Educational performance and conduct problem trajectories from childhood to adolescence: Observational and genetic associations in a Brazilian birth cohort

Abstract Background Educational difficulties are an important potential influence on both the onset and course of children's conduct problems. This study evaluated the association between school failure and children's conduct problems in Brazil, a context with high rates of both conditions, using both observational and genetic approaches. Methods Prospective, population‐based, birth cohort study in Pelotas city, Brazil. Parents reported on conduct problems four times between ages 4–15 years, and group‐based trajectory analysis was used to classify 3469 children into trajectories of childhood‐limited, early‐onset persistent, adolescence‐onset, or low conduct problems. School failure was measured as having repeated a school grade up to age 11, and a polygenic risk score (PRS) predicting educational attainment was calculated. Multinomial adjusted regression models were used to estimate the association between school failure (observational measure and the PRS) and conduct problem trajectories. To consider possible variation in effects of school failure by social context, interactions were tested with family income and school environment (using both observational and PRS methods). Results Children repeating a school grade had increased odds of being on to childhood‐limited (OR: 1.57; 95% CI 1.21; 2.03), adolescence‐onset (OR: 1.96; 95% CI 1.39; 2.75), or early‐onset persistent trajectory (OR: 2.99; 95% CI 1.85; 4.83), compared to the low conduct problem trajectory. School failure also predicted increased risk for early‐onset persistent problems versus the childhood‐limited problems (OR: 1.91; 95% CI 1.17; 3.09). Using a genetic PRS approach, similar findings were observed. Associations varied according to the school environment: school failure had larger effects on children in better school environments. Conclusion School performance, whether measured in terms of repeating school grades or genetic susceptibility, was consistently associated with trajectories of child conduct problems into mid‐adolescence. We also found a larger association for children in better school environments.


INTRODUCTION
Several, now classic, longitudinal studies in the second half of the 20th Century revealed a strong association between academic underachievement and elevated child conduct problems (see Maguin & Loeber, 1996;Murray & Farrington, 2010, for reviews). Nonetheless, methodological difficulties posed a significant challenge to identify the true impact of education on behaviour (Hinshaw, 1992), and current evidence is still inconclusive on whether a causal relation exists between poor educational performance and conduct problems, in either direction (Kulkarni et al., 2021). Moreover, despite the importance of person-environment interactions in developmental psychopathology (Rutter et al., 1997), it is unclear whether any effects of educational performance on conduct problems vary by social context.
Given the complex bio-psycho-social determination of behaviour problems, isolating the contribution of poor educational performance is a major challenge (Fairchild et al., 2019). Education and behaviour are each influenced by both multiple environmental factors and genetics that could confound any association between them (Allegrini et al., 2020). Traditionally, studies have used statistical adjustment to adjust for measured environmental confounds. Wertz et al. (2018) recently suggested the use of polygenic risk scores (PRS) as a novel approach to testing the relationship between educational performance and conduct problems, providing triangulation with results from simple statistical adjustment. Over the last decade, GWAS has permitted the analysis of thousands of genetic variants to produce what are known as PRS, predicting a particular phenotype. A PRS summarises cumulative information across the entire genome to predict a particular condition. The first major success of this endeavour in the social sciences was the creation of a PRS predicting educational attainment among 100 thousand individuals (Rietveld et al., 2013), and this has been updated in multiple cohorts (Lee et al., 2018;Okbay et al., 2022).
A PRS genetic propensity affecting education should be less associated with other confounding factors (both environmental and genetic), than a measure of actual school performance. Hence, testing whether an education PRS predicts conduct problems can help strengthen causal inference. In the only study to date using this approach, Wertz et al. (2018) constructed a PRS for educational performance in the Dunedin longitudinal study in New Zealand, and the E-Risk cohort in the UK, and related that to both poor educational attainment in mid-adolescence and measures of conduct problems through the life-course. They found that an education PRS predicted both crime and early-onset persistent conduct problems across both samples. Further tests in other samples are needed, particularly using the PRS approach to compare with findings from traditional statistical adjustment methods in the same sample.
Poor educational performance may contribute to the development of conduct problems via multiple mechanisms. According to strain theories in criminology, frustration with school failure may lead to negative emotions affecting aggressive behaviour (Agnew, 1992) and may also influence cognitive and executive function skills involved in conduct problems (Moffitt, 1993a). Moreover, social bonding theory points to the importance of attachments to prosocial institutions, including school, protecting against conduct problems (Hirschi, 1969). Given multiple possible mechanisms involving social change after poor school performance, it seems especially likely that school failure could influence conduct problems when grade retention occurs-when a rupture of social ties is implicated, in addition to more direct, psychological consequences for children having academic difficulties.
Conduct problems themselves might also contribute to poor educational performance, but current evidence is unclear on whether bidirectional effects exist (Kulkarni et al., 2021). Although educational and behavioural difficulties may reinforce each other, the nature of this relationship could also vary according to different developmental trajectories of conduct problems. According to Moffitt's influential theory (Moffitt, 1993b(Moffitt, , 2018, conduct problems are best understood when specified by their longitudinal course -according to whether they start early in childhood, whether they persist or decline thereafter, or first become elevated from adolescence onwards. Each trajectory may involve different mechanisms, distinguished either qualitatively or quantitatively (Fairchild et al., 2013). According to this model, it seems most likely that poor school performance and conduct problems become intertwined for children with elevated neuropsychological and behavioural difficulties early in life. Little evidence is available on this question, but two cohorts from New Zealand and the UK both supported the hypothesis that poor educational performance is most strongly linked to early-onset persistent conduct problems, compared to other trajectories (Odgers et al., 2008;Wertz et al., 2018). In addition to common genetic risk factors, Carlisi et al. (2020) provided initial evidence that there are differences in brain surface morphometry in individuals with life-course-persistent conduct problems, a feature which may also partly explain their educational difficulties.
Despite improvements in recent years, Brazil is a country with high levels of school failure (UNICEF Brasil, 2021) and major challenges regarding youth conduct problems leading to serious crime (Murray et al., 2015). However, is no information regarding the conduct problems trajectories for Brazilian samples and how the educational performance impact each group. In the current study, we aim to estimate the effects of educational performance on children's conduct problem trajectories in a population-based birth cohort in Brazil, using both observational and genetic approaches. Neither method is likely to be completely unbiased on its own but using both approaches in the same study provides a valuable point of triangulation to test the robustness of associations. We test the hypothesis that school failure (measured both observationally, and via a PRS for education) predicts increased conduct problems, especially manifest as early-onset and persistent from childhood to adolescence. Finally, we also investigate possible variations in these effects according to the child's family income and school environments.

