Quasi-experimental analyses of the effect of ADHD on education performance in youths across sexes and ancestry

doi:10.21203/rs.3.rs-4638880/v1

Attention-deficit hyperactivity disorder (ADHD) is prevalent in children and associated with lower education performance. This association is often obtained from observational studies, which have limited ability to identify causal relationships. Knowledge of causes is useful, because it delineates intervention avenues more clearly, but research into causation in childhood is hindered by ethical and practical limitations to randomized clinical trials. The present study uses twin study designs and polygenic risk scores (PRS) to investigate the causal relationship between ADHD symptoms and poor educational performance. We combined family-based direction of causation modeling and Mendelian randomization (MR) approaches and applied them to ABCD study data for improved causal inference. The models used necessarily make assumptions about either horizontal pleiotropy, or the level of individual-specific confounding factors. Results support the hypothesis that ADHD symptoms have a causal effect on educational performance, and vice versa. No heterogeneity related to sex was found, nor moderation of causal estimates by gender. These findings highlight the complex interplay between ADHD and educational outcomes and emphasize the need for larger sample sizes in future research.

ADHD

educational attainment

MR-DoC TENTATIVE JOURNAL: Behavior Genetics

Attention-Deficit/Hyperactivity Disorder (ADHD) is one of the most prevalent externalizing neurodevelopmental disorders, affecting from 7.2% (95%CI 6.7–7.8) (Thomas et al., 2015) to 7.6% (95%CI 6.1–9.4%) of all children and 5.6% (95%CI 4.8-7%) of adolescents (Salari et al., 2023). Inattention, hyperactivity and impulsivity are the main symptoms of ADHD, which is frequently comorbid with other psychiatric disorders, including learning disorders, autism and obsessive-compulsive disorder (Fayyad et al., 2017; Sousa et al., 2020). ADHD is also genetically correlated with other disorders like Autism spectrum disorder and depression (Demontis et al., 2019; Loughnan et al., 2022).

ADHD symptoms are highly heritable, apparently due to multiple genetic loci, each with small effects (Faraone and Larsson, 2019). Twin studies found a 9-fold increased risk of ADHD in siblings of probands than in siblings of controls, with a sibling correlation of 0.3 (Chen et al., 2008; Faraone and Larsson, 2019). Twin studies also found a heritability of 80% (Chen et al., 2017), and adoption studies suggest minimal shared environmental effects (Sprich et al., 2000). The heritability is considered similar in males and females (Faraone and Larsson, 2019), but there is a male-to-female prevalence ratio that ranges from 4:1 to 2:1 (Sousa et al., 2020). Furthermore, clinical characteristics slightly differ between sexes, with males showing more hyperactivity and impulsivity than females (Biederman et al., 1999). Additionally, there is evidence of a sex-of-proband effect, with male relatives of females probands having higher genetic risk (Taylor et al., 2016). There is no evidence of differential heritability of ADHD between males and females (Larsson et al., 2014), and the genetic correlation for ADHD between sexes is close to one (Taylor et al., 2016). Altogether these findings suggest that the genetic factors influencing ADHD are the same in males and females, but females are protected against their expression. It is currently unclear whether these differences result in differential effects on the association between ADHD and other phenotypes, or more specifically on educational performance.

ADHD correlates negatively with academic success (Boomsma et al., 2010; Demontis et al., 2019). However, these correlational studies are unable to establish a causal relationship between EA and ADHD. For causal inference, one usually needs a randomized intervention design, typically a randomized controlled trial (RCT). Such experiments are harder to execute during childhood due ethical or logistical limitations. Mendelian randomization (MR) studies offer an alternative when RCTs are not feasible. MR analyses require a genetic variant that correlates enough with the exposure to avoid weak instrument bias (Bowden et al., 2015) In its simplest form, the comparison of GWAS with outcome and GWAS with exposure, MR also requires the assumption that the variant correlates with the outcome exclusively via the exposure variable, which is also known as the no horizontal pleiotropy or exclusion-restriction assumption. If this assumption holds, the genetic variant may be considered to be an instrumental variable. Other quasi-experimental methods, such as twin or longitudinal studies, provide another alternative to RCTs, in particular the now classic Direction of Causation (DoC) model (Heath et al., 1993), which uses the cross-trait cross-twin correlations to make inferences about causal direction. Interestingly, to our knowledge there has been no previous DoC study of ADHD and educational performance or attainment.

In an attempt to provide information about a causal relationship, a group recently reported results from within-sibship GWAS and Mendelian randomization in the Dutch population registry. They reported a negative causal effect of ADHD liability on educational attainment (IVW ß= -0.38, P < 0.004; Demange et al., 2023). However, they reported results from adult ADHD, which is itself a controversial construct (Hinshaw, 2018), and might not represent the most affected children with the condition. Demange et al. (2023) further reported MR results, which suggested a bidirectional effect between EA and ADHD, the effect of EA on ADHD was IVW OR = 0.69, P < 0.004. Their approach consisted of using instruments (polygenic scores) for EA and ADHD to perform two two-stage MR analyses, one for each causal direction. The Demange et al. (2023) paper is consistent with previous findings from two-sample MR studies that also reported a bidirectional effect between EA and ADHD (Dardani et al., 2021; Michaëlsson et al., 2022).

MR studies have become very popular in situations where RCTs have limited use (Lewis and Vassos, 2020). It has four key assumptions: 1) exchangeability, which means that the variant is not associated with a confounding factor in the relationship between the exposure and outcome; 2) exclusion restriction, which states that the variant only affects the outcome through the exposure; 3) relevance, indicating that the instruments are sufficiently predictive of the exposure; and 4) there is no reverse causation from outcome to exposure. Evaluating whether the last three assumptions hold can often be challenging. First, the exclusion restriction assumption (also known as no horizontal pleiotropy) states that the genetic variants affect the outcome exclusively via the exposure variable. When the underlying biochemistry is well understood, such an assumption may be reasonable. For complex traits such as ADHD, there is a general lack of understanding of the pathways from SNP variant to outcome phenotype. It is therefore important to validate this assumption. Second, for complex traits such as ADHD that are far removed from the molecular genetic biochemistry, the SNP variants affecting the outcome trait, are expected to be very weak. Most mental disorders are highly polygenic with each variant exerting a tiny effect on an outcome phenotype (Abdellaoui and Verweij, 2021). Thus there is a high risk of encountering biases due to weak instrument biases while applying MR to complex traits. While this problem may seem to be circumvented by the use of a weighted sum of the variant’s effects, such as a polygenic score (PS), the opportunities for violation of the exclusion restriction assumption rapidly multiply.

