Participants and procedure
This study was conducted within the iBerry Study, a prospective cohort study in the Netherlands, designed to investigate the development of adolescent and adult psychiatric disorders [29]. Within a total sample of 16,736 children living in the greater Rotterdam Rijnmond area, self-reported emotional and behavioral problems in the first year of high school were assessed as part of a standard healthcare approach by completing the Strengths and Difficulties Questionnaire-Youth (SDQ-Y) [30]. All adolescents scoring in the top 15% of scores stratified by sex and a random sample of adolescents with the lowest 85% scores were approached to participate in a long-term cohort study. At baseline, 1,022 adolescents and their parent(s) provided written informed consent and visited the research center for in-depth psychiatric interviews and questionnaires. Approximately nine months after these baseline assessments, 679 adolescents (66,4% of 1,022; mean age = 14.77, SD = 0.81) completed a second questionnaire. All variables and covariates used in this study were gathered at baseline, except for the Youth Psychopathic Inventory (which was administered in youngsters approximately 9 months after baseline). Adolescents received a small monetary compensation. Researchers were blind to screening status. The Erasmus MC’s Medical Center's Medical Ethics Review Committee approved the study protocol (MEC-2015–007).
Measures
Multi-informant antisocial behavior
The Dutch version of the Achenbach System of Empirically Based Assessment (ASEBA) questionnaires was used to measure emotional and behavioral problems [31,32,33,34]. The self-report measurement was obtained by the Youth Self Report (YSR). Both parents were asked to complete the analogous Child Behavior Checklist (CBCL) whereas the teachers were asked to complete the Teacher’s Report Form (TRF). These ASEBA-questionnaires use a three-point Likert scale ranging from 0 (“not true”) to 2 (“very true”). For the present study we used the multi-informant scores indicating “Rule Breaking Behavior” (covert) and “Aggressive Behavior” (overt) from both self-report (YSR) and informant-report (CBCL and TRF) measures, following the recommendations in measuring antisocial behavior in youth [35, 36]. Internal consistency (Cronbach’s α) scores on rule breaking behavior scales (0.85, 0.81, and 0.95 on YSR, CBCL, and TRF respectively), and aggressive behavior scales (α = 0.94, 0.86, and 0.95 on YSR, CBCL, and TRF respectively) are good [32]. Composite scores were calculated using all available informants, for which at least 75% of items per scale was completed. Inter-informant assessments were sufficiently correlated (r’s ≥ 0.239, p < 0.01) to combine them into one composite index score per scale.
CU traits
The Callous-Unemotional dimension of the Dutch version of the Youth Psychopathy Traits Inventory—Short Child Version (YPI-SCV, [37]) was used to measure CU traits. This self-report measure consists of 18 items, of which six items relate to the affective scale. The items were measured on a four-point Likert scale ranging from 1 (“does not apply at all”) to 4 (“applies very well”). Baardewijk et al. [37] showed good reliability (Cronbach’s α’s of 0.85 and 0.83) and validity (respectively r = 0.95 and r = 0.93 compared to the longer version and external measures of conduct problems, [38]); in our sample the internal consistency of the affective scale appeared acceptable (α = 0.70).
Anxiety symptoms
We used the Anxious/Depressed scale of the Youth Self Report (YSR) to quantify anxiety symptoms at baseline. The ASEBA-questionnaires demonstrated adequate reliability and validity [32], with mean alpha scores for the Anxious/Depressed scale being 0.84, 0.84, and 0.86 for the YSR, CBCL, and TRF respectively.
Other outcome measures: self-reported offending
In order to pinpoint types of delinquent behavior the Dutch adaptation of the Self-Reported Early Delinquency (SRED) was used to measure self-reported violent offending and property offenses in the past 6 months [39, 40]. This interview consisted of 23 items; internal consistency in our sample was acceptable (α = 0.70). Six items were considered as violent offending based on the Statistics Netherlands classification: joining a fight, hitting someone in public, carrying a weapon, hitting someone resulting in use of medical care, robbery and fighting with a weapon. Ten items concerning property offenses were derived from the SRED, including destroying property in school and publicly, spraying graffiti, fire setting, illegal trespassing, and stealing of products worth more than 100 euros.
Other clinical correlates
Non-verbal IQ-score was assessed using two subtests of a Dutch non-verbal IQ test: Snijders-Oomen Non-verbal Intelligent Test-Revised (SON-R 6–40). Raw test scores on the subsets ‘Analogies’ and ‘Categories’ were converted into estimated IQ-scores using norms tailored to exact age and sex. Testing non-verbal intelligence is insensitive to differences in exposure to Dutch language from early childhood onwards. The standards, internal consistency (α = 0.95), concept validity and criterion validity of the SON-R are good [41].
Physical/Sexual Abuse During interview sessions at the research center current caregivers (not accompanied by the adolescent) reported on the occurrence of lifetime incidents of physical and/or sexual abuse in their children. For this study, we dichotomized these traumatic childhood experiences into ‘any’ vs. ‘none’.
Bullying victimization/perpetration Adolescents reported bullying or being bullied by peers in a questionnaire from a Dutch population-based cohort [42]. Four items concerning bullying victimization and four items concerning bullying perpetration, e.g., by insult, spitting, slapping and social exclusion, were scored on a 5-point Likert scale ranging from 0 (never) to 4 (several times a week). Both scales showed adequate internal consistency: for the 4-item bullying victimization subscale (α = 0.82) and for the bullying perpetration subscale (α = 0.72).
