The structure of psychopathology in adolescence and its common personality and cognitive correlates.

Substance misuse

Substance misuse was assessed using the Alcohol Use Disorders Identification Test (AUDIT; Saunders, Aasland, Babor, de la Fuente, & Grant, 1993 ) and the European School Survey Project on Alcohol and Drugs (ESPAD; Hibell et al., 1997 ). The AUDIT was developed and validated by the World Health Organization to provide a brief assessment of alcohol use disorders and was specifically designed for international use. It exists in all three languages, and has been validated on primary health care patients and community samples. For this study, the scale total for problematic or harmful alcohol use in the last year included feelings of guilt or remorse after drinking, being unable to remember what happened the night before because of drinking, being injured or having injured someone as a result of drinking and relevant others being concerned about their drinking and suggestions to cut down. The ESPAD items used in this study comprised tobacco use frequency and the number of drugs used over the last 12 months (see Table 1 for prevalence and correlations between all psychopathology measures at 16 years).

Personality

Personality was assessed with the self-reported Substance Use Risk Profile Scale (SURPS; Woicik, Stewart, Pihl, & Conrod, 2009 ), the NEO Five Factor Inventory (NEO-FFI; Costa Jr & McCrae, 1992 ), and the Temperament and Character Inventory (TCI; Cloninger, Przybeck, Svrakic, & Wetzel, 1994 ). The SURPS assessed the personality traits of hopelessness, anxiety sensitivity, impulsivity, and sensation-seeking. The reliability and concurrent and predictive validity of this measure has been well established in several adolescent and adult samples in different countries ( Castellanos-Ryan et al., 2013 ; Krank et al., 2011 ; Woicik et al., 2009 ). The NEO-FFI assessed five higher order personality characteristics: neuroticism, conscientiousness, extraversion, agreeableness, and openness to experience. The TCI was used to assess novelty-seeking, which is considered a good general measure of impulsive tendencies that also includes sensation-seeking (see Table S2 in supplementary material for correlations between personality and cognitive measures).

IQ and cognitive measures

Estimates of intelligence were derived from the vocabulary and similarities subtests (verbal IQ) and block design and matrix reasoning subtests (performance IQ) of the Wechsler Intelligence Scale for Children—4th edition (WISC-IV; Wechsler, 2003 ). Digit span forward and backward subtests from the WISC-IV were also administered and used to assess short-term auditory memory and auditory working memory ( Groth-Marnat & Baker, 2003 ; Reynolds, 1997 ). Poor response inhibition was measured using the number of commission errors in a go/no-go passive avoidance learning paradigm ( Newman & Wallace, 1993 ). Delay discounting was assessed with the Kirby Delay Discounting Questionnaire ( Kirby, Petry, & Bickel, 1999 ). This measure was scored as described previously by Kirby, Petry, and Bickel (1999) , with k values (an index of delay discounting) assigned according to choice patterns across the 27 items.

Spatial working memory, risky decision-making, and information processing biases for positive and negative stimuli were assessed with three tasks from the Cambridge Cognition Neuropsychological Test Automated Battery (CANTAB; Cambridge Cognition), the Spatial Working Memory (SWM), Cambridge Gambling task, and affective go/no-go task, respectively. SWM is a self-ordered test that requires retention and manipulation of visuospatial information. A modified version of the Cambridge Gambling Task (which reduced the time between stakes from 5 s to 2 s to make the task shorter to avoid boredom effects in adolescents) was used to assess risky decision-making. Finally, the affective go/no-go is a task of emotional processing, in which affectively valenced words (happy and sad) are presented one at a time on screen. Performance variables-of-interest are the target (omission) errors to positive and negative words, with an attentional bias toward negative versus positive words being assessed with a difference score between omission errors to each set of stimuli. For further details on the cognitive tasks see Supplementary Material and http://www.cambridgecognition.com/academic/cantabsuite .

