Within-person Relations Between Social Skills and Symptoms of Oppositional Defiant Disorder and Conduct Disorder from Preschool to Adolescence – A Birth Cohort Study

A vast majority of research on behavior problems has been carried out using checklists or rating scales completed by parents or teachers. Although valuable, parent and teacher reports may be subject to errors and biases (Bennett & Offord, 2001), reflected in findings showing that a group of high scorers on such questionnaires only contained 17–42% true positives of diagnosed behavioral disorders based on clinical interviews (Sveen et al., 2013). Such an arguably low sensitivity may be attributed to lacking or ambiguous onset, frequency, duration, and intensity criteria within these instruments, as well as the potentially limited experience of what is normative behavior or not among the respondent who is set to evaluate the behavior (e.g., parents). In addition, the often-used general measure of externalizing problems in the Child Behavior Checklist (CBCL; Achenbach, 1991) also contains items pertaining to impulsivity and difficulties in regulating attention and activity, which some argue may rather be symptoms of ADHD (Hay et al., 2010). Although ADHD is grouped with ODD and CD in the category of externalizing disorders it is, at its core, a neurodevelopmental disorder with a distinct genetic risk profile (Azeredo et al., 2018). Thus, the inclusion of symptoms that actually may be symptoms of ADHD could muddle the measurement of ODD and CD (Hay et al., 2010), and perhaps this could be one factor explaining the variability in previous findings. In contrast, we argue that structured clinical diagnostic interviews administered by trained interviewers provide a more valid measure of symptoms of behavioral disorders. To the best of our knowledge, this study is the first to employ diagnostic interviews to measure clinical symptoms of ODD and CD in relation to social skills.

Theory and research, to an increasing degree, underscore the differences between ODD and CD with regard to predictors, covariates, and outcomes (Burke et al., 2022; Wiesner et al., 2015). Although there is some overlap with respect to core symptoms and features, they also include behaviors that are specific and exclusive to each one (Burke et al., 2024). Their response to treatment interventions has also been found to differ (e.g., Kolko et al., 2009). It is yet to be examined whether the developmental dynamics of ODD and CD are differentially related to social skills. Since social skills training figures in many prevention and treatment interventions, and treatments are targeted at the within-person level (Hamaker et al., 2015; Speyer et al., 2024), it would be of etiological and clinical interest to illuminate the specific within-person relations from social skills to ODD and CD respectively. Although one cross-sectional study found that certain aspects of social skills were putative correlates to both ODD and CD symptoms (Wiesner et al., 2015), we are not aware of any longitudinal studies examining this.

With some exceptions (e.g., Bornstein et al., 2010; Burt et al., 2008), single-informant reports on both behavior problems and social competence have been the norm—possibly inflating associations due to single-informant biases and shared method variance (Huber et al., 2019). Furthermore, reporter effects may change during development (Merrell, 2001). This may be especially salient in terms of parental reports on social competence, as children and parents spend decreasing amounts of time together with age (Memmott-Elison & Toseeb, 2023), leaving parents with less access to their children’s social arenas and social behaviors (Renk & Phares, 2004). Teachers, however, may see children in a different social context, which arguably varies less with age. To limit common rater and context effects, we utilize reports from teachers, children, and parents.

Previous research has typically utilized regression-type or cross-lagged panel models (CLPM), which produce results that are a mixture of between and within-person variance (Berry & Willoughby, 2017; Hamaker et al., 2015). Because causality operates on the within-person level, between-person differences need to be filtered out so that the findings can better inform etiological inferences. We are aware of four studies that have applied a within-person approach to the present topic (Memmott-Elison & Toseeb, 2023; Speyer et al., 2024; Williams et al., 2024; Zondervan-Zwijnenburg et al., 2022). The results have been mixed. Memmott-Elison and Toseeb (2023) found reciprocal relations between prosocial behavior and behavior problems from 3 to 14 years (but from conduct problems to prosocial behavior only from 3 to 5 years), and Williams and colleagues (2024) found reciprocal negative relations between prosocial behavior and conduct problems from 4–15 years. Speyer and colleagues (2024), on the other hand, found that self-reported (but not teacher-reported) aggression at 11 years predicted decreased prosocial behavior at 13 years, but not from 13 to 15, and not the other way around; while Zondervan-Zwijnenburg and colleagues (2022) found some initial support for a negative within-person cross-lagged relation from behavior problems to later externalizing behavior in from 7 to 11 years, but concluded that overall there was little to no support for any predictive within-person relation between the two.

