In predicting which antisocial youth will be more persistent and severe in their antisocial behavior, scholars have identified CU traits as one of the key distinguishing factors; youth with CU traits are more likely to commit serious offenses and persist in their offending behavior into adulthood (Frick & Marsee, 2006; Frick et al., 2014; Frick et al., 2005; Frick & White, 2008; Kahn, Byrd, & Pardini, 2013; Lawing, Frick, & Cruise, 2010; McMahon, Witkiewitz, Kotler, & The Conduct Problems Prevention Research Group, 2010). CU traits are often conceptualized or characterized as a “downward” extension of adult psychopathy applied to youth, particularly the affective dimension of psychopathic features (Barry et al., 2000; Frick, 1998; Roose, Bijttebier, Decoene, Claes, & Frick, 2010); therefore, it is not surprising that CU traits in adolescence predict psychopathy in adulthood (Burke, Loeber, & Lahey, 2007).

Indeed, the affective psychopathy dimension (i.e., CU features), more so than the interpersonal, impulsive, or lifestyle dimensions, has consistently predicted which offenders are more severe and persistent (Caputo, Frick, & Brodsky, 1999). Although prevalence rates will differ by population, one multisite study estimated that 2% to 32% of community youth and 14% to 50% of clinic-referred youth meet the criteria for CU traits, depending on whether or not they are diagnosed with conduct disorder and who the informant is (Kahn, Frick, Youngstrom, Findling, & Youngstrom, 2012).

In addition to helping define a subgroup of serious offenders, CU traits are important for understanding which youth might be more resistant to intervention. Previous studies have demonstrated that parenting style is unrelated to conduct problems among youth high in CU traits (Edens, Skopp, & Cahill, 2008; Hawes & Dadds, 2005; Hipwell et al., 2007; Oxford, Cavell, & Hughes, 2003; Wootton, Frick, Shelton, & Silverthorn, 1997), and CU traits are typically associated with poorer treatment outcomes (Frick & Dickens, 2006; Waschbusch, Carrey, Willoughby, King, & Andrade, 2007), more negative behaviors in treatment (Haas et al., 2011), and punishment insensitivity (Barry et al., 2000; Fisher & Blair, 1998; Frick et al., 2003; O’Brien & Frick, 1996). Theorizing on why CU traits may be relatively stable in regards to environmental factors has pointed to the increased heritability of conduct problems among youth with CU traits (Viding, Blair, Moffitt, & Plomin, 2005; Viding, Frick, & Plomin, 2007; Viding, Jones, Frick, Moffitt, & Plomin, 2008).

CU traits also are associated with a number of emotional and cognitive deficits in youth. For example, youth high in CU traits are less concerned about their problem behavior; in fact, they are more likely to expect positive outcomes of their aggression and delinquency (e.g., peer dominance; Pardini & Byrd, 2012; Pardini, Lochman, & Frick, 2003). They also are less accurate in recognizing emotions in facial expressions (Dadds, El Masry, Wimalaweera, & Guastella, 2008; Munoz, 2009; Woodworth & Waschbusch, 2007), are less likely to make eye contact (Dadds et al., 2008) and have decreased amygdala response to distressing stimuli (e.g., fearful faces; Jones, Laurens, Herba, Barker, & Viding, 2009; Marsh et al., 2008). Given that youth with CU traits exhibit greater problem behaviors and attitudes and decreased response to intervention, it is important for researchers, clinicians, and justice officials to be able to determine efficiently which youth are displaying clinically significant levels of these traits.

Assessments of CU traits typically involve clinician rating, self-report, parent report, and/or teacher report (Forth, Kosson, & Hare, 2003; Frick, 2003; Frick & Hare, 2001; Kimonis et al., 2008). These assessments can measure the broader construct of youth psychopathy or specifically measure the affective features of psychopathy, or CU traits. Common measures of psychopathy for children and adolescents include the Psychopathy Checklist-Youth Version (PCL-YV; Forth et al., 2003), the Antisocial Process Screening Device (APSD; Frick & Hare, 2001), the Youth Psychopathic Traits Inventory (Andershed, Kerr, Stattin, & Levander, 2002), and the Child Psychopathy Scale (Lynam, 1997).

