Higher-order dimensions of psychopathology in a neurodevelopmental transdiagnostic sample.

Dimensional Models of Psychopathology

Dimensional models of psychopathology account for widespread comorbidity between disorders. Early models assumed that two or three dimensions best explained the high rates of co-occurrence between disorders in both adults and children (e.g., Achenbach & Edelbrock, 1981 ; Kendler et al., 2003 ; Krueger, 1999 ; Wright et al., 2013 ). Considerable covariation between these early dimensions led to the development of contemporary frameworks that conceptualize psychopathology as multiple hierarchically organized transdiagnostic dimensions (e.g., Caspi et al., 2014 ; Kotov et al., 2017 ; Lahey et al., 2012 ; Michelini et al., 2019 ; Patalay et al., 2015 ). These frameworks (e.g., the Hierarchical Taxonomy of Psychopathology [HiTOP]; Kotov et al., 2017 ) include a general factor of psychopathology, which sits above spectra that align with broad internalizing and externalizing factors. These spectra then become progressively more specific, breaking down into lower-order dimensions that align with subsets of traditional diagnoses or disorders that tend to co-occur ( Slade & Watson, 2006 ), and then into symptom components or individual symptoms as the lowest tier of the hierarchy.

Hierarchical Integrative Frameworks

Traditional approaches for identifying the optimal number of dimensions in transdiagnostic hierarchical frameworks rely on factor analytic methods (e.g., exploratory factor analysis [EFA] or principal component analysis [PCA]) that extract variance in higher-order factors from the lower-order factors. Simply put, the variance explained by one factor in the top level of the hierarchy is split between the factors identified in the next tier of the hierarchy. The net result being that lower-order factors in traditional models capture subtle distinctions between indicators (e.g., what makes social phobia distinct from generalized anxiety disorder), and not shared aspects within factors (e.g., what makes the two disorders more similar). Identifying what separates or distinguishes symptoms or syndromes from one another is not congruent with a transdiagnostic approach.

Relatively new empirically derived multilevel hierarchical models can address this issue (e.g., Farmer et al., 2013 ; Forbes et al., 2017 ; Kim & Eaton, 2015 ). Goldberg’s (2006) bass-ackward factor analytic method enables the sequential extraction of dimensions from the top down: it extracts maximally distinct orthogonal components at each level of the hierarchy, starting with the extraction of a single component at the highest level, two at the second level, and so on. Crucially, it maps all measures in a multidimensional space (all indicators load on all components at each level of the hierarchy), and allows each level of the hierarchy to retain all the variance in the patterns of covariation among the symptoms. In retaining what is shared between indicators at each level of hierarchy, this approach is congruent with transdiagnostic approaches that aim to understand the shared mechanisms underlying symptoms that co-occur across traditional categorical disorders (e.g., Cuthbert & Insel, 2013 ; R.-Mercier et al., 2018 ; Newby et al., 2015 ; Owen, 2014 ; Reininghaus et al., 2019 ; Sakiris & Berle, 2019 ; Titov et al., 2011 ).

A few studies have applied Goldberg’s bass-ackward ( Goldberg, 2006 ) approach to identify hierarchies of higher–order dimensions of predominantly personality pathologies in adult populations (e.g., the Alternative Model for Personality Disorders; Forbes et al., 2017 ; Morey et al., 2013 ; Tackett et al., 2008 ; Wright et al., 2012 ). To our knowledge, there has only been one attempt to apply this method to delineate only higher-order dimensions of psychopathology. Using data from the Adolescent Brain Cognitive Development study, Michelini et al. (2019) identified a hierarchical structure with a general psychopathology factor at the first level, and five specific factors (internalizing, externalizing, somatoform, detachment, and neurodevelopmental) in a community sample of children. Many of these specific factors were included in other HiTOP, with the exception of the neurodevelopmental dimension. Crucially, each level of the hierarchy identified by Michelini et al. (2019) differentially predicted different outcomes for the children. Notable findings were specific links between the p factor and the use of mental health services, and a strong relationship between the newly identified neurodevelopmental factor and children’s academic functioning. Michelini’s (2019) findings underscore the importance of including symptoms of neurodevelopmental disorders, such as inattention, in models of psychopathology, and demonstrate the validity of examining multiple levels of the hierarchy of psychopathology to characterize children’s mental health symptoms and their relevance to different aspects of clinical and educational functioning.

