We analyzed data from 10-year-old children from the ABCD (N = 7722, multisite) and Generation R (N = 2531, single-site) Studies (Supplementary file 1b). ADHD symptoms were measured with the Child Behavioral Checklist (CBCL). T₁-weighted images were obtained with 3T scanners (Casey et al., 2018; Kooijman et al., 2016). We ran vertex-wise linear regression models for ADHD with cortical surface area, volume, and thickness. Results for surface area, which constituted the main findings here, are presented in-text, while findings for volume and thickness in the figure supplements. We adjusted for demographic and study characteristics which have been generally considered by previous literature (Supplementary file 1a): age, sex, ethnicity, and study site (ABCD only). We refer to this model as model 1, as further adjustments for confounders are outlined in subsequent steps.

We found that higher ADHD symptoms were associated with less bilateral surface area in both cohorts. As shown in Figure 1, associations were widespread, as the clusters of association covered 1165.7 cm² of the cerebral cortex in the ABCD Study, and 446.1 cm² in the Generation R Study. Across both cohorts, we consistently identified clusters for surface area in the lateral occipital, postcentral, rostral middle and superior frontal, and superior parietal cortices. For cortical thickness, we found two small frontal clusters in the ABCD Study (16.1 cm²) and no clusters in the Generation R Study, which suggests that cortical thickness does not relate or does not relate strongly to ADHD, in line with prior literature (Hoogman et al., 2019; Figure 1—figure supplement 1).

Next, we considered factors that have been previously linked to ADHD and brain structure in the literature, and are thus potential confounders. To illustrate this background knowledge and the assumptions about relations between variables, we used Directed Acyclic Graphs (DAGs), a type of causal diagram (Hernan and Robins, 2020). These guide the identification (and dismissal) of covariates that may act as confounders. Of note, while assumptions may not hold, this theoretical approach is preferred to methods selecting confounders based on model statistics (Lee, 2014). The DAGs are depicted in Figure 2 and Figure 2—figure supplements 1 and 2, and the rationales for variable inclusion are explained below and in the Methods.

Based on the literature, lower SES is associated with a higher risk for ADHD (Russell et al., 2016) and with variation in cortical brain structure (Noble et al., 2015). Thus, confounding by socioeconomic factors in the relation between ADHD and brain morphology is likely. We therefore additionally adjusted for a second set of confounders (model 2) related to SES: household income, maternal education, and maternal age at childbirth.

Moreover, several factors concerning maternal behavior, pre- and postnatally, have been associated with both ADHD and brain morphology. For instance, prenatal exposure to substances is known to increase the risk of developing ADHD symptoms and has been associated with variation in cerebral volume and surface area (Eilertsen et al., 2017; Lees et al., 2020). Postnatal maternal psychopathology has been linked to higher child ADHD symptoms (Clavarino et al., 2010) and smaller brain volume in children (Zou et al., 2019). Thus, in model 3 we additionally adjusted for prenatal exposure to substance use (tobacco and cannabis), and postnatal maternal psychopathology.

Adjustments for SES (model 2) led to reductions in the spatial extent of the clusters for surface area and volume in both cohorts (Figure 3). For surface area, cluster sizes for ADHD symptoms reduced from 1165.7 cm² in model 1 to 952.8 cm² in model 2 (=−18%) in the ABCD Study, and from 446.1 to 229.6 cm² (=−49%) in the Generation R Study. Similar reductions were observed for volume and thickness (Figure 3—figure supplement 1). After adjusting for the confounders added in model 3, across both cohorts, we consistently identified clusters for surface area in the cuneus, precuneus, fusiform, inferior parietal, isthmus of the cingulate, pericalcarine, pre- and postcentral, rostral middle and superior frontal, superior temporal and supramarginal cortices.

