Analysis of structural brain asymmetries in Attention-Deficit/Hyperactivity Disorder in 39 datasets

Abstract

Objective:

Some studies have suggested alterations of structural brain asymmetry in attention-deficit/hyperactivity disorder (ADHD), but findings have been contradictory and based on small samples. Here we performed the largest-ever analysis of brain left-right asymmetry in ADHD, using 39 datasets of the ENIGMA consortium.

Methods:

We analyzed asymmetry of subcortical and cerebral cortical structures in up to 1,933 people with ADHD and 1,829 unaffected controls. Asymmetry Indexes (AIs) were calculated per participant for each bilaterally paired measure, and linear mixed effects modelling was applied separately in children, adolescents, adults, and the total sample, to test exhaustively for potential associations of ADHD with structural brain asymmetries.

Results:

There was no evidence for altered caudate nucleus asymmetry in ADHD, in contrast to prior literature. In children, there was less rightward asymmetry of the total hemispheric surface area compared to controls (t=2.1, P=0.04). Lower rightward asymmetry of medial orbitofrontal cortex surface area in ADHD (t=2.7, P=0.01) was similar to a recent finding for autism spectrum disorder. There were also some differences in cortical thickness asymmetry across age groups. In adults with ADHD, globus pallidus asymmetry was altered compared to those without ADHD. However, all effects were small (Cohen’s d from −0.18 to 0.18) and would not survive study-wide correction for multiple testing.

Conclusion:

Prior studies of altered structural brain asymmetry in ADHD were likely under-powered to detect the small effects reported here. Altered structural asymmetry is unlikely to provide a useful biomarker for ADHD, but may provide neurobiological insights into the trait.

Introduction

Attention-deficit/hyperactivity disorder (ADHD) is among the most frequently diagnosed childhood-onset mental disorders, affecting 5% of individuals worldwide (Polanczyk, de Lima, Horta, Biederman, & Rohde, 2007). ADHD is characterized by developmentally inappropriate and impairing levels of inattention and/or hyperactivity, impulsivity, and emotional dysregulation (American Psychiatric Association, 2013). At least 15% of children diagnosed with ADHD retain the diagnosis into adulthood (Faraone et al., 2015; Fayyad et al., 2017).

Left-right asymmetry (laterality) is an important feature of human brain organization (Duboc, Dufourcq, Blader, & Roussigne, 2015; Renteria, 2012; Toga & Thompson, 2003), and altered structural or functional asymmetry has been reported for a range of psychiatric conditions (Toga & Thompson, 2003). The right hemisphere is typically dominant for some aspects of attention and arousal (Heilman, Bowers, Valenstein, & Watson, 1986), and it was observed in the 1980s that people with unilateral lesions in the right hemisphere can show ADHD-like symptoms (Heilman et al., 1986). Since then, various neuropsychological and functional imaging studies have found differences between people with ADHD compared to controls (e.g. (Cortese et al., 2012)), with some pointing to a particular involvement of right hemisphere alterations (Geeraerts, Lafosse, Vaes, Vandenbussche, & Verfaillie, 2008; Hale et al., 2014; Hale et al., 2010; Langleben et al., 2001; Stefanatos & Wasserstein, 2001; Vance et al., 2007). However, not all functional data fit a primarily right-hemisphere model (Hale et al., 2009; Mohamed, Börger, Geuze, & van der Meere, 2016; Zou & Yang, 2019).

In terms of brain anatomy, several studies have reported altered asymmetry of the caudate nucleus in ADHD, although not consistently in the direction of effect (Castellanos et al., 1996; Dang et al., 2016; Filipek et al., 1997; Hynd et al., 1993; Schrimsher, Billingsley, Jackson, & Moore, 2002; Uhlikova et al., 2007). Altered asymmetry of grey matter volumes in the superior frontal and middle frontal gyri has been reported in ADHD (Cao et al., 2014), as well as decreased asymmetry of cortical convolution complexity in the prefrontal cortex (X. Li et al., 2007). Reduced hemispheric asymmetry of white matter networks has also been reported in ADHD compared to controls (D. Li et al., 2018). Douglas et al. (Douglas et al., 2018) performed the largest study of brain anatomical asymmetry in ADHD to date, including 192 cases with ADHD with a history of pharmacotherapy, 149 medication-naïve cases with ADHD, and 508 typically developing controls (ages 6-21 years), from eight separate datasets. They calculated per-subject Asymmetry Indexes (AI) for various regional grey matter volumes, AI=(Left-Right)/((Left+Right)/2) (a widely used approach in studies of brain asymmetry (Kong et al., 2018; Kurth, Gaser, & Luders, 2015; Leroy et al., 2015; Postema et al., 2019)), but did not find any significant alterations of AIs in ADHD (Douglas et al., 2018). However, in a subset of their dataset (56 cases and 48 controls), Douglas et al. (Douglas et al., 2018) analyzed diffusion tensor imaging (DTI) data, including fractional anisotropy and mean diffusivity measures, and reported alterations in the asymmetry of six white matter tracts, again not specifically driven by alterations in the right hemisphere.

In the current study, we measured cortical regional AIs in 1,978 cases and 1,917 controls from 39 datasets, and subcortical AIs in 1,736 cases and 1,654 controls from 35 datasets, made available via the ADHD working group of the ENIGMA (Enhancing NeuroImaging Genetics through MetaAnalysis) consortium. The same datasets were recently analyzed in two other studies, by Hoogman et al.(Hoogman et al., 2017; Hoogman et al., 2019), that investigated bilateral changes in subcortical volumes and cortical measures, but not alterations of asymmetry. They found that ADHD was associated with lower average volumes of various subcortical structures (Hoogman et al., 2017), as well as lower total and regional cortical surface areas (including frontal, cingulate and temporal regions), and decreased cortical thickness in fusiform gyrus and temporal pole (Hoogman et al., 2019). These effects were largest in children, and even child-specific for the cortical findings, so that for the present study of asymmetries, we followed the age-group division of Hoogman et al. (Hoogman et al., 2019) into children (<15 years), adolescents (15-21 years) and adults (>21 years), as well as performing analysis in the total combined sample to explore age-general effects. Bilateral effect sizes reported by Hoogman et al. (Hoogman et al., 2017; Hoogman et al., 2019) were small, i.e. case-control Cohen’s d values between −0.21 and 0.06. This suggests that, if associations between ADHD diagnosis and regional brain asymmetries are similarly subtle, many previous studies of anatomical asymmetries in this disorder were underpowered, and the described effects may have been unreliable. Low statistical power in a study not only reduces the chance of detecting true effects, but also the likelihood that significant results reflect true effects (Munafo & Flint, 2010). It is important for the field of neuroimaging to mature around more highly powered analyses in relation to subtle effects. The current study aimed to provide detailed information on the extent to which laterality is affected in ADHD, based on the largest ever sample size for this question, comprised of multiple independent cohorts from around the world.

