Dimensionality reduction
Parallel analysis indicated 9 factors as optimal at both timepoints. The confidence intervals of the bootstrapped factor solution indicated a stable solution with 6 factors (see Fig. 2 for a graphical representation of the factor loading). The 6-factor model explained 56.78% of variance in the childhood data and 53.39% in the adolescence data. We propose the following labels for the 6 factors according to the items associated with them: Conduct problems (C.), Hyperactivity/Impulsivity (H.), Inattention (I.), Emotional control problems (E.), Anxiety (A.), Motor problems (M.). Overviews of the items associated with the factors in the childhood and adolescence data are presented in Fig. 2. The full loading matrix and factor correlation structure can be found in the Supplementary Materials.
Hybrid hierarchical clustering of the childhood data (10 years)
The hierarchical hybrid clustering algorithm successfully converged on a solution with the childhood data (see Fig. 3). The Calinksi–Harabasz index indicated optimal cluster numbers at k = 4 and k = 7 (see Fig. 3b). To save space, we focus on the k = 7 solution here (a detailed description of the k = 4 solution can be found in the Supplement). The clustering indicated a large group without problems in any domain (C1a ‘No Problems’, see Table 2 for descriptive statistics and Fig. 3 for an illustration of the behavioural profiles). Further, groups with problems in specific domains were indicated, which comprised groups with Anxiety problems (C1b ‘Anxiety’), Inattention (C3 ‘Inattention’), Conduct problems (C4b ‘Conduct’), or Emotional problems (C4a ‘Emotion’). There were also groups with problems in multiple domains, specifically a group with Anxiety and Hyperactivity (C2a ‘Anxiety, Hyperactivity’), and a group with Motor, Hyperactivity and Emotional problems (C2b ‘Motor, Hyperactivity, Emotion’).
Table 2 Overview of size, gender ratio, cognitive and educational assessment performance and incidence of mental health problems in the clustering-defined groups in the childhood data Regarding differences in other domains, there were more girls than expected by chance in the Inattention cluster (C3 ‘Inattention’), and more boys in the Anxiety and Hyperactivity (C2a ‘Anxiety, Hyperactivity’) and the Conduct problem cluster (C4b ‘Conduct; see Table 2). Children with anxiety and emotional problems (C1b ‘Anxiety’, C3 ‘Inattention, C4a ‘Emotion’) scored lower across multiple measures of family structure and socio-economic background (see Table 2 and Supplement for a detailed analysis). Cognitive scores and educational attainment scores were higher in children with no behavioural problems (C1a ‘No Problems’) and lower for children with anxiety, inattention, or conduct problems (C1b ‘Anxiety’, C3 ‘Inattention’, C4b ‘Conduct’, see Table 2 and Supplement). Mental health problems were more frequently indicated for children with anxiety or emotional problems (C1b ‘Anxiety’, C4a ‘Emotion’).
Hierarchical hybrid clustering of the adolescence data (16 years)
Inspection of the minima of the Calinski–Harabasz index indicated optimal solutions at 4 and 6 clusters (see Fig. 4b). To save space, we focus on the k = 6 solution (see Supplement for k = 4 results). At k = 6, the statistical comparison indicated differences in factor scores between the groups (ANOVA group x factor: F(25, 33,690) = 648.96). The clustering solution identified a large group with no particular problems (A1b ‘No Problems’, see Table 3 for descriptive statistics and Fig. 4 for an illustration of the behavioural profiles) and a cluster of adolescents with Hyperactivity (A4 ‘Hyperactivity’). Further, the clustering pulled out a group with selective Inattention problems (A1a ‘Inattention’). In addition, the clustering indicated groups with combinations of problems across domains, i.e. a group with problems with Anxiety and Motor problems (A2 ‘Anxiety, Motor’), a group with Emotional problems (A3a ‘Emotion’) and a group with problems related to Anxiety, Emotion and Inattention (A3b ‘Anxiety, Emotion, Inattention’).
Table 3 Overview of size, gender ratio, cognitive and educational assessment performance and incidence of mental health problems in the clustering-defined groups in the adolescence data Regarding differences in other factors, there were more girls in the cluster without behavioural problems (A1b ‘No Problems’), in the Inattention (A1a ‘Inattention’), and mixed problems cluster (A3b ‘Anxiety, Emotion, Inattention’) than would be expected by chance (see Table 3). There were more boys in the Anxiety and Motor problems (A2 ‘Anxiety, Motor’) and the Emotional problems cluster (A3a ‘Emotion’). There was no difference in cognitive or reading scores between the groups (see Table 3 and Supplement). Scores on the arithmetic assessment were higher for adolescents in the no problems group (A1b ‘No Problems’) and lower for adolescents with problems related to Anxiety, Inattention and Emotional Problems (A3b ‘Anxiety, Emotion, Inattention’, see Table 3 and Supplement). There were no differences in the frequency of mental health problems between the groups on the school nurse report (although note that only a third of the sample had data available, see Table 3 and Supplement for details and Discussion for limitations associated with this rating).