To explore whether the results observed for associations between brain morphology and ADHD symptoms applied to children with an ADHD diagnosis, we repeated the primary analysis using the ADHD diagnostic data from the Kiddie Schedule for Affective Disorders and Schizophrenia (KSADS) in the ABCD Study. In line with our primary results, ADHD diagnosis was associated with less bilateral surface area and volume. Compared to clusters for ADHD symptoms, those associated with ADHD diagnosis were smaller, but overlapping (Figure 1—figure supplement 2). We observed similar patterns of reduction in the spatial extent of the clusters after adjusting for each set of confounders (Figure 3—figure supplement 2). For surface area, cluster sizes for ADHD symptoms covered 234.4 cm² in model 1 and reduced to 199.5 cm² in model 2 (=−15%), and 55.5 cm² in model 3 (=−72%, compared to model 2).

Surface-based studies generally focus on the spatial extent of cortical clusters associated with the phenotype, but, in this study, we also explored how confounding adjustments affected the regression coefficients for ADHD symptoms (Figure 4).

At a vertex-wise level, adjusting for socioeconomic and maternal factors (model 3) led to reductions in the beta coefficients, across the brain, for both cohorts (Figure 4—figure supplement 1). Of note, some beta coefficients also showed increases. As confounding bias may lead to under- or overestimation, it is not surprising to observe both decreases and increases in the average beta coefficients after adjustments.

At an anatomical region level, where estimates of vertices within a given Desikan–Killiany region were averaged, beta coefficients for surface area tended to decrease from model 1 to 2 by approximately 15% (Figure 4, Figure 4—figure supplement 2). Further adjustments from model 2 to 3 led to decreases in the average beta coefficients of certain regions and increases in others. Similar patterns were found for volume (Figure 4—figure supplements 3 and 4). The average beta coefficients per region correlated moderately to strongly between the ABCD and Generation R Studies for surface area (Spearman r_M1 = 0.84, r_M2 = 0.83, r_M3 = 0.83) and volume (Spearman r_M1 = 0.57, r_M2 = 0.57, r_M3 = 0.70) (Figure 4—figure supplement 5).

We considered one additional scenario which included IQ, a factor that is often adjusted for in previous studies (Supplementary file 1a). However, based on prior literature, it holds an ambiguous role in structural anatomy–ADHD relations. Previous studies found that children with ADHD scored lower on IQ than children without ADHD (Bridgett and Walker, 2006). Differential brain structure with levels of IQ has also been shown (Mcdaniel, 2005). However, the directions of causation between these variables remain unclear (Gallo and Posner, 2016). IQ may therefore be a confounder, collider, and/or mediator in the relation between brain structure and ADHD, as depicted in the DAGs in Figure 5 and Figure 5—figure supplement 1.

First, it could be argued that IQ is partly innate and precedes brain development and ADHD, making it a confounder (Figure 5A). Second, IQ may lie in the pathway between brain structure and ADHD and therefore act as a mediator (Figure 5B). It is conceivable that cognitive differences, as a consequence of subtle neurodevelopmental differences (Lee et al., 2019), could underlie ADHD. Adjusting for a mediator would lead to bias when estimating the total association between brain structure and ADHD (VanderWeele, 2016). Third, brain structure may impact intelligence scores (Lee et al., 2019), and ADHD symptoms may affect IQ test performance (Jepsen et al., 2009; Figure 5C). A variable that is independently caused by the outcome and the determinant is also known as a collider, and adjusting for it leads to (collider) bias. Here, we explored the impact of adjusting for IQ when examining the relation between brain morphology and ADHD (model 4).

After additionally adjusting for IQ, the spatial extent of the clusters associated with ADHD symptoms reduced further in both cohorts (Figure 6). For surface area, compared to model 3, clusters reduced from 760.2 to 605.1 cm² (=−20%) for the ABCD Study, and from 208.6 to 93.1 cm² for the Generation R Study (=−55%). Clusters of association for surface area in model 4 were located in the fusiform, inferior parietal, insula, lateral occipital, middle temporal, pericalcarine, pre- and postcentral, precuneus, rostral middle, and superior frontal, superior parietal and temporal, and supramarginal cortices. Findings for volume and thickness are shown in Figure 6—figure supplement 1.