All of these measures are self-report, except for the PCL-YV, which combines an interview with a file review; the APSD has caregiver and teacher report versions. Although each of these measures is relatively reliable and valid when measuring youth psychopathy (Falkenbach, Poythress, & Heide, 2003; Kosson, Cyterski, Steuerwald, Neumann, & Walker-Matthews, 2002; Munoz & Frick, 2007), there are often only a handful of items within each measure that assess CU traits. For this reason, a measure specifically designed to measure CU traits may be preferred when studying or making decisions based on the affective features of youth psychopathy.

One measure of CU traits specifically is the Inventory of Callous-Unemotional Traits (ICU; Frick, 2003). The ICU was developed based on the original six CU items of the APSD and expands on those items in more detail. It is available in self-report and parent and teacher report versions. Although the entire measure is 24 items, a previous study found 22 of the items to be valid and reliable (Kimonis et al., 2008). The ICU has yielded good internal (Byrd, Kahn, & Pardini, 2013; Essau, Sasagawa, & Frick, 2006; Kimonis et al., 2008), test-retest (Ezpeleta, de la Osa, Granero, Penelo, & Domènech, 2013; Feilhauer, Cima, & Arntz, 2012), and inter-rater reliability (Berg et al., 2013; White, Cruise, & Frick, 2009); it also has been found to have good construct, content, and criterion validity (Byrd et al., 2013; Essau et al., 2006; Kahn et al., 2013; Kimonis et al., 2008). For example, for construct validity, the ICU positively correlates with self-reported aggression and delinquency (Essau et al., 2006; Kimonis et al., 2008) and negatively correlates with empathy and emotional reactivity (Kimonis et al., 2008). For criterion validity, the ICU has demonstrated both concurrent and predictive validity because it is significantly correlated with both concurrent charges and arrests (Byrd et al., 2013) as well as later charges and arrests (Kahn et al., 2013).

The ICU has been used with a variety of samples, including both detained (Kimonis, Cross, Howard, & Donoghue, 2013; Munoz, Frick, Kimonis, & Aucoin, 2008; Pechorro, Ray, Barroso, Maroco, & Goncalves, 2014) and community samples (Byrd et al., 2013; Ezpeleta et al., 2013; Kimonis, Branch, Hagman, Graham, & Miller, 2013; Roose et al., 2010), as well as in samples with children as young as 3 years (Ezpeleta et al., 2013) to young adults as old as 25 (Byrd et al., 2013). While it does have three subscales (Uncaring, Callousness, and Unemotional) and an acceptable factor structure, the total ICU score consistently has been found to be more reliable than the subscale scores (Kimonis et al., 2008).

Previous studies have not yet examined the criterion validity of the ICU for distinguishing adjudicated and community youth because prior studies principally have administered the ICU to a distinct sample (e.g., adjudicated, clinic-referred) of youth. Therefore, one important step forward in establishing the ICU’s criterion validity would be to administer the ICU to two different samples of youth (e.g., adjudicated and community) to determine whether the ICU can reliably discriminate between the two and examine whether a particular cutoff score might be useful in predicting which youth are at greater risk of being detained. A further issue that warrants study is how best to incorporate information from multiple sources; previous studies using the ICU or APSD often use the maximum report across informants (Berg et al., 2013; Bijttebier & Decoene, 2009; Frick, Bodin, & Barry, 2000; Jones et al., 2009; Roose et al., 2010; Viding et al., 2012; White et al., 2009), but further investigation can help determine whether this is the best method across different situations and for various purposes.

