Developmental Trajectories of Internalizing and Externalizing Symptoms in Youth and Associated Gender Differences: A Directed Network Perspective

Dimensional assessment of anxious/depressed, withdrawn/depressed, somatic complaints, social problems, thought problems, attention problems, rule-breaking behavior, and aggressive behaviors syndrome scales, representing symptom clusters, were assessed with the Child Behavior Checklist Parent’s Report Form (CBCL; Achenbach, 2001). On the CBCL, anxious/depressed, withdrawn/depressed, and somatic complaints are grouped as internalizing problems, while rule-breaking and aggressive behaviors are grouped as externalizing problems. T-scores normed by sex, age, and ethnicity were used for the present study analyses. T-scores had a lower bound of 50, representing 50th percentile or below, and an upper bound of 100, representing above 99th percentile. CBCL syndrome scales demonstrated one week test–retest reliability ranging from 0.80 to 0.94 (Achenbach, 2001). These syndrome scales have demonstrated concurrent validity with clinical diagnoses of anxiety disorders, mood disorders, attention-deficit/hyperactivity disorder, oppositional-defiant disorder, and conduct disorder (Ebesutani et al., 2010; Eiraldi et al., 2000; Kasius et al., 1997; Seligman et al., 2004).

GVAR models for panel data (Panel GVAR) were used to examine longitudinal relationships between scores on CBCL syndrome scales, representing symptom clusters, across baseline, 1-year follow-up, and 2-year follow-up time points. Panel GVAR models illustrate how CBCL syndrome scale scores influence each other and themselves across time at the within-person level while controlling for between-person differences in these scales. Additionally, Panel GVAR constrains the effects of the syndrome scale scores so that they are stable across the three timepoints in order to assess for stable effects across timepoints rather than deviations between timepoints.

Panel GVAR models were computed using full information maximum likelihood estimation. First, the full model was estimated, then a model search procedure was used to maximize the Bayesian information criterion (BIC) by pruning edges that were not statistically significant at the p < 0.05 level and adding edges that were significant at the p < 0.05 level. The comparative fit index (CFI; Bentler, 1990), Tucker-Lewis index (TLI; Tucker & Lewis, 1973), and root mean square error of approximation (RMSEA; Steiger & Lind, 1980) were calculated to assess model fit. Centrality indices were then calculated for each syndrome scale score in the final model. Instrength centrality represents the degree to which syndrome scales are predicted by scores on other scales at the previous timepoint, while outstrength centrality represents to what degree syndrome scales predict scores on other scales at the next timepoint.

Similarly, bridge in-degree centrality represents the degree to which each syndrome scale in one community (i.e., internalizing or externalizing) is predicted by syndrome scales in the other community at the previous timepoint, while bridge out-degree centrality represents to what degree syndrome scales in one community predict scores on scales in the other community at the next timepoint. Gender differences in network structures were tested by estimating the Panel GVAR model as a multi-group model, constraining parameters to be equal across gender groups, and assessing significance of change in model fit. Of note, the 13 participants who identified as transgender/other or for whom gender identity was not known were not included in the multi-group model separating male and female groups, though they were included in the full Panel GVAR model consisting of all 6,414 participants. More details regarding specifics of Panel GVAR models can be found in Epskamp (2020). Given concerns regarding the stability of network models (Forbes et al., 2019), the robustness of the estimated networks was tested by applying the same model search procedure for Panel GVAR models of 1,000 non-parametrically bootstrapped samples and calculating how often each edge was included in the optimal model, with 50% inclusion probability being considered robust (see Betz et al., 2020). Stability of centrality estimates was determined using case-dropping subset bootstrap, in which 20% of the sample was randomly dropped and the model was re-estimated across 1,000 iterations, and calculating 95% bootstrapped confidence intervals (see Epskamp et al., 2018a, b). All analyses involving Panel GVAR models were conducted using the R package “psychonetrics,” version 0.8.1. R code for study analyses presented here is available on the Open Science Framework (https://​osf.​io/​fcuhm/​?​view_​only=​54cc8a191eed4da9​a5a576bc35ef8c5b​).

Descriptive statistics for each of the CBCL subscales are described in Table S2. Each of the CBCL subscales were positively skewed and leptokurtic, as expected given the CBCL subscales’ restricted lower bound.

The temporal network (i.e., the network of temporal effects of the eight CBCL syndrome scales on themselves and each other across the three timepoints) was estimated (see Fig. 1a). All significant edges included in the final model are shown in Table 1. The final model demonstrated good fit (CFI = 0.991, TLI = 0.989, RMSEA = 0.0269). All autoregressive effects, representing predictive associations between a CBCL syndrome scale score and the same syndrome scale score at the next timepoint, were significant and positive. This indicates that higher scores on one scale significantly predict higher scores on the same scale at the next timepoint when controlling for the effects of all seven other scales. Centrality estimates for each scale are represented in Fig. 2. Outstrength centrality was highest for the thought problems scale, while social and anxious/depressed problems demonstrated the greatest instrength centrality. Withdrawn/depressed problems and rule-breaking behaviors showed the greatest bridge in-degree centrality and bridge out-degree centrality, demonstrating that they accounted for the strongest predictive relationships between internalizing and externalizing problem communities.

