The ABCD Study® is a longitudinal study of brain development and child health. The study design and recruitment strategy have previously been described [34], but in brief, the study used school-based recruitment to enrol 11,875 children from 21 metropolitan sites across the USA. Children were aged between 9 and 10 years at time of enrolment, and they and their caregiver completed the baseline visit between 1 October 2016 and 31 October 2018, which consisted of questionnaires, clinical interviews, neurocognitive interviews, and a neuroimaging protocol. Exclusionary diagnoses include a current diagnosis of schizophrenia, a moderate/severe autism diagnosis, intellectual disability, or alcohol/substance use disorder.

Our study received approval from the institutional review board of the University of Southern California. The ABCD Study obtained centralised institutional review board approval from the University of California, San Diego, and each of the 21 study sites obtained local institutional review board approval. Ethical regulations were followed during data collection and analysis. Parents or caregivers provided written informed consent, and children gave written assent. Data can be accessed through registration with the ABCD study at https://nda.nih.gov/abcd. The present analyses used data from the baseline (demographic information, co-occurring psychopathology) and follow-up phase one visits (SRS). A total of 11,878 children were recruited at baseline, and of these, 11,736 participated in sMRI scanning. As the ABCD cohort contains data from siblings, measures from a random sample of 7875 unrelated individuals were used, of which 345 were excluded due to poor quality sMRI data. Of the remaining 7521 participants, 7005 had available data on autistic traits, and thus made up the present sample. During the screening process, caregivers were asked if their child had previously received a diagnosis of a mental health condition. In the present sample, a total of 107 (1.53%) children were reported to have an autism diagnosis, and 1053 (15.03%) were reported to have a diagnosis of another mental condition including ADHD, depression, bipolar disorder, anxiety, or a specific phobia.

All neuroimaging data were collected, processed, and quality checked by the ABCD Data Analysis, Informatics & Resource Center (DAIRC). Structural MRI scans were acquired at twenty-one sites across the USA using twenty-six different scanners from two vendors (Siemens and General Electric). Data were acquired when children were 9-to-10 years of age. Methods were optimised and harmonised across ABCD study sites for 3-T scanners, of which the full details have been published previously [35]. To summarise, after completion of a pre-scan assessment, a simulation session in a mock scanner, and motion compliance training, children participated in the ABCD neuroimaging protocol. T1-weighted structural scans with 1-mm isotropic resolution were collected using adult size multi-channel coils with image acquisition protocols for 3-Tesla Siemens, Phillips, and General Electric scanners, harmonised across all testing sites. Quality control procedures were based on automated means and SDs of extracted brain measures, and trained raters checked images for poor quality (i.e. motion artefacts, blurring, or ringing). Structural MRI data were processed by the ABCD DAIRC (Data Analysis, Informatics & Resource Center) team using FreeSurfer v5.3 (http://surfer.nmr.mgh.harvard.edu/) and subjected to quality control procedures [36].

FreeSurfer extracts cortical and subcortical region of interests (ROIs) based on the Desikan–Killiany atlas [37]. The automated pipeline consists of co-registration based on a template reference surface, motion correction, and averaging. Any intensity variation across the image due to magnetic field heterogeneity is corrected, and the skull stripped from the normalised intensity image. Images are then segmented using a connected components algorithm, where connectivity is not permitted across established cutting planes. Any holes within white matter are filled, producing a single volume for each hemisphere. Volumetric segmentation are used to delineate and label global (total brain volume, subcortical volume) and regional [nucleus accumbens (NAcc), amygdala, caudate nucleus, hippocampus, pallidum, putamen, and thalamus] measures [34]. A global subcortical volume was also derived, consisting of the total volume of the thalamus, caudate, putamen, pallidum, hippocampus, amygdala, NAcc, ventral DC, and substantia nigra). Subcortical volumetric measures were derived by averaging the homotopic regional volumes.

The resulting output was then visually examined by a trained DAIC technician, who rated them from zero to three in five categories: motion, intensity homogeneity, white matter underestimation, pial overestimation, and magnetic susceptibility artefact. From this, an overall “pass” or “fail” score was generated. Participants whose images failed QC were excluded from the present analyses [34].