Methods

Ethical Considerations

This study made use of 39 pre-existing datasets from around the world. For all datasets, the participating sites had obtained ethical approval from local institutional review boards, as well as informed consent to participate.

Datasets

Bilateral brain measures derived from structural MRI were available from 39 different datasets via the ENIGMA-ADHD Working Group (ST1). The 39 datasets comprised cortical data from a total of 1,933 participants with ADHD (1,392 males; median age = 15 y; range = 4 y to 62 y) and 1,829 healthy individuals (1,116 males; median age = 14 y; range = 4 y to 63 y). Subcortical data were available from 35 of the 39 datasets and comprised 1,691 cases (1,212 males, median age = 15 y; range = 5 y to 62 y) and 1,566 controls (953 males, median age = 14 y; range= 4 y to 63 y). A previous study by Douglas et al. (Douglas et al., 2018) (see introduction) included five datasets that were also analyzed in the present study (ST1).

For all but 4 of the 39 datasets, ADHD diagnosis was based on the Diagnostic and Statistical Manual of Mental Disorders 4^th Edition (DSM-IV) (American Psychiatric Association, 2000). Other instruments used were DSM5^th Edition (DSM-5), or the International Classification of Diseases (ICD)10^th Edition) (World Health Organization, 1992). For information per dataset see ST1.

In terms of age groups, for children (<15 y) there were subcortical data from 802 cases and 842 controls, and cortical data from 912 cases and 950 controls; for adolescents (15 y – 21 y) there were subcortical data from 326 cases and 232 controls, and cortical data from 408 cases and 340 controls; for adults (> 21 y) there were subcortical data from 563 cases and 492 controls, and cortical data from 613 cases and 539 controls.

Eleven additional datasets, comprising cases-only or controls-only, were excluded for the purpose of the present study (these are not listed in ST1). This was because our analysis models included random intercepts for ‘dataset’ (below), and diagnosis would be fully confounded with ‘dataset’ for case-only or control-only datasets.

MRI-based measures

Structural T1-weighted brain MRI scans had been acquired at each study site for each of the 39 pre-existing MRI datasets. MRI data within the ENIGMA consortium are typically processed separately at each participating site, due to varying restrictions on data sharing that apply to the many legacy datasets from different countries around the world. Images were obtained at different field strengths (1.5 T or 3 T: see ST1). Scanners and scanning sequences, recruitment criteria, and demographics differed between datasets, but all sites separately applied a single image processing and quality-control protocol from the ENIGMA consortium (http://enigma.ini.usc.edu/protocols/imaging-protocols), starting from their T1 image data. The harmonized processing was based on the freely available and validated software FreeSurfer (versions 5.1 or 5.3) (Fischl, 2012), with the default ‘recon-all’ pipeline (https://surfer.nmr.mgh.harvard.edu/fswiki/recon-all), which is a 29-step procedure that includes skull stripping, registration, subcortical segmentation, normalization, white matter and pial surface creation, cortical parcellation according to the Desikan-Killiany atlas, and the output of region-specific measures of volume, average thickness and surface area. This was followed by visual inspection of both internal and external segmentations (Supplementary Methods). Exclusions on the basis of these quality control steps resulted in the sample sizes given above. The present study took as its starting point the FreeSurfer-derived measures of left and right volumes of seven bilaterally paired subcortical structures, and thickness and surface area measures for each of 34 bilaterally paired cortical regions, that were generated previously by each site. The cortical regions were defined by the Desikan-Killiany atlas (Desikan et al., 2006). In addition, the average cortical thickness and total surface area per hemisphere were analyzed. Freesurfer’s measure of intracranial volume (ICV) was also considered as a covariate in sensitivity analyses (below).

The Desikan-Killiany atlas (Desikan et al., 2006) was derived from manual segmentations of reference brain images. The labeling system incorporates hemisphere-specific information on sulcal and gyral geometry with spatial information regarding the locations of brain structures (Desikan et al., 2006). Accordingly, the mean regional asymmetries in our data might be influenced by left-right differences present in the reference dataset used for constructing the atlas. Nonetheless, this approach was appropriate for our study focused on comparing relative asymmetry between groups. The use of an asymmetrical atlas has the advantage that regional identification is likely to be accurate for structures that are asymmetrical both in the atlas and, on average, in the study population.

Asymmetry indexes

Left and right data per brain region and individual participant were loaded into R (version 3.5.3), and null values were removed. An asymmetry index (AI) was calculated for each subject and each paired left-right measure using the following formula: (Left-Right)/(Left+Right). Negative AIs therefore indicate a right>left asymmetry, while positive AIs indicate a left>right asymmetry. In the AI formula, the L-R difference (numerator) is adjusted by the bilateral measure L+R (denominator), such that the AI does not scale with the bilateral measure. We did not divide the denominator by 2, in contrast to some previous formulations of AIs (see Introduction), but this makes no difference in terms of deriving Cohen’s d effect sizes and P-values for group comparisons. Distributions of each of the AIs in the total study sample are plotted in SF1.

Correlations between AI measures in the total study sample were calculated using Pearson’s R and visualized using the corrplot package in R (SF2–SF4). Most pairwise correlations between AIs were of low magnitude (median magnitude r = 0.024 for surface area AIs, 0.040 for thickness AIs, 0.091 for subcortical volume AIs), with a minimum r = −0.42 between caudal anterior cingulate surface area and superior frontal surface area, and maximum r = 0.49 between rostral middle frontal thickness and total average thickness.

Linear mixed effects random-intercept models

Main analysis

Linear mixed effects analyses were performed separately for each subcortical volume AI, cortical regional surface and thickness AI, and the total hemispheric surface area and average thickness AI, using the nlme package in R (version 3.5.3). Analyses were conducted separately within children, adolescents, and adults, as well as on the total study sample. All models included diagnosis (a binary variable; 0=control, 1=case), sex (binary; 0=female, 1=male) and age (numeric) as fixed factors, and dataset as a random factor (39 categories for cortical data, 35 categories for subcortical data):

$$AI\sim diagnosis+sex+age+random\left(\sim 1|dataset\right)$$ (1)

The Maximum Likelihood (ML) method was used to fit the models. Whenever any of the predictor variables was missing in a given subject, the subject was omitted from the analysis (method = ‘na.omit’). The ‘optim’ optimizer (lmeControl(opt=‘optim’) was used for all models. Residual plots are in SF5–SF7.

The t-statistic for the factor ‘diagnosis’ in each linear mixed effects model was derived and used to calculate Cohen’s d (Supplementary Methods). For visualization of cerebral cortical results, Cohen’s d values were loaded into Matlab (v. R2020a) and 3D images of left hemisphere inflated cortical and subcortical structures were obtained using Freesurfer-derived ply files.

Field strength was not included as a covariate because each dataset was scanned entirely at either 1.5 T or 3T (ST1), and the models included ‘dataset’ as a random intercept effect, which adjusted for differences that applied to entire datasets.