A final scenario was also included, to reflect the commonly used adjustments for head motion during scanning (Supplementary file 1a). Motion can be a large source of bias in neuroimaging studies which is important to address. While it does not meet the criteria for confounding as it is not a common cause of ADHD problems and brain morphology (Hernan and Robins, 2020), head motion can induce measurement error of brain morphology (Van de Walle et al., 1997; Figure 7). This is also referred to as information bias and can distort estimates from their true value.

The amount of measurement error in brain morphology may differ across children with versus without ADHD. In fact, children with impulsivity and inattention have been shown to move more during MRI scanning (Thomson et al., 2021; Kong et al., 2014), determining different levels of error in the brain morphology assessments (Figure 7, path from ADHD symptoms to motion to error in MRI measurement). In this scenario, adjusting for motion might lead to two situations. On one hand, since motion is a consequence of the outcome (ADHD), adjustments would lead to bias (Westreich, 2012). On the other hand, not adjusting for motion would also lead to bias because part of the observed relation between ADHD symptoms and brain structure would be due to the higher head motion (and thus the underestimation of the cortical values) of children with ADHD. In this study, we explored the effect of adjusting for motion during scanning in the relation between brain morphology and ADHD (model 5).

After additional adjustments for head motion, the spatial extent of the clusters generally increased. For surface area, compared to model 3, clusters increased from 760.2 to 936.4 cm² (=+23.2%) for the ABCD Study and from 208.6 to 239.7 cm² (=+14.9%) for the Generation R Study (Figure 8). Clusters of associations consistently found across cohorts were highly similar to the ones identified in model 3. Results for cortical volume and thickness are shown in Figure 8—figure supplement 1.

Widespread associations between surface area and ADHD symptoms were initially identified, with higher symptoms relating to smaller brain structures, in line with previous research (Hoogman et al., 2019; Gehricke et al., 2017). After adjustments for potential confounders typically overlooked by prior literature (socioeconomic and maternal behavioral factors), approximately half of the associations remained, and considerable effect size changes were observed in both the ABCD and Generation R Studies and for all cortical measures. We observed similar patterns of cluster reductions for ADHD diagnosis in the ABCD Study.

Regions that remained associated after adjustments and which were consistently identified across cohorts were the precuneus, isthmus of the cingulate, supramarginal, pre- and postcentral, and inferior parietal cortices for both area and volume. Most of these regions (e.g., supramarginal) have been previously implicated in ADHD in clinical samples (Saad et al., 2017; Lei et al., 2014; Solanto et al., 2009). However, many different brain areas have been detected in association with the disorder (Saad et al., 2020), which may have hampered prior meta-analytic efforts to identify consistent neuroanatomical correlates for ADHD.

Of note, some inconsistencies between the ABCD and Generation R Studies, both in size of the clusters and the exact location, were observed. While we used the same processing pipelines and similar quality control procedures and measures across cohorts, potential reasons for discrepancies in results must be discussed. First, the larger sample size of the ABCD Study allows for greater power to detect smaller effects, which led to larger associated areas. Second, the multisite structure of the ABCD Study may have introduced noise in the results (e.g., by different scanners, demographic differences), and determined the identification of associations which are not replicable in the Generation R Study. Third, the two studies include children from different populations. While both are very diverse samples, the ABCD Study is comprised of a more heterogeneous sample from the US population, which, for instance, is characterized by a wider variety of ethnicities and cultures, potentially permitting the discovery of more associations. Nevertheless, there was considerable overlap in the findings from the ABCD and Generation R Studies, with consistencies across cohorts indicating the most robust and generalizable associations.