CU traits significantly predict aggressive, violent, and delinquent behavior, and they are exhibited by a subgroup of antisocial youth with more severe and stable patterns of problem behavior; further, they can be assessed reliably using the ICU. However, no study has yet examined an appropriate cutoff score for the ICU that would aid in empirical, judicial, or clinical decision making regarding whether youth are exhibiting meaningful levels of CU traits, and there is disagreement about how information from multiple informants should best be used. The current study uses a mixed sample of community and detained adolescents to address two research questions: (a) Can CU traits (as measured by the ICU) reliably distinguish concurrent detained status? And (b) can CU traits predict concurrent aggressive and violent behavior as well as the number of charges and adjudicated offenses and offense seriousness and violence levels among detained youth? We also have two aims in this study: (a) to determine a suitable cutoff score for the ICU based on its criterion validity for predicting whether youth are in the community or detained and (b) how best to use and integrate information from youth, parents, and teachers.

Data for this study were drawn from the database of a larger project (see Boxer, Huesmann, Bushman, O’Brien, & Moceri, 2009) examining risk factors for violent and nonviolent antisocial behavior among high school students (n = 430) and incarcerated youth (n = 390). In addition to interviews with the youth, information was solicited from parents and guardians as well as teachers and staff. Because the focus of this analysis is on the cross-informant reliability and validity of a rating scale, participants for the current analysis were the 634 youth from the pool of 820 (77.3%), with reports on the target measure (ICU) available from all three sources. In terms of missing data by source, ICU self-reports were missing for four (0.49%) youth, parent and guardian reports were missing for 93 (11.34%) youth, and teacher and staff reports were missing for 106 (12.93%) youth; 17 (2.07%) youth were missing data from two different sources, typically from parent and guardian and teachers and staff (15 youth, 1.83%).

Each informant report on the ICU did not vary as a function of missing data on the other reports; that is, those with missing data on one report did not have significantly higher or lower ICU scores on the other reports. There were no effects of youth sex on the likelihood of reports to be missing from any of the three sources. However, White youth were more likely than were non-White youth to have data available from parents and guardians, χ²(1, 820) = 4.57, p < .05, and from teachers and staff, χ²(1, 820) = 13.05, p < .001. Further, all missing self-reports, χ²(1, 820) = 4.43, p < .05, and more missing parent and guardian reports, χ²(1, 820) = 5.64, p < .05, emanated from the adjudicated sample.

Of the 634 youth retained for analysis (mean [M] age = 16.18 years, standard deviation [SD] = 1.31), 343 (54%) were students and 291 (46%) were detained youth; overall the full sample included 376 males (59%) and 258 females (41%) and a majority of White youth (58% White, 29% Black/African American, 3% Hispanic/Latino/a, 8% multiracial, 2% other). By sample, the student group (M = 16.83, SD = .72) was significantly older (p < .001) than was the detained group (M = 15.40, SD = 1.42). Females were significantly under-represented (p < .001) in the detained group (28%) relative to the student group (51%); non-Whites were represented about equally across the two groups (39% students, 46% detained youth; p = .08).

All procedures were approved by the university’s institutional review board (IRB) regulating the implementation of the study, the state agency overseeing the state detention facilities, the federal Office of Human Research Protections, the IRB of the Centers for Disease Control, and the directors or principals of all schools and detention facilities involved in the project. Data collection occurred during 2005 through 2007. Youth were recruited from public high schools (rural, suburban, and urban) and juvenile detention centers (county and state) selected to yield a sample representing a range of risk for aggressive and violent behavior. Across all sites, parent and guardian consent rates averaged about 40% (range by site = 33.6%–48.8%), unsurprising given the length of the survey batteries and nature of the populations sampled. With only slight variations within site types (high school or detention facility), recruitment and interview procedures were conducted differently between the students and detained youth.

In high schools, informational letters and parental consent forms were mailed with stamped return envelopes to parents and guardians of 11th- and 12th-grade students; 2 weeks after the initial mailings, second mailings were sent to parents and guardians who had not responded by that time. Remaining parents and guardians who did not respond by mail to the second contact attempt were solicited by telephone. Parents and guardians could grant permission for their children to participate in writing (mailed) or over the telephone (recorded).

