Externalizing Psychopathology from Childhood to Early Adolescence: Psychometric Evaluation using Latent Variable and Network Modeling

Abstract

Applying both latent variable and network frameworks, we conducted a comprehensive psychometric evaluation of the diverse array of symptoms from three externalizing dimensions, including attention problems, aggressive behavior, and delinquency/rule-breaking of the Child Behavior Checklist (Achenbach, 1991) across six time points from childhood to early adolescence. We also examined sex differences. Participants (N=1,339) were drawn from two multisite longitudinal studies: Fast Track and the Child Development Project. Parents reported on externalizing psychopathology in kindergarten and grades 1, 2, 4, 5, and 7. Using exploratory structural equation modeling, we found almost uniformly excellent fit across time and samples. However, we also observed multiple cross-loadings and heterogeneity in terms of which symptoms cross-loaded across time points. Alternatively, using network modeling, we observed that symptoms of attention problems and aggressive behavior had stronger connections, relative to delinquency/rule-breaking, across time and samples. Significant differences in overall connectivity were found at early (kindergarten vs. grade 1, grade 1 vs. grade 2) and late (grade 5 vs. grade 7) time points for the combined sample and only late time points for the male sample. In addition, the items ‘impulsive’ and ‘lies or cheats’ consistently displayed the greatest bridge strength, i.e., symptom from one dimension that connects to symptoms from another dimension, across time and samples. Our results illustrate how two methods—latent variable and network modeling—provide important and complementary information on multidimensional constructs. Findings also inform understanding of externalizing psychopathology through childhood to early adolescence by identifying key symptoms, critical transition points, and possible transdiagnostic liabilities.

In children and adolescents, externalizing behavior problems reflect social conflicts due to inadequate emotional and behavioral control. Theoretical and structural models have tended to summarize these problems as psychiatric diagnostic categories including attention-deficit/hyperactivity disorder (ADHD), oppositional defiant disorder (ODD), and conduct disorder (CD) or as a broad externalizing spectrum (Tuvblad et al., 2009). Importantly, strong evidence has shown that these disorders commonly co-occur (Angold et al., 1999; Connor et al., 2010; Waschbusch, 2002), and co-occurrence is associated with greater symptom severity and poorer treatment response (Kessler et al., 2011). However, it is currently unclear which specific symptoms co-occur and how these associations might differ at different developmental stages. This is because most research on externalizing problems analyzes at the level of the higher-order summary scales, such as disorders, the externalizing spectrum, or other clinically relevant dimensions (e.g., aggressive behavior), which comprise a diverse range of symptoms. In contrast, testing associations between symptoms of multiple dimensions of psychopathology can identify which symptoms co-occur and establish transdiagnostic markers.

Latent Variable Theory and Modeling

Derived from the medical field, one theoretical framework for understanding the relationship between psychopathological symptoms is the common cause model. In this model, symptoms do not necessarily have direct relations, but rather all symptoms are explained by one or more shared factors. Psychometrically, this can correspond to the latent variable model, and diagnostic categories are one such example analyzed through latent variable models. Externalizing psychopathology is often described as beginning with ADHD early in childhood, ODD and CD in early and late childhood, and substance use disorder and antisocial personality disorder in adolescence and early adulthood (Beauchaine & McNulty, 2013; Beauchaine et al., 2017; Krueger et al., 2002). Research has also shown distinct manifestations of externalizing psychopathology in childhood versus adolescence (Beauchaine et al., 2010; Olson et al., 2013). For example, studies examining prevalence have found that hyperactivity and oppositional behavior are relatively prominent earlier in development (Hart et al., 1995), whereas inattention and conduct problems are more likely to peak later in development (Moffitt, 1993). With a subsample of the current sample, research has shown that the factor structure of the externalizing spectrum (comprised of ADHD, ODD, and CD) is largely stable over childhood and into adolescence (King et al., 2018).

Some findings suggest that the externalizing spectrum may be explained by a general transdiagnostic liability (King et al., 2018). Theoretically, impulsivity and emotion dysregulation contribute to externalizing heterotypic continuity, which refers to distinct behavioral manifestations of the same underlying liability (Beauchaine et al., 2010, 2017; Goulter et al., 2022; Martel, 2009; Peterson et al., 2021). However, factor analytic studies of the dimensional Child Behavior Checklist (CBCL; Achenbach, 1991)—a very commonly used measure to examine child and adolescent psychopathology—exclude impulsive and attention problems in the assessment of externalizing psychopathology during mid-childhood and adolescence. In addition, studies examining the factor structure of the CBCL have predominantly relied on confirmatory factor measurement models (e.g., Gomez & Vance, 2014; Tabet et al., 2021). A critical limitation of this factor analytic approach is the assumption that symptoms load onto only one factor (Cooke & Sellbom, 2019; Sellbom & Tellegen, 2019). This assumption can result in model misfit, particularly when complex multidimensional constructs are evaluated (Cooke & Sellbom, 2019; Sellbom & Tellegen, 2019). A promising method in psychological assessment that capitalizes on the strengths of both confirmatory and exploratory factor analyses is exploratory structural equation modeling (Asparouhov & Muthén, 2009). This approach tests a priori factor structures but also estimates factor loadings for all indicators. Several scholars have emphasized the importance of exploratory structural equation modeling, particularly in clinical and personality research contexts (Marsh et al., 2014; Sellbom & Tellegen, 2019).

Although factor analytic approaches provide critical information with regard to weighted composites, they do not provide information on the interplay or associations between lower-order symptoms. Examining associations at the symptom level has the potential to parse out heterogeneity in the expression of externalizing psychopathology, in addition to identifying which symptoms co-occur. There is currently limited understanding of the interplay of symptoms within and across externalizing dimensions (e.g., attention problems, aggressive behavior, delinquency/rule-breaking) and whether specific symptoms of impulsivity or emotion dysregulation represent transdiagnostic indicators through childhood and into adolescence. Advances in understanding this interplay might be gained from drawing on network theory and using psychometric network analyses.

Network Theory and Modeling

Embedded in systems science, network theory proposes that psychopathology is comprised of complex networks of interconnected symptoms or elements (Borsboom & Cramer, 2013). Specifically, symptoms are not reflective of an underlying factor, but rather the direct relations between symptoms constitute psychopathological problems. These symptoms reinforce each other contributing to the maintenance of psychopathology. Psychometric network analyses test the covariance of these symptoms (i.e., how symptoms are related to one another) and identify symptom clusters (Goulter & Moretti, 2021; McElroy et al., 2018). Symptoms (or items) are plotted as nodes and the inter-item relations as edges. The stronger the association between nodes, the larger the weight of their edges (graphically displayed as width and saturation of color). This type of modeling also provides indices of centrality, indicating which symptoms characterize a construct through associations with other symptoms. Although there are several indices of centrality, we focus on strength centrality because some researchers have suggested that the strength index is the most applicable to the field of psychopathology (Bringmann et al., 2019). Strength centrality is the sum of correlations between a node and other nodes (Costantini et al., 2015). Network approaches also distinguish symptoms from one dimension that connect or increase the likelihood of symptoms from another dimension, known as bridge symptoms (Cramer et al., 2010). Thus, network analyses comprising symptoms across multiple dimensions (or disorders) can facilitate the identification of transdiagnostic indicators (Borsboom et al., 2011). This is because identified clusters may share common symptoms.