In general, the no horizontal pleiotropy assumption is thought to be implausible with respect to mental disorder phenotypes. Pleiotropy is generally classified as vertical or horizontal. Vertical pleiotropy is when a variant has a causal effect on a risk factor or exposure, which then causes the outcome.. This type of pleiotropy is explicitly modeled in MR and does not bias estimation. Horizontal pleiotropy is when genetic variant has a causal effect on the outcome, without going via the exposure variable, which may happen via several mechanisms. While such effects are often considered direct (partly because they are modeled with a single regression parameter of outcome on SNP variant or PS), these alternative pathways from genetic variant to outcome phenotype would generally involve other mediating variables. To complicate things further, SNP variants are typically in linkage disequilibrium (LD) with other variants, which may have causal effects on the outcome. This type of pleiotropy is ubiquitous in nature and is better thought of as part of the genetic architecture (Jordan et al., 2019). LD is therefore highly problematic to Mendelian randomization as it potentially biases parameter estimates (Verbanck et al., 2018). Horizontal pleiotropy is the reason why it is important to triangulate MR results using methods with different assumptions.

Statistical methods for Mendelian randomization are many; they include complex two-stages least squares, wald estimator, MR-Egger, or inverse-variance weighting. However, standard MR regression can be specified and easily estimated via structural equation modeling (SEM) (Maydeu-Olivares et al., 2019). MR nevertheless has limitations, including biased estimates when the exposure is not continuous (Howe et al., 2022) or has different test-retest correlations than the genetic variant (Maxwell & Cole, 2007). Switching the estimation framework to SEM can provide benefits by facilitating the application of a latent liability threshold to categorical variables and by allowing bidirectional causal modeling when twin data are available (L. F. S. Castro-de-Araujo et al., 2023).

Attempts to provide evidence of bidirectional causation using MR are limited, as current practice often consists of running unidirectional MR analyses twice and reporting the significance of the association in each direction (Timpson et al., 2011). This approach assumes that feed-back loops do not exist, but in principle the two disorders could both be causes of each other (Burgess et al., 2021). This type of interaction cannot be detected by most current bidirectional MR methodologies. Methods that extend MR using data from relatives, such as the classical twin study provide additional information (especially the cross-trait, cross twin correlations) that make the model with the pleiotropic path identified and also allow for bidirectional causal inference (L. F. S. Castro-de-Araujo et al., 2023). In what follows, the ABCD Study twin data will be used in three such models: one that uses exclusively the cross-twin cross trait correlations to inspect a causal direction, DoC (Heath et al., 1993); one that performs MR in the presence of pleiotropy, MR-DoC (Minică et al., 2018); and a third that allows for bidirectional causal estimation MR-DoC2 (L. F. S. Castro-de-Araujo et al., 2023). All these models are specified with SEM.

This paper aims at clarifying whether ADHD causally affects educational performance after horizontal pleiotropy or reverse causation is taken into account. This will be achieved by applying MR extensions of twin designs to data from twins in the ABCD study®. Given current findings using classical MR, it is expected that there is both a significant causal effect of ADHD on educational performance and vice versa. Also, a significant sex effect is expected, considering what is known regarding sex differences in the prevalence of ADHD.

2.1. Data source

Data were obtained from the release v5.0 of the Adolescent Brain Cognitive Development (ABCD) study (Haist and Jernigan, 2023), a large-scale, longitudinal study that aims to track the development of approximately 10,000 children across the United States. The study collects data on various domains, including neuroimaging, genetics, cognitive function, psychopathology, substance use and environmental factors (Maes et al., 2022). We will be using a subcohort of twins and siblings in the ABCD study, more specifically only those for whom data were available from the third wave of the study, which included 2,642 individuals. Zygosity assignment used a combination of the calculated genetic correlation between the twins, a twin zygosity rating (Conomos et al., 2016). Conflicting cases were discussed and decided in consensus between the first author (LFSCA) and DZ. Siblings were also used in the analyses. The rationale for using the third wave was to strike a balance between the age in which symptoms are present and at the same time not having too many participants lost to attrition. ABCD's study wave 4 was particularly affected by missed visits. Given that PRSs were calculated based on summary statistics from a GWAS that included Europeans only, the MR-DoC and MR-DoC2 analyses were restricted to individuals empirically assigned as European ancestry (n = 1,660)(Peterson et al., 2017). However, since DoC can be applied to all ancestral groups, individuals of admixed American (AMR,551) and African ancestry (AFR, 431) were also included. The resulting number of families were 1411 families (EUR = 876, AMR = 305, AFR = 230), with 284 MZ pairs (EUR = 197, AMR = 50, AFR = 37), 394 DZ pairs (EUR 260, AMR = 63, AFR = 71), and 733 sib pairs (EUR = 419, AMR = 192, AFR = 122). Unfortunately, in the ABCD Study, opposite sex pairs were not ascertained. The MZ and DZ twin pair correlations are reported in Table 2, together with the means and standard deviations (SDs) for the key variables. Covariates (6 principal components, history of ADHD medication treatment, and sex) were added to the models as definition variables, which discards cases with missing values. The analyses were performed on 782 EUR, 242 AMR, and 197 AFR complete pairs. Point estimates and standard errors are reported for DoC, and MR-DoC in Tables 3 and 4. Wald type confidence intervals are reported.

2.2 Instruments and polygenic score calculation

The Child Behavior Checklist (CBCL) ADHD subscale was used to evaluate ADHD symptoms and their progression over time. This subscale, which is reported by the caregiver, includes questions regarding inattention, hyperactivity, and impulsivity (Achenbach, 2001). Widely employed for assessing child and adolescent behavior problems, the CBCL ADHD subscale reliably tracks DSM diagnosis for ADHD (Achenbach, 2013). In comparison to the reference standard of clinical evaluation by a qualified professional, the CBCL has been reported to have a sensitivity of 0.77 and specificity of 0.73 for diagnosing ADHD (Chang et al., 2016). The DSM-oriented CBCL scales generally exhibit a high Cronbach's alpha of 0.92, while the CBCL ADHD subscale has been reported to have an area under curve of 0.75 for diagnostic efficiency (Jiang et al., 2023). This variable was recoded as a 9-category ordinal for model-fitting. Collapsing some of the categorical variable levels was required as some level bins had less than 10% of the observations, and this resulted in poor resolution. In the multigroup sex analyses, the variables were further recoded as 4-category ordinals, as an adjustment to the smaller number of observations.

The second phenotype of interest was the youths’ school grades in the previous year, as reported by their parents. This variable was coded so that 1 corresponded to A + and 12 corresponded to an F. This variable was reversed, so larger values represented better performance, and included in the models recoded as a 9-category ordinal.

Polygenic scores (PRS) were calculated from the latest GWAS for each phenotype. ADHD liability PRS were estimated with summary statistics from Demontis et al., (2023). PRS for educational attainment were calculated using summary statistics from Okbay et al., 2022. Quality Control (QC) on the GWAS summary statistics were performed following standard guidelines (Choi et al., 2020) which included filtering out SNPs with minor allele frequencies (MAF) < 0.01, INFO scores < 0.8, duplicate SNPs, and those that were ambiguous (ex A-T or G-C variants). QC of genotypic data in the target ABCD sample were also performed according to standard guidelines and included filtering out SNPs with a MAF < 0.01, Hardy-Weinberg-Equilibrium Fisher’s Exact Test p-values < 10^− 6, and missingness rates > 0.01. Individual samples with genotype missingness rates > 0.01 were also excluded. We then used PRS-cs (Ge et al., 2019) to infer posterior SNP effect sizes which were then used to compute individual level PRS in the target ABCD study samples using PLINK (Purcell et al., 2007). PRS were computed in ABCD samples of European ancestry as GWAS summary statistics from non-European ancestry groups were not available.