The Current Study

The aim of this study was to use the bass-ackward method to delineate hierarchical dimensions of psychopathology in a cohort of children with learning-related problems, and to test the links between the emergent dimensions and children’s educational, social, and clinical functioning. This is because mental ill health and diminished psychosocial functioning are common among children with diagnosed neurodevelopmental disorders of learning. Multiple studies report elevated levels of internalizing symptoms among children with attention-deficit-hyperactivity disorder (ADHD) and autism spectrum disorder (ASD; e.g., Jarrett & Ollendick, 2008 ; Larson et al., 2011 ; Rodgers & Ofield, 2018 ; Sciberras et al., 2014 ; Vaillancourt et al., 2017 ). Externalizing symptoms can also feature among these groups, for example, substance misuse, aggressive behavior, or unsafe sex ( Baker et al., 2018 ; Gillberg et al., 2004 ; Vaillancourt et al., 2017 ). Children with diagnoses of language and reading disorders also experience internalizing problems (e.g., Beitchman et al., 2001 ; Boetsch et al., 1996 ; Yew & O’Kearney, 2013 ), but externalizing difficulties are reported less often. A substantial proportion of children who have not received a diagnosis, but who nonetheless are struggling at school, also experience heightened levels of mental health problems ( Arnold et al., 2005 ; Auerbach et al., 2008 ; Bryant et al., 2020 ; Francis et al., 2019 ; Greenham, 1999 ; Maughan & Carroll, 2006 ; Willcutt et al., 2013 ; Wu et al., 2014 ; Young et al., 2012 ).

Studies attempting to understand psychopathology in pediatric populations with learning-related problems typically adopt case-controlled designs that group individuals according to the presence or absence of one or more diagnosed neurodevelopmental conditions (e.g., Humphreys et al., 2012 ; Rodriguez-Seijas et al., 2020 ). However, like psychiatric disorders, neurodevelopmental conditions are characterized by high degrees of intracondition variation and intercondition commonality. This has motivated a shift toward more transdiagnostic approaches whereby individuals are characterized based on well-known cognitive ( Holmes et al., 2020 ), behavioral ( Mareva et al., 2019 ) and neurobiological ( Siugzdaite et al., 2020 ) mechanisms as opposed to their diagnostic status. To date, there have been no attempts to delineate hierarchical dimensions of psychopathology in a neurodevelopmental transdiagnostic sample.

The present study recruited a highly heterogeneous pediatric sample with a range of neurodevelopmental symptoms linked to poor learning. It adopted a functionally defined approach of enrolling individuals identified by practitioners as having difficulties in attention, learning, and/or memory. These individuals did not fit traditional categories of neurodevelopmental disorders; some had a single diagnosis, others had multiple diagnoses, but the majority were undiagnosed despite coming to the attention of a health or educational professional for experiencing difficulties that were affecting their school progress. The sample included children with relatively mild problems, who would likely not meet diagnostic thresholds for specific learning disorders, in addition to many children whose more marked problems definitely would. A considerable proportion of the sample had a diagnosis of ADHD.