Few studies have examined the network structure of multiple externalizing dimensions. Using parent-reported conduct problems among children (3–12 years), Hukkleburg (2019) identified oppositional and inattention clusters, with the oppositional item ‘gets angry when doesn’t get his/her own way’ showing the greatest strength centrality. With specific ADHD and ODD measures, Preszler and Burns (2019) found three clusters representing ADHD-inattention, ADHD-hyperactivity/impulsivity, and ODD rated by parents of children (M_age=9.04 years). Similarly, with four cohorts (preschool, aged 3–6 years; early childhood, 6–9 years; middle childhood, 10–13 years; and adolescence 13–17 years), Martel et al. (2017) identified two clusters representing ADHD and ODD in preschool and early childhood, with impulsivity symptoms (i.e., ‘often interrupts’ and ‘difficulty waiting’) as the most central symptoms in the networks. In middle childhood and adolescence, ADHD-inattention and hyperactivity/impulsivity formed two distinct clusters with ‘often defies’ and ‘often interrupts’ as the most central symptoms. Martel and colleagues provide important information with respect to network structures at distinct developmental periods suggesting that impulsivity may underlie externalizing psychopathology at multiple points in child development. However, the study’s age groups were separate cohorts, so it is unclear how externalizing networks evolve within individual children over time. Only one study (Madole et al., 2019), has examined the network structure of psychopathology using narrowband scales from the CBCL (Achenbach, 1991). Although the Madole et al. (2019) study focused on the moderating role of cold and hot cognitive control in a sample of adolescents (M_age=15.66 years), it also found that symptoms of irritability emerged as having the strongest associations with all other clusters. To date, no research has mapped the network structure of attention problems, aggressive behavior, and delinquency/rule-breaking across childhood and into adolescence. Such findings may inform understanding of transdiagnostic markers underpinning multiple externalizing dimensions throughout child development.

The Present Study

Using data from two longitudinal studies (N=1,339), the present study aimed to test the structure of externalizing psychopathology using items of the CBCL narrowband scales (attention problems, aggressive behavior, delinquency/rule-breaking) across six time points through childhood and into adolescence. We applied both psychometric latent variable and network modeling approaches. Latent variable and network analyses provide complementary but distinct information. For example, if impulsivity or other attention problems represent an underlying liability of multiple forms of externalizing psychopathology, through a latent variable approach, we may find that impulsive and attention problems cross-load on several factors. However, in latent variable modeling, high cross-loading items are sometimes treated as methodological artefacts and removing cross-loading items can be typical practice, yet these items may provide critical information established through a network approach. To illustrate, past exploratory factor analytic work on attachment removed an item (‘I don’t fully trust this person’) that cross-loaded on attachment anxiety and avoidance dimensions (Fraley et al., 2011). However, using a network approach, others have found this specific item connected these two attachment dimensions, suggesting that trust may represent a transdiagnostic factor of attachment dimensions or a mechanism by which change in one dimension could impact change in the other dimension (McWilliams & Fried, 2019). Applying a network approach in the present study, we are, thus, able to determine whether impulsive and attention problems are directly associated with specific symptoms across both aggressive and delinquent/rule-breaking behavior providing a greater level of specificity in the characterization and identification of possible mechanisms of externalizing psychopathology. In addition, the use of both latent variable and network approaches has become increasingly common in studies aiming to further delineate the structure of a given construct; however, these studies have predominantly specified confirmatory factor models (e.g., Eadeh et al., 2021). As indicated earlier, confirmatory factor analyses assume that items load only onto one factor, which can result in model misfit and inflated correlation coefficients (Sellbom & Tellegen, 2019).

Applying exploratory structural equation modeling to our data, two research questions were addressed. First, does a correlated three-factor model represent externalizing psychopathology at different time points across childhood and into adolescence? Second, do specific symptoms cross-load onto attention problems, aggressive behavior, and delinquency/rule-breaking at different time points? Using network modeling, four research questions were addressed. First, does the network structure of externalizing psychopathology change at different time points? Second, are there distinct central symptoms (as indicated by the strength centrality index) of externalizing psychopathology at different time points? Third, which symptoms provide a bridge between attention problems, aggressive behavior, and delinquency/rule-breaking at different time points? Fourth, are the network structures of externalizing psychopathology stable at each time point in development? Stability refers to whether the network structure remains the same while removing participants from the analyses (Epskamp et al., 2018). Given differences in risk and expression of externalizing psychopathology between male and female samples (King et al., 2018; Konrad et al., 2021; Lynch et al., 2021), we also examined whether there were sex differences across both latent variable and network approaches.

Method

Participants and Procedure

Participants (N=1,339) were drawn from two longitudinal studies. The Fast Track project is a longitudinal, multisite (Durham, NC; Nashville, TN; Seattle, WA; and rural Pennsylvania) investigation of the development and prevention of child conduct problems (Conduct Problems Prevention Research Group, 2020). In 1991–1993, 9,594 kindergarteners across three cohorts were screened for classroom conduct problems by teachers using the Teacher Observation of Classroom Adaptation-Revised authority acceptance score (Werthamer-Larsson et al., 1991), and a subset were screened for home behavior problems by parents using a 22-item instrument based on the CBCL and similar measures (Achenbach, 1991). Teacher and parent screening scores were standardized within site and summed to yield a total severity-of-risk screen score. Children were selected for inclusion into the high-risk sample based on this screen score, moving from the highest score downward until desired sample sizes were reached within sites, cohorts, and groups. This multi-stage screening procedure resulted in 891 children divided into control (n=446) and intervention (n=445) samples. In addition to the high-risk sample of 891, a stratified normative sample of 387 children was identified to represent the population normative range of risk scores and was followed over time. The present study used data from the high-risk control (65% male; 44% Black, 51% white, 5% other race/ethnicity) and normative (51% male; 42% Black, 51% white, 7% other race/ethnicity) samples; the intervention sample was not included in the present analyses. Seventy-nine of the participants recruited for the high-risk control group were included as part of the normative sample; thus, the total final sample included 754 participants.