Data Analysis

A series of structural equation models on computer-generated scaled likelihood of diagnosis scores and self-reported substance use were analyzed using MPlus version 7.11 ( Muthén & Muthén, 2010 ). Based on previously reported theoretical models and analyses (e.g., Carragher et al., 2016 ; Caspi et al., 2014 ; Castellanos-Ryan et al., 2014 ; Lahey et al., 2012 ; Lahey et al., 2015 ; Tackett et al., 2013 ), several models were assessed for goodness of fit: (a) a single “psychopathology” factor loading on all indicators; (b) two correlated factor models, where variables assessing internalizing (INT) psychopathology (generalized anxiety, depression, social phobia, panic and other phobias, OCD and eating disorders) 1 and CD, ODD, ADHD, and substance use (SU) loaded on two specific INT and EXT factors (with SU variables loading on the EXT factor; Model 2a) or loaded on three specific INT, externalizing (EXT) and SU factors (Model 2b). Model 2a and 2b allowed subfactors to covary. Lastly (c), two bifactor (or general-specific) models were assessed, in which a general psychopathology factor (P) was added at the same level as the specific factors from the previous (Models 2a and 2b) models (Models 3a and 3b). In these last models, factors were not allowed to covary (i.e., they were constrained to zero), consistent with a classic bifactor model, but because many recent studies present modified versions of bifactor models (e.g., Carragher et al., 2016 ; Caspi et al., 2014 ), in which the specific factors are allowed to covary, two final revised bifactors models (Models 3a′ and 3b′) that allowed the specific factors to correlate were also examined (see Figures S1–S7 in supplementary material for a graphic representation of all models tested). In all models, gender and language (English vs. other) were entered as covariates (at the level of observed variables). In addition, all models were fit using a complex random effects design to control for testing site as a cluster variable, and used robust maximum likelihood estimation (MLR). MLR has been shown to perform well when modeling low prevalent behaviors or nonnormal data ( Asparouhov & Muthén, 2005 ). Full information maximum likelihood was used to handle missing data.

Once the best fitting model was established, two sets of correlates (personality and cognitive indices) were each entered into the model separately. That is, unadjusted associations were examined by entering the personality and cognitive variables and the psychopathology factors into the same model and allowing them to correlate. Adjusted associations were examined in four separate models in which the psychopathology factors were regressed onto (a) all SURPS subscales; (b) all NEO subscales; (c) novelty seeking (on its own); and (d) all cognitive variables entered together in the same model. The Benjamini-Hochberg procedure ( Benjamini & Hochberg, 1995 ) was used to correct for multiple testing. Once p values are sorted in ascending values, the Benjamini-Hochberg procedure allows one to calculate the false discovery rate (FDR) for each of the p values (i.e., at each “position” in the sorted list of p values, it will indicate what proportion of those are likely to be false rejections of the null hypothesis). This procedure to control for multiple testing has been shown to be less stringent and have more power than Bonferroni correction or other types of familywise error rate corrections (see online Supplementary Material for further description of the sample, measures and analytic approach).

How Stable is This Structure From Early (14 Years) to Middle (16 Years) Adolescence?

At 14 years the two specific factor bifactor model of psychopathology (Model 3a) also fit the data well, χ 2 (42) = 53.30, CFI = 1.00, RMSEA = .011; SRMR = .016; BIC = 61037.87; Adj BIC = 60793.22, resulting in very similar loadings to those found at 16 years, with just slightly lower loadings for most internalizing symptoms on the P factor (see table S4 in supplementary material). Correlations between factors at 14 years and 16 years showed that factors were largely stable over 2 years, with nonsignificant or only small longitudinal correlations across factors (see bottom of Table 3 ). However, although largely stable across time, these factors were not found to be metrically invariant over time. That is, when factor loadings were constrained to be equal across time (i.e., weak factorial invariance) the model fit was significantly worsened relative to when they were freely estimated (χ 2 diff = 146.68, DFdiff = 24, p < .001). Thus, after an inspection of the factor loading at 14 and 16 years, the loadings that did not demonstrate configural invariance (i.e., that differed across time) were freed, to test whether partial factorial invariance could be met. The model in which loadings for all internalizing indicators, except for eating disorders, number of drugs used and tobacco use were allowed to be freely estimated over time did not differ significantly from the base, freely estimated model (χ 2 diff = 17.13, DFdiff = 10, p = .072), indicating that the loadings for ADHD, CD, ODD, drinking problems, and eating disorder did not differ across time. Taken together, these results suggest that while scores within factors were stable across time and very little longitudinal association existed across factors, and the P factor bifactor structure fits well at both 14 and 16 years, the size of the contribution of psychopathology symptoms or indicators to the P factor may vary across development, with internalizing symptoms and drug and tobacco use becoming stronger with increasing age.

