 1,412) prospective (M � 4.90 years) recidivism study. The sample size for this study is nearly 3 times the size of all other prospective PAI recidivism studies combined, and the
Marcus T. Boccaccini, Samuel W. Hawes, Amber Simpler, and Jeremy Johnson, Department of Psychology, Sam Houston State University; Daniel C. Murrie, Institute of Law, Psychiatry, and Public Policy, Univer- sity of Virginia.
Additional findings from this study are in Boccaccini, Murrie, Caperton, and Hawes (2009).
Correspondence concerning this article should be addressed to Marcus T. Boccaccini, Box 2447, Department of Psychology, Sam Houston State University, Huntsville, TX 77341. E-mail: [email protected]
Psychological Assessment © 2010 American Psychological Association 2010, Vol. 22, No. 1, 142–148 1040-3590/10/$12.00 DOI: 10.1037/a0017818
142
T hi
s do
cu m
en t i
s co
py ri
gh te
d by
th e
A m
er ic
an P
sy ch
ol og
ic al
A ss
oc ia
tio n
or o
ne o
f i ts
a lli
ed p
ub lis
he rs
. T
hi s
ar tic
le is
in te
nd ed
s ol
el y
fo r t
he p
er so
na l u
se o
f t he
in di
vi du
al u
se r a
nd is
n ot
to b
e di
ss em
in at
ed b
ro ad
ly .
follow-up time is nearly 4 times longer than the follow-up time in existing studies.
ANT, AGG, and DOM on the PAI
The ANT and AGG scales are the most studied PAI measures with respect to recidivism, criminal behavior, and misconduct. Morey (1991) developed the ANT scale to measure features of antisocial personality disorder and psychopathy, including egocen- tricity, antisocial attitudes, and sensation seeking (Morey, 2007). He developed the AGG scale as a measure of aggressive attitudes and behaviors, including physical and verbal aggression (Morey, 2007). Only two published studies have examined the ability of AGG and ANT scores to predict recidivism. Walters and Duncan (2005) found that ANT (area under curve [AUC] � .65) and AGG (AUC � .68) scores were statistically significant predictors of postrelease (minimum 1-year follow-up) arrests among 91 male offenders. Salekin, Rogers, Ustad, and Sewell (1998) also found that ANT (point biserial correlation [rpb] � .26) and AGG (rpb � .29) scores were small but statistically significant predictors of postrelease arrests (minimum 1-year follow-up) in a sample of 78 female jail inmates. In an unpublished study described in the PAI manual (Morey, 2007), ANT (AUC � .66) and AGG (AUC � .62) scores were statistically significant predictors of self-reported postrelease arrests (1-year follow-up) among 326 former jail in- mates (Hastings, Stuewig, & Tangney, 2006).
Researchers have also studied ANT and AGG as predictors of institutional infractions and violence, with most studies reporting significant predictive effects in the AUC � .60 to .70 range (see Buffington-Vollum, Edens, Johnson, & Johnson, 2002; Caperton, Edens, & Johnson, 2004; Edens, Buffington-Vollum, Colwell, John- son, & Johnson, 2002; Walters, 2007; Walters, Duncan, & Geyer, 2003). Although ANT and AGG appear to be consistent small- to medium-sized predictors of recidivism and disciplinary infractions, some research suggests that these effects depend on whether inmates responded openly or defensively on the PAI. Edens and Ruiz (2006) found that ANT scores were stronger predictors of infractions for male offenders who responded openly (i.e., low Positive Impression scores) than for those who responded defensively.
Morey (2007) developed the DOM scale to measure the extent to which individuals tend to assume a controlling (high scores) or submissive (low scores) role in relationships. Morey (2007) de- scribed high scorers as “domineering” and as tending to “have little tolerance for those who disagree with their plans” (p. 47). Researchers have found that DOM scores are small but statistically significant predictors of physically aggressive disciplinary infrac- tions among male inmates (AUC � .60; Edens, 2009) and general disciplinary infractions among female inmates (r � .25; Skopp., Edens, & Ruiz, 2007).
VPI
Unlike ANT and AGG, Morey (1996) created the VPI specifically to assess risk of violence. VPI scores are based on the presence (1 point) or absence (0 points) of 20 different PAI profile features, including scores on ANT and AGG subscales. For example, an offender would receive 1 point if his ANT-A (Antisocial Behaviors) subscale score was greater than 70T. Although the VPI is not as well-studied as ANT or AGG, the existing research suggests that
predictive effects for VPI scores tend to be in the same range as those from ANT and AGG for predicting postrelease arrests (Hastings et al., 2006) but somewhat smaller for predicting institutional violence (Ca- perton et al., 2004; Skopp et al., 2007).