Here, we discerned associated areas likely subject to confounding bias from areas robust to socioeconomic and maternal behavioral factors, and replicable across two large cohorts. Comparisons with prior findings should be made with caution due to differences in study design, samples (clinical vs. population-based), and analytical methods. Importantly, we highlighted the opportunity for future studies to include covariates that go beyond age and sex, can help refine associations, and can be readily collected. Future studies may want to consider other confounding factors, depending on their research question, design, and assumed causal relations.

Avoiding bias from adjusting for variables that are not confounders is as important as identifying sources of confounding. Adjusting for mediators or colliders of the ADHD–brain structure relation would induce bias. Here, when adjusting for IQ, which plays an unclear role in brain structure–ADHD associations, cluster sizes reduced considerably in both the ABCD and Generation R Studies. This could indicate that IQ is a confounder, in which case adjustments would be necessary, or that IQ is a mediator or collider, in which case adjustments must be avoided.

First, based on previous literature and this study, the association between ADHD and IQ is relatively weak (Dennis et al., 2009) (r_ABCD = −0.11, r_GENR = −0.14), but this does not necessarily make it a weak confounder as the strength of confounding is due to a variable’s relation with the exposure and outcome. Second, if brain structure and ADHD symptoms both cause cognitive changes, adjusting for IQ could induce collider bias, although this is also dependent on when IQ is measured relative to the exposure and outcome (Hernan and Robins, 2020). Third, if brain structure determines cognitive functioning, which in turn affects ADHD symptoms (mediation by IQ), adjustments would also induce bias (VanderWeele, 2016).

Given these scenarios, we recommend moving away from routinely adjusting for IQ in ADHD neuroimaging studies, and we highlight the need to carefully consider the causal model for a specific research question to determine whether IQ may confound associations.

Adjustments for neuroimaging covariates, such as head motion, are often run to reduce confounding bias. However, head motion, rather than inducing confounding bias, creates measurement error (information bias). When adjusting for head motion during scanning, we observed increases in the spatial extent of the clusters. This might indicate a reduction or an increase in bias. First, bias might have been reduced by adjusting for the fact that children with ADHD will have more error in their cortical measures. Second, bias might have also been increased because we conditioned for head motion, which is a consequence of ADHD.

Overall, the role of head motion in the relation between brain structure and ADHD is complex and warrants the utmost care. Adjusting or not would both lead to bias, meaning that considering which bias might be strongest is necessary. Moreover, methods aiming to reduce information bias should be leveraged (Lash et al., 2021); however, further developments are needed for their application to the neuroimaging field.

Similar considerations should be applied to other neuroimaging parameters which often are indicators of information bias rather than confounders in brain structure–ADHD relations (e.g., time of the day, scanner). For instance, time of the day has been shown to influence morphometric values (Nakamura et al., 2015; Trefler et al., 2016), leading to information bias. If we expect for ADHD symptoms to influence the time of the day in which children with ADHD versus without come to the scanner (differential information bias), the same considerations for head motion would apply. If instead, children with ADHD and without come to the MRI at similar times of the day (non-differential information bias), then adjustments for time of the day would be appropriate and reduce the measurement error.

Our considerations on confounding likely generalize to the psychiatric neuroimaging field, as several confounders considered here (e.g., SES) also relate to brain function and other psychiatric disorders (Biazoli et al., 2020; Kivimäki et al., 2020; Apter et al., 2017). Similarly, other psychiatric disorders are also characterized by complex relations with IQ (Der et al., 2009). Moreover, mental health problems featuring state anxiety, like internalizing and externalizing symptoms, have also been related to increased head motion during scanning (Eijlers et al., 2021).