The robustness of the final model to sampling variation was examined by applying the same model search procedure to 1,000 bootstrapped samples and calculating the number of times each edge was included in the final model. All robust autoregressive effects and edges that were included in more than 50% of the models are shown in Fig. 1b. Inclusion probabilities of all robust autoregressive effects and edges are reported in Table 1. Of the 20 significant edges and autoregressive effects in the temporal network from the final model derived from the original sample, 18 were found to be robust in the models derived from the 1,000 bootstrapped samples. Two significant edges were not robust, representing potential false positives. Additionally, the Panel GVAR models of the bootstrapped samples produced three edges that were robust, but they were not included in the final model derived from the original sample, representing potential false negatives.

Two separate temporal networks of the eight CBCL syndrome scales were estimated for participants who identified as female (n = 3,035) and male (n = 3,366). The temporal network for girls is shown in Fig. 1a, and all significant edges included in the final model are shown in Table 2. The final model for girls demonstrated good fit (CFI = 0.988, TLI = 0.985, RMSEA = 0.0309). All autoregressive effects were significant and positive. Centrality estimates for each scale included in the temporal network for girls are represented in Fig. 2. Outstrength centrality was highest for thought problems, while instrength centrality was highest for the anxious/depressed scale. Anxious/depressed problems showed the greatest bridge in-degree centrality, while aggressive behaviors demonstrated the greatest bridge out-degree centrality.

The temporal network for boys is shown in Fig. 1a, and all significant edges included in the final model are shown in Table 2. The final model for boys demonstrated good fit (CFI = 0.988, TLI = 0.986, RMSEA = 0.0315). All autoregressive effects were significant and positive, with the exception of the autoregressive effect of anxious/depressed problems on itself at the next timepoint, which was not significant. Centrality estimates for each scale included in the temporal network for boys are represented in Fig. 2. Outstrength centrality was highest for rule-breaking behaviors, while attention problems demonstrated the lowest outstrength centrality with no significantly predictive relationships on other scales at later timepoints. Somatic problems demonstrated the highest instrength centrality. Internalizing symptom clusters showed overall greater bridge in-degree centrality than externalizing symptom clusters, while rule-breaking behaviors demonstrated the highest out-degree bridge centrality.

Robustness of the final models estimated from female and male participants were examined using the same bootstrapping strategy described above. All robust autoregressive effects and edges that were included in more than 50% of the models for girls and boys are shown in Fig. 1b. Inclusion probabilities of all autoregressive effects and edges that were robust in the female and/or male temporal networks are reported in Table S3. Of the 24 significant edges and autoregressive effects in the temporal network from the final model derived from the female participants, 21 were robust in the models derived from the 1,000 bootstrapped samples. Three significant edges were not robust, representing potential false positives. Additionally, bootstrapped models produced two edges that were robust but not included in the final model derived from the original female sample, representing potential false negatives.

Of the 22 significant edges and autoregressive effects in the temporal network from the final model derived from the male participants, 21 were robust in the models derived from the 1,000 bootstrapped samples. One edge was not robust, representing a potential false positive finding. Additionally, bootstrapped models produced one autoregressive effect and six edges that were robust but not included in the final model derived from the original male sample, representing potential false negatives.

Of the 35 total autoregressive effects and edges that were robust in either the bootstrapped female or male temporal networks, only eight autoregressive effects and nine edges were robust for both girls and boys (see Table S3). Additionally, three edges that were robust for both girls and boys differed in directionality (i.e., positive or negative) between boys and girls, specifically the predictive relationships of rule-breaking behavior on anxious/depressed problems, rule-breaking behavior on thought problems, and aggressive behaviors on somatic complaints.

When estimating the Panel GVAR model as a multi-group model, constraining all auto-regressive effects and edges to be equal across gender groups led to a significant drop in model fit (Δχ²(64) = 139.97 [p < 0.0001]), indicating that temporal network structure is unlikely to be equivalent overall. One auto-regressive effect and 10 edges were found to lead to a significant drop in model fit when individually constrained to be equal across gender groups (see Table 2). Notably, predictive relationships of externalizing problems (i.e., rule-breaking and aggressive behaviors) with other symptom clusters were unlikely to be equivalent across groups. Further examination of these edges across genders shows that, while externalizing problems tended to negatively predict other symptom cluster for girls, externalizing problems positively predicted other symptom clusters for boys (see Table 2). This is also apparent in comparing bridge out-degree centrality for rule-breaking behavior in which the magnitude of predictive relationships of rule-breaking behaviors on internalizing symptoms is substantially higher for boys than girls (see Fig. 2c).