Current Study
In this study, we examine the predictive validity of ANT, AGG, DOM, and VPI with respect to postrelease arrests in a sample of 1,412 sex offenders who had anywhere from 2.25 years to 7.50 years of opportunity to reoffend (M � 4.90 years, SD � 1.48). Each offender was screened for commitment as an SVP but was not committed. Most offenders completed the PAI upon entry to a sex offender treatment program. Evaluators who conduct SVP evaluations in Texas receive information about PAI scores as part of the offender’s institutional records. Thus, these PAI scores were available to inform SVP risk assessments by forensic evaluators.
To help gauge the usefulness of scores on these PAI measures, we examined their incremental validity over easily obtained offender characteristics (e.g., age at release, prior offenses). For sexually vio- lent recidivism, we also examined their incremental validity over scores from an actuarial measure specifically designed to predict sexual recidivism (STATIC-99; Hanson & Thornton, 1999). In a larger but partially overlapping sample of sex offenders, the STATIC-99 was the most effective actuarial measure for predicting sexually violent recidivism (AUC � .58) and the combination of violent and sexually violent recidivism (AUC � .57; Boccaccini, Murrie, Caperton, & Hawes, (2009)). Although the predictive effects for the STATIC-99 were relatively small in size, they provide a useful point of comparison for the PAI and suggest that factors other than those measured by actuarial instruments (e.g., scores from self-report measures) may aide in prediction.
Method
Participants and Procedure
The men in this study were inmates in the Texas Department of Criminal Justice (TDCJ) at the time they completed the PAI. TDCJ staff administered the PAI as part of routine procedure upon admission to a sex offender treatment program, usually about 18 months prior to release. Offenders being considered for commit- ment who were not enrolled in the treatment program were asked to complete the PAI about 16 months prior to release. PAI scores are available to a multidisciplinary team that makes decisions about which offenders warrant referral for further evaluation as part of the state’s SVP commitment process. This team also has access to scores from risk measures designed to predict sexual recidivism (e.g., STATIC-99). The experts (usually psychologists) who conduct the subsequent evaluations for commitment receive information about PAI and risk measure scores as part of a larger packet of institutional records.
TDCJ allowed the research team to access an existing data- base of offender data (e.g., release date, age, and race) and risk instrument scores for the first 1,983 offenders screened for commitment (between 1999 and 2004). The only information about PAI scores in the database was a column of text summa- rizing the overall PAI profile, so we obtained PAI scale scores from paper copies of computer generated PAI reports in of-
143SEX OFFENDER RECIDIVISM
T hi
s do
cu m
en t i
s co
py ri
gh te
d by
th e
A m
er ic
an P
sy ch
ol og
ic al
A ss
oc ia
tio n
or o
ne o
f i ts
a lli
ed p
ub lis
he rs
. T
hi s
ar tic
le is
in te
nd ed
s ol
el y
fo r t
he p
er so
na l u
se o
f t he
in di
vi du
al u
se r a
nd is
n ot
to b
e di
ss em
in at
ed b
ro ad
ly .
fender files. The files contained PAI scores for 1,532 (77.3%) of the 1,983 offenders. Those with PAI scores did not differ from those without PAI scores in terms of age at release (d � 0.08), STATIC-99 scores (d � 0.08), or any recidivism cate- gory (odds ratios � 1.30, p � .25). Those without PAI scores had significantly more prior arrests (M � 5.37, SD � 4.02) than did those with PAI scores (M � 4.86, SD � 4.02), although this difference was small (d � 0.12). Offenders identified as His- panic were more likely to be missing PAI scores (missing for 30%) than were those identified as White (missing for 18%) or Black (missing 24%), �2(2, N � 1972) � 33.13, p � .01. The TDCJ database provided an explanation for some missing scores, identifying 57 offenders as too mentally ill to be tested, 51 as refusing the evaluation, and 26 as producing invalid scores. It is unclear whether the PAI was given to the remaining 317 offenders with no PAI scores in their files.
We removed 7 offenders with PAI scores from the dataset because we could not obtain any information about their pre- or postrelease arrests. We removed 16 additional offenders who were civilly committed as SVPs because they did not have the same opportunity to reoffend as released offenders.1 Civilly committed offenders had somewhat higher ANT (d � 0.46), AGG (d � 0.28), and VPI (d � 0.37) scores than did noncommitted offenders, and lower DOM (d � �0.17) scores, although these differences were not large enough to reach statistical significance given the small number of committed offenders (n � 16).