Significance and detectable effect sizes

Significance was assessed based on the P-value for the diagnosis term within each model. Separately within each age group, and again within all age groups combined, we applied False Discovery Rate (FDR) correction (Benjamini & Hochberg, 1995) for multiple testing, separately across the seven subcortical structures, the 35 cortical surface area AIs (i.e. 34 regional AIs and one hemispheric total AI), and again for the 35 cortical thickness AIs, each time with an FDR threshold of 0.05. Therefore twelve separate FDR corrections were done. We also applied an additional FDR correction for the total combined analysis across all age groups and AIs of different types.

As each linear model included multiple predictor variables, the power to detect an effect of diagnosis on AI could not be computed exactly, but we obtained an indication of the effect size that would be needed to provide 80% power had we been using simple t-tests and Bonferroni correction for multiple testing, using the pwr command in R (Supplementary Methods). For this purpose, a significance level of 0.0071 (i.e., 0.05/7) or 0.0014 (i.e. 0.05/35) was set in the context of multiple testing over the seven subcortical volumes, or the regional and total cortical surface areas (N = 35) or thicknesses (N = 35). This showed that, in the total study sample, a case-control effect size of roughly Cohen’s d=0.12 (subcortical), or d=0.13 (cortical), would be detectable with 80% power. For the analyses in the different age groups, this was, respectively, d=0.16 and d=0.19 in children, d=0.26 and d=0.30 in adolescents, and d=0.21 and d=0.24 in adults.

Directions of asymmetry changes

For any AIs showing nominally significant effects (i.e., unadjusted P<0.05) of diagnosis in any of the primary analyses, post hoc linear mixed effects modelling was also performed on the corresponding L and R measures separately, to understand the unilateral changes involved. The models included the same terms as were used in the main analysis of AIs (i.e., diagnosis, age and sex as fixed factors, and dataset as random factor). Again, the Cohen’s d effect sizes for diagnosis were calculated based on the t-statistics. The raw mean AI values were calculated separately in controls and cases, to describe the reference direction of healthy asymmetry in controls, and whether cases showed lower, higher, or reversed asymmetry relative to controls.

Sensitivity analyses

The relationships between AIs and age appeared roughly linear across all age groups combined (SF8–10). Therefore, no polynomials for age were incorporated in the main model (Supplementary Methods). However, analyses were repeated (only for all age groups combined) using an additional non-linear term for age, to check whether this choice had affected the results. The variables age and age² are inevitably highly correlated. To include linear and non-linear effects of age in the same model, we made use of the poly()-function in R for these two predictors, which created a pair of uncorrelated variables related to age (so-called orthogonal polynomials) (Chambers & Hastie, 1992), where one variable was linear and one non-linear:

$$AI\sim diagnosis+\mathrm{poly}\left(age,2\right)+sex+age+\mathrm{random}\left(\sim 1|dataset\right)$$ (2)

Note that we were not interested to measure the effects of age or age-squared, but simply to correct for linear and non-linear effects related to age, as we measured the effects of diagnosis on brain asymmetry.

No AI outliers were removed for the main analysis, but to confirm that results were not dependent on outliers, the main analysis was also repeated (for all age groups combined) after having winsorized using a threshold of k= 3, for each AI measure separately in the total combined dataset.

Associations between brain asymmetries and IQ, comorbidity, ADHD severity and psychostimulant medication

Within the ADHD participants only (all age groups combined), brain asymmetries were tested in relation to several potentially associated variables (IQ, comorbidity, severity, medication use; see SF11, SF12), using separate models in which each variable was considered as a fixed effect:

$$AI\sim variable+age+sex+\mathrm{random}\left(\sim 1|dataset\right)$$ (3)

See Supplementary Methods for the derivation of these variables. For binary variables, datasets were removed if they had < 1 subject per category, to avoid the random variable ‘dataset’ being fully confounded with the binary variable for any datasets. Depending on the availability of each specific AI, data for testing association with IQ were available for up to 1,719 ADHD individuals (exact numbers per AI depended on image quality control for that region and can be found in the relevant results tables, see below). For the presence/absence of comorbidities, four different binary variables were constructed: mood disorder (up to 179 yes, 384 no), anxiety disorder (up to 82 yes, 503 no), oppositional defiant disorder (ODD; up to 80 yes, 151 no), and substance use disorder (SUD; up to 77 yes, 335 no). For ADHD symptom severity, two continuous variables were used: hyperactivity/impulsivity (up to 1,009 ADHD participants) and inattention (1,006 ADHD participants). For psychostimulant medication use, two binary variables were constructed: lifetime use (up to 337 yes, 188 no), and current use (i.e. at the time of scanning, up to 361 yes, 377 no) (see SF12 for the distributions, and supplementary methods for more explanation).

IQ was also examined in controls only (all age groups combined) to explore the relationships between IQ and brain asymmetries in typically developing individuals. IQ and AI data were available for up to 1,663 controls. The model for each AI was:

$$AI\sim IQ+age+sex+random\left(\sim 1|dataset\right)$$

IQ, handedness and intracranial volume as covariates in disorder case-control analysis

See the supporting information for the derivation of IQ and handedness measures, and above for ICV. Distributions are in SF11. We did not adjust for IQ, handedness or ICV as covariate effects in our main, case-control analysis (above). This was because, a priori, there are various possible causal relations linking these traits to ADHD and brain asymmetry and other, possibly underlying factors shared between some or all of them. In this context, it is important not to bias associations between ADHD and brain asymmetry through correcting for these factors as covariates in primary analysis, as they may be colliders (Cole et al., 2010) (see the Discussion for more on this issue). However, we included a set of additional, secondary models to test for case-control effects in the presence of these variables as covariates:

$$AI\sim diagnosis+age+sex++handedness+random\left(\sim 1|dataset\right)$$

$$AI\sim diagnosis+age+sex++handedness+handednes{s}^{\ast }diagnosis+random\left(\sim 1|dataset\right)$$

$$AI\sim diagnosis+age+sex+IQ+random\left(\sim 1|dataset\right)$$

$$AI\sim diagnosis+age+sex+IQ+I{Q}^{\ast }diagnosis+random\left(\sim 1|dataset\right)$$

$$AI\sim diagnosis+age+sex++ICV+random\left(\sim 1|dataset\right)$$

$$AI\sim diagnosis+age+sex+ICV+IC{V}^{\ast }diagnosis+random\left(\sim 1|dataset\right)$$

The analyses were also repeated after winsorization of outliers, as above.

Results

Associations of brain asymmetry with ADHD

Results for all AIs across the different age groups, and for all age groups combined, are listed in the supplement (ST2–ST13), and are also available as supplementary comma-delimited text files.