The Child Development Project is an investigation of children’s social development and adjustment (Dodge et al., 1990). In 1987 and 1988, a sample of children was identified at the time of kindergarten pre-registration and then followed over time. Participants were recruited in each of two annual cohorts at each of three geographic sites (Nashville, TN; Knoxville, TN; and Bloomington, IN). Within each site, federally subsidized lunch rates and neighborhood housing patterns were used to identify schools that served a full demographic range of the communities. Parents registering their children were approached at random by research staff and asked to participate in a longitudinal study of child development. About 75% agreed. Interested parents were then visited by research staff who explained the project in detail and obtained parents’ informed consent. Overall, 585 children participated (52% male; 81% white, 17% Black, 2% other race/ethnicity).

Across both studies, legal guardians provided consent and participants assented to procedures. Parents and participants were provided monetary compensation. All procedures were approved by the relevant Institutional Review Board for each site. Data and code for this study are available by emailing the corresponding author. This study was not preregistered.

Measures

Externalizing Psychopathology.

Externalizing psychopathology was assessed in kindergarten and grades 1, 2, 4, 5, and 7 using parent report on the CBCL (Achenbach, 1991) narrowband scales of attention problems (20 items), aggressive behavior (11 items), and delinquency/rule-breaking (13 items). Items are scored on a 3-point scale (0 ‘not true’, 1 ‘somewhat or sometimes true’, 2 ‘very or often true’). Internal consistencies, computed across the combined dataset, for attention problems (α=.77–.83; ω=.81–.85), aggressive behavior (α=.89–.91; ω=.90–.93), and delinquency/rule-breaking (α=.66–.78; ω=.71–.84) were good across time points.

Analytic Approach

Latent Variable Modeling

Does a correlated three-factor model represent externalizing psychopathology at different time points?

We used exploratory structural equation modeling with oblique Geomin rotation in Mplus 8 (Muthén & Muthén, 2017) to examine factor structure. We used the Weighted Least Squares Mean and Variance Adjusted (WLSMV) estimator, which handles data as categorical. Relative fit indices (e.g., Akaike Information Criterion and Bayesian Information Criterion) are not available for models using the WLSMV, and thus, we relied on absolute model fit indices, including the Root Mean Square Error of Approximation (RMSEA), Comparative Fit Index (CFI), and Tucker-Lewis Index (TLI) to determine the model fit. Models with RMSEA values <.05 are typically considered excellent and values <.08 as acceptable (Brown, 2014; Little, 2013); models with CFI and TLI values >.95 are considered excellent and values >.90 as acceptable (Brown, 2014; Little, 2013).¹

Do specific symptoms cross-load onto attention problems, aggressive behavior, and delinquency/rule-breaking at different time points?

To determine whether specific items cross-loaded onto multiple dimensions, we relied on recommendations that loadings .32 or higher on two or more factors represent cross-loading (Costelle & Osborne, 2005).

Are there sex differences in the latent structures of externalizing psychopathology?

Analyses were repeated with male and female participants separately.

Network Modeling

Does the network structure of externalizing psychopathology change at different time points?

We constructed network visualizations in R version 3.6.1 using R Studio (R Core Team, 2016). We estimated missing data using imputation with all items across all time points in line with other longitudinal network studies (McElroy et al., 2018). Current network approaches do not have the WLSMV estimator. Items are plotted as nodes and the inter-item relations as edges. The weight, or correlation, between items is represented by the thickness and saturation of the edges. Our network analyses are graphically represented by partial-correlation Gaussian Models with the Least Absolute Shrinkage Operator (Epskamp et al., 2012). Attention problem items are depicted in white, aggressive behavior items are depicted in blue, and delinquency/rule-breaking behavior items are depicted in gray. We also calculated overall connectivity or global strength, which is the sum of all edge weights in a network (van Borkulo et al., 2016). Global strength provides information on whether the networks are more or less strongly connected over time. We used the NetworkComparisonTest package (van Borkulo et al., 2016) to test invariance in global strength of networks at adjacent time points (i.e., kindergarten vs. grade 1, grade 1 vs. grade 2, grade 2 vs. grade 4, grade 4 vs. grade 5, grade 5 vs. grade 7).

Are there distinct central symptoms of externalizing psychopathology at different time points?

To determine whether central items changed across development, we calculated strength centrality at each time point using the qgraph package (Epskamp et al., 2012). This centrality index indicates the sum of correlations between a node and other nodes as a z-score (Costantini et al., 2015). Other centrality indices (e.g., closeness and betweenness) were not calculated in the present study due to recent recommendations that these indices are more difficult to interpret in the context of psychopathology (Bringmann et al., 2019). To test for significant differences between items on strength centrality, we also conducted bootstrapped difference tests using the bootnet package (Epskamp & Fried, 2015).

Which symptoms bridge between attention problems, aggressive behavior, and delinquency/rule-breaking at different time points?

To determine whether specific items from the three narrowband scales (attention problems, aggressive behavior, delinquency/rule-breaking) acted as bridge symptoms between the narrowband dimensions, we estimated bridge strength using the networktools package (Jones et al., 2017). Bridge strength indicates the strength of a node’s connectivity with another dimension as a z-score. We report bridge strength as at least 2 SD below or above the mean.

Are the network structures of externalizing psychopathology stable at each time point in development?

To determine whether the network structures were stable at each time point, we estimated the stability of strength centrality using the bootnet package (Epskamp & Fried, 2015). These analyses also produce a correlation stability (CS) coefficient, which at .70 reflects the maximum number of participants that can be dropped to retain 95% probability between the calculated centrality and that of the subsets. The authors suggest that the CS should be above .50 and not below .25 for accurate interpretability (Epskamp et al., 2018).

Are there sex differences in the network structures of externalizing psychopathology?

Finally, analyses were repeated with male and female participants separately. We also used the NetworkComparisonTest (van Borkulo et al., 2016) to test invariance in overall connectivity between male and female networks at each time point.

Results

Descriptives

Descriptive statistics and the percentage of participants endorsing clinically significant levels of externalizing for the combined sample, male sample, female sample, and the higher-risk Fast Track sample are reported in Supplementary Table S1.

Latent Variable Modeling

Does a correlated three-factor model represent externalizing psychopathology at different time points?

Fit indices from the correlated three-factor exploratory structural equation models are shown in Supplementary Table S2. Across all time points, all models were associated with acceptable to excellent levels of fit. However, some models were qualified by nonpositive definite matrices, which are very common when ordinal response formats are applied (Lorenzo-Seva & Ferrando, 2021). In addition, some symptoms—predominantly from the delinquency/rule-breaking scale—had zero or low variability (e.g., ‘alcohol/substance use’, ‘truant’); these were removed and modified three-factor models were respecified. All respecified models were also associated with acceptable to excellent levels of fit. Latent correlation coefficients among factors are reported in Supplementary Table S3. Across all time points, all factors were positively correlated with each other (with the exception of factors 1 and 3 in grade 1).