Of the 1,509 remaining offenders, we considered scores from 97 (5.0%) to be invalid due to either a high likelihood of random responding (Inconsistency scale [INC] � 73; n � 82), overreport- ing of psychopathology (Negative Impression scale [NIM] � 92; n � 13), or both (n � 2). Scores from these offenders were not used in the study analyses.2 The average age at release for the 1,412 offenders with valid PAI scores was 42.84 years (SD � 11.93). Almost half of the offenders were identified by the TDCJ database as White (n � 733, 51.9%), whereas others were identi- fied as Hispanic (n � 377, 26.7%), Black (n � 293, 20.8%), or other (n � 9, 0.6%). The length of time between release and collection of recidivism data ranged from 2.25 years to 7.50 years (M � 4.90, SD � 1.48).
Measures
PAI. The PAI (Morey, 1991, 2007) is a 344-item self-report measure designed to measure critical clinical variables, such as psychopathology and treatment motivation. In this study, we ex- amined the predictive validity of the PAI AGG, ANT, and DOM scales, as well as the VPI. Scores on AGG (� � .90), ANT (� � .86), and DOM (� � .82) all demonstrated acceptable levels of internal consistency in the PAI normative clinical sample (Morey, 2007). Internal consistency values for scores on these scales have been generally consistent across samples, although the internal consistency of DOM scores has been below .70 in some studies (see Morey, 2007, p. 135). The PAI manual does not report reliability data for VPI scores.
STATIC-99. The STATIC-99 is a 10 item actuarial mea- sure designed to predict sexual recidivism (Hanson & Thornton, 1999). Examples of items scored on the STATIC-99 include age at release, prior convictions for nonsexual violence, history of offending against unrelated victims, and number of prior of-
fenses. Each of the items except number of prior offenses is scored as 1 (present) or 0 (absent), and scores can range from 0 to 12. The STATIC-99 is the most frequently used (Jackson & Hess, 2007) and well-studied measure for predicting sexual recidivism, with a recent meta-analysis reporting a mean effect size of d � 0.67 for STATIC-99 scores across 63 samples (Hanson & Morton-Bourgon, 2009). Hanson and Morton- Bourgon (2009) reported a median rater agreement value of .90 for the STATIC-99 scores across 12 samples. The STATIC-99 was included in this study as a point of comparison for the PAI measures. Detailed recidivism findings for the STATIC-99 and the Minnesota Sex Offender Screening Tool (Epperson, Kaul, & Hesselton, 1998) in larger and partially overlapping sample are available elsewhere (Boccaccini et al., 2009).
Pre- and postrelease arrests. Information about pre- and postrelease arrests was provided by the Texas Department of Public Safety. The Department of Public Safety listed a Na- tional Crime Information Center code number and arrest date for each arrest. We used National Crime Information Center scores to group arrests into four main (nonoverlapping) cate- gories: violent sexual (e.g., rape, sexual assault, kidnapping of minor to sexually assault), violent nonsexual (e.g., murder, assault, robbery), nonviolent and nonsexual (e.g., weapons pos- session, substance related charges, probation violation), and sex offender registry violations. We also created an arrest category that identified offenders who had either a violent sexual or a violent nonsexual rearrest because this may be a more sensitive indicator of sexually motivated offending than sexual offenses alone (see Rice, Harris, Lang, & Cormier, 2006) and because the STATIC-99 provides norms for this combined recidivism category (see Helmus, Hanson, & Thornton, 2009).
We analyzed recidivism arrests as a dichotomous variable be- cause of the low base rate for several of the recidivism categories (see Table 1). Because all offenders had prior arrests, we analyzed prior arrests as a continuous (count) variable. The analyses re- ported in this study focus on the total number of prior arrests as a measure of offense history.3 The total number of prior arrests ranged from 1 to 55, with a mean of 4.83 (SD � 4.05).
1 Although Texas uses outpatient commitment for SVP offenders, we removed them from the study because committed offenders are subject to intensive monitoring and restrictions that are qualitatively distinct from those of noncommitted offenders.
2 We did not use Positive Impression (PIM) scores to identify invalid protocols (defensive responding) because we planned to use PIM as a potential moderator of predictive effects (see Edens & Ruiz, 2006).
3 We examined predictive effects for specific types of prior arrests (e.g., sexually violent, violent, nonviolent) and found that total number of prior arrests performed similarly to or outperformed these more specific prior offense variables. We also examined predictive effects separately for offenders who were discharged and those released on parole or mandatory supervision because, in a partially overlapping sample of offenders, actu- arial measures were more effective for predicting recidivism among dis- charged offenders (Boccaccini et al., 2009). Predictive effects for the PAI measures were nearly identical in both subgroups of offenders.