Children

There were no associations of diagnosis with AIs that had FDR <0.05 in children (T1–T3, ST2–ST4). The children showed nominally significant associations (unadjusted P<0.05) of diagnosis with the AIs of total hemispheric surface area (t=2.10, P=0.036), medial orbitofrontal cortex surface area (t=2.7, P=0.006), and paracentral lobule surface area (t=−2.16, p=0.031) (Table 2, ST3). The Cohen’s d for these effects were 0.11, 0.13 and −0.10 respectively (Figure 1, SF13, ST3). Post hoc analysis showed that the effects on total hemispheric and medial orbitofrontal surface area asymmetries both involved relatively greater reductions on the right-side than left-side in ADHD compared to controls (ST14). The effect on paracentral lobule surface area asymmetry was driven by a larger decrease of left compared to right-hemispheric surface area in this region (ST14).

Table 1.

Linear mixed model results for subcortical volume AIs.

Subcortical volume AI	Children only		Adolescents only		Adults only		Total study sample
	P1	d 2	P1	d 2	P1	d 2	P1	d 2
Accumbens	0.26	−0.06	0.36	−0.08	0.90	0.01	0.32	−0.03
Amygdala	0.78	−0.01	0.72	0.03	0.69	−0.03	0.61	−0.02
Caudate Nucleus	0.60	0.03	0.88	0.01	0.45	0.05	0.41	0.03
Globus Pallidus	0.65	−0.02	0.39	−0.08	0.004	−0.18	0.01	−0.09
Hippocampus	0.84	−0.01	0.09	0.15	0.46	0.05	0.62	0.02
Putamen	0.54	−0.03	0.87	−0.02	0.52	−0.04	0.26	−0.04
Thalamus*	0.42	0.04	0.28	0.10	0.48	0.04	0.15	0.05

¹Uncorrected P-values for diagnosis are indicated, with in bold those that are significant at the uncorrected level (P < 0.05), and in bold-italic those that survive multiple testing correction within the particular analysis indicated (see text).

²Cohen’s d value for the effect of diagnosis..

Table 3.

Linear mixed model results for the cortical thickness AIs.

Cortical thickness AI	Children only		Adolescents only		Adults only		Total study sample
	P1	d 2	P1	d 2	P1	d 2	P1	d 2
Banks of superior temporal sulcus	0.05	−0.10	0.54	−0.05	0.64	−0.03	0.06	−0.06
Caudal anterior cingulate cortex	0.25	0.05	0.60	−0.04	0.06	0.11	0.11	0.05
Caudal middle frontal cortex	0.04	0.10	0.09	0.13	0.73	0.02	0.03	0.07
Cuneus	0.69	0.02	0.04	−0.15	0.06	0.11	0.56	0.02
Entorhinal cortex	0.12	−0.08	0.79	0.02	0.65	−0.03	0.26	−0.04
Frontal pole	0.27	0.05	0.20	−0.10	0.19	0.08	0.34	0.03
Fusiform gyrus	0.56	−0.03	0.98	0.002	0.79	0.02	0.94	−0.003
Inferior parietal cortex	0.96	0.00	0.59	−0.04	0.51	0.04	0.81	0.01
Inferior temporal gyrus	0.24	−0.05	0.79	0.02	0.84	−0.01	0.69	−0.01
Insula	0.05	−0.09	0.32	−0.08	0.94	−0.004	0.05	−0.06
Isthmus cingulate cortex	0.81	−0.01	0.22	0.09	0.35	−0.06	0.91	0.00
Lateral occipital cortex	0.76	0.01	0.40	−0.06	0.03	0.13	0.41	0.03
Lateral orbitofrontal cortex	0.75	−0.01	0.51	0.05	0.14	0.09	0.42	0.03
Lingual gyrus	0.34	−0.04	0.85	0.01	0.59	−0.03	0.29	−0.04
Medial orbitofrontal cortex	0.06	−0.09	0.31	0.08	0.04	0.12	0.97	0.001
Middle temporal gyrus	0.75	−0.02	0.62	−0.04	0.01	−0.17	0.11	−0.05
Paracentral lobule	0.15	−0.07	0.12	0.12	0.77	−0.02	0.53	−0.02
Parahippocampal gyrus	0.07	−0.09	0.09	−0.13	0.39	0.05	0.12	−0.05
Pars opercularis of inferior frontal gyrus	0.80	0.01	0.39	0.07	0.89	−0.01	0.45	0.02
Pars orbitalis of inferior frontal gyrus	0.36	0.04	0.95	−0.004	0.37	0.05	0.30	0.03
Pars triangularis of inferior frontal gyrus	0.67	−0.02	0.36	0.07	0.90	−0.01	0.92	0.003
Pericalcarine cortex	0.92	−0.004	0.98	−0.002	0.004	0.17	0.15	0.05
Postcentral gyrus	0.94	−0.004	0.92	−0.01	0.01	−0.15	0.11	−0.05
Posterior cingulate cortex	0.57	−0.03	0.47	−0.05	0.87	−0.01	0.35	−0.03
Precentral gyrus	0.02	0.11	0.32	−0.08	0.17	0.08	0.05	0.06
Precuneus	0.73	0.02	0.22	0.09	0.69	0.02	0.36	0.03
Rostral anterior cingulate cortex	0.92	−0.004	0.06	0.15	0.36	0.06	0.21	0.04
Rostral middle frontal gyrus	0.68	0.02	0.78	−0.02	0.34	−0.06	0.85	−0.01
Superior frontal gyrus	0.77	0.01	0.10	0.13	0.64	0.03	0.30	0.03
Superior parietal cortex	0.98	−0.001	0.47	−0.06	0.85	0.01	0.77	−0.01
Superior temporal gyrus	0.06	0.09	0.42	0.07	0.36	−0.06	0.28	0.04
Supramarginal gyrus	0.18	−0.06	0.51	−0.05	0.93	−0.005	0.19	−0.04
Temporal pole	0.56	0.03	0.77	0.02	0.62	−0.03	0.77	0.01
Transverse temporal gyrus	0.66	0.02	0.65	0.03	0.34	−0.06	0.98	−0.001
Total average thickness	0.92	−0.005	0.78	0.02	0.75	0.02	0.78	0.01

¹Uncorrected P-values for the effects of diagnosis are indicated, with in bold those that are significant at the uncorrected level (P < 0.05). None of the associations with diagnosis survived multiple testing correction.

²Cohen’s d value for the effect of diagnosis.

Table 2.

Linear mixed model results for the cortical surface area AIs.