Do specific symptoms cross-load onto attention problems, aggressive behavior, and delinquency/rule-breaking at different time points?

Factor loadings are shown in Table 1. Although factor 1 tended to represent attention problems, factor 2 aggressive behavior, and factor 3 delinquency/rule-breaking, there were multiple item cross-loadings across factors. Furthermore, across time points, there was high variability in the items that cross-loaded. In kindergarten, ‘impulsive’, ‘nervous’, and ‘loud’ cross-loaded on factors 1 and 2; ‘destroys own things’ cross-loaded on factors 1 and 3; and ‘bullying’, ‘gets in many fights’, and ‘runs away’ cross-loaded on factors 2 and 3. In grade 1, in addition to ‘impulsive’, ‘nervous’, and ‘destroys own things’, several other items also cross-loaded on factors 1 and one or more of the other factors including ‘destroys others’ things’, ‘can’t sit still’, ‘sets fires’, ‘steals at home’, ‘steals outside of the house’, and ‘vandalism’. Grade 2 had fewer cross-loadings altogether, but a few items still cross-loaded on factors 1 and 2 including ‘talks too much’ and ‘loud’. In grade 3, ‘nervous’, ‘poor at school work’, ‘demands attention’, ‘disobedient at school’, ‘showing off’, and ‘loud’ all cross-loaded onto factors 1 and either factor 2 or 3, in addition to several items cross-loading on factors 2 and 3. Grade 5 had very few cross-loadings, but ‘impulsive’ and ‘nervous’ cross-loaded again on factors 1 and 2. Finally, in grade 7, ‘poor at school work’, ‘demands attention’, and ‘talks too much’ cross-loaded onto factors 1 and either factor 2 or 3, in addition to several items cross-loading on factors 2 and 3.

Table 1.

Externalizing Symptoms Abbreviation, Item Description, and Factor Loadings from Three-Factor Exploratory Structural Equation Models for the Combined Sample

		Kindergarten			Grade 1			Grade 2			Grade 4			Grade 5			Grade 7
Abbreviation	Item Description	F1	F2	F3	F1	F2	F3	F1	F2	F3	F1	F2	F3	F1	F2	F3	F1	F2	F3
Attention Problems
1. Young	Acts too young	.42	.08	.04	.44	.13	−.02	.38	.10	.11	.39	.16	.03	.38	.24	.06	.45	.08	.11
2. Conc	Can’t concentrate	.61	.24	.02	.58	.23	−.01	.70	.01	.17	.74	−.11	.26	.67	−.03	.43	.69	.00	.24
3. Sitstil	Can’t sit still	.53	.29	.06	.39	.42	−.10	.63	.18	.05	.64	.04	.24	.51	.13	.31	.64	.10	.14
4. Conf	Confused	.65	.00	.10	.70	.06	−.03	.81	−.11	.11	.75	−.01	.04	.66	.11	.11	.80	−.11	.07
5. Daydm	Day dreams	.58	.03	−.11	.60	.06	−.08	.79	−.18	.01	.75	.01	−.10	.74	.07	−.01	.82	−.09	−.03
6. Impul	Impulsive	.33	.44	.08	.35	.52	−.07	.41	.29	.19	.40	.31	.22	.32	.39	.20	.29	.37	.27
7. Nerv	Nervous	.44	.35	−.10	.33	.33	.01	.55	.28	−.08	.45	.36	−.02	.38	.45	−.06	.56	.25	−.06
8. Twitch	Twitches	.52	.08	−.03	.58	−.02	.07	.58	.14	−.05	.45	.19	.07	.41	.26	.04	.49	.17	.04
9. Schwrk	Poor school work	.58	−.02	.19	.55	.09	.26	.48	.02	.25	.46	−.03	.36	.30	.01	.48	.36	.02	.42
10. Clum	Clumsy	.45	.15	.09	.53	.03	−.03	.61	.05	.08	.58	.13	−.06	.43	.29	.06	.45	.13	.07
11. Blank	Stares blankly	.60	−.04	.10	.67	.01	.14	.74	−.06	−.03	.72	.13	−.20	.67	.22	−.02	.75	−.14	.02
Aggressive Behavior
12. Argue	Argues	−.01	.65	.05	−.12	.74	−.01	−.02	.70	.05	.04	.61	.14	.03	.78	−.06	−.01	.67	.07
13. Brag	Brags	.02	.47	−.05	−.04	.58	−.09	.02	.49	.14	.09	.42	.22	.06	.52	.13	.09	.55	.09
14. Bully	Bullying	.02	.45	.36	.01	.63	.22	.02	.59	.29	−.08	.56	.36	−.15	.62	.31	−.03	.53	.38
15. Atten	Demands attention	.27	.54	−.05	.00	.73	−.17	.30	.43	.07	.36	.35	.15	.27	.56	.00	.38	.46	−.00
16. Down	Destroys own things	.32	.03	.58	.44	.30	.33	.38	−.01	.64	.25	−.01	.67	.19	.00	.69	.26	.16	.52
17. Doth	Destroys others’ things	.24	−.01	.76	.43	.32	.39	.31	−.00	.74	.11	.09	.75	.11	.09	.68	.19	.19	.60
18. Dhme	Disobedient at home	.13	.50	.18	−.07	.71	.12	.09	.53	.23	.05	.44	.38	.07	.56	.31	.12	.44	.37
19. Dsch	Disobedient at school	.31	.20	.31	.21	.46	.17	.28	.26	.37	.32	.03	.56	.04	.17	.70	.15	.24	.57
20. Jeal	Easily jealous	.25	.49	−.07	.02	.63	−.14	.19	.48	.02	.20	.53	−.01	.07	.63	−.09	.23	.49	−.00
21. Fight	Gets in many fights	.00	.44	.38	.04	.61	.30	−.01	.50	.35	−.00	.46	.46	−.14	.48	.43	.09	.30	.43
22. Attac	Physically attacks others	−.10	.54	.30	.00	.58	.35	−.01	.52	.25	−.17	.52	.45	−.17	.73	.17	−.03	.50	.40
23. Srea	Screams	.02	.55	.13	−.18	.80	.02	.16	.59	−.01	.05	.72	.02	.00	.79	−.05	.03	.64	.07
24. Show	Showing off	.17	.57	.00	.09	.71	−.18	.17	.52	.10	.35	.34	.16	.16	.44	.20	.17	.50	.19
25. Sull	Sullen	.03	.75	−.06	−.26	.87	−.01	.10	.68	−.05	.10	.80	−.11	.01	.95	−.20	.08	.66	.01
26. Mood	Sudden changes in mood	.18	.55	−.01	−.12	.80	.01	.29	.52	−.06	.24	.70	−.14	.03	.80	.12	.16	.58	.00
27. Talk	Talks too much	.24	.48	−.07	.11	.65	−.35	.35	.43	−.15	.40	.26	.03	.25	.47	−.01	.33	.50	−.22
28. Tease	Teases a lot	.04	.51	.21	−.09	.77	.01	.06	.65	.03	.14	.54	.12	−.02	.63	.12	.05	.64	.12
29. Tantr	Temper tantrums	−.03	.73	.05	−.27	.91	.02	.05	.73	−.02	−.00	.83	−.04	−.04	.86	−.10	−.01	.74	.14
30. Threat	Threatens people	−.10	.63	.28	−.04	.68	.23	−.09	.73	.23	−.12	.72	.31	−.18	.81	.21	−.10	.54	.49
31. Loud	Loud	.32	.49	.01	.04	.80	−.29	.32	.55	−.02	.32	.47	.08	.17	.63	.03	.29	.60	−.09
Delinquency/Rule-Breaking
32. Guilt	Doesn’t feel guilty about misbehavior	.18	.26	.31	.11	.53	.03	.19	.30	.35	.11	.40	.26	.02	.39	.33	.06	.33	.37
33. Troub	Friends get in trouble	.22	.24	.31	.16	.43	.19	.16	.35	.32	.20	.15	.48	−.03	.10	.66	.04	−.00	.73
34. Lies	Lies or cheats	.16	.28	.45	.18	.58	.19	.11	.31	.54	.18	.21	.54	.08	.27	.58	.18	.17	.60
35. Oldk	Prefers older kids	.01	.39	.03	.09	.42	−.04	.18	.29	.01	.13	.28	.20	.02	.27	.21	.11	.23	.23
36. Runs	Runs away	−.38	.47	.43	−.00	.29	.59	−.06	.38	.28	-	-	-	.20	.11	.25	.09	−.02	.67
37. Fires	Sets fires	−.01	−.16	.75	.52	−.14	.56	.25	.01	.44	.13	.22	.43	.08	−.04	.70	.03	.01	.75
38. Stealh	Steals at home	.02	.00	.79	.36	.20	.51	−.01	.09	.80	.08	−.08	.78	−.00	−.02	.76	.15	−.03	.73
39. Stealo	Steals outside the home	.06	−.01	.78	.55	.00	.61	.05	−.04	.83	.00	−.13	.95	−.04	−.03	.82	.07	−.05	.85
40. Swear	Swears	−.07	.32	.43	−.03	.55	.31	.12	.42	.24	−.08	.48	.41	−.15	.39	.41	−.06	.35	.50
41. Sex	Thinks about sex too much	−.14	.41	.18	−.04	.41	.19	−.06	.50	.21	.08	.32	.23	.06	.45	.19	.16	.23	.29
42. Truan	Truant	-	-	-	.27	−.01	.66	-	-	-	-	-	-	−.06	.14	.61	−.00	−.16	.91
43. Drug	Alcohol/substance use	-	-	-	.72	−.13	.25	-	-	-	-	-	-	-	-	-	−.20	−.02	.94
44. Vand	Vandalism	.10	.28	.49	.50	.09	.51	−.06	.49	.46	−.10	.04	.90	.12	.08	.66	−.11	.11	.86