2.3 Covariates

Sex was added as covariate, except in the sex heterogeneity test. History of ADHD medication (Concerta, Daytrana, Focalin, Metadate, Ritalin, Amphetamine, Adderall, Vyvanse) use at baseline was also included as covariate. Also, in order to control for population stratification, the first six principal components (PCs) were included as covariates. These were within-ancestry PCs calculated in-house for the ABCD study® (Fan et al., 2023).

2.4. Data analysis

We used R version 4.2.2 (https://cran.r-project.org/) in all further analyses, including the OpenMx and umx packages used for SEM (structural equation modeling) analyses (Bates et al., 2019; Neale et al., 2016). Parameter estimation used full information maximum likelihood (FIML) with the NPSOL optimizer. We verified that the models were identified using the OpenMx utility mxCheckIdentification (Hunter et al., 2021). The high computational costs of analyzing both phenotypes as ordinal categorical variables required the use of 23 CPUs from the VIPBG cluster (vipbg.vcu.edu/resources/computational-facilities). All loaded R packages and times for each run can be found in the repository for this paper (github.com/lf-araujo/2023-mrdoc2-EA-ADHD). Descriptive information of the data is reported in Table 1. All tests considered an alpha of 0.05, although we focus more on effect size than significance, due to the large sample size of ABCD (Dick et al., 2021). A full list of estimates for all parameters in each ancestry can be seen in the linked repository. We report the full list of estimates of MR-DoC2 in Table 3.

2.5. Model Specification

Three statistical models are used here. The Direction of Causation (DoC) model (Fig. 1, A), a now classic model that uses cross-twin cross-trait correlations to assess direction of causation between two phenotypes (Heath et al., 1993), and its extensions: MR-DoC, which adds a Mendelian randomization component (Minică et al., 2018); and MR-DoC2, which extends MR-DoC by allowing bidirectional causal inference in the presence of background confounding (L. F. S. Castro-de-Araujo et al., 2023). Figure 1 shows path diagrams of all three models. MR-DoC allow for causal estimation without bias from horizontal pleiotropy, when one genetic locus directly affects both exposure and outcome, a common limitation from classic MR. MR-DoC2 accounts for indirect horizontal pleiotropy, when one genetic locus affects both exposure and outcome via linkage disequilibrium.

The DoC model is designed for data collected from relatives, and when these are MZ and DZ twin pairs it allows typical ACE decomposition. The phenotypic variances are decomposed into components associated with additive genetic (A), shared environmental (C) and unique environmental (E) variances. This research design cannot jointly estimate dominance genetic variance (D), which is confounded with (C). Adoption and twin studies have found that there is little shared environmental variance in ADHD (Sprich et al., 2000), so it is often modeled as an AE phenotype. However, ACE modeling in these data found all three variance components to be non-zero (Maes et al., 2022) (Maes et al 2023) for this reason we kept the full model in this analysis. For identification reasons this model can only estimate three of the five paths (g1, g2, covA, covC, and covE); the specification used here drops g2 (reciprocal causal path) and covE (specific environmental confounding or correlated errors of measurement); Fig. 1, A).

DoC models have been extended with PRSs to allow for Mendelian randomization with various combinations of horizontal pleiotropy and background confounding. Minica et. al (2018) proposed MR-DoC (Fig. 1, B), which is an extended direction-of-causation model for twin data that uses polygenic scores as instrumental variables in MR. This model improves causal inference, which in some versions is robust to horizontal pleiotropy. The b2 path (Fig. 1, B) models the direct effect of the PRS on the outcome, therefore controlling for horizontal pleiotropy, whereas the path g1 is the causal path. The model is identified with five of the six paths (b1, b2, g1, covA, covC, and covE); the specification used in these analyses retains b2, as we are interested in controlling for horizontal pleiotropy, but dropped covE. Dropping covE when its true value is not zero biases the estimates of the causal path (g1) in the cases where covE does not equal zero. However, this bias is small (L. Castro-de-Araujo et al., 2023).

The MR-DoC2 model is an extension to MR-DoC in which both traits have a PRS, thus allowing estimation of bidirectional causal paths with cross-sectional data. Each PRS works as an instrumental variable for each of the phenotypes in the model (Ph1, Ph2; Fig. 1C). Pleiotropy is partially addressed by the model, but instead of the more direct specification of the b2 path in MR-DoC, it is specified indirectly as the paths b1*rf and b3*rf. Additionally, MR-DoC2 is identified with full ACE background confounding control (covA, covC, and covE; Fig. 1C). The causal paths are g1 and g2.

Finally, ordinal variables were specified using the Mehta-Neale approach of fixing the first two thresholds (Mehta et al., 2004) and all covariates were included in the model using the definition variables approach (Neale et al., 2016). All specification was done in algebra.

The OpenMx package does not currently provide a full list of fit indices for models that also includes definition variables, which is the case of this analysis. The first author wrote a custom function (see linked repository for code) to extract the CFI in these cases, the fit indices for DoC model were − 2lnL 9031.141, AIC 3295.141, CFI 0.981. The CFI suggests the model fits well to the data. The point estimate for the causal path (g1, Fig. 2) was − 0.268 (95%CI -0.412, -0.124). This estimate should be interpreted as a negative effect of ADHD on educational performance. The interpretation follows similar logic in all subsequent models. The direction of the causal effect was similar in EUR, AMR (-0.259; 95%CI: -0.333, -0.184) and AFR (-0.07; 95%CI: -0.261, 0.122), however it was not significant in the latter. The reverse causal relationships (from education to ADHD scores) presented good fit to the data and resulted causal negative estimates and significant: EUR − 0.731 (95%CI -1.11, -0.35), AMR − 0.975(95%CI: -1.11, -0.839), and AFR − 0.058 (95%CI: -0.449, 0.334). See Fig. 2 for a range-in-plot including point estimates of the causal paths in all models.

The PRSs regressed on the phenotypes for the full data set used in the analyses (Table 1) shows an adjusted R² of 0.02 with an F-statistic of 32.5 for ADHD and 0.01 and F-statistic 17.7 for EA/EP.

MR-DoC was fit to the EUR data to assess the change in parameter estimates given the added polygenic score information. Indices for MR-DoC in the ADHD to education performance direction were − 2lnL 12541.883, AIC 4103.883, CFI 0.825, showing a slightly worse fit to the data according to the CFI (Table 4). However, these two models are not nested and MR-DoC includes the extra PRS data, preventing direct comparison to the DoC model. The point estimate is weaker than in the DoC model: g1 = -0.2523 (95%CI -0.3451,-0.159421). The reverse MR-DoC (education to ADHD) resulted in g1 = -0.693 (95%CI -0.9512,-0.4351; Fig. 2).