144 BOCCACCINI, MURRIE, HAWES, SIMPLER, AND JOHNSON
T hi
s do
cu m
en t i
s co
py ri
gh te
d by
th e
A m
er ic
an P
sy ch
ol og
ic al
A ss
oc ia
tio n
or o
ne o
f i ts
a lli
ed p
ub lis
he rs
. T
hi s
ar tic
le is
in te
nd ed
s ol
el y
fo r t
he p
er so
na l u
se o
f t he
in di
vi du
al u
se r a
nd is
n ot
to b
e di
ss em
in at
ed b
ro ad
ly .
Results and Discussion
Univariate Predictors of Recidivism
We calculated two univariate effects for each predictor: area under the receiver operating characteristic curve (AUC) and Co- hen’s d (see Table 1). In the context of recidivism research, an AUC value indicates the likelihood that a randomly selected re- cidivist will score higher on the measure than will a randomly selected nonrecidivist. Cohen’s d is an effect size for comparing mean scores, reflecting the number of pooled standard deviation units between means for the two groups being compared (e.g., recidivists vs. nonrecidivists).
Scores from several PAI measures were statistically significant predictors of multiple types of recidivism (see Table 1). The most consistent predictor of recidivism was AGG, which was a signif- icant predictor of all types of recidivism except sexually violent recidivism. AGG scores also had the strongest (in terms of abso- lute value) predictive effects in the study (violent recidivism AUC � .63, sex offender registry violation AUC � .65). ANT, DOM, and VPI scores were generally small but statistically sig- nificant predictors of nonsexual recidivism. These effects for AGG scores are consistent with those from prior PAI research examining postrelease recidivism in smaller samples (Hastings et al., 2006; Salekin et al., 1998) and tend to support the construct validity of scores on these PAI subscales, particularly AGG.
Overall, the PAI measures appeared to be most effective for pre- dicting sex offender registry violations, that is, criminal charges for failing to register (or update) one’s status with local authorities when required to do so. Although registry violations are not violent offenses (and therefore not directly related, conceptually, with AGG or VPI), they may reflect an antisocial style of nonconforming, reckless, or irresponsible behavior. The effect sizes in these analyses reveal that the PAI measures are far from perfect predictors of registry violations. However, the effects also hint that an offender’s self-reported antiso- cial attitudes may bear some relation to that offender’s performance under community management or supervision. In future research, this possibility should be more directly examined.
No PAI measure was a strong predictor of sexually violent recid- ivism or the combination of violent or sexually violent recidivism, which are the two categories most relevant to SVP laws and public safety concerns. Across all offenders, the best performing PAI mea- sure for these outcomes was the DOM scale. DOM was a small (AUC � .59) but statistically significant predictor of the combination of violent and sexually violent recidivism and the only PAI measure positively (i.e., AUC � .50) associated with sexually violent recidi- vism. Conceptually, the relationship between DOM and sexual reoff- ense appears reasonable and offers some support for the construct validity of DOM. However, the small effect size probably means that the practical value of DOM for predicting violent sexual recidivism is limited.
Table 1 Comparison of PAI Scores Between Recidivists and Nonrecidivists
Recidivism type: Predictor
Recidivist Nonrecidivist
d AUC SEM SD M SD
Sexually violent base rate � 2.8% PAI: ANT 54.03 7.81 55.82 8.75 �0.21 .45 0.05 PAI: AGG 46.69 9.65 48.06 9.59 �0.14 .46 0.05 PAI: DOM 52.82 8.65 50.72 9.08 0.23 .56 0.05 PAI: VPI 2.13 2.28 2.69 2.70 0.21 .43 0.05
Violent nonsexual base rate � 6.4% PAI: ANT 58.13 10.52 55.61 8.57 0.29 .57� 0.03 PAI: AGG 52.51 11.22 47.71 9.39 0.50 .63�� 0.03 PAI: DOM 53.53 9.50 50.59 9.02 0.32 .59�� 0.03 PAI: VPI 3.47 3.16 2.62 2.64 0.32 .59�� 0.03
Violent or sexually violent base rate � 8.8% PAI: ANT 57.10 10.30 55.64 8.58 0.17 .54 0.03 PAI: AGG 50.66 11.17 47.78 9.38 0.30 .58�� 0.03 PAI: DOM 53.44 9.27 50.52 9.02 0.32 .59�� 0.03 PAI: VPI 3.11 3.03 2.64 2.65 0.18 .55 0.03
Sex offender registry violation base rate � 11.2% PAI: ANT 59.75 9.94 55.27 8.43 0.52 .63�� 0.02 PAI: AGG 52.68 10.80 47.44 9.26 0.55 .65�� 0.02 PAI: DOM 52.83 9.82 50.53 8.95 0.25 .57�� 0.02 PAI: VPI 3.80 3.55 2.54 2.53 0.48 .60�� 0.02
Nonviolent nonsexual base rate � 18.3% PAI: ANT 58.35 9.93 55.19 8.33 0.37 .59�� 0.02 PAI: AGG 50.64 11.23 47.43 9.08 0.34 .58�� 0.02 PAI: DOM 52.38 9.31 50.42 8.99 0.22 .56�� 0.02 PAI: VPI 3.33 3.07 2.53 2.57 0.30 .58�� 0.02
Note. Cohen’s d is an effect size for comparisons of scale scores for those who were and were not rearrested. Asterisks indicate whether the AUC value is significantly difference from chance. PAI � Personality Assessment Inventory; ANT � Antisocial Features scale; AGG � Aggression scale; DOM � Dominance scale; VPI � Violence Potential Index; AUC � area under the receiver operating characteristic curve. � p � .05. �� p � .01.