After their children were interviewed, parents, guardians, and youths’ teachers (usually social studies teachers) were given survey booklets to complete. Parents and guardians had the option of completing surveys over the telephone, as we have done previously in field research (Boxer et al., 2009). Teachers completed surveys by paper and pencil. Youth interviews were conducted via paper-and-pencil Scantron survey forms in small groups ranging typically from about 10 to 15 students depending upon availability and were led by at least two trained staff for every 10 to 15 students.

In detention facilities, at the start of data collection, the facilities sent informational letters and consent forms typically to parents and guardians of all youth housed, as well as to the parents and guardians of any new admissions to the facilities over the period of data collection. In these facilities, we were permitted to make the follow-up telephone calls without first sending a second mailing. As with the students, after a detained youth completed his or her interview, we mailed a survey to parents and guardians and provided a survey to staff. Again, parents and guardians had the option to complete their surveys over the telephone, and staff completed surveys by paper and pencil. Trained staff via laptop computers individually conducted youth interviews. Most youth interviews with students and detained youth took approximately one hour.

Across data collection sites, all individuals who provided data were compensated financially in some manner, primarily gift certificates to local merchants, with variations from site to site due to agency regulations or extraneous factors. All high school students received $20, except those in a school collaborating with our research team on another investigation necessitating compensation of $40. All detained youth received $10 compensation due to agency restrictions. Parents and guardians of high school students received $25; parents and guardians of detained youth received $50. All teachers and staff received $5 per completed survey, although this was distributed differently by site due to school or agency policies (e.g., teachers typically received cash but staff had their compensation put into a common fund for staffwide rewards such as appreciation lunches).

Table 2 shows descriptive statistics for study variables, separately by sample. We computed a series of two-way analyses of variance (ANOVAs) to examine differences by sex and sample (students vs. detained youth) on ICU scores from youths, parents and guardians, and teachers and staff. These analyses showed generally that detained youth received significantly higher scores than did students, across all informants (all ps < .001; partial η² estimates = .07 [youth], .23 [parents and guardians], .12 [teachers and staff]). For youth and teachers and staff, these main effects were qualified somewhat by modest sex by sample interactions; high school students were rated lower than detained youth and females were rated lower than males, with a larger gender difference for students than detained youth (both ps < .05; both partial η² estimates = .01). Exploratory t-test analyses indicated no difference in youth and parent and guardian ICU scores as the function of race (White vs. non-White); teachers and staff rated non-White youth higher on the ICU compared to White youth (p < .001). Youth age was modestly to moderately negatively correlated with ICU scores (r values −.17 to −.34, p < .001).

Paired samples t-tests indicated that youth underreported ICU scores compared to parents (p < .001) and teachers (p < .001), but parents and teachers reported relatively similar mean ICU scores (p = .116). Separate paired t tests by sample indicated that informer agreement differed by sample; although youth and parents agreed more for the high school sample (p = .649), youth in the adjudicated sample underreported symptoms compared to their parents (p < .001). Youth underreported symptoms compared to teachers in both samples (p < .001). For the high school sample, teachers reported more symptoms than parents (p < .001), but the reverse was true for the adjudicated sample (p = .024).

Table 3 shows the bivariate correlations among the manifest behavioral criterion measures and the ICU ratings, across informants, separately by sample. Cross-informant correlations on the ICU were modest in magnitude for both samples and slightly smaller for the sample of detained youth. Intra-rater correlations generally are higher than are cross-informant correlations. Still, one typically does not expect great consistency across informants for youth behavior ratings, necessitating the aggregation of cross-informant reports to summarize adequately a target youth’s behavioral status. For subsequent analyses, we therefore relied on our cross-informant latent composite scores indicating violence and general aggression.