Note. Bolded items represent cross-loadings.

Are there sex differences in the latent structures of externalizing psychopathology?

Fit indices from the correlated three-factor exploratory structural equation models are shown in Supplementary Table S2. There was greater symptom endorsement for the male sample relative to the female sample, and thus, several further modifications had to be made to the female sample. Other problematic symptoms for the female sample included ‘runs away’, ‘sets fires’, ‘thinks about sex too much’, and ‘vandalism’. All respecified models were also associated with acceptable to excellent levels of fit. In both the male and female samples, the majority of factors were positively correlated with each other (see Supplementary Table S3).

Factor loadings are shown in Supplementary Tables S4 and S5 for the male and female samples, respectively. Similar to the combined sample, there was extensive heterogeneity in item cross-loadings. In brief, for the male sample, ‘can’t sit still’, ‘impulsive’, and ‘nervous’ cross-loaded onto factors 1 and 2 at multiple time points. However, several other items cross-loaded onto factors 1 and 2 only at single time points, including ‘easily jealous’, ‘showing off’, ‘bullying’, ‘sudden changes in mood’, ‘threatens people’, and ‘demands attention’. For the female sample, ‘can’t concentrate’, ‘can’t sit still’, ‘confused’, ‘nervous’, and ‘poor at school work’ cross-loaded onto factors 1 and either factor 2 or 3 at multiple time points. Across both male and female samples, there were also several items that cross-loaded onto factors 2 and 3.

Network Modeling

Does the network structure of externalizing psychopathology change at different time points?

Network plots for each time point are shown in Figure 1. In comparison to latent variable modeling, no items were removed in network modeling. In the following description, ‘–’ is used to represent network edges. In kindergarten, attention problems were strongly clustered together as indicated by the thickness and saturation of edges. In particular, ‘nervous’–‘twitches’, ‘clumsy’–‘poor school work’ and ‘day dreams’–‘stares blankly’ represented the strongest symptom pairs. A cluster of aggressive behavior items also showed strong connections, including ‘disobedient at home’–‘disobedient at school’–‘sullen’–‘temper tantrums’–‘screams’. In grade 1, similar attention problems were strongly clustered together, including ‘nervous’–‘twitches’ and ‘can’t concentrate’–‘can’t sit still.’ In addition, the attention problems symptom ‘impulsive’ showed a strong association with the aggressive problems symptom ‘showing off.’ Out of all networks, grade 2 appeared the least densely connected with lower edge weights overall. However, ‘can’t concentrate’–‘can’t sit still’ still demonstrated relatively high connectivity. Similarly, for grade 4, ‘can’t concentrate’–‘can’t sit still’ showed strong associations. In grade 5, ‘can’t concentrate’–‘can’t sit still’ and ‘day dreams’–‘stares blankly’–‘confused’ were strongly associated with each other, in addition to several aggressive behavior associations, ‘talks too much’–‘loud’–‘screams’ and ‘easily jealous’–‘demands attention’. Finally, in grade 7, ‘brags’– ‘showing off’’–‘teases a lot’ and ‘talks too much’–‘loud’ represented strong connections. Across all time points, ‘destroys own things’–‘destroys others’ things’, and ‘steals at home’–‘steals outside the home’ showed strong symptom pair associations. Overall, across all time points, aggressive behavior and delinquency/rule-breaking symptoms were quite integrated, but attention problems remained as a distinct cluster of symptoms.

Figure 1.