Unadjusted Personality and Cognitive Correlates of Psychopathology

All predictor models with covariates showed good model fit (see note under Table 4 ). Table 4 presents correlations between covariates and the bifactor model of psychopathology from Model 3a (associations with factors from Model 3B’ can be found in Table S5 in supplementary material). Results showed that after controlling for multiple testing, common variance across psychopathology (P factor) was significantly associated with high levels of impulsivity, novelty-seeking, neuroticism, hopelessness, sensation-seeking, and extraversion, and low levels of agreeableness and conscientiousness. The P factor was also associated with high delay-discounting, low verbal and performance IQ, low working memory (spatial and verbal), poor response inhibition and risk-taking. Unique variance for EXT (SU and low ADHD) symptoms was significantly associated with high novelty-seeking, sensation-seeking, and extraversion, high verbal IQ and high risk-taking. In contrast, unique variance for INT symptoms was associated with high neuroticism, hopelessness and anxiety sensitivity, low novelty-seeking and extraversion, high conscientiousness, high attention (as measured by digit-span forward), and an attentional bias toward negatively valenced verbal stimuli.

Adjusted Associations Between Personality, Cognitive Correlates, and Psychopathology

In order to examine whether associations between personality, cognitive correlates, and psychopathology factors remained once the effects of other personality and cognitive characteristics were adjusted for, different path analyses were conducted where correlations between correlates and factors were substituted by regression paths in the models. That is, for example, one model was conducted where all cognitive characteristics were entered as predictors and the psychopathology factor being regressed on all cognitive correlates simultaneously. Because of high correlations between some personality traits across measures (e.g., r = .47 between neuroticism and hopelessness, see supplementary table S2) and for ease of comparability with previous findings, separate models were conducted in which psychopathology factors were regressed on personality traits from each personality measure. Adjusted associations between correlates at 14 years and psychopathology factors at 16 years (see the second part of Table 4 ), showed that the P factor was predicted by high impulsivity, novelty-seeking, extraversion, hopelessness, and neuroticism, and low agreeableness, conscientiousness, and anxiety sensitivity. The P factor was also predicted by low spatial IQ, high delay discounting, and poor response inhibition. The specific EXT (SU and low ADHD) factor was predicted by high novelty-seeking, sensation seeking, and extraversion, and high verbal IQ and high risk-taking (gambling task). Finally, the specific INT factor was associated with high neuroticism, hopelessness, conscientiousness, anxiety sensitivity, and low novelty-seeking and extraversion. The INT factor was also associated with an attentional bias toward negatively valenced verbal stimuli. In these models, personality traits explained 8% to 15% of the variance of the P factor, 3% to 7% of the variance of the EXT (SU low ADHD) factor and 4% to 14% of the variance of the INT factor. Cognitive correlates explained 6%, 2%, and 5% of the variance of the P factor, EXT (SU and low ADHD) factor, and INT factor, respectively.

Discussion

The first objectives of the current study were to model the structure of psychopathology in a community sample of European adolescents, and to examine the stability of psychopathology symptoms from early to middle adolescence (14 to 16 years). Findings demonstrated that a general psychopathology (P) factor can be modeled in this community adolescent sample, as well as either (a) two specific externalizing and internalizing factors or (b) three specific ADHD/CD/ODD, substance use, and internalizing factors, providing further support for a spectrum and latent trait model of psychopathology (e.g., Caspi et al., 2014 ; Lahey et al., 2012 , 2015 ; Murray et al., 2016 ). This study contributed to the literature by extending previous bifactor models to include eating disorders and a wider range of substance use symptoms. This study also showed that a bifactor structure of psychopathology was stable across early-to-middle adolescence, a period characterized by substantial change and the onset of many disorders, but that the contributions made by different psychopathology symptoms to the P factor changed across development. Indeed, longitudinal factorial invariance analyses conducted in the present sample showed that loadings for internalizing symptoms, as well as drug use and tobacco use, became stronger with age.