We used a set of logistic regression models to predict detained status, or whether or not youth were detained in a facility, from ICU scores and composites. For each model, age, sex, and race (White/non-White) are included as covariates, and one of the ICU scores (youth, parent, teacher) or composites (mean and max composite of all scores, mean and max composite of parent and teacher scores) is included as a predictor, generating logistic regression models. An eighth model also is included, in which all three ICU scores are entered as predictors. Because non-nested models are being compared, Bayesian information criterion (BIC) and Akaike’s information criterion (AIC) scores are reported along with Wald chi-square tests. Cluster-robust standard errors were computed with data collection site as the cluster variable. Because of two cases that were missing data on race/ethnicity, the number of cases drops from 634 to 632 for these models. The results of these logistic regression models are shown in Table 4.

All of the logistic regression models significantly predicted detained status: the model with youth report, Wald χ²(4) = 100.91, p < .001, BIC = 614.91, AIC = 592.67, McFadden’s R² = .33, Tjur’s D = .40; the model with parent report, Wald χ²(4) = 145.28, p < .001, BIC = 540.05, AIC = 517.81, McFadden’s R² = .42, D = .49; the model with teacher report, Wald χ²(4) = 46.56, p < .001, BIC = 605.50, AIC = 583.26, McFadden’s R² = .34, D = .41; the model with all three reports as separate predictors, Wald χ²(6) = 444.91, p < .001, BIC = 536.13, AIC = 504.99, McFadden’s R² = .44, D = .51; the model with the mean composite of all three reports, Wald χ²(4) = 176.01, p < .001, BIC = 535.99, AIC = 513.75, McFadden’s R² = .42, D = .50; the model with the max composite of all three reports, Wald χ²(4) = 180.19, p < .001, BIC = 541.97, AIC = 519.72, McFadden’s R² = .41, D = .49; the model with the mean composite of parent and teacher reports, Wald χ²(4) = 88.35, p < .001, BIC = 532.07, AIC = 509.83, McFadden’s R² = .43, D = .50; and the model with the max composite of parent and teacher reports, Wald χ²(4) = 82.87, p < .001, BIC = 544.36, AIC = 522.11, McFadden’s R² = .41, D = .48.

Because these models are not nested, BIC and AIC can be used to compare the relative fit of these models, with smaller values representing better fit and parsimony for the model. A difference in BIC of less than 2 provides weak evidence that the model with the smaller BIC is a better fit, while a difference of 2 to 6 provides positive evidence, a difference of 6 to 10 is strong evidence, and a difference greater than 10 provides very strong evidence (Raftery, 1995). Using these criteria, the best-fitting model according to BIC appears to be the mean composite of parent and teacher reports because it has the smallest BIC value by 3.93, which provides positive support. However, AIC indicates that the model with all three reports as separate predictors performs the best, with a difference from the next best model of 4.84, which again provides support for this model as the best-fitting model.

Further, Akaike weights, calculated by taking the difference of each model’s AIC and the minimum AIC, indicate that the probability for the model with all three reports as separate predictors has a probability of 90.59, while the other models have a probability of 8.05 or less. The fact that the AIC chose the model with all three reports and the BIC did not is not surprising because this model appeared to perform relatively well compared to the others, and BIC has a greater penalty for model complexity (e.g., more regressors in the model).

To further examine which ICU scores and composites provide a greater advantage in predicting detained status, we conducted receiver operating characteristic (ROC) analyses that plot sensitivity against the inverse specificity (1 = specificity), and we calculated the area under the curve (AUC) for each ROC plot. A significantly greater AUC suggests a measure with a better trade-off between specificity, or the ability to identify positive cases, and sensitivity, or the ability to not identify negative cases. The AUCs are included in Table 4, and Figure 1 shows the ROC curves plotted from each model, except for the covariates-only model (not shown).

The covariates-only model (sex, age, and race) produced an ROC curve (not shown in Figure 1) with an AUC of .85, which was significantly lower than the AUCs of all subsequent models (p < 0.05). The remaining models all included these covariates, as well as additional regressors. The next model included the self-report ICU, and produced an ROC curve with an AUC of .86, meaning that a randomly chosen adjudicated participant has an 86% probability of having a higher self-reported ICU score than a randomly chosen high school participant. The remaining AUCs are as follows: for parent report ICU, .90; for teacher report ICU, .86; for all three (youth, parent, and teacher) reports, .90; for the mean composite, .90; and for the max composite, .90.