Network structure of externalizing psychopathology from kindergarten through grade 7 with the combined sample (items from the attention problems narrowband scale are depicted in white, items from the aggressive behavior narrowband scale are depicted in blue, and items from the delinquency/rule-breaking behavior narrowband scale are depicted in gray).

Overall connectivity or global strength statistics (i.e., sum of all correlations within a network) were similar across time with the exception of grade 1 and grade 7, which showed higher global strength relative to all other time points (kindergarten=18.06; grade 1=19.49; grade 2=18.45; grade 4=18.75; grade 5=18.77; grade 7=19.85). Findings from NetworkComparisonTests indicated a significant difference in global strength between kindergarten and grade 1 (p=.003), grade 1 and grade 2 (p=.045), and between grade 5 and grade 7 (p=.009). There were no other significant differences between adjacent time points.

Are there distinct central symptoms of externalizing psychopathology at different time points?

Strength centrality statistics (i.e., sum of correlations between a node and other nodes) and bootstrapped significant differences are displayed in Supplementary Figures S1 and S2, respectively. In sum, for kindergarten, ‘destroys others’ things’ showed the greatest strength centrality, whereas for grade 1 the ‘impulsive’ symptom was strongest. In grade 2, ‘loud’ and ‘destroys others’ things’ showed the greatest strength centrality. Conversely, for grade 4, ‘can’t sit still’, ‘lies or cheats’, and ‘impulsive’ represented higher strength. There was some similarity across the last two time points, with ‘can’t concentrate’ depicting the greatest strength in grade 5, and ‘temper tantrums’ and ‘can’t concentrate’ in grade 7. Across all time points, ‘alcohol/substance use’ showed the lowest strength centrality, with the exception of grade 7, in which ‘clumsy’ showed the lowest strength.

Which symptoms bridge between attention problems, aggressive behavior, and delinquency/rule-breaking at different time points?

Across all time points, ‘impulsive’ and ‘lies or cheats’ showed the greatest bridge strength, and ‘alcohol/substance use’ showed the lowest bridge strength (see Supplementary Figure S3).

Are the network structures of externalizing psychopathology stable at each time point in development?

Stability of strength centrality is displayed in Supplementary Figure S4. The CS indicated that all time points showed adequate stability of strength centrality (i.e., CS>.50; kindergarten=.67; grade 1=.67; grade 2=.67; grade 4=.75; grade 5=.75; grade 7=.75).

Are there sex differences in the network structures of externalizing psychopathology?

Network plots for the male and female samples for each time point are shown in Supplementary Figures S5 and S6, respectively. In brief, for the male sample, global strength statistics were as follows: kindergarten=17.52; grade 1=18.80; grade 2=18.27; grade 4=18.80; grade 5 = 18.45; grade 7 = 19.60. Invariance testing indicated a significant difference in global strength between grade 5 and grade 7 (p=.013). For the female sample, global strength statistics were as follows: kindergarten=18.08; grade 1=19.17; grade 2=18.54; grade 4=18.17; grade 5=18.49; grade 7=19.43. There were no significant differences between adjacent time points. These findings suggest that networks were mostly comparably connected at different ages. Invariance testing also indicated that there were no significant differences across the sex-specific samples at equivalent time points (e.g., male kindergarten vs. female kindergarten).

With regard to strength centrality, for the male sample, whereas ‘lies or cheats’ and ‘impulsive’ mostly represented the greatest strength in kindergarten through grade 4, ‘can’t concentrate’ and ‘temper tantrums’ showed the greatest strength in grade 5 and grade 7, respectively (see Supplementary Figures S7 and S9). For the female sample, whereas ‘can’t sit still’, ‘loud’, and ‘destroys others’ things’ typically represented the greatest strength in kindergarten through grade 4, ‘lies or cheats’ showed the greatest strength in grade 5 and grade 7 (see Supplementary Figures S8 and S10). Similar to the full sample, ‘alcohol/substance use’ showed the lowest strength centrality across most time points.

With regard to symptoms providing the greatest bridge to other dimensions, similar to the full sample, ‘impulsive’ and ‘lies or cheats’ showed the greatest bridge strength for both the male and female samples (see Supplementary Figures S11 and S12).

Finally, for the male sample, within sample stability of strength centralities were as follows: kindergarten=.36; grade 1=.44; grade 2=.44; grade 4=.59; grade 5=.59; grade 7=.59. For the female sample, stability of strength centralities are as follows: kindergarten=.28; grade 1=.44; grade 2=.44; grade 4=.44; grade 5=.52; grade 7=.52 (see Supplementary Figures S13 and S14).

Discussion

Latent variable and network perspectives represent two, complementary theoretical and statistical frameworks for understanding psychopathology. By applying both perspectives to symptoms comprising attention problems, aggression, and rule-breaking in the same sample followed across an important era of development, we gained a more comprehensive view of the structure of externalizing psychopathology. Latent variable modeling, using an exploratory structural equation modeling procedure, supported a correlated three-factor model representing externalizing psychopathology. We found acceptable to excellent levels of model fit across all time points; however, there were multiple cross-loadings and heterogeneity in terms of which items cross-loaded across time points, i.e., high factor loadings with a factor other than an item’s primary factor, such as an item that helps define the attention problems dimension also loading on aggressive behavior or delinquency/rule-breaking.

Network modeling, using network visualizations, global strength invariance, strength centrality, bridge strength, and stability of strength centrality, gave a complementary picture of how externalizing psychopathology is structured. In general, network plots depicted symptoms comprising attention problems and aggressive behavior with stronger associations (i.e., greater edge weights) relative to delinquency/rule-breaking symptoms. In addition, there were significant differences in overall connectivity (i.e., global strength) at early (kindergarten vs. grade 1, grade 1 vs. grade 2) and later (grade 5 vs. grade 7) time points. Strength centrality predominantly supported symptoms of attention problems and aggressive behavior as central to externalizing psychopathology over time, and ‘impulsive’ and ‘lies or cheats’ consistently displayed the greatest bridge to other externalizing dimensions over time. We also identified high stability of the network structures at each time point for the full sample, which suggests that the network solutions were robust in the face of variations in which subsets of participants’ data were modeled. We next comment on important details of findings from the latent variable and the network models, as well as on sex differences in aspects of the models.

Latent Variable Theory and Modeling

We applied a correlated three-factor measurement model built around the standard dimensions of attention problems, aggressive behavior, and delinquency/rule-breaking. Our model provided almost uniformly excellent fit across time points and samples. However, consistent with expectations based on children’s behavioral repertoires as they develop from age 5 or 6 to age 12 or 13, some items from the delinquency/rule-breaking scale had zero or low variability (e.g., ‘alcohol/substance use’) and were removed from further analyses. Symptoms of oppositionality and aggression are salient from early childhood, while some symptoms of delinquency become salient only in adolescence (Hart et al., 1995; Moffitt, 1993). Our dataset included only early adolescence. Also worth noting is our finding that the three factors were correlated with one another across most time points and samples, which supports our supposition that attention problems be considered in the general domain of externalizing behavior, even though attention problems are often considered as a distinct dimension to externalizing psychopathology in the age 5–13 era.