Psychotic experiences The Prodromal Questionnaire (PQ-16) was used to measure self-reported subclinical psychotic symptoms [43]. It consists of 16 items, of which 14 concern positive symptoms and 2 concern negative symptoms to which the adolescent can agree (1) or disagree (0). Higher scores are indicative of more psychotic symptoms. In this sample internal consistency was sufficient (α = 0.77).
ADHD-symptoms We used the DSM-oriented scale of Attention Deficit/Hyperactivity of the CBCL (6–18) [32], which consists of 7 items. In our sample internal reliability was good (α = 0.84).
Substance use The Dutch adaptation of the Self-Reported Early Delinquency (SRED) was used to measure substance use. Two separate items were asked on respectively alcohol and illicit drug use in the past 6 months. Results were dichotomized into no use versus any use.
Internet gaming addiction problems The Video Game Addiction Test (VAT) [44] is derived from the Compulsive Internet use Scale and is a measure for Internet Gaming Disorder. The self-report scale consists of 14 items like loss of behavioral control, interpersonal conflict, preoccupation, gaming as a mood stabilizer and withdrawal symptoms when not playing. Each item can be scored on a 3-point scale (1 = never; 2 = sometimes; 3 = often). In our sample the VAT demonstrated adequate internal reliability (α = 0.87).
Covariates
Gender, age, ethnic origin, household income, and educational level were identified as possible confounders, based on known correlations of antisocial behavior and environmental factors [45, 46]. Background characteristics were determined at enrollment. The adolescent was classified as of non-Dutch origin if one of the parents or the child itself was born abroad. The country of birth of the child, or the country of birth of the mother decided on the categorization into Dutch, other-Western and non-Western background, according to the Dutch standard classification criteria of Statistics Netherlands [47]. Contrary to this classification, the iBerry study considers Japan and Indonesia as non-western countries. Household income was categorized into a total monthly income of < €1599, €1600—€2399, €2400—€4399 and > €4400. Educational level of the adolescent was coded as special needs/pre-vocational education, higher general secondary education, pre-university education, or combined education level.
Statistical analyses
For response analyses, we compared baseline characteristics of adolescents for which information on the YPI questionnaire was available (n = 679) to baseline characteristics of adolescents with missing data on the YPI (n = 343), by using Mann–Whitney U tests (non-normally distributed variables) and Chi-square tests (categorical variables).
Differences in baseline characteristics between boys and girls were explored. Correlations between covariates (gender, age, ethnic origin, household income, and level of education) with all variables of interest for the LPA were performed using bivariate Pearson's or Spearman’s correlation. Variables of interest for the LPA were self-reported CU traits (YPI), self-reported anxiety symptoms (YSR, subscale Anxious/Depressed), multi-informant composite scores of rule-breaking behavior (YSR, CBCL, TRF subscale Rule Breaking Behavior), and the multi-informant composite score of aggressive behavior (YSR, CBCL, TRF subscale Aggressive Behavior). Mplus 8.0 [48] was used to conduct latent profile analysis (LPA, [49]) in order to identify meaningful subgroups of antisocial adolescents with similar patterns of CU traits and anxiety symptoms and to examine the associations between these subgroups and distal outcomes and predictors. LPA is used to identify homogenous subgroups based on participants’ patterns of response to continuous indicators (i.e., CU traits, anxiety symptoms). We specifically conducted an LPA with the manual BCH method, which is the recommended option for models with both predictors and continuous distal outcomes [49, 50]. This method has several advantages over 1-step LPA or 3-step LPA, including not having to re-calculate estimations of LPA when including covariates or distal outcomes while also avoiding shifts in latent class in the final stage that the 3-step method is susceptible to [49]. In the first step of this method a LPA is estimated using only the latent class indicator variables, with the other variables defined as auxiliary, and the BCH weights are saved. In step 2, the model with covariates and distal outcomes is specified, while controlling for the saved BCH weights from step 1. To identify the optimal number of latent classes, the Bayesian information criterion (BIC), the adjusted Lo-Mendell-Rubin likelihood ratio test (LMR-LRT), sufficient class counts to perform secondary analyses with adolescent gender, and theoretical interpretation were employed. Previous studies showed that LMR-LRT and BIC are good indicators to select the optimal number of classes [51, 52]. A smaller BIC indicates a better fit of the model, while p-values below 0.05 on the LMR-LRT indicate that the k-class model is statistically better than the k-1 class model. Except for indicators of the latent classes (i.e., CU traits and anxiety symptoms), the following other variables were entered as auxiliary variables in the model: gender (covariate), multi-informant aggressive behavior, and multi-informant rule-breaking behavior (distal outcomes) [49]. Logistic regression, (i.e., predicting class membership from covariates and outcome variables) was used in step 2 to establish between-group differences in gender, aggressive behavior and rule breaking behavior. For further class characterization with other correlates of CU variants, classes were exported to IBM SPSS Statistics Version 26. This was done because modeling all outcomes at once in the LPA would yield a too complex model for the BCH method [53]. All classes were compared on distal outcomes by using a series of one-way ANCOVA with ethnic origin and household income as dummy-coded categorical covariates and age as continuous covariate. For dichotomous outcome measures, binary logistic regression was used. All statistical tests were considered significant at the p < 0.05 level, post hoc comparisons for ANOVA and ANCOVA procedures were Bonferroni corrected.
Missing data resulted from declined interviews or unreturned questionnaires. For baseline demographics missings ranged from 0% to 6.1%, for all between-groups analyses on found classes in LPA missings ranged from 0% to 4.1%. The study sample was selected on those with complete data on the YSR anxious/depressed subscale and YPI scores, as these variables were necessary indicators to determine the CU variants. For additional analyses incomplete data on item level were handled by mean-value substitution. When > 25% of items were missing on a scale, cases were excluded pairwise from the analysis.