The present study shows that outstrength centrality is relatively high for thought problems. These results differ from past findings that depression and worry are most predictive of other symptoms at future timepoints (Funkhouser et al., 2021). Funkhouser et al. (2021) as well as Speyer et al. (2022) found attention problems to be highly predictive of other symptoms at future timepoints, a finding that was not replicated in the present study, with attention problems having the lowest outstrength centrality. Interestingly, Speyer et al. (2022) theorized that behavioral difficulties in ADHD, specifically hyperactivity and inattention, may lead to increased struggle with peer interactions and exclusion by peers, potentially leading to antisocial behavior and conduct problems. Our results provide evidence of effects in the opposite direction, with aggressive problems positively predicting later attention problems, particularly for boys. Importantly, Speyer et al. (2022) analyzed children with a broader age range, and it is possible that the results of the current study are indicative of a pattern unique to the age range of the current sample, nine to twelve year-olds.

Bridge symptoms in our temporal networks are represented by symptom clusters that show predictive relationships between internalizing and externalizing problems. Previous literature investigating cascade models has yielded mixed results with some studies showing bidirectional relationships between internalizing and externalizing problems (e.g., Achenbach & Rescorla, 2000; Mesman et al., 2001) and some work showing more restricted unidirectional effects of externalizing problems on future internalizing problem (e.g., Masten & Cicchetti, 2010; D’urso & Symonds, 2022). Results of the present study are generally supportive of a more unidirectional effect; externalizing problem symptom clusters had overall greater bridge out-degree centrality than internalizing problem symptom clusters, while internalizing problems had greater bridge in-degree centrality than externalizing problems. However, a significant and robust positive predictive relationship of withdrawn/depressed problems on rule-breaking behaviors at later timepoints was also found. Additionally, these findings are consistent with past research that has found depression to be a bridge symptom between internalizing and externalizing disorders (McElroy et al., 2018a, b).

Several differences in the patterns of predictive relationships between symptom clusters across female and male participants were noted. For symptom centrality, anxious/depressed problems showed the highest outstrength centrality for girls while showing low outstrength centrality for boys. Additionally, rule-breaking behaviors showed the highest outstrength centrality for boys but relatively lower outstrength centrality for girls. Another key difference was that externalizing problems negatively predicted other symptom clusters at later timepoints for girls, while externalizing problems positively predicted other symptom clusters at later timepoints for boys. The finding of a negative longitudinal relationship between externalizing problems and other symptom domains in girls runs contrary to the adjustment erosion hypothesis (Moilanen et al., 2010), which theorizes that early externalizing problems interfere with academic progress and peer relations leading to development of internalizing problems. Interestingly, Panayiotou and Humphrey (2018) found a similar pattern of initial externalizing problems predicting less internalizing problems later on, though this pattern was true for both boys and girls. They hypothesized that engagement in externalizing behaviors, particularly with delinquent peers, may lead to greater self-esteem by alleviating internalizing symptoms or that externalizing behaviors may trigger more attention from teachers which acts as a protective factor. It is possible that these hypothesized mechanisms may function differently between male and female identifying participants in the present study, leading to differences in effects of early externalizing behaviors. For example, rule-breaking behaviors positively predicted later social problems, specifically for boys but not for girls in our sample, and social problems positively predicted later internalizing problems. Therefore, male identifying participants in our sample showed a pattern consistent with the adjustment erosion hypothesis: early externalizing problems, specifically rule-breaking behaviors, predicted greater social problems, including problems with peer relations, that in turn positively predicted later internalizing problems, specifically anxious/depressed problems. In contrast, female identifying participants showed an opposite pattern from what would be expected by the adjustment erosion hypothesis: early externalizing problems, specifically aggressive and rule-breaking behaviors, predicted fewer future social and internalizing problems. This suggests that the mechanisms hypothesized by Panayiotou and Humphrey (2018) are more influential for girls rather than boys, leading to observed differences in effects of early externalizing behaviors, though factors such as self-esteem and attention from teachers were not measured, limiting conclusions.