Confounding control is paramount to studies examining determinants of a phenotype, like ADHD. However, even in these studies, one may be tempted to conduct correlational research with limited confounding adjustments, and then speculate about biological causal mechanisms (Hernán, 2018; Grosz et al., 2020). Rather, we suggest leveraging prior literature and expert knowledge to identify and adjust for key confounders. This can help eliminate the influence of alternative mechanisms (to the ones hypothesized) on the relation of interest (Hernan and Robins, 2020). Charting the assumed (causal) structures to identify confounders can be done through the use of tools such as DAGs (Hernan and Robins, 2020). Naturally, the plausibility of such assumptions should be evaluated. To facilitate the minimization of confounding bias in psychiatric neuroimaging, we propose a workflow in Figure 9.

Despite leveraging two large samples with similar characteristics and assessments, this study presents several limitations. First, there is always potential for residual confounding through unmeasured confounders and misclassification of measured confounders. For example, given that genetic factors influence both ADHD and brain morphology and that there is a genetic correlation between ADHD risk and intracranial volume (Jansen et al., 2015; Klein et al., 2019; Klein et al., 2017), certain genetic risk variants may be unmeasured confounders. However, we aimed to illustrate plausible confounding bias scenarios for ADHD and brain structure, and not to provide an exhaustive list of potential confounders, which may vary depending on the study population and research question. Future studies should also consider bias analyses to assess the impact that residual confounding may have on the study results (Lash et al., 2021). Bias analyses can help understand the minimum association strength an unmeasured confounder needs to have with the determinant and outcome to fully explain away the findings (VanderWeele and Ding, 2017). Developments may be needed, however, for their adaptation to the neuroimaging field.

Second, due to our cross-sectional design, deliberately chosen to correspond to most neuroimaging studies, we must assume all confounders precede our determinant and outcome. This is a plausible assumption for the Generation R Study as, being a prospective birth cohort, we could ensure that the confounders here considered temporally preceded both ADHD and neuroanatomical assessments. However, this was not possible for the ABCD Study, which started sampling at child ages 9–10 years. Future research on the temporal relations between potential confounders, ADHD, and brain structure will aid the minimization of confounding bias when investigating the structural substrates of ADHD.

Third, while we leveraged both symptom-level and diagnostic data for ADHD, this was done within population-based studies. Our results cannot, therefore, be generalized to a clinical population. Future research could examine the extent to which associations between brain structure and ADHD change after adjustments for likely confounders in clinical samples.

In conclusion, leveraging an empirical example from two large studies on neuroanatomy and ADHD symptoms, we highlighted the opportunity for future studies to consider further key confounders. These can be identified based on prior literature and causal diagrams as well as be readily collected, offering a feasible venue for future research. Adjusting for these potential confounders helped identify more refined cortical associations with ADHD symptoms, robust to the influence of demographic and socioeconomic factors, pregnancy exposures, and maternal psychopathology. We also evaluated the potential role of IQ, which could be a mediator, collider, and/or confounder. While adjusting for IQ led to reductions in associations, these would, however, likely not be attributable to reduced confounding bias. Lastly, we explored how head motion reflects information, rather than confounding bias. We discussed the generalizability of these considerations on confounding bias to psychiatric neuroimaging, and suggest a workflow that can be followed to minimize confounding bias in future studies.

We analyzed data from two independent population-based cohorts: the ABCD Study and the Generation R Study. The ABCD Study is conducted across 21 study sites in the US and recruited since 2015 children aged 9–10 at baseline (Garavan et al., 2018). The Generation R Study is based in Rotterdam, the Netherlands, with data collection spanning from fetal life until early adulthood, and started in 2002 (Kooijman et al., 2016). Details of the sampling rationale, recruitment, methods, and procedures have been described elsewhere (Kooijman et al., 2016; Garavan et al., 2018). Research protocols for the ABCD Study were approved by the institutional review board of the University of California, San Diego (#160091), and the institutional review boards of the 21 data collection sites, while the design of the Generation R Study was approved by the Medical Ethics Committee of the Erasmus MC (METC-2012-165). For both studies, written informed consent and assent from the primary caregiver or child were obtained.