At this point, we have results that suggest a bidirectional relationship, but none of the fitted models explicitly specified bidirectional causation. We then proceeded to fit MR-DoC2 to the EUR data. The fit indices were − 2lnL 16256.539, AIC 5094.539, and CFI 0.896; therefore, we observed a fit improvement in relation to MR-DoC but not in relation to the DoC model. The point estimates were g1 = -0.856 (95%CI -0.9742,-0.7370) and the reverse g2 = -0.5 (95%CI -0.5835,-0.4158). So the causal effect from ADHD to education performance increased, whereas the reverse direction decreased in comparison to DoC, and MR-DoC (Table 3, Fig. 2).

The EUR set was further split into male and female subgroups and a multiple group model by sex was fit to the data. The number of twin pairs in the male group was 306 and 300 in the female. The causal estimates were: g1 = -0.2488 (95%CI -0.50485, 0.007328), g2= -0.3954 (95%CI -0.55838, -0.232480) for males; and g1 = -0.5123 (95%CI -0.72800, -0.296623), g2= -0.5079 (95%CI -0.70970, -0.306039) for females.

The results support the hypothesis that there is a causal effect from ADHD symptomatology on poor educational performance and vice-versa. A likelihood ratio test where either g1, g2, or both causal paths are fixed to zero and compared to the full model reveals that the models without g1 or both g1 and g2 are significantly worse (Table 4). The unconstrained model had the lowest AIC (16359). Thus confirming that the feedback-loop is significant in this model.

To assess sex differences, the MR-DoC2 model was fitted to EUR males and females separately, and twice the negative log-likelihood (-2lnL) was extracted for each model. The sum of the − 2lnLs for each sex (males − 2lnL = 6144; females − 2lnL = 5717) is equal to the − 2lnL of the model fitted to the combined male and female data set (total − 2lnL = 11861). This test has 41 degrees of freedom, the number of parameters in the model. This overall heterogeneity test of male-female differences suggests no important sex effect. This finding was surprising given the known sex differences in ADHD. A follow-up LRT test to assess whether sex has a specific effect on the causal paths (g1 and g2) was performed by equating g1 and g2 across sexes in a multiple group analysis. The test revealed a Δ Fit of 0.203 (2 df), which is not significantly different from the model without the equality constraint, so the hypothesis that the degree of causation differs between the sexes is not rejected.

This quasi-experimental study examined the relationship between ADHD symptoms and poor educational performance using a twin design and polygenic risk scores (PRS). Results are consistent with a causal relationship of ADHD symptoms on poor educational performance, with a causal path estimate of -0.856 from ADHD to education performance and − 0.5 in the reverse direction, both significant, when using a PRS-informed bidirectional causal inference twin and siblings analysis. Sex differences were also examined, which revealed no overall sex difference in means or variances or both.

The findings of this study provide further evidence for a causal relationship where ADHD symptoms adversely affect educational performance. This is consistent with previous findings from analysis using MR (Dardani et al., 2021; Demange et al., 2023; Michaëlsson et al., 2022), however in our study the effects of ADHD in children on their grades were tested using a model that includes horizontal pleiotropy. Better education performance significantly decreased ADHD symptomatology according to the CBCL scale, also confirming previous MR findings (Dardani et al., 2021; Michaëlsson et al., 2022).

Although the ABCD study includes individuals of diverse ancestries, the PRSs were trained only in European GWAS, and therefore we only were able to use the data on individuals of European ancestry. This limitation highlights the need for larger sample sizes and more robust methodologies to fully use data from diverse ancestries and to further understand the complex interplay between ADHD and educational performance across all human populations.

Potential sex differences in the relationship between ADHD and educational performance were assessed. An omnibus test for sex differences revealed no significant effect. A follow-up analysis also did not identify sex moderation of the causal path from ADHD to educational performance. The absence of overall sex heterogeneity was surprising given the known sex differences in ADHD prevalence (Hinshaw, 2018; Sousa et al., 2020) and the known sex-of-proband effect (Taylor et al., 2016). Unfortunately, in the ABCD Study, opposite sex pairs were not ascertained which hindered further sex moderation analyses.

There are several limitations that should be taken into account when considering this study. Basic ACE models assume that dominance, epistasis, assortative mating, gene-environment covariation and interaction are all absent and that the genetic effects are additive.. However, it is plausible that there is significant gene-environment covariance for psychiatric conditions (Abdellaoui and Verweij, 2021; Demange et al., 2021). Models for latent gene-environment interaction effects are becoming possible in recent additions to structural equation modeling (Boker et al., 2023). Marital correlations for educational attainment are substantial, although this may be for reasons of social stratification rather than direct phenotypic assortment. Stratification and assortment would inflate estimates of shared environment variance. Another limitation is that the classical twin design involves the equal environment assumption (EEA) that non-elicited environmental effects are equally correlated for MZ and DZ twin pairs.

The CBCL used in this study is not a diagnostic tool, although it is highly correlated with structured diagnostic instruments (Achenbach, 2013). Study wave 3 was entirely performed during the Covid-19 pandemic, and coincided with ages at which ADHD typically reaches peak prevalence (Sousa et al., 2020). The pandemic started at the end of wave 2 and may have increased attrition in wave 4. The pandemic could potentially have affected students’ grades in the previous year and thus may have added error variance to the study’s findings.

Several environmental factors are associated with ADHD. Kim et al., (2020) report in an umbrella review that childhood eating disorder, low weight or preterm birth, low parental educational attainment, and exposure to lead (among others) are environmental risk factors associated with ADHD. Parental educational attainment also predicts offspring educational attainment (Jami et al., 2021). Therefore, the educational environment that parents provide for their children may influence their children’s ADHD and education performance, which is a potential source of environmental confounding. Classic MR cannot account for this type of confounding, nor can MR-DoC (provided one intends to control for horizontal pleiotropy, retaining path b2), but it is controlled in MR-DoC2. A stark advantage of our approach.

Causal studies of ADHD in youth are uncommon. Instead, studies have used data from ADHD diagnosis in adults to extrapolate results to children. This situation may have arisen simply from the lack of data in children, which in turn may be due to ethical considerations associated with this vulnerable population. The present study adds to extant findings suggesting a causal relationship between ADHD and education. Its strength comes from using FIML for ordinal variables, the model's partial robustness to the horizontal pleiotropy assumption, dynastic and assortment effects. Our findings suggest that intervening both in ADHD treatment and in improving educational performance will generate better results, as these conditions act in a feedback-loop.

Funding

NIMH T32 MH-20030 supported Dr. Araujo, and R01-DA049867 supported Dr. Neale. The ABCD Study® is supported by the National Institutes of Health and additional federal partners under award numbers U01DA041048, U01DA050989, U01DA051016, U01DA041022, U01DA051018, U01DA051037, U01DA050987, U01DA041174, U01DA041106, U01DA041117, U01DA041028, U01DA041134, U01DA050988, U01DA051039, U01DA041156, U01DA041025, U01DA041120, U01DA051038, U01DA041148, U01DA041093, U01DA041089, U24DA041123, U24DA041147. A full list of supporters is available at https://abcdstudy.org/federal-partners.html.