The abovementioned studies differ from the present in that they have all utilized rating scale measures of externalizing behavior problems or aggression, as opposed to diagnostic interviews geared towards symptoms of diagnosable disorders. Furthermore, they all focus on the concept of prosocial behavior. Prosocial behavior is generally defined as voluntary helping, sharing, and supporting behaviors. Although a significant part of social skills, the latter encompasses a broader range of social behaviors, such as communication, cooperation, assertion, engagement, and self-control (Gresham et al., 2011; Little et al., 2017). Existing theories on behavior disorders suggests that these subskills, or lack thereof, are related to ODD and CD (e.g., Gilmour et al., 2004; Wiesner et al., 2015; Zhang et al., 2023). These additional subskills are also often targeted in some of the widely used psychosocial treatments of behavioral disorders, which involve training children in communication skills, emotional regulation, social information processing, and conflict resolution (e.g., Beelmann & Lösel, 2021; Goertz-Dorten et al., 2017; Webster-Stratton & Reid, 2018). Thus, examining the longitudinal within-person relations between changes in the broader concept of social skills and symptoms of ODD and CD separately could aid in providing a more precise description of how social skills are related to these two disorders across development. Such knowledge could aid in choosing and designing disorder-appropriate interventions. Furthermore, both Memmott-Elison and Toseeb (2023), Speyer et al. (2024), and Williams et al. (2024) measured both behavior problems and prosocial behavior using subscales from one questionnaire only, as well as using the same reporter on both prosocial behavior and behavior problems—issues that may have led to inflated prospective relations. With regards to Zondervan-Zwijnenburg et al. (2022), as the authors themselves profess, there is some uncertainty attached to the findings due to the sample size being below the recommended cut-off for use with random intercept cross-lagged panel models (RI-CLPM) (Masselink et al., 2018).

The Trondheim Early Secure Study (TESS) consists of children of the 2003 and 2004 birth cohorts in Trondheim, Norway (N = 3,456) (Steinsbekk & Wichstrøm, 2018). A letter of invitation, along with the Strengths and Difficulties Questionnaire (SDQ) 4–16 version (Goodman, 1997), was sent to their homes, and parents brought the completed SDQ to the routine 4-year health check-up (n = 3,358). Parents were informed about TESS using procedures approved by the Regional Committee for Medical and Health Research Ethics Mid-Norway, and written consent was obtained.

To increase statistical variance and thus power, we oversampled children with emotional or behavioral problems based on their SDQ total difficulties score and then divided the sample into four strata (cut-offs: 0–4, 5–8, 9–11, and 12–40). Using a random number generator, drawing probabilities to participate increased with increasing SDQ scores (0.37, 0.48, 0.70, and 0.89). Of the 1,250 children randomly drawn, 1,007 were enrolled at T1. The drop-out rate after consent did not differ according to SDQ score [(t = 0.17, df = 1, p = 0.86) or gender (χ² = 1.02, df = 1, p = 0.31)]. The mean age at the first assessment was 4.7 years (SD = 0.30, 49.9% males). Retesting occurred at 6 years (T2): M_age = 6.7 years, SD = 0.25; 8 years (T3): M_age = 8.8, SD = 0.24; 10 years (T4): M_age = 10.5 years, SD = 0.16; 12 years (T5): M_age = 12.5, SD = 0.15;14 years (T6): M_age = 14.4 years, SD = 0.16, and 16 years (T7): M_age = 16.9, SD = 0.31. Overall, 1,079 participants had information from at least one wave of data collection and comprised the analytical sample.

