1 INTRODUCTION

The Global Burden of Disease Studies (Ezzati, Lopez, Rodgers, & Murray, 2004) have been critical in developing methods to study health outcomes and have proven invaluable when advocating for health equity, both cross-nationally and across diseases. The 2010 update (Lozano, Naghavi, Foreman, et al., 2012) systematically quantified prevalence of 1,160 sequelae of 289 diseases and injuries across 21 geographical regions. Results for specific diseases and impairments have highlighted the high rates of disability from mental disorders (particularly depression) and substance use disorders (Whiteford et al., 2013). Alcohol and cannabis are among the most widely abused substances globally, and second only to tobacco use in terms of frequency of use (Degenhardt et al., 2013; Ritchie & Roser, 2019).

Harmful alcohol use and dependence have long been associated with cognitive impairments in multiple neuropsychological domains, including evidence from the very earliest case–control studies employing standardised test batteries, such as the Luri-Nebraska Battery or the Wechsler Adult Intelligence Scale (Chmielewski & Golden, 1980; Miller & Orr, 1980). However, a long-standing challenge in these and successive studies is the multifactorial aetiology of such impairments, including pre-existing variation, foetal alcohol effects, and both direct alcohol-induced excitotoxicity in the brain and indirect toxicity related to factors such as impaired nutrition (e.g. thiamine deficiency; Joyce, 1994), head injury, liver disease, psychiatric comorbidity, and complex interactions between these factors (Tarter & Alterman, 1984). Cognitive deficits have also been recognised in social drinkers (Parsons & Nixon, 1998), with the suggestion that there is a continuum of deficits related to intensity of alcohol use, including impairments in verbal and non-verbal performance, learning, memory, abstract reasoning, and speed of information processing and efficiency (Parsons, 1998). Finally, recent studies have focused on more selective and aetiologically relevant impairments in emotion and reward processing (Kornreich et al., 2001; Townshend & Duka, 2003).

Multiple structural imaging studies have shown generalised cortical atrophy particularly for measures of grey (Fein et al., 2002) and white matter integrity (Gallucci et al., 1989) in alcohol dependence (Sullivan & Pfefferbaum, 2005). Other reports of brain-related abnormalities include the disruption of white matter tracts (Pfefferbaum & Sullivan, 2005) and abnormal functional activity (Rosenbloom, Sullivan, & Pfefferbaum, 2003). Functional imaging studies in alcohol dependence have identified lower cerebral metabolism in frontal brain regions that show a correlation with executive neuropsychological deficits (Wang et al., 1993). Studies in the alcohol-related Wernicke–Korsakoff Syndrome have shown prominent fronto-striatal impairment (Reed et al., 2003).

Recent attempts to reconcile inconsistent findings across neuroimaging studies with alcohol dependent patients have led to the proposal that, despite alcohol's widespread acute effects on brain function, brain deficits due to chronic and severe alcohol consumption might be limited to fronto-cerebellar and meso-cortico-limbic circuitry, while other brain circuits might be spared or might even undergo compensatory changes (Chanraud, Pitel, Müller-Oehring, Pfefferbaum, & Sullivan, 2012). However, prior studies may simply not have been sufficiently powered to reliably detect brain-related impairments and compensatory processes due to small sample sizes, variability across neuroimaging methods, and failure to distinguish disease-specific neurological abnormalities. Standard meta-analysis across neuroimaging studies has proven difficult, mainly due to heterogeneity in the methods used to collect, quality control and parcellate brain data. These limitations also apply to any potential comparisons with other psychiatric conditions. Structural brain abnormalities have also been observed in adults with heavy or problematic cannabis use (Batalla, Bhattacharyya, Yuecel, et al., 2013). Heavy cannabis users have been shown to have lower grey matter density in the right parahippocampus and greater grey matter density in the precentral gyrus and right thalamus (Matochik, Eldreth, Cadet, & Bolla, 2005), and anterior cerebellum (Cousijn et al., 2012). Other studies reported cannabis use and misuse associated with bilateral volumetric reductions in the hippocampus (Ashtari et al., 2011; Matochik et al., 2005; Yücel et al., 2008) and amygdala (Cousijn et al., 2012; Yücel et al., 2008). However, findings have not always been replicated, so further studies are needed to confirm these observations (Nader & Sanchez, 2018).