Conflicts of interest

Authors report no conflicts of interest

Ethics approval

Not applicable

Consent for publication

Not applicable

Availability of data and material

Data used in the preparation of this article were obtained from the Adolescent Brain Cognitive Development (ABCD) Study (https://abcdstudy.org), held in the NIMH Data Archive (NDA). This is a multisite, longitudinal study designed to recruit more than 10,000 children age 9-10 and follow them over 10 years into early adulthood. A listing of participating sites and a complete listing of the study investigators can be found at https://abcdstudy.org/consortium_members/.

Code availability

Scripts are available in a repository for replication (github.com/lf-araujo/2023-mrdoc2-EA-ADHD). The scripts used for genetic ancestry assignment and genetic principal component analysis are available upon request

Author contributions

ABCD consortium investigators designed and implemented the study and/or provided data but, apart from Dr. Neale, did not participate in the analysis or writing of this report. This manuscript reflects the views of the authors and may not reflect the opinions or views of the NIH or ABCD consortium investigators. LA performed all analyses. MS, SK, RL, supported in the analyses. DZ calculated the PRS. All authors contributed in writing and reviewing the manuscript.

Abdellaoui A, Verweij KJH (2021) Dissecting polygenic signals from genome-wide association studies on human behaviour. Nat Hum Behav 5:686–694. https://doi.org/10.1038/s41562-021-01110-y
Achenbach TM (2013) DSM-Oriented Guide for the Achenbach System of Empirically Based Assessment (ASEBA®)
Achenbach TM (2001) Manual for ASEBA school-age forms & profiles. Univ. Vt. Res. Cent. Child. Youth Fam
Bates TC, Maes H, Neale MC (2019) umx: Twin and Path-Based Structural Equation Modeling in R. Twin Res. Hum Genet 22:27–41. https://doi.org/10.1017/thg.2019.2
Biederman J, Faraone SV, Mick E, Williamson S, Wilens TE, Spencer TJ, Weber W, Jetton J, Kraus I, Pert J, Zallen B (1999) Clinical correlates of ADHD in females: findings from a large group of girls ascertained from pediatric and psychiatric referral sources. J Am Acad Child Adolesc Psychiatry 38:966–975. https://doi.org/10.1097/00004583-199908000-00012
Boker SM, von Oertzen T, Pritikin JN, Hunter MD, Brick TR, Brandmaier AM, Neale MC (2023) Products of Variables in Structural Equation Models. Struct Equ Model Multidiscip J 30:708–718. https://doi.org/10.1080/10705511.2022.2141749
Boomsma DI, Saviouk V, Hottenga J-J, Distel MA, de Moor MHM, Vink JM, Geels LM, van Beek JHDA, Bartels M, de Geus EJC, Willemsen G (2010) Genetic Epidemiology of Attention Deficit Hyperactivity Disorder (ADHD Index) in Adults. PLoS ONE 5:e10621. https://doi.org/10.1371/journal.pone.0010621
Bowden J, Davey Smith G, Burgess S (2015) Mendelian randomization with invalid instruments: effect estimation and bias detection through Egger regression. Int J Epidemiol 44:512–525. https://doi.org/10.1093/ije/dyv080
Burgess S, Swanson SA, Labrecque JA (2021) Are Mendelian randomization investigations immune from bias due to reverse causation? Eur J Epidemiol 36:253–257. https://doi.org/10.1007/s10654-021-00726–8
Castro-de-Araujo L, Singh M, Zhou Y, Vinh P, Maes HHM, Verhulst B, Dolan C, Neale MC (2023) Power, measurement error, and pleiotropy robustness in twin-design extensions to Mendelian Randomization. https://doi.org/10.21203/rs.3.rs–3411642/v1
Castro-de-Araujo LFS, Singh M, Zhou Y, Vinh P, Verhulst B, Dolan CV, Neale MC (2023) Behav Genet 53:63–73. https://doi.org/10.1007/s10519-022-10122-x. MR-DoC2: Bidirectional Causal Modeling with Instrumental Variables and Data from Relatives
Chang L-Y, Wang M-Y, Tsai P-S (2016) Diagnostic Accuracy of Rating Scales for Attention-Deficit/Hyperactivity Disorder: A Meta-analysis. Pediatrics 137:e20152749. https://doi.org/10.1542/peds.2015–2749
Chen Q, Brikell I, Lichtenstein P, Serlachius E, Kuja-Halkola R, Sandin S, Larsson H (2017) Familial aggregation of attention-deficit/hyperactivity disorder. J Child Psychol Psychiatry 58:231–239. https://doi.org/10.1111/jcpp.12616
Chen W, Zhou K, Sham P, Franke B, Kuntsi J, Campbell D, Fleischman K, Knight J, Andreou P, Arnold R, Altink M, Boer F, Boholst MJ, Buschgens C, Butler L, Christiansen H, Fliers E, Howe-Forbes R, Gabriëls I, Heise A, Korn-Lubetzki I, Marco R, Medad S, Minderaa R, Müller UC, Mulligan A, Psychogiou L, Rommelse N, Sethna V, Uebel H, McGuffin P, Plomin R, Banaschewski T, Buitelaar J, Ebstein R, Eisenberg J, Gill M, Manor I, Miranda A, Mulas F, Oades RD, Roeyers H, Rothenberger A, Sergeant J, Sonuga-Barke E, Steinhausen H-C, Taylor E, Thompson M, Faraone SV, Asherson P (2008) DSM-IV combined type ADHD shows familial association with sibling trait scores: a sampling strategy for QTL linkage. Am J Med Genet Part B Neuropsychiatr Genet Off Publ Int Soc Psychiatr Genet 147B:1450–1460. https://doi.org/10.1002/ajmg.b.30672
Choi SW, Mak TS-H, O’Reilly PF (2020) Tutorial: a guide to performing polygenic risk score analyses. Nat Protoc 15:2759–2772. https://doi.org/10.1038/s41596-020-0353–1
Conomos MP, Reiner AP, Weir BS, Thornton TA (2016) Model-free Estimation of Recent Genetic Relatedness. Am J Hum Genet 98:127–148. https://doi.org/10.1016/j.ajhg.2015.11.022
Dardani C, Riglin L, Leppert B, Sanderson E, Rai D, Howe LD, Davey Smith G, Tilling K, Thapar A, Davies NM, Anderson E, Stergiakouli E (2021) Is genetic liability to ADHD and ASD causally linked to educational attainment? Int J Epidemiol 50:2011–2023. https://doi.org/10.1093/ije/dyab107