Participants
The 2004  3469 participants (82.0% of the original cohort) were included in the final analyses of both approaches, after excluding participants with missing conduct problem trajectories (n = 293) and genetic information (n = 759). The cohort methodology has been described in detail elsewhere (Santos et al., 2014) and details are also included in Supporting Information S1.

Measures of conduct problem trajectories
Conduct problem trajectories from ages 4-15 years were previously estimated, using repeated measures of conduct problems reported by the mother and transformed to z-scores. Trajectories could be constructed for the vast majority of the cohort (n = 3938; 51.9% male and 48.1% female), with valid conduct problem information at 4, 6, 11 and 15 years (ns of 3750, 3580, 3563, and 1942, respectively). Note that, only around half of the participants were assessed at the age of 15 years. At age 4 years, conduct problems were measured using the parent-rated Child Behaviour Checklist (CBCL) (Achenbach, 1991), which has been validated for use with Brazilian children (Bordin et al., 1995). The CBCL consists of 118 behavioural and emotional items, which are divided into eight subscales. The aggressive behaviour and rule-breaking behaviour subscales were summed to derive a composite measure of conduct problems (ranging from 0 to 52). At ages 6-15 years, we used the parent-rated Strengths and Difficulties Questionnaire (SDQ; ranging from 0 to 10) (Goodman, 2001). The SDQ has been validated for use in Brazil and correlates strongly with CBCL conduct problems (Saur & Loureiro, 2012). Group-based trajectory modelling, a semi-parametric approach (Nagin, 2005), was used to previously estimate conduct problems trajectories from 4 to 15 years (Martins-Silva et al., under review).

