Childhood Behavioural Problems and Adverse Outcomes in Early Adulthood: a Comparison of Brazilian and British Birth Cohorts

Purpose Examine associations between childhood behavioural problems with criminal behaviour, emotional disorders, substance use and unemployment in early adulthood in two birth cohorts from a middle- and high-income country. Methods Data were utilised from large, prospective birth cohorts in Brazil (1993 Pelotas Birth Cohort; N = 3939) and the UK (Avon Longitudinal Study of Parents and Children; ALSPAC; N = 5079). Behavioural problems were reported on by parents at age 11 years (including disobeys, temper, lies, fights, steals). Outcomes (assessed with youth between ages 22 and 24 years) included criminal behaviour, emotional disorders, substance use and NEET (not in education, employment or training). Results In both cohorts, children with ‘conduct problems’ (those with increased probability of all five behaviours at age 11), were at higher risk of criminal behaviour, emotional disorders and NEET in adulthood compared to those with ‘low problems’. Associations for ‘conduct problems’ were stronger in Pelotas compared to ALSPAC for hazardous alcohol use [Pelotas: risk ratio = 1.39, 95% CI = 1.14–1.70; ALSPAC: risk ratio = 0.76, 95% CI = 0.57–1.02] and illegal drug use [Pelotas: risk ratio = 1.32, 95% CI = 1.16–1.50; ALSPAC: risk ratio = 1.05, 95% CI = 0.91–1.20], whereas associations for criminal behaviour [Pelotas: risk ratio = 1.92, 95% CI = 1.29–2.86; ALSPAC: risk ratio = 2.75, 95% CI = 2.04–3.73] and NEET [Pelotas: risk ratio = 1.38, 95% CI = 1.13–1.70; ALSPAC: risk ratio = 3.04, 95% CI = 1.99–4.65] were stronger in ALSPAC. Conclusions Childhood conduct problems were associated with criminal behaviour, emotional disorders and unemployment in adulthood in both Brazil and the UK. Additional associations were found for substance use in Brazil. Electronic supplementary material The online version of this article (10.1007/s40865-019-00126-3) contains supplementary material, which is available to authorized users.

The three-class model was supported by the fit indices and was also theoretically meaningful In step 1 an omnibus test of measurement invariance was performed. Two nested models were estimated (see Online Resource 1 Figure 2A and Online Resource 1 Figure 2B). a much better fit to the data than model M1.0 (no DIF), suggesting that study is a source of DIF for at least one of the five latent class indicators in at least one of the latent classes.
In step 2, individual indicator tests for nonuniform DIF were performed. Biasadjusted three-step methods (Vermunt 2010) were used (using the modal classes from the original unconditional three-class model and the estimated average classification errors for the modal class assignment) to estimate two models for each of the five latent class indicators (10 models in total). Online Resource 1 Figure 2C and Online Resource 1 Figure 2D Figure 2E).
ST2 (step 3) shows that model M3.0 provided a better fit to the data than model M1.0 (no DIF) and did not provide a worse fit to the data (according to the aBIC) than model M1.1 (all DIF).
In step 4, three latent class MIMIC models were estimated. In each of these models, the direct effect from study to one latent class indicator was constrained to be class-invariant (uniform DIF) while the remaining direct effects were class-varying (nonuniform DIF).
Online Resource 1 Figure 2F shows a model testing for uniform DIF for 'disobeys' (model M4.1). Online Resource 1 Figure 2G shows a model testing for uniform DIF for 'fights' (model M4.2). Online Resource 1 Figure 2H shows a model testing for uniform DIF for 'steals' (model M4.3). These models were each compared to model M3.0. As can be seen in Online Resource 1 The decision that the latent class profiles were comparable across study was based on the number and magnitude of the estimated DIF effects, and the evaluation of the study specific profile plots (shown in Figure 2, main text). It was also considered whether the DIF effects made theoretical sense. The differences seen could be due to cultural differences, for example if punishment is more severe in Brazil compared to the UK, children may be less likely to disobey parents, even when they display behavioural problems. The DIF effects seen for 'fights' and 'steals' could be a consequence of having a latent class that represents both aggressive and non-aggressive conduct problems, with young people in the UK being more likely to display non-aggressive conduct problems and young people in Brazil being more likely to display aggressive conduct problems (Murray et al. 2015). Additionally, the particulary large DIF effect for 'fights' could be due to translation. In ALSPAC, the question asks about "fighting or bullying other children", whereas in Pelotas the translation of the Brazilian Portuguese is "fighting or frightening other children".
In step 7, the association between study and class membership was evaluated by comparing model M5.0 (relabeled as model M7.1 in Online Resource 1  Figure 2 (main text).  (Goodman 1997). Associations between the validators and the latent classes (with the 'low' class as the reference class) are shown in Online Resource 1 Table 3.
Online Resource 1 Table 3. Validation of the latent classes of behavioural problems across study using related constructs assessed at age 11 years; showing multinomial odds ratio (95% confidence interval) with 'low problems' as the reference class