Demange PA, Boomsma DI, van Bergen E, Nivard MG (2023) Evaluating the causal relationship between educational attainment and mental health. medRxiv 2023.01.26.23285029 https://doi.org/10.1101/2023.01.26.23285029
Demange PA, Malanchini M, Mallard TT, Biroli P, Cox SR, Grotzinger AD, Tucker-Drob EM, Abdellaoui A, Arseneault L, van Bergen E, Boomsma DI, Caspi A, Corcoran DL, Domingue BW, Harris KM, Ip HF, Mitchell C, Moffitt TE, Poulton R, Prinz JA, Sugden K, Wertz J, Williams BS, de Zeeuw EL, Belsky DW, Harden KP, Nivard MG (2021) Investigating the genetic architecture of noncognitive skills using GWAS-by-subtraction. Nat Genet 53:35–44. https://doi.org/10.1038/s41588-020-00754–2
Demontis D, Walters GB, Athanasiadis G, Walters R, Therrien K, Nielsen TT, Farajzadeh L, Voloudakis G, Bendl J, Zeng B, Zhang W, Grove J, Als TD, Duan J, Satterstrom FK, Bybjerg-Grauholm J, Bækved-Hansen M, Gudmundsson OO, Magnusson SH, Baldursson G, Davidsdottir K, Haraldsdottir GS, Agerbo E, Hoffman GE, Dalsgaard S, Martin J, Ribasés M, Boomsma DI, Soler Artigas M, Roth Mota N, Howrigan D, Medland SE, Zayats T, Rajagopal VM, Nordentoft M, Mors O, Hougaard DM, Mortensen PB, Daly MJ, Faraone SV, Stefansson H, Roussos P, Franke B, Werge T, Neale BM, Stefansson K, Børglum AD (2023) Genome-wide analyses of ADHD identify 27 risk loci, refine the genetic architecture and implicate several cognitive domains. Nat Genet 55:198–208. https://doi.org/10.1038/s41588-022-01285–8
Demontis D, Walters RK, Martin J, Mattheisen M, Als TD, Agerbo E, Baldursson G, Belliveau R, Bybjerg-Grauholm J, Bækvad-Hansen M, Cerrato F, Chambert K, Churchhouse C, Dumont A, Eriksson N, Gandal M, Goldstein JI, Grasby KL, Grove J, Gudmundsson OO, Hansen CS, Hauberg ME, Hollegaard MV, Howrigan DP, Huang H, Maller JB, Martin AR, Martin NG, Moran J, Pallesen J, Palmer DS, Pedersen CB, Pedersen MG, Poterba T, Poulsen JB, Ripke S, Robinson EB, Satterstrom FK, Stefansson H, Stevens C, Turley P, Walters GB, Won H, Wright MJ, Consortium (PGC), A.W.G. of the P.G., Consortium, E.L.& G.E. (EAGLE), Team, 23 and Me Research, Andreassen OA, Asherson P, Burton CL, Boomsma DI, Cormand B, Dalsgaard S, Franke B, Gelernter J, Geschwind D, Hakonarson H, Haavik J, Kranzler HR, Kuntsi J, Langley K, Lesch K-P, Middeldorp C, Reif A, Rohde LA, Roussos P, Schachar R, Sklar P, Sonuga-Barke EJS, Sullivan PF, Thapar A, Tung JY, Waldman ID, Medland SE, Stefansson K, Nordentoft M, Hougaard DM, Werge T, Mors O, Mortensen PB, Daly MJ, Faraone SV, Børglum AD, Neale BM (2019) Discovery of the first genome-wide significant risk loci for attention-deficit/hyperactivity disorder. Nat. Genet. 51, 63. https://doi.org/10.1038/s41588-018-0269–7
Dick AS, Lopez DA, Watts AL, Heeringa S, Reuter C, Bartsch H, Fan CC, Kennedy DN, Palmer C, Marshall A, Haist F, Hawes S, Nichols TE, Barch DM, Jernigan TL, Garavan H, Grant S, Pariyadath V, Hoffman E, Neale M, Stuart EA, Paulus MP, Sher KJ, Thompson WK (2021) Meaningful associations in the adolescent brain cognitive development study. NeuroImage 239:118262. https://doi.org/10.1016/j.neuroimage.2021.118262
Fan CC, Loughnan R, Wilson S, Hewitt JK, Agrawal A, Dowling G, Garavan H, LeBlanc K, Neale M, Friedman N, Madden P, Little R, Brown SA, Jernigan T, Thompson WK, ABCD Genetic Working Group (2023) Genotype Data and Derived Genetic Instruments of Adolescent Brain Cognitive Development Study® for Better Understanding of Human Brain Development. Behav Genet 53:159–168. https://doi.org/10.1007/s10519-023-10143–0
Faraone SV, Larsson H (2019) Genetics of attention deficit hyperactivity disorder. Mol Psychiatry 24:562–575. https://doi.org/10.1038/s41380-018-0070–0
Fayyad J, Sampson NA, Hwang I, Adamowski T, Aguilar-Gaxiola S, Al-Hamzawi A, Andrade LHSG, Borges G, de Girolamo G, Florescu S, Gureje O, Haro JM, Hu C, Karam EG, Lee S, Navarro-Mateu F, O’Neill S, Pennell B-E, Piazza M, Posada-Villa J, Have M ten, Torres Y, Xavier M, Zaslavsky AM, Kessler RC (2017) The descriptive epidemiology of DSM-IV Adult ADHD in the World Health Organization World Mental Health Surveys. Atten. Deficit Hyperact. Disord. 9, 47–65. https://doi.org/10.1007/s12402-016-0208–3
Ge T, Chen C-Y, Ni Y, Feng Y-CA, Smoller JW (2019) Polygenic prediction via Bayesian regression and continuous shrinkage priors. Nat Commun 10:1776. https://doi.org/10.1038/s41467-019-09718–5
Haist F, Jernigan TL (2023) Adolescent Brain Cognitive Development Study (ABCD) - Annual Release 5.0. https://doi.org/10.15154/8873-ZJ65
Heath AC, Kessler RC, Neale MC, Hewitt JK, Eaves LJ, Kendler KS (1993) Testing hypotheses about direction of causation using cross-sectional family data. Behav Genet 23:29–50. https://doi.org/10.1007/BF01067552
Hinshaw SP (2018) Attention Deficit Hyperactivity Disorder (ADHD): Controversy, Developmental Mechanisms, and Multiple Levels of Analysis. Annu Rev Clin Psychol 14:291–316. https://doi.org/10.1146/annurev-clinpsy–050817–084917
Howe LJ, Tudball M, Davey Smith G, Davies NM (2022) Interpreting Mendelian-randomization estimates of the effects of categorical exposures such as disease status and educational attainment. Int J Epidemiol 51:948–957. https://doi.org/10.1093/ije/dyab208
Hunter MD, Garrison SM, Burt SA, Rodgers JL (2021) The Analytic Identification of Variance Component Models Common to Behavior Genetics. Behav Genet 51:425–437. https://doi.org/10.1007/s10519-021-10055-x
Jami ES, Hammerschlag AR, Bartels M, Middeldorp CM (2021) Parental characteristics and offspring mental health and related outcomes: a systematic review of genetically informative literature. Transl Psychiatry 11:1–38. https://doi.org/10.1038/s41398-021-01300–2