145SEX OFFENDER RECIDIVISM
T hi
s do
cu m
en t i
s co
py ri
gh te
d by
th e
A m
er ic
an P
sy ch
ol og
ic al
A ss
oc ia
tio n
or o
ne o
f i ts
a lli
ed p
ub lis
he rs
. T
hi s
ar tic
le is
in te
nd ed
s ol
el y
fo r t
he p
er so
na l u
se o
f t he
in di
vi du
al u
se r a
nd is
n ot
to b
e di
ss em
in at
ed b
ro ad
ly .
Multivariate Analyses
We used logistic regression to answer two questions.4 First, do PAI scores demonstrate incremental validity over easily obtained offender background characteristics (age at release, total number of prerelease arrests)5 for predicting recidivism? We focused these analyses on violent recidivism and sex offender registry violations because the PAI scores were somewhat stronger predictors of these offense categories than others. For these analyses, we entered age at release and total number of prerelease arrests into the first regression model. We then added AGG scores in the second model because AGG was the strongest (absolute value) univariate pre- dictor of both types of recidivism. In the third model, we entered DOM and ANT scores. We did not use VPI scores in the regres- sion models because VPI is not statistically independent from AGG, ANT, or DOM (i.e., several VPI items are based on AGG, ANT, or DOM scores).
Findings from the first set of logistic regression analyses are sum- marized at the top of Table 2. Because the third model (adding ANT and DOM) failed to reach statistical significance for both recidivism categories, we present results from the second model in Table 2. In the second regression models, AGG scores demonstrated incremental validity over both age at release and the total number of prerelease arrests for predicting both types of recidivism. These findings are consistent with those from Walters (2007) and colleagues (Walters & Duncan, 2005; Walters et al., 2003) in showing that AGG scores demonstrate incremental validity over background characteristics and other PAI scores for predicting offender outcomes.
The second question we attempted to answer using logistic regres- sion was whether PAI scores demonstrate incremental validity over the STATIC-99. The intended use of the STATIC-99 is to predict sexual recidivism or the combination of violent and sexual recidivism (Helmus et al., 2009). Because no PAI measure was a significant predictor of sexually violent recidivism, we focused on the combined category of violent or sexually violent recidivism to examine this issue. The STATIC-99 was a statistically significant predictor of the combination of violent or sexually violent recidivism (see Table 2) but was not a significant predictor when age at release and prerelease arrests were also entered into the model.6 Thus, we examined the incremental validity of PAI scores over the STATIC-99 without offender background characteristics in the model. For this analysis, we entered STATIC-99 scores as a single predictor in the first model and added scores from the two significant univariate predictors of the combination of violent or sexually violent recidivism (AGG and DOM) in the second model. Both AGG and DOM scores demon- strated incremental validity over the STATIC-99 in the second model (see Table 2). Nevertheless, this model did not improve classification accuracy beyond simply predicting that nobody would reoffend. The effects for DOM and AGG were small. For example, when the odds ratio for DOM is recalculated to provide information about 10-point increases, the odds of committing a violent or sexually violent offense are only 1.43 times greater for an offender with a DOM score of 70, compared with an offender with a score of 60.
Conclusion
With more than 1,400 offenders and an average of nearly 5 years of follow-up time, this study represents, to our knowledge, the largest and longest prospective recidivism study of the PAI. In
earlier studies, no more than 326 offenders were examined (Hast- ings et al., 2006), and offenders were followed for about one year. Each of the PAI measures that we identified as conceptually relevant to recidivism was indeed a statistically significant predic- tor of at least one category of recidivism, providing support for the construct validity of scores on these measures. The AGG scale appeared to be the strongest and most consistent predictor of recidivism, both in terms of effect size and incremental validity over age at release and prerelease arrests.