The attrition rate was lower for those with higher teacher-rated social skills at the previous assessment point for age 6 (OR = 0.49, p = 0.003), age 8 (OR = 0.46, p = 0.017), and age 12 (OR = 0.29, p < 0.001). Even so, the combined effects of the predictors of attrition were small, varying between Cox and Snell’s Proxy R² 0.017–0.023. Little’s MCAR test (Li, 2013) was χ² = 5286.06, df = 4904, p < 0.001, supporting that data was not missing completely at random (MCAR). However, as the normed test was 1.08, data was likely missing at random (MAR).

Analyses were conducted using structural equation modeling (SEM) in Mplus 8.5, employing a robust maximum likelihood estimator. Missingness was handled with a full information maximum likelihood (FIML) procedure under the assumption that the data was MAR. To arrive at correct population estimates due to the stratification, data were weighted with a factor corresponding to the number of children in the stratum divided by the number of participating children in that stratum. As the number of ODD and CD symptoms was expected to be heavily right-skewed, we used the natural log.

The RI-CLPM allows for the separation of variance into a stable between-person part (consisting of three latent random intercepts loading on social skills, symptoms of ODD, and symptoms of CD) and a within-person part, which estimates changes from one’s own mean level in a variable (e.g., social skills) as a function of changes in that variable at the previous measurement point (autoregression) and in the predictor (i.e., ODD or CD symptoms; a cross-lagged effect). Parent-rated and teacher-rated social skills were analyzed in separate models.

To examine whether cross-lagged paths differed between ages, a model where the cross-lagged paths in question were allowed to vary across time was compared to a model where paths were set to be identical using the Satorra-Bentler scaled Chi-square test (Bryant & Satorra, 2012). As we conducted a high number of comparisons, we adjusted for the false discovery rate (Benjamini & Hochberg, 2000). To ease comparison with prior studies using a blend of within- and between-person information, the analyses were also run with an ordinary CLPM.

The RI-CLPM involving parent-rated social skills revealed a somewhat sub-optimal fit to the data, χ² = 304.94, df = 129, p < 0.001, Comparative Fit Index (CFI) = 0.937, Tucker-Lewis Index (TLI) = 0.897, Root-mean-square error of approximation (RMSEA) = 0.036, 90% CI: 0.030–0.041. However, a model with 4-year autoregressive lags included did fit the data adequately, χ² = 204.30, df = 114, p < 0.001, CFI = 0.968, TLI = 0.940, RMSEA = 0.027, 90% CI: 0.021–0.033.

Setting all the paths from social skills to ODD and to CD at ages 4–16 to be similar did deteriorate model fit slightly, Δχ² = 18.45, df = 10, p = 0.048. Inspecting the freely estimated model suggested that the paths from social skills to ODD at ages 4 and 6 might be stronger than at later ages. We thus set these two paths to be similar and the remaining paths to be similar with no change in model fit compared to the original freely estimated model, Δχ² = 7.89, df = 9, p = 0.546, and this constrained model fit the data well, χ² = 212.30, df = 123, p < 0.001, CFI = 0.968, TLI = 0.945, RMSEA = 0.026, 90% CI: 0.020–0.032, and was considered our final and parsimonious model (Fig. 1).