Jiang Z, Wang D, Xu H, Zhang A, Zhao Q, Yan J, Wang X, Zhang W, Li Y, Yang K, Hu S, Cui Y, Li Y (2023) Diagnostic efficiency and psychometric properties of CBCL DSM-oriented scales in a large sample of Chinese school-attending students aged 5–16. Asian J Psychiatry 88:103724. https://doi.org/10.1016/j.ajp.2023.103724
Jordan DM, Verbanck M, Do R (2019) HOPS: a quantitative score reveals pervasive horizontal pleiotropy in human genetic variation is driven by extreme polygenicity of human traits and diseases. Genome Biol 20:222. https://doi.org/10.1186/s13059-019-1844–7
Kim JH, Kim JY, Lee J, Jeong GH, Lee E, Lee S, Lee KH, Kronbichler A, Stubbs B, Solmi M, Koyanagi A, Hong SH, Dragioti E, Jacob L, Brunoni AR, Carvalho AF, Radua J, Thompson T, Smith L, Oh H, Yang L, Grabovac I, Schuch F, Fornaro M, Stickley A, Rais TB, de Pablo S, Shin G, Fusar-Poli JI, P (2020) Environmental risk factors, protective factors, and peripheral biomarkers for ADHD: an umbrella review. Lancet Psychiatry 7:955–970. https://doi.org/10.1016/S2215–0366(20)30312–6
Larsson H, Chang Z, D’Onofrio BM, Lichtenstein P (2014) The heritability of clinically diagnosed attention deficit hyperactivity disorder across the lifespan. Psychol Med 44:2223–2229. https://doi.org/10.1017/S0033291713002493
Lewis CM, Vassos E (2020) Polygenic risk scores: from research tools to clinical instruments. Genome Med 12:44. https://doi.org/10.1186/s13073-020-00742–5
Loughnan RJ, Palmer CE, Makowski C, Thompson WK, Barch DM, Jernigan TL, Dale AM, Fan CC (2022) Unique prediction of developmental psychopathology from genetic and familial risk. J Child Psychol Psychiatry 63:1631–1643. https://doi.org/10.1111/jcpp.13649
Maes HHM, Lapato DM, Schmitt JE, Luciana M, Banich MT, Bjork JM, Hewitt JK, Madden PA, Heath AC, Barch DM, Thompson WK, Iacono WG, Neale MC (2022) Genetic and Environmental Variation in Continuous Phenotypes in the ABCD Study®. Behav Genet. https://doi.org/10.1007/s10519-022-10123-w
Maydeu-Olivares A, Shi D, Rosseel Y (2019) Instrumental Variables Two-Stage Least Squares (2SLS) vs. Maximum Likelihood Structural Equation Modeling of Causal Effects in Linear Regression Models. Struct Equ Model Multidiscip J 26:876–892. https://doi.org/10.1080/10705511.2019.1607740
Mehta PD, Neale MC, Flay BR (2004) Squeezing interval change from ordinal panel data: latent growth curves with ordinal outcomes. Psychol Methods 9:301–333. https://doi.org/10.1037/1082–989X.9.3.301
Michaëlsson M, Yuan S, Melhus H, Baron JA, Byberg L, Larsson SC, Michaëlsson K (2022) The impact and causal directions for the associations between diagnosis of ADHD, socioeconomic status, and intelligence by use of a bi-directional two-sample Mendelian randomization design. BMC Med 20:106. https://doi.org/10.1186/s12916-022-02314–3
Minică CC, Dolan CV, Boomsma DI, de Geus E, Neale MC (2018) Extending Causality Tests with Genetic Instruments: An Integration of Mendelian Randomization with the Classical Twin Design. Behav Genet 48:337–349. https://doi.org/10.1007/s10519-018-9904–4
Neale MC, Hunter MD, Pritikin JN, Zahery M, Brick TR, Kirkpatrick RM, Estabrook R, Bates TC, Maes HH, Boker SM (2016) OpenMx 2.0: Extended Structural Equation and Statistical Modeling. Psychometrika 81:535–549. https://doi.org/10.1007/s11336-014-9435–8
Peterson RE, Edwards AC, Bacanu S-A, Dick DM, Kendler KS, Webb BT (2017) The utility of empirically assigning ancestry groups in cross-population genetic studies of addiction. Am J Addict 26:494–501. https://doi.org/10.1111/ajad.12586
Purcell S, Neale B, Todd-Brown K, Thomas L, Ferreira MAR, Bender D, Maller J, Sklar P, de Bakker PIW, Daly MJ, Sham PC (2007) PLINK: A Tool Set for Whole-Genome Association and Population-Based Linkage Analyses. Am J Hum Genet 81:559–575. https://doi.org/10.1086/519795
Salari N, Ghasemi H, Abdoli N, Rahmani A, Shiri MH, Hashemian AH, Akbari H, Mohammadi M (2023) The global prevalence of ADHD in children and adolescents: a systematic review and meta-analysis. Ital J Pediatr 49:48. https://doi.org/10.1186/s13052-023-01456–1
Sousa A, de Isabel F, Castanho C, Polanczyk J, Rohde G, L (2020) Attention-defficit hyperactivity disorders. In: Rey J (ed) E-Textbook of Child and Adolescent Mental Health. International Association for Child and Adolescent Psychiatry and Allied Professions, Geneva
Sprich S, Biederman J, Crawford MH, Mundy E, Faraone SV (2000) Adoptive and biological families of children and adolescents with ADHD. J Am Acad Child Adolesc Psychiatry 39:1432–1437. https://doi.org/10.1097/00004583-200011000-00018
Taylor MJ, Lichtenstein P, Larsson H, Anckarsäter H, Greven CU, Ronald A (2016) J Am Acad Child Adolesc Psychiatry 55:504–512e2. https://doi.org/10.1016/j.jaac.2016.04.004. Is There a Female Protective Effect Against Attention-Deficit/Hyperactivity Disorder? Evidence From Two Representative Twin Samples
Thomas R, Sanders S, Doust J, Beller E, Glasziou P (2015) Prevalence of attention-deficit/hyperactivity disorder: a systematic review and meta-analysis. Pediatrics 135:e994–1001. https://doi.org/10.1542/peds.2014–3482
Timpson NJ, Nordestgaard BG, Harbord RM, Zacho J, Frayling TM, Tybjærg-Hansen A, Smith GD (2011) C-reactive protein levels and body mass index: Elucidating direction of causation through reciprocal Mendelian randomization. Int. J. Obes. 2005 35, 300–308. https://doi.org/10.1038/ijo.2010.137
Verbanck M, Chen C-Y, Neale B, Do R (2018) Detection of widespread horizontal pleiotropy in causal relationships inferred from Mendelian randomization between complex traits and diseases. Nat Genet 50:693–698. https://doi.org/10.1038/s41588-018-0099–7