Cortical surface area AI	Children only		Adolescents only		Adults only		Total study sample
	P1	d 2	P1	d 2	P1	d 2	P1	d 2
Banks of superior temporal sulcus	0.80	−0.01	0.53	−0.05	0.81	0.01	0.48	−0.02
Caudal anterior cingulate cortex	0.75	−0.01	0.29	−0.08	0.71	0.02	0.64	−0.02
Caudal middle frontal cortex	0.41	0.04	0.55	−0.05	0.22	0.07	0.19	0.04
Cuneus	0.16	0.07	0.92	−0.01	0.07	−0.11	0.74	−0.01
Entorhinal cortex	0.95	0.003	0.42	−0.06	0.10	−0.10	0.34	−0.03
Frontal pole	0.05	−0.09	0.22	0.09	0.25	−0.07	0.10	−0.05
Fusiform gyrus	0.17	−0.06	0.35	0.07	0.11	−0.10	0.15	−0.05
Inferior parietal cortex	0.27	0.05	0.98	−0.002	0.89	−0.01	0.44	0.03
Inferior temporal gyrus	0.57	0.03	0.84	0.02	0.25	0.07	0.25	0.04
Insula	0.10	0.08	0.56	0.04	0.64	−0.03	0.28	0.04
Isthmus cingulate cortex	0.95	−0.003	0.19	−0.10	0.49	0.04	0.75	−0.01
Lateral occipital cortex	0.59	−0.02	0.96	−0.004	0.05	0.12	0.48	0.02
Lateral orbitofrontal cortex	0.18	−0.06	0.54	−0.05	0.42	−0.05	0.06	−0.06
Lingual gyrus	0.88	−0.01	0.14	−0.11	0.50	0.04	0.92	−0.003
Medial orbitofrontal cortex	0.01	0.13	0.27	0.08	0.72	−0.02	0.03	0.07
Middle temporal gyrus	0.15	0.07	0.45	−0.06	0.89	−0.01	0.38	0.03
Paracentral lobule	0.03	−0.10	0.96	−0.004	0.28	−0.06	0.03	−0.07
Parahippocampal gyrus	0.37	0.04	0.25	−0.09	0.13	−0.09	0.73	−0.01
Pars opercularis of inferior frontal gyrus	0.88	0.01	0.19	0.10	0.58	0.03	0.34	0.03
Pars orbitalis of inferior frontal gyrus	0.20	0.06	0.02	0.18	0.55	0.04	0.02	0.08
Pars triangularis of inferior frontal gyrus	0.32	0.05	0.14	0.11	0.57	−0.03	0.24	0.04
Pericalcarine cortex	0.30	0.05	0.13	−0.12	1.00	0.00	0.94	0.002
Postcentral gyrus	0.44	0.04	0.29	0.08	0.98	0.00	0.39	0.03
Posterior cingulate cortex	0.62	−0.02	0.46	−0.06	0.84	0.01	0.59	−0.02
Precentral gyrus	0.85	0.01	0.09	−0.13	0.05	−0.12	0.09	−0.06
Precuneus	0.29	0.05	0.47	−0.06	0.65	0.03	0.46	0.02
Rostral anterior cingulate cortex	0.97	−0.002	0.98	0.002	0.36	−0.05	0.51	−0.02
Rostral middle frontal gyrus	0.10	−0.08	0.77	−0.02	0.60	−0.03	0.11	−0.05
Superior frontal gyrus	0.28	0.05	0.09	0.13	0.11	−0.09	0.55	0.02
Superior parietal cortex	0.09	0.08	0.33	0.07	0.68	−0.02	0.27	0.04
Superior temporal gyrus	0.09	0.08	0.87	0.01	0.19	−0.08	0.62	0.02
Supramarginal gyrus	0.86	0.01	0.25	−0.09	0.21	−0.07	0.24	−0.04
Temporal pole	0.65	0.02	0.69	0.03	0.34	−0.06	0.97	0.001
Transverse temporal gyrus	0.66	−0.02	0.44	0.06	0.94	0.005	0.93	0.003
Total average surface area	0.04	0.10	0.73	0.03	0.23	−0.07	0.54	0.02

¹Uncorrected P-values for diagnosis are indicated, with in bold those that are significant at the uncorrected level (P < 0.05). None survived multiple testing correction.

²Cohen’s d value for the effect of diagnosis.

Figure 1.

Cohen’s d effect sizes of the associations between ADHD diagnosis and AIs of subcortical volumes, cortical surface areas and cortical thicknesses for (A) children, (B) adolescents, (C) adults, and (D) all age groups combined. Positive values (red) indicate mean shifts towards greater leftward or reduced rightward asymmetry in ADHD, and negative values (blue) indicate mean shifts towards greater rightward asymmetry or reduced leftward asymmetry in ADHD.

The children also showed nominally significant associations of diagnosis with four regional cortical thickness AIs, which were the banks of the superior temporal sulcus (t=−2.0, P=0.047; increased rightward asymmetry in ADHD), caudal middle frontal cortex (t=2.1, P=0.037; increased leftward asymmetry), precentral gyrus (t=2.4, P=0.019; increased leftward asymmetry) and insula (t=−2.0, P=0.047, decreased leftward asymmetry) (T2, ST14).

Adolescents

There were two nominally significant associations between diagnosis and AIs in adolescents, but none with FDR <0.05 (Table 1–Table 3, ST5–ST7). These involved the pars orbitalis of inferior frontal gyrus surface area (t=2.4, P=0.017), which showed lower rightward asymmetry in ADHD compared to controls, due to a smaller left than right sided decrease (ST14), and cuneus thickness (t=−2.0, P=0.043), which showed greater rightward asymmetry in ADHD compared to controls, due to an increase in right- and a decrease in left-hemispheric thickness (ST14).

Adults

In adults, the globus pallidus AI was significantly associated with ADHD diagnosis with FDR <0.05 (t=−2.9, P=0.004, uncorrected) (T1, ST8). The Cohen’s d effect size for this association was −0.18 (T1, Figure 1, SF13). This effect involved a decrease in leftward asymmetry in ADHD compared to controls, driven by a larger reduction of left-side volume than right-side volume in ADHD compared to controls (ST14). Note this association was only significant in the context of FDR correction for 7 subcortical AIs within adults specifically. (No effects were significant at FDR-corrected P<0.05 when the correction was done across all age groups and AIs of different types, see below).

There were other nominally significant associations of AIs with diagnosis in adults: lateral occipital cortex surface area (t=2.0, P=0.049; increased leftward) (T2, ST9, ST14) and thickness (t=2.2, P=0.026; decreased rightward) (T3, ST10, ST14), medial orbitofrontal cortex thickness (t=2.0, P=0.045; increased leftward), middle temporal gyrus thickness (t=−2.6, P=0.009; increased rightward), pericalcarine cortex thickness (t=2.9, P=0.004; decreased rightward), and postcentral gyrus thickness (t=−2.5, P=0.013; decreased leftward). The corresponding unilateral effects are shown in ST14.

All age groups combined

When combining all age groups, there were nominally significant associations of AIs with diagnosis for the medial orbitofrontal cortex surface area (t=2.2, P=0.029; decreased rightward), paracentral lobule surface area (t=−2.2, P=0.029; increased rightward), pars orbitalis of inferior frontal gyrus surface area (t=2.3, P=0.021; decreased rightward), caudal middle frontal thickness (t=2.2, P=0.027; increased leftward), insula thickness (t=−2.1, P=0.040; decreased leftward), as well as the volume of the globus pallidus (t=−2.6, P=0.010; decreased leftward) (T1–T3, ST11–ST13). The corresponding unilateral effects are shown in ST14.