One approach used by researchers to see beyond inconsistent findings in human substance dependence is the meta-analysis. Recent meta-analyses on alcohol use disorder (AUD) observed grey matter abnormalities in corticostriatal-limbic circuits, such as prefrontal cortical regions, thalamus, striatum and hippocampus (Klaming et al., 2019; Xiao et al., 2015; Yang, Tian, Zhang, et al., 2016). Even if several meta-analyses were performed to determine general impacts of a single drug on the adult brain, to our knowledge no meta-analyses compared two widely consumed drugs to other major psychiatric conditions to evaluate the relative volumetric variations observed between cases and controls among psychiatric disorders. Similarly, a recent meta-analysis observed that the hippocampus and the orbitofrontal cortex were most consistently identified as having structural alterations in regular cannabis users (Lorenzetti, Chye, Silva, Solowij, & Roberts, 2019a).

1.1 Quantifying brain anomalies across disorders

There is a recent trend in psychiatric epidemiology to create standardised metrics for the purpose of cross-jurisdiction and cross-disorder comparisons (Murray, Barber, Foreman, et al., 2015; Whiteford et al., 2013) However, until recently, such harmonised approaches have not been used in the field of psychiatric neuroimaging. The development of such methods would help to quantify the subtle and specific brain-related correlates of major psychiatric conditions, which may contribute to, or reflect, an individual's specific symptom profile, general quality of life, and disability. More objective measures of brain impairment may help to identify disorder-specific processes and quantify brain-related impairment and may also help to reduce the effect of social stigma when evaluating disease-related impairment, or when addressing gaps in access to services. In the case of substance use disorders, brain structural measures have been linked to the duration and severity of the disorder, as well as likelihood of relapse (Zahr, 2014). However, research on other psychiatric conditions, such as mood disorders, psychosis, and attention-deficit/hyperactivity disorder (ADHD) has reported abnormalities in similar brain structures (Hoogman et al., 2017; Schmaal, Hibar, Sämann, et al., 2017; van Erp, Hibar, Rasmussen, et al., 2016). Despite potential similarities in brain-related outcomes, public health policies on how these conditions are managed dramatically differ across cultures and across disorders. AUDs are detrimentally under-treated in most parts of the world, relative to other psychiatric conditions (Kohn, Saxena, Levav, & Saraceno, 2004). However, it is not clear if the treatment gaps observed across disorders can be justified based on objective indicators of impairment.

The current study aims to capitalise on the methods developed in the Enhancing Neuroimaging Genetics through Meta-Analysis (ENIGMA) consortium to use standardised protocols for quantifying and comparing the brain-related correlates of various psychiatric conditions for the purpose of promoting health equity. Using standardised protocols to harmonise neuroimaging data and conduct meta-analyses for cross-disorder comparisons (Hibar, Stein, Renteria, et al., 2015; Stein, Medland, Vasquez, et al., 2012; Thompson, Stein, Medland, et al., 2014), we report, for the first time, effect sizes derived from meta-analysis involving alcohol dependent and cannabis dependent cases and age- and sex-matched controls worldwide. Results of this analysis will be compared to effects found in large-scale international studies of major depressive disorder (MDD), schizophrenia (SCZ), bipolar disorder (BD) and ADHD using identical brain parcellation methods and co-variates (Hoogman et al., 2017; Schmaal et al., 2017; van Erp et al., 2016).

The ENIGMA network (Thompson et al., 2014) was formed to address the need for replicability and increased sample sizes and to increase power for genome-wide association studies of brain measures. With the development of standard anatomical templates and coordinate-based reference systems, researchers worldwide can now relate their new findings to previous results in a consistent way. The pooling of datasets across sites and clinical samples now allows us to study uncommon or complex phenomena and compare findings across disorders. The most commonly used statistical approach in ENIGMA is a meta-analysis, in which evidence for association is combined using effect sizes for each separate site, which are weighted to adjust for each site's sample size and error variance.