Measures of school failure and education-PRS
School failure was measured at age 11 years, asking carers whether children had ever failed a grade (yes or no). A PRS for poor educational performance (hereafter called education-PRS) was based on DNA information collected and analysed from saliva (see details in Supporting Information S1). The education-PRS was based on a recent GWAS meta-analysis of educational attainment based on 1,131,881 individuals of European ancestry, for whom 1271 independent SNPs were identified (Lee et al., 2018). Although we examined failure at school up to age 11 years, rather than later educational attainment, two cohort studies have found that a similar education-PRS predicts both (Wertz et al., 2018).
To identify the most appropriate education-PRS, we used the approach based on the clumping and thresholding methods due to its wide usage (Ni et al., 2021). We calculated four possible PRSeducation for each individual in the target sample as the sum of the SNPs weighted by their associations with education in the discovery study (Lee et al., 2018). Different P-value cut-off thresholds (P T ) were used to define whether an SNP was included in the four PRS scores: P T = 5 � 10 −8 ; P T = 5 � 10 −6 ; P T = 0.05 and P T = 0.5 (Euesden et al., 2014). We adjusted for multiple comparisons by using the false discovery rate (FDR) to control possible Type I error. All four PRS-education scores were reverse-coded and transformed into z-scores-so that higher scores indicate a higher genetic risk for poor educational performance. The resulting education-PRS with the strongest association with school failure in this cohort was used in the main analysis (see Table S2).

Covariates
When analysing carer-reports of school failure as a predictor of conduct problem trajectories, we adjusted for child sex (observed at birth; male or female), and the following covariates measured at age 48 months: maternal schooling (0-4, 5-8, or ≥9 years), family income (tertiles), maternal depression, poor child development, and child attention problems. Maternal depression was measured using the Edinburgh Postnatal Depression Scale (EPDS) (Cox et al., 1987).
The EPDS is a self-report questionnaire, asking about depressive symptoms over the past 7 days and has been cross-culturally adapted and validated for use in Brazil (Santos et al., 2007). The 10 items are scored on a 3-point scale (ranging from 0 to 30).
Battelle's Development Inventory (BDI) (Newborg, Stock, & Wnek, 1988) is a screening tool for child development comprising 96 items. We dichotomised the BDI scores to define a low child development group using the 10th percentile as the cut-off point (i.e. children with low child development belonging to the first decile), given the relevance of this group in prior research in this sample (Barros et al., 2010). We also used the attention problems subscale of the CBCL (ranging from 0 to 16) (Achenbach, 1991). At age 6 years, the intelligence quotient (IQ) was measured, using the Wechsler Intelligence Scale for Children-III (Barros et al., 2010).
Low IQ was defined as a score of ≤70.
At age 11, we used information about the school environment as both a covariate and effect modifier. The school environment was assessed using 11 questions concerning the child's school and their EDUCATIONAL PERFORMANCE AND CONDUCT PROBLEM TRAJECTORIES -3 of 11 relationships with classmates and teachers (including both positive and negative aspects; e.g., 'Kids in my classroom push and shove each other a lot' and 'I feel safe at my school'). Items were chosen and adapted from the Peace Zone questionnaire (Prothrow-Stith et al., 2001; For more details see Supporting Information S1). Items on school violence were reverse-coded, so the final score (ranging from 0 to 33) indicated a more supportive, less violent school environment. As a covariate, the continuous score was used, and in moderator analysis, a categorical variable (in tertiles) was used.
When analysing education-PRS, we include child sex and the first 10 principal components analysis (PCA) of ancestry. The PCA was based on the whole genomic dataset and was performed using PLINK1.9 (Purcell et al., 2007).

Statistical analysis
First, missing data analyses were conducted comparing children included in the current analyses versus children lost to follow-up on covariates, using Pearson's chi-square tests and t-tests. Second, we examined the association between carer-reported school failure (repeating a school grade) and conduct problem trajectories, using adjusted multinomial logistic regression. Results are presented as odds ratios with 95% confidence intervals. All variables were entered simultaneously in the adjusted analyses. In sensitivity analyses, we re-ran the models to see if associations remained when also adjusting for pre-school behaviour problems, measured on the CBCL at age 4 years.
Next, we examined associations between education-PRSs (based on different P T values) and carer-reported school failure, using logistic regression analysis. To identify the variance explained by each education-PRS, a pseudo-R-squared (R 2 ) was estimated by calculating the difference between the full model (including the relevant education-PRS and covariates) and the null model (covariates only).
Using the best education-PRS identified in our sample, we examined its associations with conduct problem trajectories, using adjusted multinomial logistic regression. In the adjusted model we included family income, maternal schooling, the school environment score, sex of the child, and 10 first principal components for genetic ancestry.
Finally, we tested whether associations between educational performance and conduct problem trajectories varied according to the following possible modifiers: family income and school environment, using the likelihood ratio test-comparing logistic regression models with and without relevant interaction terms. Analyses were performed with PRSice 2.2.1, STATA 16.1, and the RStudio software, and the code used is available at: https://github.com/ ThaisMartins-Silva/Educational-performance-and-conduct-problemtrajectories-code.git.