With regards to teacher-reported social skills, a freely estimated model did evince a sub-par fit to the data (χ² = 251.84, df = 114, p < 0.001, CFI = 0.925, TLI = 0.875, RMSEA = 0.033, 90% CI: 0.028–0.039), whereas a model where 4-year autoregressions were included showed adequate fit (χ² = 181.06 df = 100, p < 0.001, CFI = 0.956, TLI = 0.916, RMSEA = 0.027, 90% CI: 0.021–0.034). A model where the paths from social skills to ODD and CD were set to be equal over time did worsen the fit compared to the freely estimated model, Δχ² = 35.06, df = 10, p < 0.001. However, setting the paths from social skills to ODD at ages 4–12 to be similar and letting the last cross-lagged paths vary, and setting the paths from social skills to CD to be identical across ages except for the path from age 6 to 8, did not deteriorate model fit, Δχ² = 7.49, df = 8, p = 0.485. Moreover, comparing this model to a model where the paths from social skills to ODD were freely estimated (also at ages 8–12), did not change fit, Δχ² = 1.73, df = 2, p = 0.227, indicating that the paths to ODD in middle childhood did not differ in strength and were stronger than the age 4 to 6, 10 to 12 and the age 14 to 16 years effects (Fig. 2). This model had favorable model fit, χ² = 188.86, df = 108, p < 0.001, CFI = 0.956, TLI = 0.923, RMSEA = 0.026, 90% CI: 0.020–0.032.

In the above models, increased parent-rated social skills predicted reduced symptoms of CD (β = −0.07 to −0.15, p < 0.010) and ODD (β = −0.10 to −0.21, p < 0.001) at all time points. In the case of teacher-reported social skills, a slightly different picture emerged, where increased social skills only predicted reduced symptoms of CD at one time point, from 6 to 8 years (β = −0.22, p < 0.001). Regarding ODD, though, increased social skills predicted reduced symptoms of ODD at all time points (β = −0.13 to −0.14, p < 0.001), except for one, from age 14 to age 16 (β = −0.06, p = 0.465).

The model incorporating parent reports showed no predictive relations between increased symptoms of CD and later changes in social skills. The same held true for ODD, with the exception of one time point, from age 6 to 8 (β = −0.10, p = 0.028), but this path was rendered just beyond significance when adjusting for the false discovery rate (p = 0.052) Regarding teacher report, increased symptoms of CD predicted reduced social skills from age 6 to 8 years (β = −0.16, p = 0.001) only, while increased symptoms of ODD predicted reduced social skills from age 8 to 10 years (β = −0.12, p = 0.015) only.

At the between-person level, the random intercepts of ODD and CD, respectively, were not correlated with the random intercept of parent-rated social skills (r = 0.13, p = 0.628 and r = 0.19, p = 0.601). Teacher ratings correlated with ODD symptoms (r = −0.40, p = 0.001) but not with CD symptoms (r = −0.37, p = 0.097). At the within-person level, changes in CD at each time point were negatively related to changes in parent-rated social skills at that time point (r = −0.11, p = 0.041 to r = −19, p < 0.001), except at age 6 (r = −0.07, p = 0.148). Changes in ODD were consistently negatively related to parent ratings (r = −0.17, p = 0.001 to r = −26, p < 0.001). Regarding teacher ratings, only one significant relation was observed in the case of CD (age 8, r = −0.12, p = 0.041), while the rest were insignificant (r = −0.03, p = 0.663 to r = −08, p = 0.227). For ODD, a similar picture emerged, where only one relation was significant (age 8, r = −0.12, p = 0.039), while the rest were not (r = −0.03, p = 0.572 to r = −08, p = 105).

Separate models were run for boys and girls, and the strengths of the paths were compared. Taking the final constrained models as vantage points, the scaled chi-square test for teacher-rated social competence revealed no gender difference (Δχ² = 51.73, df = 42, p = 0.144). The model for parent-rated social skills did not converge and could not be estimated.