No effects were significant at FDR-corrected P<0.05 when the correction was done across all age groups and AIs of different types.

The addition of non-linear effects of age to the model had negligible influences on the six nominally significant associations with diagnosis, all of which remained nominally significant except insula thickness (now P=0.050). Likewise, winsorizing outliers (using a threshold k=3, see Methods) also had little influence on the results (the effect on insula thickness asymmetry was no longer nominally significant, P=0.061) (ST15–ST17).

Associations brain asymmetries with comorbidity, ADHD severity, psychostimulant medication, and IQ

Analyses in this section were carried out in all age groups combined.

When testing associations of comorbidity, ADHD severity, psychostimulant medication, or IQ with brain asymmetries within ADHD individuals (ST18–29), only one significant association was found (FDR <0.05 within the particular type of AI and age-defined group), namely between comorbid mood disorder and the rostral middle frontal gyrus thickness AI (P=0.0002, t=3.70) (ST26). Furthermore, various nominally significant (P<0.05) associations were observed: ADHD severity was associated with the AI of the entorhinal cortex surface area (t=2.12, P=0.034; hyperactivity/impulsivity) (ST19). ADHD severity was also associated with four regional cortical thickness asymmetries: the caudal anterior cingulate thickness AI (t=2.66, P=0.008; hyperactivity/impulsivity), pars opercularis of the inferior frontal gyrus thickness AI (t=2.12, P=0.034; hyperactivity/impulsivity, and t=2.04, P=0.04; inattention), and pericalcarine cortex thickness AI (t=2.04, P=0.04; hyperactivity/impulsivity) (ST20).

Current psychostimulant medication use was associated with two cortical regional surface area asymmetries, i.e., precuneus (t=−2.25, P=0.025) and transverse temporal gyrus (t=−2.34, P=0.020) (ST22), and with two thickness asymmetries, i.e., inferior parietal cortex (t=−2.33, P=0.020) and precentral gyrus (t=−2.16, P=0.031) (ST23). Lifetime psychostimulant medication use was associated with three cortical surface area asymmetries (insula (t=−2.03, P=0.043), supramarginal gyrus (t=−2.08, P=0.038), and rostral anterior cingulate cortex (t=1.97, P=0.049) (ST22), and the thickness asymmetry of the paracentral lobule (t=2.15, P=0.032) (ST23). Among the AIs which showed nominally significant associations with medication use, one had also shown a nominally significant association with diagnosis in all age groups combined, i.e., the AI of paracentral lobule surface area (see above). The direction of medication effect was positive, i.e. the opposite to the effect of diagnosis on this AI (see above).

For mood disorder, associations were observed with six thickness AIs (i.e., entorhinal cortex, pars triangularis of inferior frontal gyrus, pericalcarine cortex, precuneus, rostral middle frontal gyrus, and transverse temporal gyrus), and two surface area AIs (i.e., inferior temporal gyrus, and rostral anterior cingulate cortex), of which the association with rostral middle frontal thickness AI survived multiple testing correction (FDR < 0.05) (ST25, 26). Anxiety Disorder was associated with thickness AIs of the cuneus and lateral occipital cortex (ST26). For ODD, associations were found with the AIs of medial orbitofrontal thickness (ST26) and temporal pole surface area (ST25). Additionally, SUD was associated with the thickness AIs of the cuneus and paracentral lobule (ST26), and with surface area AIs of the postcentral gyrus and supramarginal gyrus (ST25). None of these regions showed a nominally significant effect of diagnosis in the main analysis of all age groups combined.

Finally, within ADHD individuals, IQ was nominally associated with the accumbens volume AI (t=2.16, P=0.031), hippocampus volume AI (t=−2.06, P=0.039) (ST27) and lateral occipital cortex surface area AI (t=−2.17, P=0.030) (ST28). Within controls, IQ was associated with the middle temporal gyrus surface area AI (t=−2.52, P=0.012) (ST28), rostral anterior cingulate thickness cortex AI (t=2.47, P=0.014), and supramarginal gyrus thickness AI (t = −2.55, P=0.011) (ST29).

Including IQ, handedness or intracranial volume as covariates in case-control analysis

We carried out secondary analyses in which IQ, handedness or intracranial volume were included as covariates in case-control analysis, with or without interaction terms for these variables with diagnosis (i.e. case-control status) (see Methods for the models used). These extra models identified a small number of main effects of diagnosis, or interactions with diagnosis, that survived multiple testing correction at FDR<0.05 within the specific subset of AIs and ages being analyzed (but would not survive further correction for multiple testing). However, after winsorization of outliers (see Methods), only the diagnosis term for globus pallidus volume AI remained significant, in the model AI ~ diagnosis + age + sex + ICV + random (~1|dataset), when analyzed in the total study sample (P=0.005, t=−2.75), or when analyzed in adults only (P=0.0035, t=−2.93). Complete model results from all of these secondary analyses can be found in supplementary comma-delimited text files.

Discussion

We conducted the largest study to date of associations between anatomical brain asymmetries and ADHD. Linear mixed effects model mega-analyses were carried out separately in children, adolescents, and adults, following previous ENIGMA ADHD working group studies of bilateral brain differences that showed contrasting effects in these age groups (Hoogman et al., 2017; Hoogman et al., 2019). We also analyzed the total study sample for age-general effects. All statistical effects of diagnosis on asymmetries were very small, with Cohen’s d ranging from −0.18 to 0.18. Only one of these associations was significant with a false discovery rate <0.05 within the specific subset of AIs and age-defined subjects in which it was found (globus pallidus asymmetry in adults), and this effect was not significant in analysis of all age groups combined, with FDR correction across all AIs. Therefore, all effects remain tentative, even in this unprecedented sample size. The small effect sizes mean that altered brain asymmetry is unlikely, in itself, to be a useful biomarker or clinical predictor of ADHD. In addition, our results suggest that significant effects reported in prior studies, based on much smaller samples, may have been unrealistically large. As noted in the Introduction, low power not only reduces the chance of detecting true effects, but also increases the likelihood that statistically significant results do not reflect true effects (Munafo & Flint, 2010). There were some notable associations of diagnosis with cortical asymmetry that reached nominal significance in our study. Among these, children with ADHD showed reduced rightward asymmetry of total hemispheric surface area, and medial orbitofrontal surface area. In a recent ENIGMA consortium study of autism spectrum disorder (ASD), medial orbitofrontal cortex surface area asymmetry was altered in the same direction, and to a similar extent, as in the present study (Postema et al., 2019). ADHD and ASD often co-occur (Leitner, 2014) and are known to share genetic influences (Ghirardi et al., 2019; Stergiakouli et al., 2017), such that the two diagnostic labels are likely to capture a partly overlapping spectrum of related disorders (Demopoulos, Hopkins, & Davis, 2013; van der Meer et al., 2012). Studies that aimed to identify shared brain structural traits between ADHD and ASD have found mixed results (Nevena V. Radonjić MD, 2019; Premika S.W. Boedhoe et al., 2019), with perhaps the greatest overlap involving regions of the ‘social brain’, including orbitofrontal cortex (Baribeau et al., 2019). However, laterality has not been specifically studied in this regard, so that our finding of reduced rightward medial orbitofrontal cortex surface area in both disorders may be a new insight into shared neurobiology between ADHD and ASD. Altered lateralized neurodevelopment may play a causal role in disorder susceptibility, or else may arise as a correlated trait due to other underlying susceptibility factors, or even be a downstream consequence of having the disorder (Bishop, 2013). Some aspects of brain asymmetry are partly heritable (Guadalupe et al., 2016; Kong et al., 2018), so that future gene mapping studies for brain asymmetry and disorder susceptibility may help to resolve causal relations underlying their associations.