RESULTS
Conduct problem trajectories identified were shown in Figure 1 and descriptive statistics are presented in Table 1. We identified four groups of conduct problems: early-onset persistent, in which conduct problems emerge in childhood and persist into adolescence; childhood-limited, in which conduct problems remit in the transition from childhood to adolescence; adolescence-onset, in which conduct problems first emerge in adolescence; and low conduct problems (see Figure 1 and Table S1). Almost one-fourth of children (24.2%) failed a school grade by age of 11 years and nearly one-fifth (17.9%) were classified as having childhood-limited conduct problems, whereas a smaller proportion had adolescence-onset (7.1%) and early-onset persistent conduct problems (3.9%). Children in the analytic sample were less likely to have low IQ and had slightly lower levels of attention problems compared to those excluded (see Table 1).

Education-PRS and school failure
We examined four different education-PRS (based on different P T thresholds) in relation to school failure. All were associated with school failure, even after FDR correction. As expected from previous GWAS studies, the higher (inverted) education-PRS, the greater the risk of school failure. The education-PRS that had the strongest association with reported school failure (P T = 0.05, 67,668 SNPs) explained 1% of the variance (pseudo-R 2 = .0102, OR: 1.43; 95% CI: 1.27; 1.60) and was used in the main analyses (see Table S2).  were not significant when comparing the early onset-persistent to the adolescence-onset group. When behaviour problems at age 4 years were added to the model (Model B), these associations remained significant, with slightly larger effect sizes.

Are there different effects of school failure and education-PRS in different social environments?
Next, we examined whether associations between school failure and education-PRS and conduct problem trajectories varied according to family and school environments. This was done separately for each indicator of education in relation to a broad measure of the family income and school environment.
The most consistent and strong variations in the association between both school failure and education-PRS and conduct problems trajectories were observed for the school environment variable (see Table 3). First, using the carer-reported measure of school failure, the associated risk for childhood-limited, adolescence-onset, and early-onset persistent groups (compared to the low group) was significantly larger among children in more favourable versus less favourable school environments (1st vs. 3rd tertile). For example, among children in the less favourable school environment, repeating a grade increased the odds for early-onset persistent However, interaction results for family income showed little evidence. Considering education-PRS, its association with childhoodlimited problems was higher in mid-income compared to low-income families (2nd vs. 1st tertile; p = .024), but no other interaction for family income was found either for observed or education-PRS measures (see Table S3).

DISCUSSION
This population-based, longitudinal study in Brazil showed one in four children experienced school failure (repeating a grade) up to age 11 years and school failure was strongly associated with conduct problems persisting through childhood into adolescence. Smaller associations were also observed with childhood-limited and adolescence-onset groups. These findings were consistent when analysed using two different approaches, considering both school failure measured via self-report and an education-PRS, with triangulation across these methods helping to strengthen causal inference. Interestingly, the association between school failure and conduct problems trajectories was larger for children in better school environments; no variation in effects was found according to family socioeconomic factors.