For researchers and test developers, these findings hint at the possibility that self-report personality measures may someday en- hance prediction beyond the simple historical variables on measures like the STATIC-99. But from the perspective of forensic evaluators, the small size of the predictive effects suggests that PAI scores may be of only limited practical value for improving risk assessments at present. From the perspective of evaluators who conduct SVP eval- uations (especially in Texas), findings from this study provide impor- tant information about what they can and cannot say about PAI scores as predictors of recidivism. First, it does not appear that PAI scores are predictive of arrests for offenses that are clearly sexual in nature. Second, although PAI scores are not strong predictors of violent or nonviolent recidivism, they are not entirely irrelevant to questions about risk for these types of offenses among sex offenders (as we have seen attorneys assert in court). Indeed, Cohen’s d effect sizes for PAI scores for these types of recidivism ranged from .17 to .55, similar to those from measures specifically designed to predict recidivism among sex offenders (ds range from �.05 to .49 for the STAITC-99 and �.17 to .41 for the Minnesota Sex Offender Screening Tool— Revised; see Boccaccini et al., 2009). Finally, neither PAI scores nor risk-specific measure scores were strong predictors of recidivism in this sex offender sample, a finding that adds to a growing research literature of relatively weak predictive effects for risk-relevant mea- sures in U.S. sex offender samples (Boccaccini et al., 2009).
Several important aspects of the study design may limit the gen- eralizability of findings from this sample. For example, each of the offenders included in this study had been convicted of more than one sexual offense. Although findings may generalize to other samples of repeat sex offenders, they may generalize less to first-time sex of- fenders or to nonsex offenders. In addition, we excluded from the study 16 committed offenders who may have been at an exceptionally
4 We also used hierarchical logistic regression to examine whether the relation between PAI scores and recidivism depended on Positive Impres- sion (PIM) scores (i.e., defensive responding; see Edens & Ruiz, 2006) and found that they did not. With each model, we examined whether the product of the PAI scale (ANT, AGG, DOM, or VPI) and PIM added incremental validity in prediction over a model including the PAI scale and PIM. Neither PIM nor the interaction term was a statistically significant predictor in any of the models.
5 We also examined predictive effects separately for offenders who were discharged and those released on parole or mandatory supervision because, in a partially overlapping sample of offenders, actuarial measures were more effective for predicting recidivism among discharged offenders than those on parole/mandatory supervision (Boccaccini et al., 2009). Predictive effects for the PAI measures were nearly identical for both subgroups of offenders.
6 See Boccaccini et al. (2009) for detailed analyses of the incremental validity of the STATIC-99 over offender age and other background vari- ables in a partially overlapping sample.
146 BOCCACCINI, MURRIE, HAWES, SIMPLER, AND JOHNSON
T hi
s do
cu m
en t i
s co
py ri
gh te
d by
th e
A m
er ic
an P
sy ch
ol og
ic al
A ss
oc ia
tio n
or o
ne o
f i ts
a lli
ed p
ub lis
he rs
. T
hi s
ar tic
le is
in te
nd ed
s ol
el y
fo r t
he p
er so
na l u
se o
f t he
in di
vi du
al u
se r a
nd is
n ot
to b
e di
ss em
in at
ed b
ro ad
ly .
high risk for reoffending. Excluding committed offenders, who tended to have higher PAI scores, may have led to attenuated predictive effects.7 Finally, we sought, but could not locate, PAI scores or information for 317 offenders. Offenders without PAI scores had, on average, a slightly higher number of prerelease arrests compared with those who had PAI scores, and a disproportionate number of offend- ers identified as Hispanic were missing PAI scores. The impact of these missing data on the predictive validity of the PAI measures is unclear. One clear limitation of using self-report measures to predict recidivism is that offenders can refuse to complete them. Related limitations of self-report measures are that offenders may produce invalid protocols or may be too mentally ill to be tested.
Overall, this study’s findings underscore some of the possibilities and limitations of efforts to assess risk of recidivism among offenders. The significant effects and incremental validity for scores from some PAI measures, particularly as compared with the STATIC-99, under- score that self-report measures may have more to offer in assessing recidivism risk than is typically assumed (see Walters, 2006) and hint that self-report approaches may ultimately supplement actuarial mea- sures of historical variables. On the other hand, the significant effects tended to be small and underscore the observation that— despite recent advances in violence risk assessment—prediction will probably remain far from perfect (Buchanan, 2008).