Of note, although like other studies we found that a bifactor model of psychopathology, with either two or three specific factors fit the data well, our findings differ from previously reported P factor models in the following ways: (a) externalizing symptoms loaded more strongly on the P factor in this study, rather than internalizing symptom, which has been the case for many studies modeling the P factor (e.g., Caspi et al., 2014 ; Lahey et al., 2012 ); and (b) when only two specific EXT and INT factors were modeled, the EXT factor included significant positive loadings for SU variables but negative loadings for ADHD and ODD. These discrepancies could reflect differences across sample demographics, measures used, and/or developmental stage. Future studies on this sample could examine the structure of psychopathology using different indicators (e.g., symptom scores) and test for factorial invariance across countries to aid in confirming these hypotheses. That said, this study joins the fast growing literature confirming that, regardless of symptoms/disorders measured, there is substantial variance shared among these that can be captured by a general psychopathology factor.

Interestingly, the P factor in this study accounted for the common variance across externalizing symptoms that was previously attributed to a general externalizing factor in another IMAGEN study focusing specifically on the structure of externalizing symptoms ( Castellanos-Ryan et al., 2014 ). This finding highlights that modeling the structure of psychopathology based on a broad range of symptoms may clarify the nature, antecedents, and implications of liabilities to multiple psychiatric problems, which may have been incompletely captured by narrower analyses (e.g., modeling just the EXT or INT spectrums). Our results suggest that nonsubstance-related externalizing problems (i.e., CD, ADHD, and ODD) may not have more in common with each other and with substance use problems than the general liability to psychopathology shared with internalizing problems.

Another objective was to identify some of the common and unique personality and cognitive correlates of general and specific psychopathology factors. Findings showed that after controlling for multiple testing, common variance across psychopathology was generally related to most personality measures, with the exception of openness to experience and anxiety sensitivity, in theoretically expected ways and in line with previous findings. Namely, the P factor was associated positively with neuroticism, hopelessness, impulsivity, novelty-seeking, and negatively with agreeableness and conscientiousness ( Carragher et al., 2016 ; Caspi et al., 2014 ; Tackett et al., 2013 ). These associations remained largely unchanged after adjusting for other personality traits in the model, suggesting that the general liability to psychopathology may be characterized by a dysregulated personality profile involving high negative affect, low positive affect and poor behavioral control.

In terms of cognitive correlates, unadjusted findings also replicate the pattern of results suggesting that the P factor was associated with poor general cognitive functioning ( Caspi et al., 2014 ). Adjusted associations showed that high-delay discounting, poor response inhibition, and low performance IQ were uniquely associated with the general psychopathology factor in this sample of adolescents. These findings are consistent with those of Caspi et al. (2014) and Lahey et al. (2015) identifying low performance IQ and poor executive function as important correlates of a general psychopathology factor. Delay discounting has not previously been examined as a correlate of a general liability to psychopathology within a spectrum or bifactor methodology framework, but this finding echoes studies identifying poor delay discounting as an important nonspecific risk factor for psychopathology (or transdisease mechanism; e.g., Bickel, Koffarnus, Moody, & Wilson, 2014 ; Castellanos-Ryan et al., 2014 ).

Correlates of specific factors were also identified, with high sensation-seeking, high verbal IQ, and risk-taking being related to variance specific to substance misuse, and high neuroticism, hopelessness, anxiety-sensitivity, conscientiousness, and agreeableness but low novelty seeking and extraversion, as well as an attentional bias toward negatively valenced verbal stimuli being associated with variance specific to internalizing disorders. These findings of dissociation in the cognitive profiles of specific substance use factors from general externalizing, or in this case psychopathology factors, are consiste