For the sake of comparison, we ran a traditional CLPM. All paths were set to be equal, and as in the RI-CLPM, we found that in the case of parent-rated social skills, increased social skills predicted a reduced number of both ODD- (β = −0.06 to −0.13, p > 0.001) and CD-symptoms (β = −0.04 to −0.07, p < 0.002) at all time points. Regarding the opposite direction, increased symptoms of ODD at age 6 predicting reduced social skills at age 8 (β = −0.09, p = 0.036), emerged as the only path from symptoms to later social skills. In the teacher-ratings model, the paths from 4 to 6, 12 to 14, and 14 to 16 years were set to be equal in the case of ODD, whilst for CD, all paths but those from 6 to 8 years were set to be equal. The CLPM results for teachers coincided with those of parents in terms of CD, increased social skills predicted decreased symptoms of CD at all time points (β = −0.04 to −0.25, p < 0.010). A slightly different picture emerged for ODD, as increased social skills only predicted increased symptoms of ODD at ages 6 to 8, 8 to 10, and 10 to 12 years (β = −0.11 to −0.21, p < 0.010). Regarding symptoms predicting social skills, increased symptoms of CD predicted reduced social skills only from age 6 to 8 years (β = −0.15, p < 0.001), whilst symptoms of ODD predicted reduced social skills from 8 to 10 years only (β = −0.09, p = 0.028). Thus, the CLPM and RI-CLPM findings for parents were similar in terms of social skills predicting later changes in disorder symptoms. For teachers, the CLPM showed a consistent relation between social skills and symptoms of CD that was not present in the RI-CLPM. Relations between social skills and ODD also differed, as the RI-CLPM showed a predictive relation from social skills to symptoms of ODD at all timepoints except the last, while the CLPM only produced the same relations in middle childhood. (Figures S2 and S3 in online supplement).

The results partly align with previous longitudinal RI-CLPM studies chronicling a predictive relation between prosocial behavior and behavioral problems (e.g., Memmott-Elison & Toseeb, 2023; Williams et al., 2024), but contrast those finding no relation (e.gSpeyer et al., 2024; Zondervan-Zwijnenburg et al., 2022) however, we go beyond this by showing that this relation is present for a wide set of social skills beyond prosociality, and for DSM-defined symptoms of ODD and (partly for) CD. Although identifying the mechanisms underlying the observed relations is outside the scope of this article, we will draw attention to some putative alternatives. First, it is conceivable that increasing social skills may lead the way for a positive developmental trajectory, which in turn leads to a downstream reduction in symptoms of ODD and CD. This dovetails with the fact that treatments focused on social skills training have led to a reduction in symptoms of ODD/CD in studies, also at follow-up (e.g., Giudice et al., 2022; van Manen et al., 2004). For example, increased socially skilled behavior may increase access to prosocial peers who will support and foster prosocial behavior. In contrast, children who have poor social skills are often rejected by socially skilled peers and are left to socialize with other aggressive children (Kendler et al., 2008; Patterson et al., 2000), who in turn may model and reinforce disruptive behavior (Patterson et al., 2000). Increased child social skills may also result in positive interactions with significant adults, perhaps leading to less conflict and harsh discipline—factors which may contribute to later reductions in symptoms of ODD and CD (e.g., Larsson et al., 2009). In that way, enhanced social skills could pave the way for continued prosocial relations as well as elicit positive responses and acceptance from prosocial others that preserve this behavior, and thereby contribute to a reduction in symptoms of ODD and CD. Thus, being accepted into prosocial groups, establishing friendships with peers, and eliciting positive responses from significant adults, could lead to a lasting positive effect that works to reduce or prevent behavior problems, evidenced by the longitudinal effects. Still, identifying such mediational effects awaits further inquiry.