7 We examined the potential effect of omitting committed offenders by assuming that all 16 committed offenders had been released and committed both a violent and a sexually violent offense. Area under the curve (AUC) values for PAI scores in these analyses were virtually identical to those from those in Table 1, with no effect larger than .63 and none differing from the value in Table 1 by more than five tenths of a point (e.g., AUC of .43 vs. .48).
References
Amenta, A. (2005). The assessment of sexual offenders for civil commit- ment proceedings: An analysis of report content (Unpublished doctoral dissertation). Sam Houston State University, Huntsville, TX.
Boccaccini, M. T., Murrie, C. D., Caperton, J. A., & Hawes, S. (2009). Field validity of the STATIC-99 and MnSOST-R among sex offenders evaluated for civil commitment as sexually violent predators. Psychol- ogy, Public Policy, and Law, 15, 278 –314.
Buchanan, A. (2008). Risk of violence by psychiatric patients: Beyond the “actuarial versus clinical” assessment debate. Psychiatric Services, 59, 184 –190.
Buffington-Vollum, J., Edens, J. F., Johnson, D. W., & Johnson, J. K. (2002). Psychopathy as a predictor of institutional misbehavior among sex offenders: A prospective replication. Criminal Justice and Behavior, 29, 497–511.
Caperton, J. D., Edens, J. F., & Johnson, J. K. (2004). Predicting sex offender institutional adjustment and treatment compliance using the Personality Assessment Inventory. Psychological Assessment, 16, 187– 191.
Edens, J. F. (2009). Interpersonal characteristics of male criminal offend- ers: Personality, psychopathological, and behavioral correlates. Psycho- logical Assessment, 21, 89 –98.
Edens, J. F., Buffington-Vollum, J. K., Colwell, K. W., Johnson, D. W., & Johnson, J. K. (2002). Psychopathy and institutional misbehavior among incarcerated sex offenders: A comparison of the Psychopathy Check- list—Revised and the Personality Assessment Inventory. International Journal of Forensic Mental Health, 1, 49 –58.
Edens, J. F., & Ruiz, M. A. (2006). On the validity of validity scales: The importance of defensive responding in the prediction of institutional misconduct. Psychological Assessment, 18, 220 –224.
Epperson, D. L., Kaul, J. D., & Hesselton, D. (1998). Minnesota Sex Offender Screening Tool—Revised (MnSOST–R): Development, perfor-
Table 2 Summary of Logistic Regression Analyses Examining Incremental Validity of PAI Measures of Offender Background Characteristics and the STATIC-99
Recidivism: Predictor B SE OR Lower Upper �2
Incremental validity of PAI over background characteristics
Violent recidivism Age at release �0.068�� .012 0.94 0.91 0.96 67.31��
Prerelease arrests 0.092�� .024 1.10 1.05 1.15 AGG 0.030�� .010 1.03 1.01 1.05 Constant �2.031�� .698
Sex offender registry violations Age at release �0.050�� .009 0.95 0.94 0.97 91.27��
Prerelease arrests 0.099�� .020 1.10 1.06 1.15 AGG 0.037�� .008 1.04 1.02 1.06 Constant �2.433�� .554
Incremental validity of PAI over STATIC-99
Violent or sexually violent recidivism Model 1
STATIC-99 0.110� .049 1.17 1.01 1.23 4.94�
Constant �2.666�� .184 Model 2
STATIC-99 0.086 .050 1.09 0.99 1.20 24.57��
DOM 0.036�� .011 1.04 1.02 1.06 AGG 0.026�� .009 1.03 1.01 1.05 Constant �5.733 .731
Note. PAI � Personality Assessment Inventory; AGG � Aggression scale from the PAI; DOM � Dominance scale from the PAI; OR � odds ratio. � p � .05. �� p � .01
147SEX OFFENDER RECIDIVISM
T hi
s do
cu m
en t i
s co
py ri
gh te
d by
th e
A m
er ic
an P
sy ch
ol og
ic al
A ss
oc ia
tio n
or o
ne o
f i ts
a lli
ed p
ub lis
he rs
. T
hi s
ar tic
le is
in te
nd ed
s ol
el y
fo r t
he p
er so
na l u
se o
f t he
in di
vi du
al u
se r a
nd is
n ot
to b
e di
ss em
in at
ed b
ro ad
ly .
mance, and recommended risk cut scores. St. Paul, MN: Iowa State University and Minnesota Department of Corrections.
Gookin, K. (2007). Comparison of state laws authorizing involuntary commitment of sexually violent predators: 2006 update, revised. (Doc- ument No. 07– 08 –1101). Olympia, WA: Washington State Institute for Public Policy.