One functional imaging study (94 cases, 85 controls) reported lower rightward lateralization in medial orbitofrontal cortex in ADHD compared to controls, based on temporal variability during resting-state (Zou & Yang, 2019). Furthermore, a study of 218 participants with ADHD and 358 healthy controls reported that orbitofrontal cortex thickness, but not surface area, showed a left>right asymmetry in childhood controls that switched to right>left asymmetry by late adolescence, while this change did not occur to the same extent in ADHD (Shaw et al., 2009). However, in the present study, there was no effect of diagnosis on thickness asymmetry of this region in children or adolescents, while in adults, ADHD was associated with a relatively rightward shift of asymmetry compared to controls, i.e. opposite to what might be expected according to Shaw et al. For other cortical asymmetries too, our findings in this large-scale study were discrepant with what might have been expected from previous reports in smaller samples (see references in the Introduction). For example, a prior study reported reversed grey matter volume asymmetry (i.e., leftward instead of rightward) of the superior frontal gyrus in ADHD (Cao et al., 2014), but we saw no clear evidence of this in the present study.

The most often reported alteration of brain asymmetry in ADHD has involved the caudate nucleus, although the direction of the effect has not been consistent (Castellanos et al., 1996; Dang et al., 2016; Filipek et al., 1997; Hynd et al., 1993; Schrimsher et al., 2002; Uhlikova et al., 2007). We did not find evidence for altered asymmetry of caudate nucleus volume in the present study, again suggesting that prior findings were false positives in smaller samples. As mentioned above, we found a tentative association with ADHD for another regional asymmetry of the basal ganglia, namely of the globus pallidus, in adults-only. The globus pallidus is involved in movement and reward processing (Munte et al., 2017), both of which are involved in the symptomatology of ADHD. A previous meta-analysis comprising data from a total of 114 participants with ADHD (or a related disorder) and 143 control participants, noted a significantly lower average right putamen and right globus pallidus volumes in ADHD (Ellison-Wright, Ellison-Wright, & Bullmore, 2008), although asymmetry was not quantified in that study. Regardless, our finding of lower leftward asymmetry seems discrepant with this earlier report.

We have already remarked on the limited statistical power of previous studies as a likely explanation for their findings being discrepant with the current study. Low sample sizes in relation to subtle effects can result in poor reproducibility (Button et al., 2013; Munafo & Flint, 2010). Here, we had 80% power to detect case-control Cohen’s d effect sizes as low as roughly 0.12, or as high as 0.3 in the smallest subset by age (see Methods). In addition to limited sample sizes, there are various other possible explanations for discrepancies with previous studies. Methodological differences in hardware, software, and data processing pipelines can influence results (Biberacher et al., 2016), although our focus on asymmetry through use of the AI is likely to have reduced the impact of heterogeneity factors that affect both hemispheres equally. In contrast to some previous studies mentioned above, we did not consider gyral/sulcal patterns or cortical grey matter volumes as such. Rather, we studied regional cortical thicknesses and surface areas as distinct measures, which together drive grey matter volumetric measures. Since area and thickness have been shown to vary relatively independently (Panizzon et al., 2009), separate analyses are advisable, although cortical thickness measures are particularly prone to effects related to site-, scanner- or protocol differences (Chung et al., 2017; Fortin et al., 2018). Likewise, the choice of brain atlas can influence results, as each atlas has its own properties that impact brain segmentation (Yaakub et al., 2020). In addition, the approach we used is based on hemisphere-specific definitions of regional anatomy, because each hemisphere has its own atlas, based on its own average distribution of features (Desikan et al., 2006). Correspondence between hemispheres is then achieved at the regional level, based on expert neuroanatomical regional segmentation that was adapted to each hemisphere’s distinct features when constructing the atlas. However, future studies using higher-resolution atlases, or vertex-based analysis using hemispheric co-registration (Kang, Herron, Cate, Yund, & Woods, 2012; Maingault, Tzourio-Mazoyer, Mazoyer, & Crivello, 2016), may identify restricted regions showing stronger associations between ADHD and cortical asymmetry than we report here. Furthermore, for subcortical volume asymmetries, discrepancies between the findings of our study and previous studies could be due to differences in parcellation methods, which can perform with varying accuracy (Guadalupe et al., 2014; Pardoe, Pell, Abbott, & Jackson, 2009; Perlaki et al., 2017).

The conceptualization of laterality can also differ across studies. In terms of AIs, our cortical results are largely in line with a previous report based on measuring grey matter volume asymmetries in 192 participants with ADHD and 508 controls (Douglas et al., 2018), insofar as no FDR-significant results were found (five of those datasets were in common with the present study, see Methods). However, the authors of that study also calculated the unsigned magnitudes of the AIs (i.e., absolute degrees of asymmetry, regardless of directions). They reported significant differences in absolute asymmetry for various cortical and subcortical structures (Douglas et al., 2018). In the present study, we did not calculate absolute AIs, in order not to compound multiple testing, and because these measures are highly non-normal with a floor effect at value zero, which would violate the assumptions of the modelling that we applied. It is not clear whether this issue may have affected the results in the earlier study (Douglas et al., 2018). Future studies may consider the unsigned magnitude of brain asymmetry indexes further in ADHD, but it will be necessary to use statistical methods that can account for non-normal distributions.

Discrepancies with earlier studies may also be due to differences in clinical features of the disorder that arise from case recruitment and diagnosis, for example with respect to medication use (which has been suggested to partly normalize brain structural abnormalities, although the previous ENIGMA studies of bilateral changes in ADHD did not support this) (Nakao, Radua, Rubia, & Mataix-Cols, 2011; Pretus et al., 2017), comorbidities (Reale et al., 2017), symptom severity, and/or IQ. Some asymmetries showed tentative associations with some of these clinical variables in the present study, although none of these results survived correction for multiple testing, apart from mood disorder with the rostral middle frontal thickness AI. Also, some of the clinical variables (medication, comorbidity) were missing for many ADHD individuals in this study. Regardless, it remains possible that certain subsets of ADHD might be associated more strongly with altered brain asymmetry than was apparent in our large-scale analysis of average changes over many datasets, comprising many and varied collections of ADHD individuals and controls.