Differences in behavioural problem profiles between childhood and adolescence
Of the 6744 adolescents included in the analysis, 5487 participants were also included in the childhood data. Statistical comparison indicated no disproportionate loss per group either defined in the childhood or adolescence data (defined at 6/7 cluster level; childhood grouping: Χ2(49) = 56, p = 0.229; adolescence grouping: Χ2(25) = 30, p = 0.224). The ratings of behavioural difficulties shifted between the assessments at childhood and adolescence. Statistical comparison indicated that the average parent rating of behavioural problems across all items decreased (childhood: mean = 1.31, SE = 0.008; adolescence: mean = 1.27, SE = 0.007; paired-sample t test: t (5791) = 11.04, p < 0.001, d = 0.15). The pattern of differences differed by the behavioural domain. The largest change was observed for items related to Hyperactivity with lower ratings for adolescence compared to childhood (10 yo: mean = 1.41, SE = 0.009; 16 yo: mean = 1.24, SE = 0.007; paired-sample t test: t (5791) = 26.61, p < 0.001, d = 0.35). Smaller decreases were found for Anxiety (childhood: mean = 1.46, SE = 0.011; adolescence: mean = 1.36, SE = 0.007; t = 16.37, p < 0.001, d = 0.22) and Inattention (childhood: mean = 1.43, SE = 0.009; adolescence: mean = 1.36, SE = 0.008; t = 10.13, p < 0.001, d = 0.13). There were smaller decreases for items related to Motor and Conduct Problems (Motor: d = 0.05; Conduct: d = 0.10). In contrast, significant increases were found for problems related to Emotion (childhood: mean = 1.34, SE = 0.009; adolescence: mean = 1.41, SE = 0.009; t = − 11.39, p < 0.001; d = -0.15). In summary, the overall incidence of behavioural problems and the nature of behavioural problems changed between the childhood and the adolescence assessment.
The changing nature of parent-reported behavioural problems was also reflected in the hybrid hierarchical clustering solution in childhood and adolescence. Comparison of the Fowlkes–Mallows index at k = 4 indicated that the groups were not similar (two-sided permutation-p = 0.034, mean = 0.36, variance = 4.861 × 10–10), and was significantly lower in adolescence compared to a bootstrapped sample of the same index based on random permutations of the childhood dendrogram (two-sided permutation-p < 0.001, 1,000 permutations, (see Fig. S3 in the Supplementary Materials). When comparing the dendrograms at the 6-group (childhood) and 7-group cut-off (adolescence), the clustering solution was also dissimilar (mean = 0.25, variance = 5.944 × 10–6) and lower than can be expected by chance (two-sided permutation-p < 0.001, 1000 permutations, see Supplementary Materials). In short, the profiles of behavioural problems were different in adolescence and childhood.
Differences between the clustering at childhood and adolescence were reflected in transitions of individuals between the behavioural difficulty profiles (see Fig. 5). Across all profiles in childhood, a significant proportion transitioned to the low behavioural problems group in adolescence (A1b, see Fig. 5b). There were also transitions to profiles with behavioural problems. Specifically, a larger-than-expected proportion transitioned from no behavioural problems in childhood to Emotion problems (A3a ‘Emotion’: see Fig. 5), while a transition to Inattention (A1a ‘Inattention’) or Hyperactivity (A4 ‘Hyperactivity’) was less common. For children with Anxiety problems (C1b ‘Anxiety’), a large proportion showed Emotional problems in adolescence (A3a ‘Emotional’). Children with mixed problems related to Emotion, Motor and Hyperactivity (C2b ‘Motor, Hyperactivity, Emotional’) mostly transitioned to problems related to Inattention in adolescence (A1a ‘Inattention’). Children with Anxiety problems (C3) tended to show problems with Emotion in adolescence (A3a, ‘Emotion’). A large number of individuals with Conduct problems in childhood (C4b ‘Conduct’) transitioned to mixed problems with Anxiety, Emotion and Inattention (A3b ‘Anxiety, Emotion, Inattention’) in adolescence.
Next, we investigated if any of the demographic factors in childhood predicted common transitions between childhood and adolescence. In each case, we compare a group with a particular profile of behavioural problems, e.g. A1a, to the rest of the sample, e.g. A1ac. The assessments at childhood were used to assess their impact on problems in adolescence. Children who transitioned to the no behavioural problems cluster in adolescence (A1b, ‘No Problems’) from childhood clusters with some indication of problems, i.e. any of the childhood clusters except C1a (‘No Problems’), came from families with higher income (meanA1b = 1.38, SEA1b = 0.007; meanA1bc = 1.34, SEA1bc = 0.006, t (5060) = 4.34, p < 0.001), more educated parents (meanA1b = 2.17, SEA1b = 0.021; meanA1bc = 2.08, SEA1bc = 0.018, t (5371) = 2.91, p = 0.018) and higher social class (meanA1b = 3,36, SEA1b = 0.024; meanA1bc = 3.55, SEA1bc = 0.019, t (5325) = − 6.21, p < 0.001). Children who overcame behavioural problems by age 16 also had better cognitive scores in childhood (meanA1b = 0.21, SEA1b = 0.020; meanA1bc = 0.09, SEA1bc = 0.018, t (5480) = 4.76, p < 0.001). Immigration background and the number of siblings were not related to the transition to low-behavioural problems (all p > 0.1).
Transition to the cluster with problems with Emotions (A3a ‘Emotion’) from any of the childhood clusters was predicted by lower family income (meanA3a = 1.32, SEA3a = 0.012; meanA3ac = 1.37, SEA3ac = 0.005, t (5066) = − 3.52, p = 0.002), a greater number of siblings (meanA3a = 2.65, SEA3a = 0.0,039; meanA3ac = 2.49, SEA3ac = 0.015, t (5452) = − 2.26, p < 0.001) and lower social class (meanA3a = 3.64, SEA3a = 0.041; meanA3ac = 3.43, SEA3ac = 0.017, t (5066) = − 3.52, p = 0.002). Immigration background, highest parental education and childhood cognitive scores were not related to the transition to Emotion problems in adolescence (all p > 0.1).