Significance tests indicated that the self-reports and teacher reports were similar and significantly lower than all other models (besides the covariates-only model; p < 0.05), while the parent, max, and mean composites were all similar and significantly higher than the self and teacher reports (p < 0.05). The model that included all three reports was significantly higher than all other models (p < 0.05), except for the max composite model, from which it was not significantly different. Thus, while BIC values suggest that the model with the mean composite provides the best balance between fit and simplicity, and the AIC values suggest that the model with all three reports provide the most information, the AUC values suggest that the model with the max composite and the model with all three reports both strike the best balance between sensitivity and specificity.

To answer our second research question, we used regression models to determine whether CU traits were significantly associated with measures of aggressive and violent behavior, counts of charges and adjudicated offenses, and seriousness and violence level of crimes, among a sample of detained youth. For each of the four dependent variables, we ran a series of eight regression models with the same set of predictors as in the logistic regressions. The results of these series of models are presented in Table 4 and described below. The number of possible observations for these models drops to 289; out of the 291 detained youth with full ICU information across all informants, two were missing data on race/ethnicity. Again, calculated standard errors are robust in regards to the data collection site. For these models, raw coefficients are reported because standardized coefficients cannot be computed for the count models; however, it should still be easy to compare coefficients across models because all ICU scores and composites are measured on the same scale. The frequencies of charges and adjudicated offenses are reported in Table 5, and the coefficients, significance levels, and BIC values for all models are reported in Table 6.

Because aggressive behavior, violent behavior, and seriousness and violence level of charges were all continuous, normally distributed variables, they are modeled using ordinary least squares regression. For the number of charges and adjudicated offenses, we used negative binomial regression models to account for the overdispersion in the distributions. A zero-truncated negative binomial model is used to predict charges because each of the adjudicated youth necessarily had to have a charge, but this was not the case for adjudicated offenses, for which a regular negative binomial model is used.

In all eight models predicting aggressive behavior (n = 289), each of the CU predictors was significant at the p < .05 level. According to the BIC values, the three equally best models are the one with all three reports as separate predictors and the mean and max composites of all three reports. The BIC values cannot distinguish among these three because the difference between any two of them is less than two; however, they are preferred to the other models because the next best model has a larger BIC by at least 27.52. According to the Akaike weights, the model with the max composite of all three models is most likely, with a probability of .51, followed by the model with all three reports at .29 and the mean composite model at .20. In all eight models predicting violent behavior (n = 289), each of the CU predictors was significant at the p < .01 level. According to both BIC and AIC, the best model is the max composite of all three reports. The max composite has the lowest BIC by at least 13.50, and has an Akaike weight probability that rounds to 1.

In the models predicting charges (n = 245), none of the individual ICU scores was significant at the p < .05 level; however, all four ICU composites were significant at the p < .001 level. The BIC values could differentiate only the max composite of all three reports as performing worse than the other composites, while Akaike weights indicated that the two mean composites were equally probable and about three times more probable than either of the max composites. In the models predicting adjudicated offenses (n = 245), the only individual ICU score to be significant at the p < .05 level was the parent report; however, just as in the models predicting charges, all four composites were significant at the p < .001 level.

When considering five models (the model with parent report and the four models with composites as predictors), BIC values indicated that the parent report performed worse than most composites, but the composites performed relatively similarly, and Akaike weights indicated that the max composite of parent and teacher reports was most probable, with a probability of .31, although the mean composites were close behind at .26 and .25 for all three reports and parent and teacher reports, respectively. When predicting crime seriousness and violence (n = 245), none of the CU predictors across all eight models was significant at the p < .05 level.