Although our latent measurement models provided support for three distinct but correlated factors, multiple cross-loadings were observed. The identification of multiple cross-loadings extends past research that has used confirmatory factor analytic approaches (e.g., Gomez & Vance, 2014; Tabet et al., 2021) and demonstrates the added value of exploratory structural equation modeling for a more nuanced study of multidimensional constructs. Although it is possible to include cross-loadings in confirmatory models, inclusion is limited and strong a priori hypotheses are needed (Sellbom & Tellegen, 2019). Despite the fact there was diversity with regard to which symptoms cross-loaded on factors and we did not identify a developmental pattern, there was some commonality in cross-loading items in the combined sample and male and female samples. For example, several items that typify attention problems, such as ‘can’t sit still’, ‘impulsive’, and ‘nervous’, cross-loaded onto factors represented by aggressive behavior and delinquency/rule-breaking. In addition, several items that are subsumed under aggressive behavior on the CBCL, such as ‘demands attention’, ‘showing off’, ‘talks too much’, and ‘loud’, cross-loaded onto the attention problems scale. Using recommendations by statisticians, several of these items meet loading cutoffs for possible deletion or respecification in factor analytic modeling (Costelle & Osborne, 2005; Matsunaga, 2010). However, as described in the next section on network theory and modeling, many of these items that may be removed through factor analytic methods were identified as most central to the networks and/or bridged dimension clusters (e.g., ‘impulsive’); thus, representing possible critical items in the emergence of childhood externalizing psychology. Our cross-loading findings also lend support to theory and research suggesting that attention problems may play a role in the expression of externalizing problems at multiple points in early development (Beauchaine et al., 2010, 2017).

Network Theory and Modeling

A particular advantage of network modeling of symptoms is that it allows for the identification of associations between symptoms. Overall, there was a similar pattern in the network plots across time points. Relative to delinquency/rule-breaking symptoms, symptoms comprising attention problems and aggressive behavior were more strongly associated within clusters as illustrated by more densely connected nodes and higher weight loadings of the edges. To illustrate, the attention problems symptoms ‘can’t concentrate’–‘can’t sit still’ showed strong associations, as did specific aggressive behavior symptom associations such as ‘disobedient at home’–‘disobedient at school’–‘sullen’–‘temper tantrums’–‘screams’. These findings provide a greater level of specificity for understanding how childhood externalizing psychopathology is expressed. In addition, the attention problems symptom ‘impulsive’ showed a strong association with the aggressive problems symptom ‘showing off.’ This finding may suggest a pathway through which impulsivity leads to aggressive behavior among children. Although the present study focused on delineating the structure of externalizing psychopathology, we also identified specific symptom associations that could reflect internalizing psychopathology processes. For example, ‘nervous’–‘twitches’ and ‘day dreams’–‘stares blankly’–‘confused’ showed strong associations. Further research is needed to understand how these symptom associations may promote both externalizing and internalizing psychopathology in children and adolescents. Such evidence of interplay between specific symptoms provides unique information beyond what we see with latent variable modeling.

We found significant differences in overall connectivity at earlier (kindergarten vs. grade 1, grade 1 vs. grade 2) and later (grade 5 vs. grade 7) time points. These findings underscore the importance of a developmental understanding of externalizing psychopathology across childhood and into adolescence. Significant differences at these time points may reflect important transition points. The kindergarten to grade 1 transition is a formative developmental stage characterized by changes in children’s environment, biological, social-emotional, and behavioral functioning (Goulter et al., 2021; Rimm-Kaufman & Pianta, 2000). Similarly, the transition from grade 5 through grade 7 represents the earliest phases of adolescence, with its own changes in children’s functioning (Steinberg & Morris, 2001). It would be important for future research to extend current findings by investigating the network structure throughout adolescence and into early adulthood representing additional important transition stages. Invariance in overall connectivity was identified at all other adjacent time points, supporting some research suggesting that the externalizing spectrum is largely stable over childhood and into adolescence (King et al., 2018). Importantly, with the exception of grade 5 versus grade 7 in the male sample, there were no significant differences between adjacent time points in the overall connectivity among the items in either the male or the female subsamples.

Although we did not identify a clear developmental pattern of central symptoms over time, there was consistency in symptoms showing greatest strength centrality. Several symptoms of attention problems (‘impulsive’, ‘can’t sit still’, ‘can’t concentrate’) and aggressive behavior (‘destroys others’ things’, ‘loud’, ‘temper tantrums’) were identified as central to externalizing psychopathology over time, with one delinquency/rule-breaking symptom (‘lies or cheats’) also showing high importance. Conversely, ‘alcohol/substance use’ consistently showed the lowest strength centrality—a finding that is also likely related to developmental stage. Our findings extend past cross-sectional network studies that have also identified attention problems and aggressive behavior as central to externalizing psychopathology during childhood (Hukkleburg, 2019; Martel et al., 2017). In addition, we found some sex-specific differences in strength centrality, such that in the male sample there was greater covariation of attention problems (e.g., ‘impulsive’) and delinquency/rule-breaking (e.g., ‘lies or cheats’) with other symptoms in early years and aggressive behavior (e.g., ‘temper tantrums’) in later years, whereas in the female sample there was greater covariation of attention problems (e.g., ‘can’t sit still’) and aggressive behavior (e.g., ‘loud’) with other symptoms in early years and delinquency/rule-breaking (e.g., ‘lies or cheats’) in later years. These findings support past research suggesting that certain externalizing symptoms may be perceived as more normative for specific sexes (King et al., 2018).

‘Impulsive’ was identified as a consequential bridging symptom across multiple time points and samples. Trait impulsivity has many definitions in the literature; however, it is commonly viewed as an individual difference expressed as deficient self-control, a preference for immediate over delayed reward, or a failure to plan ahead (Beauchaine et al., 2017). Characterized by well-defined neurobiological substrates (including high genetic heritability and an underresponsive mesolimbic dopamine system; Neuhaus & Beauchaine, 2017), impulsivity has been identified as a transdiagnostic liability for multiple externalizing psychopathologies (Beauchaine et al., 2017). Our findings provide further support for impulsivity as a transdiagnostic symptom possibly predisposing children to externalizing psychopathology. However, impulsivity is multifaceted, and further network research with more comprehensive measures assessing a larger array of impulsivity symptoms is needed (e.g., UPPS; Whiteside et al., 2005). Although highly heritable, certain pathogenic environmental experiences such as maltreatment have also been shown to compromise cortical functioning involved in impulse-control and emotion regulation, thereby predicting further development in externalizing psychopathology (Beauchaine et al., 2017). Thus, treatments targeting both temperamental trait impulsivity and adverse environmental factors have the potential to mitigate the development of multiple dimensions of externalizing psychopathology.