In this cross-sectional study, we leveraged data from the baseline assessment of the ABCD Study (release 2.0.1) and the 10-year assessment of the Generation R Study. Both waves included behavioral and neuroimaging measures. We included children with data on ADHD symptoms and T₁-weighted MRI images. Participants were excluded if (1) they had dental braces, (2) incidental findings, (3) their brain scans failed processing or quality assurance procedures, or (4) they were twins or triplets. Of note, excluding children with dental braces is unlikely to determine selection bias by SES in either the ABCD or the Generation R Study as the former cohort covered the costs of dental braces removal for all children who enrolled, while dental care is insured for all children in the Netherlands. Within the Generation R Study, a small set of participants were additionally excluded because they had a different scan sequence. Finally, for each non-twin sibling set, one was randomly included to minimize shared method variance bias. Flowcharts for participant inclusion and exclusion are available in Figure 10. The final samples consisted of 7722 and 2531 children from the ABCD and Generation R Studies, respectively.

Similar covariates were grouped into confounding sets to minimize the number of tested models while including relevant confounders. Factors included in model 1 related to demographic and study characteristics (age, sex, ethnicity, and study site [for ABCD only]). Age and sex were selected as these have been previously adjusted for in previous neuroimaging studies of ADHD (Supplementary file 1a). Ethnicity was used as a proxy for differential health risk exposure among people of different ethnic groups. The study site was incorporated to account for location and scanner differences in the ABCD Study.

Further potential confounders were selected based on previous literature and with the aid of DAGs, as described in the Results. In model 2, variables indicating socioeconomic factors were included (parental education, household income, maternal age at childbirth). Household income and parental education are generally considered to measure childhood SES in health research (American Psychological Association, 2021). Maternal age at childbirth can additionally inform on the SES of the child by capturing part of the variance unexplained by income and education (e.g., younger mothers facing higher occupational challenges, highly educated mothers delaying childbirth Heck et al., 1997). In model 3, maternal factors from the prenatal and postnatal period were grouped (tobacco and cannabis use during pregnancy and maternal psychopathology) to measure early life exposures which may impact a child’s brain and psychiatric development.

The R statistical software (version 4.1.0) was used for all analyses. Missing data on covariates were imputed with chained equations using the mice R package (Buuren and Groothuis-Oudshoorn, 2011). Linear vertex-wise analyses were performed with the QDECR R package (Lamballais and Muetzel, 2021), with surface area/volume/thickness and ADHD symptoms as variables of interest. Correction for multiple testing was applied by using cluster-wise corrections based on Monte Carlo simulations with a cluster forming threshold of 0.001, which yields false-positive rates similar to full permutation testing (Greve and Fischl, 2018). A Bonferroni correction was applied to adjust for analyzing both hemispheres separately (i.e., p < 0.025 cluster-wise).

Our analyses for aim 1 involved three vertex-wise linear regression models, which progressively expanded to adjust for additional confounding factors. The first model focused on demographic covariates, the second on socioeconomic ones, and the third on maternal behavioral variables related to psychopathology and pregnancy exposures. These models were run for ADHD symptoms (in both the ABCD and Generation R Studies) and ADHD diagnosis (in the ABCD Study only, as sensitivity analysis). For the results, we primarily reported cluster sizes, that is, the area of the cerebral cortex that was statistically significantly associated with ADHD symptoms. The sizes of the clusters were reported in cm².

Two additional models, building upon model 3, were run to address aim 2, to illustrate the consequences of adjusting for factors with a complex and unclear relation with brain structure and ADHD symptoms: IQ and head motion (analyses ran separately). Of note, given that IQ and ADHD were weakly correlated (r_ABCD = −0.11, r_GENR = −0.14), multicollinearity was not expected.

Code used to conduct this project is publicly available at https://github.com/LorenzaDA/ADHD_brainmorphology_confounding (Copy archived at swh:1:rev:95c01381fc7fad9bedb3b5918fb80b02b1dcbdfa) (Dall’Aglio, 2022) under CC by 4.0.