Finally, two different methods were used to determine the optimal cutoff scores for each measure of ICU, and the results of both are displayed in Table 7. The first method uses predicted probabilities of detained status from each of the models; the predicted probabilities from each model are plotted in Figure 2. Predicted probabilities below .5 indicate a prediction of nondetained status, while predicted probabilities above .5 indicate a prediction of detained status. Therefore, the point at which each measure’s marginal effect exceeds .5 can be used as the optimal cutoff score for that measure, above which the likelihood of detained status increases.

Based on these measures, the optimal cutoff scores are as follows: for youth report ICU, 28; for parent report ICU, 30; for teacher report ICU, 33; for the model with all three reports, youth is 40, parent is 30, and teacher is 36; for mean Y/P/T composite, 26.97; for max Y/P/T composite, 42; for mean P/T composite, 28.84; and for max P/T composite, 39. Scores are reported to two decimal places for the mean composites because they are the only scores that have noninteger values. If a more stringent cutoff score is desired to reduce the false positive rate, then the probability cutoff can be increased. Therefore, we also have calculated the cutoff scores for a predicted probability of detained status of .75 as follows: for youth report ICU, 58; for parent report ICU, 48; for teacher report ICU, 61; for the model with all three reports, there is no score available for youth and teacher report because these reports did not produce predicted probabilities beyond .57 and .66, respectively, while the cutoff for parent report ICU is 51; for mean Y/P/T composite, 36.88; for max Y/P/T composite, 56; for mean P/T composite, 40.64; and for max P/T composite, 54.

The other method used to determine appropriate cutoff scores for the ICU was Youden’s index, or the sum of each cutoff score’s sensitivity and specificity minus one, in which we identified cutoff scores that maximized Youden’s index (Youden, 1950). Based on this index, the optimal cutoff scores were as follows: for youth report ICU, 37 (Youden’s J = .60); for parent report ICU, 41 (J = .66); for teacher report ICU, 21 (J = .61); for the model with all three scores, the cutoffs are 11 for youth report, 21 for parent report, and 25 for teacher report (J = .69); for mean Y/P/T composite, 26.00 (J = .68); for max Y/P/T composite, 46 (J = .66); for mean P/T composite, 29.50 (J = .68); and for max P/T composite, 44 (J = .65).

We then classified youth based on these cutoffs and compared AUC values to determine whether there was a difference in using the score from the predicted probabilities or from the Youden’s index for each report. Using the cutoff scores generated from the predicted probabilities resulted in more stable and valid results, as indicated by generally greater AUC values, particularly for youth and parent report, the two max composites, and the P/T mean composite. When comparing different scores and composites for the probability cutoffs, parent report outperformed both youth and teacher report, and both mean composites outperformed both max composites. Parent report and mean composites performed relatively similarly.

Results from this study offer some insight about the relative contributions of different informants to the assessment of CU traits and how to best integrate this information from multiple sources, but it is not without limitations. First, the data are cross-sectional in nature; therefore, we can provide evidence of only concurrent, not temporal, predictive validity. It is possible that administering the ICU to youth, parents and guardians, and teachers and staff members before the youth were incarcerated would have led to weaker predictions, presumably because the knowledge of a youth as a “delinquent” would shape perceptions of that youth and interpretations of their behavior, leading to inflated ICU scores for the detained youth and an overstated relationship between the two variables.

The cross-sectional design also limits inferences of causality. For example, is it the case that youth high in CU traits are more aggressive and antisocial and therefore end up in detention facilities at higher rates? Or is it instead the case that youth who have been detained in a facility develop higher CU traits as a response to their environment, compared to youth in the community? Prospective designs following youth over time would lend better insight into this critical issue. Low response rate for the surveys also was an issue and could potentially bias some of the results. Finally, although the ICU captures the four different types of symptoms included in the “with limited prosocial emotions” specifier in the DSM-5, we did not attempt to look at each type of symptom specifically to ensure that youth met two or more of the criteria. Future studies would benefit from doing this type of in-depth analysis to aid in approximating the diagnostic criteria, or it could alternatively employ traditional diagnostic assessments and examine their association with “real-world” criteria, such as adjudication status or number or type of offenses.