Prior longitudinal studies have found inconsistent and often
weak results for the relationship between educational performance and conduct problems when adjusting for child IQ or language skills (Kulkarni et al., 2021). All these studies were conducted in North T A B L E 1 Description of sample and comparison between participants included and not included in the analysis Adjusted by sex of the child, family income, maternal schooling, maternal depression, school environment score, and child neurocognitive indicators (low child development, low IQ, and attention problems). Model B: Adjusted by sex of the child, family income, maternal schooling, maternal depression, school environment score, child neurocognitive indicators (low child development, low IQ, and attention problems), and behaviour problems at age 4-years. Model C: Adjusted by family income, maternal schooling, school environment score, sex of the child, and 10 first principal components for genetic ancestry.
Abbreviations: 95% CI, 95% confidence interval; education-PRS, education polygenic risk score; OR, odds ratio; P T , P-value threshold for the PRS. trajectories have specific psychosocial determinants (Fairchild et al., 2013;Moffitt, 2018). Indeed, two cohorts in the UK and New Zealand, which did examine longitudinal trajectories as an outcome (and also used education-PRS), also found that education indicators (including genetic) were most strongly related to early-onset persistent group, rather than other trajectories (Wertz et al., 2018).
We emphasise again that, in this and our study, the association could reflect a cumulative impact of poor school performance and conduct behaviours reinforcing each other, as well as unidirectional effects.
The current study showed that the association between educational performance and conduct problem trajectories varied by school context. Pelotas, like the rest of Brazil, is a city with considerable socioeconomic inequality and markedly different school environments. Perhaps surprisingly, repeating a grade (and having a high education-PRS) was associated with conduct problems most strongly for children in better schools (in terms of social support and less violence). These results could be explained by negative labelling effects that override any protective influence of more favourable school environments on conduct problems, when grade repetition occurs. According to labelling theory, escalating behaviour problems arise when individuals internalise social reactions that communicate, they have deviated from social norms and expectations (Farrington & Murray, 2014). This labelling hypothesis was also suggested by another study identifying poor academic performance as a determinant of conduct problems, in which the authors considered that different expectations by teachers, students, and parents tended to follow school failure, and children then became less motivated to engage in school, and more prone to behaviour problems (Eisenberg et al., 2009). However, we cannot rule out the possibility that our were considered at risk of failing at school (Schweinhart et al., 1985).
These findings, and those from other studies (Heckman et al., 2010), suggest a key strategy tackle both educational and behavioural challenges is to invest in preschool enrichment programmes. Adjusted by sex of the child, family income, maternal schooling, maternal depression, and child neurocognitive indicators (low child development, low IQ, and attention problems). c Adjusted by family income, maternal schooling, sex of the child, and 10 first principal components for genetic ancestry.

of 11 -
problems can complement support for parents to manage behavioural difficulties at home. The evidence-based Incredible Years programme, which includes modules for both parents and teachers is particularly relevant in this regard (Webster-Stratton et al., 2008).
As far as we know, this is the first study examining the association between educational attainment and conduct problem trajectories across different school environments using a PRS approach.
Education-PRS as a point of triangulation to compare with results from traditional analyses and more clearly meets the first criterion for causality of temporal precedence before the outcome. It is also plausible that an education-PRS is less confounded than a measure of actual educational performance for analysing conduct problems. However, education-PRS could also influence conduct problems through pathways other than educational performance, meaning that associations may not reflect a true causal effect of education. We addressed this in sensitivity analyses, by adjusting for children's baseline behaviour, and found similar results. The consistency of results based on the education-PRS and regression models analysing carer-reported school failure helps bolster confidence in the findings. The education-PRS in this study explained a small (1.0%) amount of variance in observed school performance, which is similar to in previous education-PRS studies (Wertz et al., 2018). Nonetheless, the education-PRS was still associated with adolescent-limited and early-onset persistent conduct problems (compared to low-conduct problems). Like in the observational analyses based on actual school failure, the education-PRS was associated with a higher risk for early-onset persistent problems compared to adolescent-onset problems; however, this was not significant, perhaps because of low power for the education-PRS which explained just 1% of the variance in educational performance. suggesting the robustness of the score, and we adjusted for ancestry in the analyses. The fact of gene-environment correlation (education-PRS being associated with actual school failure) was key to our genetic study design, to consider if educational failure contributed to conduct problems. We also adjusted for pre-school conduct problems to reduce the chance that any association with conduct problems reflects other paths between education-PRS and the outcome.
However, we cannot rule out that the education-PRS was associated either with other genetic variants related to conduct problems or that the education-PRS has evocative or active correlations with other environmental factors influencing conduct problems that were not adjusted for in this study.
In conclusion, this multi-method study suggests that educational performance and conduct problems are intimately entwined, particularly when children demonstrate behavioural problems spanning from childhood into adolescence. This is the first study to suggest that school context matters, such that repeating a grade in a 'better'

CONFLICT OF INTEREST
The authors have declared that they have no competing or potential conflicts of interest.
EDUCATIONAL PERFORMANCE AND CONDUCT PROBLEM TRAJECTORIES -9 of 11 DATA AVAILABILITY STATEMENT Data available on request from the authors.

ETHICAL CONSIDERATIONS
The 2004