The hierarchical model was derived using subscales from three inventories: the Revised Child Anxiety and Depression Scale - Parent Version (RCADS; Chorpita et al., 2000 ), the Conners-3 Parent Short Form (CPSF; Conners, 2008 ) and the Strengths and Difficulties Questionnaire (SDQ; Goodman, 1997 ). These were selected for two reasons. First, they provide a broad sweep of internalizing and externalizing behaviors. Second, they are widely used in clinical practice to determine symptom severity, enhancing the translational relevance of our findings. The measures selected for inclusion maximize the breadth of the internalizing and externalizing symptoms while avoiding including very similar measures of the same symptom. Data reduction methods such as the one used here identify the main axes of variation in selected measures, representing the largest amounts of the variance within the dataset. Including multiple indicators of some symptoms and not others can influence the dimensions that emerge (for example, if three indicators of hyperactivity were included and only one of inattention, a “hyperactivity” dimension might emerge as distinct from inattention simply because the input was weighted more heavily toward capturing hyperactivity). To avoid this, a single indicator of each symptom was included from the subscales (see Method for details about subscale selection). Subscales from the SDQ and CPSF that were not included in the model were included as clinical outcomes alongside measures from the Behavior Rating Inventory of Executive Function (BRIEF; Gioia et al., 2000 ) to test whether levels of the hierarchy differentially predicted different aspects of clinical function. The Learning Problems subscale from the CPSF was held out of the model to test how the different levels of the hierarchy predicted concurrent education performance. This was of particular interest given the children were referred primarily for experiencing learning difficulties. There were no predictions about the specific dimensions that would emerge as the data analysis was exploratory. Nonetheless, a general p factor and two higher-level spectra (internalizing and externalizing) were expected given they are well-established dimensions of psychopathology in both traditional and newer hierarchical models of psychopathology (e.g., Caspi et al., 2014 ; Kotov et al., 2017 ; Michelini et al., 2019 ). We modeled the associations between the emergent dimensions and children’s learning, clinical, and social functioning. These analyses were also exploratory.

Materials and Procedure

Children aged 5 to 18 years were referred to CALM by health and education professionals for problems in attention, learning and/or memory. Children completed a 4-hr assessment of learning and cognition, and parents/legal guardians/carers completed questionnaires measuring the child’s behavior and mental health. All procedures complied with the ethical standards of the national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008. All procedures involving human participants were approved by the National Health Service (REC: 13/EE/0157). Parents/caregivers provided written consent and child verbal assent was obtained.

Three measures used widely in clinical practice and capture internalizing and externalizing symptom severity were selected: the RCADS, CPSF, and SDQ. All six subscales from the RCADS were included in the hierarchical dimensional model: Separation anxiety disorder, Social phobia, Generalized anxiety disorder, Panic disorder, Obsessive–compulsive disorder, and Depression. The following subscales from the CPSF were also included: Inattention, Hyperactivity/impulsivity, Executive function, Aggression, and Peer relations. Two subscales from the SDQ were also included: Prosocial behavior (reverse coded to reflect low prosocial behavior) and Conduct problems. The measures selected for inclusion in the model were chosen to maximize the range of symptoms captured, while including only a single indicator of any particular symptom from all subscales. Including multiple indicators of some symptoms and not others can bias the dimensions that emerge. For this reason a single indicator of each symptom was selected. Subscales from the RCADS were chosen as the primary input to capture multiple symptoms of internalizing difficulties; this was preferable to using the Emotion subscale from the CPSF that combines symptoms of anxiety and depression into a single measure. Next, subscales from the CPSF that did not overlap with the RCADS were included to capture both externalizing symptoms and symptoms associated with neurodevelopmental difficulties. The CPSF was prioritized over the SDQ at this point as it captures more symptoms of neurodevelopmental problems. Subscales from the SDQ were then used to supplement the model with additional externalizing domains not captured by the RCADS and CPSF.

Predicted Outcomes

The predictive value of different levels of the model were estimated using a limited number of variables from the CALM dataset. Some of these assessed concurrent clinical outcomes: Emotional symptoms, Peer relationship problems, and Hyperactivity and Inattention from the SDQ, and the Global executive function composite from the BRIEF. Concurrent educational performance was measured using the Learning problems subscale from the CPSF. Finally, the IMD was used as a proxy for socioeconomic status.

Analysis Plan

First, principal components analysis was used to extract and rotate (with geomin) factor solutions for the measures of psychopathology. The maximum number of factors to extract was determined by parallel analysis. All factor structures from one to the maximum number were considered. Second, the hierarchical structure was derived by correlating factor scores on adjacent levels of the hierarchy using Goldberg’s bass-ackward hierarchical method ( Goldberg, 2006 ). This is the only available method to delineate multiple hierarchical levels using an exploratory approach. It allows factors to be correlated across levels without statistically removing variance shared with a general factor. To test the predictive value of the dimensions, factor scores from each level were entered into a series of regression models with the predicted outcomes as dependent variables. The incremental predictive value of each level of the hierarchy was examined by testing the significance of changes in F and R 2 (Δ F and Δ R 2 ) between models with different numbers of factors as predictors. All analyses were conducted in R Version 4.0.3 using the Psych package 2.0.12.