We did, however, fail to see the same pattern regarding teacher-rated CD, where only one path from social skills to later symptoms emerged. One possible factor in the observed findings could be the effect of single-rater bias, as parents reported on both social skills and symptoms, which could arguably inflate the associations; something that is less likely for teachers that only reported on social skills. Furthermore, it should be noted that the social skills that parents and teachers reported on are not exactly the same, as teachers responded to the subscales cooperation, assertiveness and self-control, while parents responded to those three in addition to the subscale responsibility. The fact that we did observe a relation between teacher-rated social skills and symptoms of ODD, though, suggests that there is a difference in the social skills perceived by teachers versus those perceived by parents, specifically pertaining to the relation between social skills and later symptoms of CD. Furthermore, this difference is likely tied to situational or relational factors specific to the school or classroom environment. Social skills are context-specific, and different contexts require different skills (Gresham et al., 2011). Parents mostly view their child’s social skills as they unfold within the home, in familiar relations within the family, or with close friends. Teachers observe skills in larger group settings in a structured environment, which require participating in both hierarchical interactions with non-family adults as well as horizontal relations with a larger number of peers. It is plausible that these two situations require the use of different skills, and the teachers will likely be the ones to notice the presence or lack of these skills. With regards to this only pertaining to the social skills ➔ CD relation, the difference between ODD and CD has been getting an increasing amount of recognition, as previously mentioned. Perhaps such disorder-specific differences are reflected in these findings. Alternatively, it could be that the social skills items tap into skills that are related more so to everyday behavior difficulties that are more in line with ODD (e.g., cooperating in obeying, self-control in anger) than they are with the more serious behaviors of CD, that involve behaviors such as aggression towards others, destruction of property, deceitfulness, or serious violations of rules.

Changes in the number of symptoms of ODD and CD did not forecast later parent-rated social skills at any time point, and for only one timepoint for ODD and one for CD, with regards to teacher-rated social skills. This runs counter to findings from both some traditional cross-lagged (Burt & Roisman, 2010; Chen et al., 2010; Obradovic & Hipwell, 2010; Obsuth et al., 2015; Padilla-Walker et al., 2018) and within-person analyses (e.g., Memmott-Elison & Toseeb, 2023; Speyer et al., 2024; Williams et al., 2024; Zondervan-Zwijnenburg et al., 2022). Why our findings are at odds with other within-person results requires some scrutiny. First, despite conducting multiple statistical comparisons, none of them corrected for the false discovery rate, thus risking false positives (Benjamini & Hochberg, 2000). Secondly, the use of one single reporter on both measures may also have introduced a common-method bias in the case of Memmott-Elison and Toseeb (2023), Williams et al (2024) and Speyer et al (2024). Finally, the result may not be directly comparable to ours due to the aforementioned differences in measurements, as no previous study has specifically measured clinical symptoms of ODD and CD respectively, in conjunction with social skills. We will await further within-person studies that examine the longitudinal relations between social skills and symptoms of ODD and CD, in hopes of further elucidation of the observed findings.

In line with previous research, we did not find any gender differences in predictions (e.g., Burt & Roisman, 2010; Obsuth et al., 2015; Padilla-Walker et al., 2018) Thus, despite girls being rated more socially competent than boys across cultures, ages, and informants (Sørlie et al., 2008), whereas boys have been found to have more ODD and CD (e.g., Diamantopoulou et al., 2011), the within-person longitudinal relations do not appear to differ across gender, suggesting that reduced social skills are equally important to boys’ and girls’ development of ODD and CD.

Despite using a large birth cohort sample spanning 12 years and utilizing sound measures and state-of-the-art analyses that allow for the examination of within-person variance, this study has some limitations. Although we have taken steps to minimize common-method variance, parents reported on symptoms of ODD, CD and social skills, potentially inflating the longitudinal relations. Still, this was not the case for the model including teacher-reported social skills. Furthermore, findings from parent and teacher ratings were, with the exception of the relation between social skills and CD, in agreement, which is not always the case (Renk & Phares, 2004). We believe this adds confidence to the results. However, as Norwegian children generally have a lower prevalence of ODD and CD than children in many other countries (Heiervang et al., 2007), generalization to other contexts and cultures should be done with utmost care. It also needs to be mentioned that even though the RI-CLPM, by way of the random intercepts, captures all unmeasured time-invariant confounding effects, time-varying effects are not adjusted for in the same manner. However, the RI-CLPM is a very power-demanding approach, and even with only two variables included, sample sizes approaching 1,000 are typically recommended (Masselink et al., 2018). Including even more variables further reduces the power. We therefore opted not to include time-varying covariates because we would reduce the risk of reporting false negative results. However, this comes with the risk of portraying false positive effects.