In addition to ‘impulsive’, ‘lies or cheats’ was also identified as a symptom that bridges externalizing dimensions at many time points and across samples. Although understudied, developmental researchers have suggested that childhood dishonesty can be used to understand cognitive, social, and moral competencies (Talwar & Crossman, 2012). Compared with typically developing children, children with disruptive behavior disorders (e.g., ODD, CD) are more likely to engage in lying (Burke et al., 2002) and this is more likely for personal gain in early childhood and to conceal wrongdoing in later childhood (Mugno et al., 2019). Lying is also predictive of future delinquent and criminal behavior (Frick, 2012). Further, children characterized as cheaters were rated by parents and teachers as displaying high levels of externalizing problems and impulsivity (Callender et al., 2010). Current findings suggest that in addition to impulsivity, dishonest behavior may also be a critical factor underpinning multiple forms of childhood externalizing psychopathology and should be routinely considered in assessment and treatment.

In the full sample, network structures were stable at each time point in development, demonstrating that the network solutions were robust across subsets of participants. However, in the male sample only the later time points were stable, and in the female sample only one time point reached the recommended cutoff (CS >.50) for stability. Past research examining the network structure of broader externalizing psychopathology has not examined sex differences, and thus, our study represents the first effort to test whether the network structure of attention problems, aggressive behavior, and delinquency/rule-breaking generalizes across sex-specific samples. Our findings, however, do support other research that also identified lower stability in sex-specific samples for other externalizing dimensions, including CD and callous-unemotional traits (Goulter & Moretti, 2021). Deficient network stability may be related to heterogeneity in endorsement of symptoms within sex-specific samples, although we suggest caution in the interpretation of these findings until replicated.

Theoretical and Empirical Implications

The aim of the present study was to comprehensively assess the structure of childhood externalizing symptoms at multiple time points by testing two conceptual frameworks: latent variable (or common cause) and network. Throughout this paper we have compared these perspectives; however, as others have recently highlighted, these approaches need not be mutually exclusive (Bringmann & Eronen, 2018; Roefs et al., 2022). For example, as described by Roefs et al. (2022), proximal causes or direct mechanisms in the development and maintenance of psychopathology may be represented by a specific node or edge in the network, whereas other indirect mechanisms external to the symptom network (e.g., genetic and environmental factors) could be characterized as distal or common causes. Importantly, from a psychometric perspective, we examined each time point separately, and thus, we have avoided causal language throughout. However, future work using time series data combining both latent variable/common cause and network frameworks to map the dynamic causal connections of complex networks hold promise for advancing our understanding of the development of psychopathology (Epskamp et al., 2017; Fried, 2020).

Network theory focuses on lower-order parts (e.g., symptoms, traits, or other elements). This symptom-oriented approach has also been emphasized in recent research initiatives for advancing the mental health field, including the Hierarchical Taxonomy of Psychopathology (HiTOP; Kotov et al., 2017, 2021). HiTOP is an empirically derived descriptive system arranging continua of psychopathology from normative to maladaptive in a hierarchy, including superspectra, spectra, subfactors, syndromes, and symptoms/traits (Kotov et al., 2017, 2021). As currently constructed in HiTOP, externalizing psychopathology encompasses two spectra: antagonistic and disinhibited (Krueger et al., 2021). The antagonistic externalizing spectrum includes the propensity to navigate interpersonal situations with conflict, callousness, and hostility. The disinhibited externalizing spectrum includes tendencies to act impulsively with little regard for consequences. Both externalizing spectra also comprise antisocial behavior, such as aggression. Our approach of including impulsive and attention problems within externalizing psychopathology aligns with HiTOP, and we add to the HiTOP literature with our focus on childhood. HiTOP is currently based on the adult literature; however, the Developmental Workgroup plans to form a youth-adapted version.

Strengths and Limitations

Strengths of the present study include its multisite, longitudinal design and its use of both factor analytic and network analytic methods with multiple timepoints to infer the structure of externalizing psychopathology. One limitation of the study is that data are from only parent reports. Further research using multiple informants, including teacher- and youth-, in addition to parent-report data will provide more comprehensive models of the development of psychopathology. A second limitation is that our sample, although representing considerable demographic diversity, does not well represent the full array of cultural and economic diversity in our population. We are particularly interested in future research that is able to explore possible differences in the structure of externalizing in lower- versus higher-risk samples.² Examining the structure of externalizing psychopathology in a higher-risk population is particularly important given our sample largely comprised population-representative community children and showed low endorsement of some symptoms (e.g., ‘alcohol/substance use’). Elucidating the psychometric structure and specific item associations of these more severe symptoms would be important for identifying those children at higher risk of a severe and persistent externalizing psychopathology developmental pathway. A third limitation to bear in mind is that network psychometrics has only recently begun to be applied to psychopathology and there is ongoing debate among some researchers on best methodolgies (Borsboom et al., 2017; Forbes et al., 2017a, 2017b).

Conclusion

To conclude, the present findings extend previous externalizing research on higher-order factors. Our results illustrate how two methods—latent variable and network modeling—provide important and complementary information on key multidimensional constructs. Findings from exploratory structural equation modeling add to current understanding on the CBCL factor structure that has typically excluded attention problems. Results also showed that multiple items had high cross-loadings at every assessment point, suggesting that factor analytic approaches may be neglecting critical information for conceptualizations of psychopathology. Our network analytic approach provides new insights into the structure of childhood externalizing psychopathology. Network analyses provide easy to interpret visualizations of data and greater specificity in associations between symptoms. We observed specific symptom associations within and across the a priori dimensions of attention problems, aggression, and rule-breaking/delinquency; thus, providing unique information beyond latent variable modeling. We also found that whereas critical transition points (kindergarten → elementary school; childhood → adolescence) characterized times of change in the expression of externalizing psychopathology, attention problems and aggressive behavior were central throughout this developmental period. Finally, symptoms representing impulsivity and dishonest behavior may confer transdiagnostic vulnerability for broad externalizing problems across early development.

Public Health Significance:

Externalizing psychopathology in childhood and adolescence contributes significant societal burden. In this article, we applied both latent variable and network modeling to determine the structure of attention problems, aggressive behavior, and delinquency/rule-breaking, in addition to examining the associations between symptoms. Our findings reveal key symptoms, critical transition points, and possible transdiagnostic liabilities.

Data and Code Availability:

Data and code for this study are available by emailing the corresponding author. This study was not preregistered.