Results

The sample profile is summarized in Table 1 . Scores were in the age-typical range for Prosocial behavior and all subscales of RCADS, except Depression, which was borderline elevated. Aggression as rated on the CPFS, and Conduct problems, Emotional symptoms, and Peer relationship problems from the SDQ were also borderline elevated. Scores were in the clinical/abnormal range on the remaining CPSF subscales, Hyperactivity and Inattention (SDQ), and the Global executive composite (BRIEF).

The maximum number of factors to extract was determined with parallel analyses (extraction was stopped when eigenvalues fell within the 95% confidence interval (CI) of eigenvalues from simulated data). This indicated that up to three factors could be extracted (see Figure 1 ). For completeness, a four-factor solution was also considered (see Supplemental Materials Table 8 ). This produced a model with a single indicator on the fourth factor, and eigenvalues that were outside the acceptable 95% CI. With fewer than three indicators on the fourth factor it was not possible to interpret ( Fabrigar et al., 1999 ; Velicer & Fava, 1998 ), indicating the maximum number of factors that could be extracted was three.

The factor loadings, extracted using principal components, are presented in Table 2 . Note that although components were extracted, the term “factor” is used from here on for ease of interpretability and because the two terms are used interchangeably in studies adopting the same bass-ackward methods (e.g., Michelini et al., 2019 ). Also, the term factor is more synonymous with dimensional approaches in the wider literature. The one, two and three factor models were tenable and interpretable; see Figure 2 for the hierarchical structure.

The one factor solution reflected a general psychopathology p factor, and the two factor solution broad internalizing and broad externalizing factors. In the three-factor solution, the broad externalizing factor split into narrower neurodevelopmental (inattention, executive function, and hyperactivity/impulsivity) and social maladjustment factors (aggression, conduct problems, low prosocial behavior, and peer relations, with a lower cross-loading for hyperactivity/impulsivity). The latter factor was moderately associated with the more general broad internalizing factor. The broad internalizing factor was fully represented by a specific internalizing factor in the three-factor solution. This encompassed symptoms of generalized anxiety, panic and obsessive–compulsive disorders, social phobia, separation anxiety, and depression. Symptoms of social phobia were additionally, negatively and more weakly, correlated with the broader social maladjustment factor.

As the sample included more boys than girls, sex differences were explored by comparing boys and girls across the factors (see Table 3 ). Significant sex differences were observed on the broad externalizing and social maladjustment factors, with boys expressing greater difficulties on both dimensions.

Predictive Value

A series of linear regressions were performed for each of the following predictors: Emotional symptoms, Peer relationship problems, Hyperactivity and Inattention, Global executive function, Learning Problems, and IMD. For each outcome, three separate regression models were calculated. In the first, factor scores from the one factor model were entered as predictors. In the second, factor scores from both factors in the two-factor model were entered, and in the final model, the three factor scores for the three-factor models were entered. The contributions of the different factor models to each validator were compared using change in R 2 and F. The models were compared in pairs (Model 1 vs. Model 2, and Model 2 vs. Model 3) to test whether there was a significant change for a more complex versus a simpler structure. The results are summarized below and shown in Figure 3 .

The one, two and three factor models significantly predicted emotional symptoms ( Supplemental Materials Table 1 ). The p factor explained 42% of variance, F (1, 380) = 270, p < .001. When both factors from the two factor model were entered, 53% of variance was explained, F (2, 379) = 212.8, p < .001, but only the broad internalizing factor was a significant predictor. The specific internalizing, social maladjustment and neurodevelopmental factors from the three-factor model were significant predictors, explaining 54% of variance, F (3, 378) = 145.1, p < .001. The addition of more differentiated factors significantly increased the amount of variance explained: the two-factor model accounted for significantly more variance than the one-factor model, Δ F (1, 379) = 91.34, p < .001, Δ R 2 = .11, and the three factor more than the two factor model, Δ F (1, 378) = 5.06, p = .025, Δ R 2 = .01. The more complex two-factor structure explained 11% more variance than p-alone, but the change from the two to three factor models was minimal (.6%), despite being significant.