Hanson, R. K., & Morton-Bourgon, K. E. (2009). The accuracy of recid- ivism risk assessments for sexual offenders: A meta-analysis. Psycho- logical Assessment, 21, 1–21.
Hanson, R. K., & Thornton, D. (1999). Static-99: Improving actuarial risk assessments for sex offenders (User Report 99 – 02). Ottawa, Ontario, Canada: Department of the Solicitor General of Canada.
Hare, R. D. (2003). The Hare Psychopathy Checklist Revised (2nd ed.). Toronto, Ontario, Canada: Multi-Health Systems.
Hastings, M. E., Stuewig, J., & Tangney, J. (2006, March). Relative efficacy of the PAI, PCL:SV, and VRAG in predicting institutional misconduct and short-term recidivism. Paper presented at the annual meeting of the American Psychology–Law Society, St. Petersburg, FL.
Helmus, L., Hanson, R. K., & Thornton, D. (2009) The forum: Newsletter for the Association for the Treatment of Sexual Abusers, 21, 38 – 45.
Jackson, R. L., & Hess, D. T. (2007). Evaluation of civil commitment of sex offenders: A survey of experts. Sexual Abuse: A Journal of Research and Treatment, 19, 425– 448.
Miller, H. A., Amenta, A. E., & Conroy, M. A. (2005). Sexually violent predator evaluations: Empirical evidence, strategies for professionals, and research directions. Law and Human Behavior, 29, 29 –54.
Morey, L. C. (1991). Personality Assessment Inventory: Professional man- ual. Odessa, FL: Psychological Assessment Resources.
Morey, L. C. (1996). An interpretive guide to the Personality Assessment Inventory (PAI). Odessa, FL: Psychological Assessment Resources.
Morey, L. C. (2007). Personality Assessment Inventory: Professional man- ual (2nd ed.). Odessa, FL: Psychological Assessment Resources.
Morey, L. C., Warner, M. B., & Hopwood, C. J. (2007). The Personality Assessment Inventory: Issues in legal and forensic settings. In A. M.
Goldstein (Ed.), Forensic psychology: Emerging topics and expanding roles (pp. 97–126). Hoboken, NJ: Wiley.
Mullen, K. L., & Edens, J. F. (2008). A case law survey of the Personality Assessment Inventory: Examining its role in civil and criminal trials. Journal of Personality Assessment, 90, 300 –303.
Rice, M. E., Harris, G. T., Lang, C., & Cormier, C. (2006). Violent sex offenses: How are they best measured from official records? Law and Human Behavior, 30, 525–541.
Salekin, R. T., Rogers, R., Ustad, K. L., & Sewell, K. W. (1998). Psy- chopathy and recidivism among female inmates. Law and Human Be- havior, 22, 109 –128.
Skopp, N. A., Edens, J. F., & Ruiz, M. A. (2007). Risk factors for institutional misconduct among incarcerated women: An examination of the criterion-related validity of the Personality Assessment Inventory. Journal of Personality Assessment, 88, 106 –117.
Walters, G. D. (2006). Risk-appraisal versus self-report in the prediction of criminal justice outcomes: A meta-analysis. Criminal Justice and Be- havior, 33, 279 –304.
Walters, G. D. (2007). Predicting institutional adjustment with the Life- style Criminality Screening Form and the Antisocial Features and Ag- gression scales of the PAI. Journal of Personality Assessment, 88, 99 –105.
Walters, G. D., & Duncan, S. A. (2005). Use of the PCL-R and PAI to predict release outcome in inmates undergoing forensic evaluation. The Journal of Forensic Psychiatry and Psychology, 16, 459 – 476.
Walters, G. D., Duncan, S. A., & Geyer, M. D. (2003). Predicting disci- plinary adjustment in inmates undergoing forensic evaluation: A direct comparison of the PCL-R and the PAI. Journal of Forensic Psychiatry and Psychology, 14, 382–393.
Received July 21, 2009 Revision received September 14, 2009
Accepted September 16, 2009 �
148 BOCCACCINI, MURRIE, HAWES, SIMPLER, AND JOHNSON
T hi
s do
cu m
en t i
s co
py ri
gh te
d by
th e
A m
er ic
an P
sy ch
ol og
ic al
A ss
oc ia
tio n
or o
ne o
f i ts
a lli
ed p
ub lis
he rs
. T
hi s
ar tic
le is
in te
nd ed
s ol
el y
fo r t
he p
er so
na l u
se o
f t he
in di
vi du
al u
se r a
nd is
n ot
to b
e di
ss em
in at
ed b
ro ad
ly .