Several disease-specific working groups have been formed, focusing on performing meta-analysis of case–control disease differences of measures extracted using the ENIGMA protocols. This approach also allows for the unification of both case–control and cohort datasets with a standard protocol and allows for the largest imaging studies of the human brain to be performed while focusing on a particular disease process. To date, ENIGMA working groups such as MDD, BD and SCZ have reported small to moderate effect sizes in terms of brain-related abnormalities (Hoogman et al., 2017; Schmaal et al., 2017; van Erp et al., 2016), such as alterations in hippocampus (except for ADHD), amygdala (except for MDD) and thalamus (except for ADHD and MDD).

The combined datasets contain participants with structural brain data and addiction phenotyping and involves both cohort and case–control designs, representing individuals with ages ranging from 12–60 years and a number of different addiction phenotypes. For the purpose of the current study, we omit cohort studies and focus on case–control samples representing AUD and cannabis use disorder (CUD) since these substances are among the most consumed worldwide (Degenhardt et al., 2013; Peacock, Leung, Larney, et al., 2018; Ritchie & Roser, 2019). To be able to compare effect sizes for addiction subgroups to published results from other ENIGMA disease groups, we focused our first set of analyses on structural measures of subcortical and cortical brain regions (Hoogman et al., 2017; Schmaal et al., 2017; van Erp et al., 2016).

ENIGMA provides quality control procedures for harmonising neuroimaging and genetic data, available online through http://enigma.loni.ucla.edu/ongoing/gwasma-of-subcortical-structures. While many working groups within ENIGMA are now moving towards the ‘mega-analysis’ strategy, where all phenotypic and genotypic data are sent to a centralised site for pooling and analysis, the first ENIGMA studies used a ‘meta-analysis’ approach. Meta-analyses circumvent barriers associated with data sharing across sites and countries and allow sites to maintain responsibility for their data and its integrity. This approach also has advantages when the purpose of the study is to make cross-disorder comparisons. ENIGMA-Addiction recently evaluated the subject-specific volumetric variations in drug-specific groups using innovative methods (Chye, Mackey, Gutman, et al., 2019; Mackey et al., 2018). These approaches observed subcortical and cortical variations in alcohol-dependent participants but not in cannabis-dependent participants, which is interesting given the fact that a recent meta-analysis observed an association between cannabis use and reduced volumes in subcortical regions and thinning of cortical thickness in the orbitofrontal cortex (Lorenzetti, Chye, Silva, Solowij, & Roberts, 2019b). Similar variations were observed in ADHD, MDD, SCZ and BD but have never been compared to drug-specific variations using common metrics.

2 SUBJECTS AND METHODS 2.1 Samples

The ENIGMA-Addiction Working Group includes international samples with neuroimaging and clinical data from substance dependent patients and healthy controls. This is an ongoing study and new research groups continue to join the consortium regularly. Inclusion criteria for study enrolment were that sites must agree to their data being processed using the ENIGMA scripts and basic information on dependence criteria and patterns of early use are available for cases and controls. Demographic details for seven international samples of alcohol-dependent subgroups and seven international samples of cannabis use are presented in Table 1. The number of participants included vary from the sample of the mega-analysis performed by Mackey et al. (2018) because sites that did not include a control group could not derive a comparable effect size. Most participants were classified as having substance use disorder, or not, using validated structured diagnostic interviews that conform to Diagnostic and Statistical Manual of Mental Disorders-Fourth-TR Edition (American Psychiatric Association, 1994) criteria. One site included chronic cannabis users for whom DSM criteria could not be confirmed (Hester, Nestor, & Garavan, 2009). Similarly to the mega-analysis of the ENIGMA Addiction working group, subjects with a lifetime history of neurological disease and/or a current DSM-IV axis I diagnosis (other than depressive and anxiety disorders) were excluded from the analyses. AUD and CUD cases were mostly lifetime dependence cases (Mackey et al., 2018). Participants with a co-occurring substance use disorder were removed from the analyses. All control participants were confirmed with similar interviews to be free of any substance use disorder. All participating sites obtained approval from local institutional review boards and ethics committees. All study participants provided written informed consent at their local institution for the local study. CHU Ste Justine provided ethical approval for this meta-analysis. This study includes 435 cases with primary AUD and 363 matched healthy controls; and 200 cases of CUD and 247 controls.