Table 1

Main demographic characteristics of the ABCD sample, stratified by sex. The CBCL ADHD score and the students' grades in the previous year were treated as count variables with 8 levels. Abbreviations: n, number of observations; SD, standard deviation; NA, missing; F, female; M, Male; p, p-value associated with the groupwise comparisons. Chi-squared tests are performed in the categorical cases and t-tests for continuous cases. Not adjusted for multiple comparisons.

	Overall	F	M	p
n	2642	1305	1337
Zygosity (%)				0.674
DZ	764 (28.9)	387 (29.7)	377 ( 28.2)
MZ	558 (21.1)	276 (21.1)	282 ( 21.1)
Sibs	1320 (50.0)	642 (49.2)	678 ( 50.7)
Age (mean (SD))	10.01 (0.66)	9.98 (0.66)	10.04 (0.65)	0.014
Ancestry (%)				0.084
AFR	431 (16.3)	231 (17.7)	200 ( 15.0)
AMR	551 (20.9)	256 (19.6)	295 ( 22.1)
EUR	1660 (62.8)	818 (62.7)	842 ( 63.0)
Grades last yr (%)				< 0.001
1	62 ( 2.3)	20 ( 1.5)	42 ( 3.1)
4	63 ( 2.4)	17 ( 1.3)	46 ( 3.4)
5	84 ( 3.2)	27 ( 2.1)	57 ( 4.3)
6	160 ( 6.1)	56 ( 4.3)	104 ( 7.8)
7	146 ( 5.5)	70 ( 5.4)	76 ( 5.7)
8	652 (24.7)	294 (22.5)	358 ( 26.8)
10	367 (13.9)	187 (14.3)	180 ( 13.5)
11	705 (26.7)	399 (30.6)	306 ( 22.9)
12	340 (12.9)	207 (15.9)	133 ( 9.9)
NA	63 ( 2.4)	28 ( 2.1)	35 ( 2.6)
CBCL ADHD scores (%)				< 0.001
0	894 (33.8)	516 (39.5)	378 ( 28.3)
1	344 (13.0)	185 (14.2)	159 ( 11.9)
2	233 ( 8.8)	107 ( 8.2)	126 ( 9.4)
3	296 (11.2)	126 ( 9.7)	170 ( 12.7)
5	91 ( 3.4)	41 ( 3.1)	50 ( 3.7)
6	78 ( 3.0)	26 ( 2.0)	52 ( 3.9)
7	61 ( 2.3)	24 ( 1.8)	37 ( 2.8)
8	55 ( 2.1)	13 ( 1.0)	42 ( 3.1)
11	37 ( 1.4)	9 ( 0.7)	28 ( 2.1)
NA	553 (20.9)	258 (19.8)	295 ( 22.1)
ADHD medication use (mean (SD))	0.11 (0.31)	0.07 (0.25)	0.15 (0.36)	< 0.001

Table 2

ABCD Study® sample means, standard deviations (SD), and pair correlations for MZ and DZ twin pairs at wave 3 including all three ancestries. Mean age 9.9 (SD = 0.62).

Variable	Mean	SD	rMZ (284)	rDZ (394)
Grades last year	6.51	1.94	0.79 (0.02)	0.48 (0.04)
CBCL ADHD	2.98	2.17	0.55 (0.05)	0.25 (0.05)
ADHD meds lifetime	0.07	0.26	0.68 (0.03)	0.22 (0.05)
EA PRS	0	1	1 (0)	0.5 (0.05)
ADHD PRS	0	1	0.99 (0)	0.48 (0.05)

Table 3

Parameter estimates for the MR-DoC2 model. Fit: -2lnL 16256.539, AIC 5094.539, and CFI 0.896. Observations: 782. Parameter definitions can be found in Fig. 1. Extra parameters tx2, ty2, and covT are the sibling variances for ADHD (x), education performance (y), and their covariance respectively. Parameters ending in dev# represent standard normal deviations from one threshold to the next under the threshold liability model. Parameters starting with b represent the beta regression estimates for each of the definition variables (sex, ADHD medication lifetime use and PCs 1:6).

Parameter	Estimate	Std.Error
g1	-0.856	0.060
g2	-0.500	0.043
b1	0.059	0.016
b3	0.065	0.017
ax2	0.463	0.041
covA	0.419	0.048
ay2	0.610	0.046
rf	-0.077	0.027
cx2	-0.059	0.034
covC	-0.020	0.029
cy2	0.001	0.034
ex2	0.319	0.034
covE	0.257	0.025
ey2	0.255	0.029
σ3	0.993	0.009
σ4	0.982	0.009
mean_Ph1	0.111	0.036
mean_Ph2	0.090	0.068
mean_prs1	0.016	0.031
mean_prs2	-0.035	0.030
tx2	0.045	0.031
covT	0.096	0.024
ty2	0.211	0.030
cbcl_adhd_dev3	0.326	0.026
cbcl_adhd_dev4	0.481	0.036
cbcl_adhd_dev5	0.248	0.031
cbcl_adhd_dev6	0.253	0.036
cbcl_adhd_dev7	0.289	0.046
cbcl_adhd_dev8	0.413	0.075
sag_grade_type_dev3	0.299	0.045
sag_grade_type_dev4	0.302	0.038
sag_grade_type_dev5	0.271	0.032
sag_grade_type_dev6	0.858	0.040
sag_grade_type_dev7	0.411	0.026
sag_grade_type_dev8	0.990	0.043
bX_sex	-0.185	0.031
bY_sex	0.112	0.037
bX_meds_lifetime	1.135	0.102
bY_meds_lifetime	-0.574	0.084
bX_C1	-19.331	2.669
bY_C1	47.305	5.584
bX_C2	-33.602	4.741
bY_C2	38.569	5.371
bX_C3	-0.173	0.150
bY_C3	-23.780	2.670
bX_C4	-0.055	0.149
bY_C4	-19.392	2.461
bX_C5	8.371	1.243
bY_C5	1.246	0.186
bX_C6	6.727	0.988
bY_C6	-0.531	5.112

Table 4

Assessing bidirectional effects. The MR-DoC2 model was fitted to the ABCD data on ADHD and Educational attainment. To evaluate whether a unidirectional or bidirectional model is preferred, we performed a series of likelihood ratio tests (LRT), fixing either of the causal paths (No g1, and No g2 models) and both paths (no loop model). Since these models are nested in relation to MR-DoC2 it is possible to compare their fit. Models did not differ significantly, however the No loop model presented the smallest AIC. This indicates that a bidirectional causal relationship cannot be rejected. EP, estimated parameters; Δ Fit, difference in -2lnL; df, degrees of freedom; p-value associated with the LRT; AIC, Akaike Information Criterion.

Model	EP	Δ Fit	Δ df	p	AIC	Compare with Model	Fit units
MR-DoC2	51				16359		-2lnL
No g1	50	8.282	1	0.004	16365	MR-DoC2	-2lnL
No g2	50	3.255	1	0.071	16360	MR-DoC2	-2lnL
No loop	49	11.941	2	0.003	16367	MR-DoC2	-2lnL

No competing interests reported.

Quasi-experimental analyses of the effect of ADHD on education performance in youths across sexes and ancestry

Status:

Version 1

Abstract

Figures

1. Introduction

2. Methods

2.1. Data source

2.2 Instruments and polygenic score calculation

2.3 Covariates

2.4. Data analysis

2.5. Model Specification

3. Results

4. Discussion

Declarations