Post-processed FreeSurfer derived phenotypes from the ABCD cohort have been widely used in studies assessing predictors of interest with brain morphology outcomes, including those examining subcortical ROIs [35, 36, 38, 39].

Autistic traits were assessed using the SRS, which is primarily used to assess the severity of social difficulties across the full range of severity in both autistic and non-autistic children [24]. Statistical properties of the SRS have previously been evaluated in a UK population-based sample of 5-to-8-year-old children [38]. In the ABCD sample, parents answered an 11-item abridged version of the questionnaire which has previously shown strong loadings on the first unrotated factor of a principal components analysis of the SRS in a paediatric sample [39]. Parents were asked to rate statements on a four-point Likert scale; 0 (not true); 1 (sometimes true); 2 (often true); and 3 (almost always true). It encompasses the three DSM-IV autism domains, with items relating to reciprocal social behaviour (e.g. “Has difficulty making friends, even when trying his or her best”.), stereotyped and repetitive behaviours (e.g. “Has more difficulty than other children with changes in his or her routine”.), and communication impairments (e.g. “Has trouble keeping up with the flow of normal conversation”.). Total raw summary scores from participants were calculated (mean 3.49 SD = 0.49, range = 0–39).

Potential confounders of the exposure-outcome relationship were defined a priori based on the previous literature. A minimum set of confounders required to adequately account for confounding were defined as: age, sex, ethnicity, cognition score, and a measure of socioeconomic status (family-level income). To assess the impact of co-occurring psychopathology, separate models were conducted with the inclusion of T-scores of externalising and internalising symptoms extracted from the Child Behaviour Checklist (CBCL) [40].

Demographic information (child sex, age at time of MRI, ethnicity, and total household income) was extracted from a demographics survey answered by the child’s main caregiver. Child cognitive ability was assessed using the NIH Toolbox® cognition measures (http://www.nihtoolbox.org) [41]. The toolbox consists of seven tasks that cover episodic memory, executive function, attention, working memory, processing speed, and language abilities and is used to generate a total cognitive score composite. The composite score demonstrates good test re-test reliability and validity in children [42].

The CBCL parent report was used to measure internalising and externalising symptoms in participants [43]. This is a well-established parent-completed measure of emotional, behavioural, and social problems in children and adolescents [44]. Composite scores of internalising and externalising problems were used for these analyses. Raw scores were converted to standardised t-scores, scaled so that fifty was average for child age and sex, with a standard deviation (SD) of ten points. Higher scores indicate increased behavioural and emotional problems.

Differences in covariates across tertials of SRS scores were examined using χ2-testing for categorical variables, and univariate regression modelling for continuous variables. The association of covariates with total subcortical volume was assessed using univariate regression modelling.

The association between SRS and brain morphology outcomes (volumes of the thalamus, caudate, putamen, pallidum, amygdala, hippocampus, and NAcc) was modelled using seemingly unrelated regression (SUR). The SUR system allows for a single model containing a number of linear equations, permitting correlation among the error terms. As SUR is used to analyse correlated outcomes, correlations between measures were assessed as a preliminary step in the analysis.

Models were conducted in three steps to assess the impact of confounding variables. Model 1: adjustment for child age, sex, ethnicity, family income, and ABCD recruitment site. Model 2: model 1 with the addition of cognition score (Aim 1). Model 3: model 2 with the addition of externalising symptoms and internalising symptoms (Aim 2). Raw p-values were adjusted for multiple testing by using Holm correction. All models were conducted with the inclusion of ICV to explore whether differences in subcortical volume were explained by differences in global brain size. Analyses were conducted in Stata v16.0 [45], with the -sureg command utilised to conduct SUR. Correlation between individual ROIs was assessed using the -pwcorr command.

A number of sensitivity analyses were conducted. Firstly, to explore whether associations were lateralised, analyses were replicated using homotopic ROIs. Secondly, to examine whether sex was a moderator of any observed associations, analyses were conducted with the addition of an interaction term. Finally, to explore the specific impact of dimensions of co-occurring mental health conditions, additional analyses were conducted controlling for internalising, externalising, and attention problems separately.