TABLE 1. Demographic details for each site Substance of dependence Number of studies Groups N Female Age All 14 Case 635 188 28.12 Control 610 208 29.23 Alcohol 7 Case 435 127 32.43 Control 363 136 33.58 IRC (Sinha & Li, 2007, Li et al., 2009, Seo et al., 2011) Case 43 11 28.05 Control 84 21 37.49 Effects of heavy alcohol abuse on adolescent brain structure and function (Fein et al., 2013) Case 60 34 14.81 Control 56 31 14.94 NIAAA (Senatorov et al., 2015, Grodin et al., 2013, Momenan et al., 2012) Case 212 57 31.11 Control 140 67 38.48 Neuro-ADAPT (Korucuoglu et al., 2017) Case 18 6 19.35 Control 23 11 18.72 NESDA-AD (Sjoerds et al., 2014) Case 42 19 48.6 Control 20 6 48.29 ADPG study (Jansen et al., 2015, van Holst et al., ,2014) ) Case 28 0 43.43 Control 24 0 37.17 TrIP study (Schmaal et al., 2014) Case 32 0 41.69 Control 16 0 39.94 Cannabis 7 Case 200 61 23.80 Control 247 72 24.88 Trinity-THC (Hester et al., 2009) Case 15 2 23.27 Control 15 4 22.4 Orr (Orr et al., 2013) Case 13 1 16.00 Control 14 1 16.77 Cannabis prospective (Cousijn et al., 2012, 2013, 2014) Case 38 12 21.85 Control 40 15 21.39 ADS Case 7 6 18.96 Control 93 44 19.00 Chronic cannabis users (Barcelona) (Batalla et al., 2014, Blanco-Hinojoet al., 2017, Pujol et al., 2014, ) Case 16 1 35.00 Control 18 2 38.98 Chronic cannabis (Yücelet al., 2008, Solowij et al., 2011, 2013, Lorenzetti et al., 2015) Case 81 39 30.47 Control 38 6 33.21 Chronic cannabis-memory (Zalesky et al., 2012, Harding et al., 2012, Jakabek et al., 2016, Yücel et al., 2016) Case 30 0 21.03 Control 29 0 22.41 2.2 Image processing and analysis

Structural T1-weighted magnetic resonance imaging brain scans were acquired at each site. Using the fully automated and validated segmentation software FreeSurfer (Fischl et al., 2002), the segmentations of seven subcortical grey matter regions (nucleus accumbens, amygdala, caudate, hippocampus, pallidum, putamen and thalamus), 34 cortical regions, lateral ventricles, and total intracranial volume (ICV) were derived following standardised protocols designed to facilitate harmonised image analysis across multiple sites (http://enigma.ini.usc.edu/protocols/imaging-protocols). Image acquisition parameters and software descriptions for each sample are similar to those in previous ENIGMA studies (Hoogman et al., 2017; Schmaal et al., 2017; van Erp et al., 2016), to facilitate the between-disorder comparison. A majority of the datasets were prepared using CBRAIN, a network of high-performance computing facilities in Canada (Sherif, Rioux, Rousseau, et al., 2014). Sites followed the ENIGMA protocols for quality control http://enigma.ini.usc.edu/protocols/imaging-protocols/. The detection of outliers and visual inspection were performed in a series of standard planes to avoid the inclusion of poorly segmented and mislabelled structures. Quality control procedures at each site were conducted according to standardised protocols to minimise potential site effects. Additional visual inspection was performed on a randomly selected sub-sample of participants centrally at the University of Vermont to ensure uniformity of quality control across sites.