In general, between-centre heterogeneity (in terms of scanning setup, patient subgroups, demographics) may result in reduced statistical power to detect effects that are specific to certain subgroups of datasets, or to individual datasets, when tested in mega-analysis over all datasets. For example, harmonization of scanning protocols might lead to stronger effects being found, as heterogeneity of this aspect would be reduced. Here we used random-intercept models to adjust for heterogeneity between datasets. This was a strong correction for cross-dataset heterogeneity, as it removed mean differences between datasets, although between-dataset heterogeneity that affected model coefficients within datasets would not be fully accounted by this approach. While the random-intercept model cannot fully rescue power in the case that effects are truly specific to certain subsets, no single centre has been able to collect such a large ADHD-control sample alone. Our large sample size yields accurate estimates of effect sizes with respect to the overall case-control population, as represented across many research centres. In this way, the findings from multi-centre studies such as ours can be considered more generalizable than single-centre studies (Costafreda, 2009). In any case, as long as researchers publish separate papers based on many single, smaller datasets, collected in particular ways, the field overall has the same issue of heterogeneity. Of note, the ENIGMA consortium previously showed that using the random intercept approach to account for dataset heterogeneity is similar to random effect meta-analysis across datasets, but preferable because it produces lower standard errors and narrower confidence intervals than meta-analysis (Boedhoe et al., 2018).

Although not a longitudinal study, our data spanned a wide age range from childhood through to older adulthood, which allowed us to study different age groups separately, as the disorder may be neurobiologically distinct in different age groups (Alexander & Farrelly, 2018; Hoogman et al., 2019). The previous ENIGMA study of bilateral cortical differences in ADHD found children to be most affected, particularly in frontal, cingulate, and temporal regions, as well as the total hemispheric surface area, which was lower in ADHD (Hoogman et al., 2019). In the children-only analysis in our present study of asymmetries, we also found associations with diagnosis for some frontal and temporal regions (including the caudal middle frontal cortex thickness, precentral gyrus thickness, medial orbitofrontal cortex surface area, banks of the superior temporal sulcus thickness), as well as a change in the asymmetry of total hemispheric surface area, driven by a greater decrease of area in ADHD on the right-side than the left-side. These findings offer a more nuanced description of brain changes in childhood ADHD, which may involve altered lateralized neurodevelopment.

However, considering all brain asymmetry measures, the effect sizes in the present study were not stronger in children as compared to adolescents or adults. Furthermore, bilateral case-control differences are not necessarily a good guide to case-control differences in asymmetry, since a difference in asymmetry can arise, for example, from a simultaneous left-sided increase and right-sided decrease in a brain measure, which can involve no change at all in the bilateral measure. Hence, we took a screening approach to the present study, rather than constraining our search on prior bilateral findings. It is also not entirely clear how/whether to statistically adjust the test for total hemispheric surface asymmetry, in the context of also testing multiple sub-regions, and also with respect to study-wide multiple testing. Therefore, we present all P-values unadjusted, while also being mindful of the tentative nature of these findings in the context of our survey across many brain asymmetry measures. Together with the corresponding effect size estimates, this mapping information should be useful for the field.

We did not include handedness, IQ or brain size as covariates in our primary analysis, in order to avoid possible collider bias (Cole et al., 2010), as there are various plausible causal relations linking these traits with ADHD, brain asymmetry and other, possibly underlying factors shared between some or all of them. For example having the disorder for other underlying reasons may lead to altered asymmetry and brain size, or altered asymmetry and brain size may contribute to having the disorder. A priori, altered asymmetry may not be associated with the disorder, but be associated with brain size, which can be associated with disorder. In this latter case, correcting for brain size can induce spurious associations between asymmetry and disorder. Collider bias is under-appreciated in the field, perhaps because it is not intuitive. Alternatively, including brain size as a covariate in case-control analysis might have reduced the power to detect an association of diagnosis with asymmetry. This would occur if underlying susceptibility factors contribute both to altered asymmetry and reduced brain size, as part of the ADHD phenotype. Regardless, our primary interest was to detect associations of diagnosis with asymmetry regardless of other brain features such as overall size. We have made available, in supplementary csv files, the results from secondary analyses in which we included handedness, IQ or intracranial volume as covariates, with or without interaction terms with case-control status. As regards handedness specifically, previous studies of subcortical and cortical anatomical asymmetry in over 15,000 subjects from healthy control and population datasets, also performed by the ENIGMA consortium (Guadalupe et al., 2016; Kong et al., 2018), found no significant effects of handedness.

Our study was limited to macro-anatomical asymmetries of cortical grey matter and subcortical volumes. It is possible that altered brain asymmetry in ADHD will be more apparent in different structural or functional modalities, or at different scales. For example, cortical thickness measures can correlate with the degree of myelination (Natu et al., 2019), such that quantitative neuroimaging methods that are sensitive to microstructural tissue content may reveal further alterations in ADHD. At a larger scale, asymmetries of white matter tracts (Wu et al., 2020) may also benefit from the large-scale approach that we have used here. Asymmetries of functional asymmetry, particularly linked to attentional tasks, may also reveal stronger case-control differences than the structural effects we observed (see Introduction).

Conclusion

We carried out the largest case-control study of structural brain asymmetry in ADHD. We described average changes of asymmetry that are small, but helpful towards a more complete description of brain anatomical changes in this disorder. Results were largely discrepant with earlier, inconsistent findings from smaller-scale studies, which illustrates the value of taking a large-scale approach to human clinical neuroscience. The small effects that we found remain statistically tentative in the context of multiple testing, even in this unprecedented sample size. Future longitudinal and genetic studies may probe causative relations between ADHD and brain asymmetry, focused on measures defined from this study, such as total hemispheric surface area asymmetry, medial orbitofrontal area asymmetry, or globus pallidus volume asymmetry.

Key points and relevance.

• The extent to which brain anatomical asymmetry is altered in ADHD has remained unclear. Previous studies of brain asymmetry in ADHD were based on small sample sizes, so that findings may have been unreliable.

• We carried out the largest-ever study of brain anatomical asymmetry in ADHD. Average case-control differences of asymmetry were very small, and the regions implicated were largely discrepant with earlier findings based on smaller samples.

• This study illustrates the value of a large-scale approach to human clinical neuroscience. The findings provide an improved description of brain anatomical changes in ADHD.

• Of itself, altered anatomical asymmetry is not likely to be a useful biomarker for ADHD. Future longitudinal and genetic studies may probe causative relations between ADHD and asymmetry of the total hemispheric surface area, medial orbitofrontal area, and globus pallidus volume.