Additionally, within externalizing problems, out-degree bridge centrality of aggressive behaviors was greater than rule-breaking behaviors for girls, but the opposite was true for boys. Additionally, internalizing problems, specifically withdrawn/depressed problems, positively predicted future externalizing problems, specifically rule-breaking behaviors in both female and male participants. However, the positive predictive relationship of withdrawn/depressed problems on rule-breaking behaviors was not robust for female participants. Taken together, these results indicate that developmental cascades between internalizing and externalizing problems may be more bidirectional for boys but more unidirectional for girls. Several explanations exist for mechanisms through which externalizing behaviors can lead to internalizing problems. For example, the irritable depression model hypothesizes that depressed mood can manifest as irritability which leads to conduct problems over time (Wolff & Ollendick, 2006). Interestingly, there is some evidence that irritability presents more commonly as a symptom of depression in both adult men and boys compared to women and girls (Khesht-Masjedi et al., 2017; Winkler et al., 2006). This suggests that the mechanistic pathway proposed by the irritable depression model may be more applicable to boys compared to girls, potentially explaining the presence of a robust positive longitudinal effect of internalizing problems on future externalizing problems in boys, but not in girls, in our sample.

Though Panel GVAR models can accurately describe predictive relationships in panel data with as little as three timepoints (see Epskamp, 2020), the inclusion of additional timepoints beyond the 2-year duration of the current study is necessary to test the consistency of the patterns of predictive relationships found in the present study. Similarly, the low number of measurement occasions may explain the suboptimal stability of centrality estimates when applying case-drop subset bootstrapping. Additionally, there is evidence that patterns of developmental cascades change during development (Moilanen et al., 2010), and GVAR models’ assumptions of equal relations over time limit ability to detect changes in developmental patterns. Therefore, though GVAR models in this study found effects between symptom clusters that were consistent across all three timepoints, these effects may not be generalizable beyond the age group considered by this study. As the ABCD study continues gathering more data and participants continue developing, it will be important to examine whether patterns of relationships among symptom clusters change during different stages of preadolescence and adolescence. Additionally, although the focus on the transdiagnostic symptom clusters represented by the CBCL syndrome scales allows for conclusions about broad patterns in development of psychopathology, the lack of granularity in terms of specific symptoms limits conclusions about which specific symptoms within a given symptom cluster serve as central symptoms or bridge symptoms. The use of parent-report on the CBCL is also a limitation as parent-report may not reflect participants’ symptomatology, particularly given low parent–child concordance on internalizing symptoms (Rey et al., 1992). Finally, past reported psychopathology and gender effects are only two contributors to psychopathology trajectories, and as past literature has stated, consideration of how protective and risk factors also contribute to the development of psychopathology in youth is necessary (Fried et al., 2017; Goh & Martel, 2021). Possible protective and risk factors include school engagement and prosocial behaviors as well as abuse or neglect.

Though recruitment for the ABCD study was intended to be nationally representative, there are certain limitations to the generalizability of this sample to the general U.S. population (Heeringa & Berglund, 2020). Notably, when compared to the American Community Survey (ACS) conducted by the U.S. Bureau of Census, 40.7% of the ABCD sub-sample included in the present study had an annual combined family income greater than $100,000 compared to 17.3% in the ACS. As previously reported, the ABCD sample also underrepresents children attending schools in rural settings (Heeringa & Berglund, 2020). Given these limitations, caution should be exercised in generalizing these findings to the broader U.S. population.

Symptom clusters with high outstrength centrality and bridge out-degree centrality can represent causal mechanisms in the development of psychopathology in youth (Borsboom & Cramer, 2013; Cramer et al., 2010). Therefore, focused assessment of these symptom clusters can help identify youth who are at risk for developing future psychopathology, and treatment of these symptom clusters may prevent development of future comorbid problems. Based on the total sample panel GVAR model, anxious/depressed problems and aggressive behaviors represent highly influential symptom clusters, while withdrawn/depressed problems and rule-breaking behavior represent bridge symptoms between externalizing and internalizing problems. Early screening and intervention for these symptom clusters may be important for prevention of future psychopathology. However, female participants demonstrated a different pattern, where anxious/depressed problems and aggressive behavior were still influential symptom clusters, but aggressive behavior negatively predicted future increases in other symptom clusters. Additionally, only withdrawn/depressed problems represented a bridge symptom between externalizing and internalizing problems for female participants. Male participants also demonstrated a different pattern, where rule-breaking behaviors were the influential symptom cluster rather than anxious/depressed problems and aggressive behavior. However, rule-breaking behavior and withdrawn/depressed problems represented bridge symptoms between externalizing and internalizing problems for male participants. Our results imply that screening and early interventions targeting anxious/depressed problems and withdrawn/depressed problems for girls, and rule-breaking behavior and withdrawn/depressed problems in boys, on average, could have broad effects in the reduction and/or prevention of psychopathology. Lastly, the symptom trajectories yielded here, and associated gender differences may provide promising pathways for understanding disorder (dis)continuity and co-occurrence, although longer-term longitudinal data are needed to test pathway stability and divergence across development. Still, caution should be exercised to avoid over-interpretation of these results. Present results only reflect within-person temporal relationships for the average participant, and individualized within-person longitudinal modeling will yield more precise information on symptom centrality and causal structure that is better suited for the treatment of an individual patient (Epskamp et al., 2018; Goh & Martel, 2021).