2.3 Statistical framework of meta-analysis

Consistent with other ENIGMA working groups, multiple linear regression analyses derived case versus control group differences within each site-specific sample using the mean volume of bilateral subcortical region of interests (ROIs) ([left + right]/2) along with left and right thickness and surface area for each cortical ROI as the outcome measures and control/case as the binary grouping independent variable. In order to make results comparable with those of yielded from the SCZ, MDD, BD, and ADHD working groups, models for subcortical ROIs covary for age, sex and total ICV, and models for cortical ROIs covary for age and sex. In order to be consistent with the reported findings of the other working groups, we did not covary for past 30-day alcohol and nicotine use in the analyses. Furthermore, because past 30-day substance use is so highly correlated with dependence scores, co-varying for this variable in meta-analysis would likely lead to underestimation of main differences between cases and controls. Regression models were fit for each site separately and t-statistics were used to estimate effect sizes. A Cohen's d-effect size estimate was obtained using an inverse variance-weighted random-effect meta-analysis model in R (metafor package; Viechtbauer, 2010). Uncorrected and false discovery rate (FDR) corrected p values are reported and, are indicated as significant effect sizes by an asterisk in figures below. I-square indices were calculated to provide a measure of heterogeneity. Significance level was determined with a pFDR < .05 for all regions of interest. Differences between effect sizes were considered significant if confidence intervals (CIs; or SEs) did not overlap, which is appropriate when Cohen's d is derived based on sample sizes above 20 (Lakens, 2013). We performed a post hoc sensitivity analysis on the meta-analytic results for the alcohol and cannabis subgroups excluding the two adolescent sites using a leave-one-out approach (Viechtbauer, 2010). The inclusion of adolescent sites in the AUD and CUD analyses might lead to greater variability and inconsistent findings due to the volumetric variations that occur during the adolescence (Lenroot, Gogtay, Greenstein, et al., 2007). Because the CIs for the adolescent sites overlap with the adult-only sites for all regions of interest, it did not significantly differ from the rest of the sample. The Results section includes significant variations when the adolescent sites are included since structural effect sizes remained consistent for subcortical volumes and cortical thickness (Supporting Information Figures 1–4). The Cohen's d values were obtained after adjustment for age, the adolescent sites were included in final AUD and CUD sample sizes in order to increase the sample size.

3 RESULTS

Figure 1 shows the forest plot with effect sizes and 95% CIs for subcortical volumes when comparing AUD cases to their matched controls: FDR corrected significant differences are shown in the thalamus (d = −0.23, CI = [−0.42, −0.04]), putamen (d = −0.27, CI = [−0.45, −0.08]), hippocampus (d = −0.50, CI = [−0.76, −0.24]), amygdala (d = −0.39, CI = [−0.63, −0.16]) and the accumbens (d = −0.30, CI = [−0.49, −0.12]). The caudate (d = −0.04, CI = [−0.22, 0.15]) and the pallidum (d = −0.10, CI = [−0.24, 0.04]) were not significant following the FDR correction.

Forest plot with effect sizes and confidence intervals for bilateral subcortical volume for the alcohol use disorder versus controls comparison controlling for age, sex when females were included, and intracranial volume. Error bars represent 95% confidence intervals. The caudate and the pallidum were not significant following the false discovery rate correction. ROI, region of interest

Figure 2 presents the case control comparisons for CUD, showing non-significant case–control differences after correcting for multiple comparisons.

Forest plot with effect sizes and confidence intervals for bilateral subcortical volume for the cannabis use disorder versus controls comparison controlling for age, sex when females were included, and intracranial volume. Error bars represent 95% confidence intervals. All results are non-significant following false discovery rate correction. ROI, region of interest

Figure 3 presents a comparison of subcortical results for AUD and CUD and previously published ENIGMA meta-analyses on MDD, SCZ, BD and ADHD. If CIs overlap between groups for a ROI, no significant difference was observed. The magnitude of effect sizes for case–control AUD comparisons appear larger than effect sizes reported for MDD, BD, and ADHD except for the caudate and the pallidum. However, these effect sizes do not differ significantly from the other psychiatric conditions since the CIs overlap. In comparison to effects reported for SCZ, CIs overlapped between AUD and SCZ on all subcortical ROIs. When comparing other subcortical regions across AUD and MDD, CIs did not overlap for the putamen, hippocampus, amygdala and accumbens, with AUD associated with significantly smaller volumes in these regions. When comparing AUD and BD, CIs overlap for all regions but the amygdala, with significantly greater differences evident in AUD case control comparison (AUD associated with smaller volume). The effects reported for CUD in the amygdala and accumbens appear comparable to those reported for AUD and SCZ, but due to large CIs, these effects were not shown to significantly differ from non-effect line and the other disorders. CUD observations also overlapped with AUD and SCZ findings. These CUD results suggest considerable heterogeneity or variability in the CUD studies (Supporting Information Figure 5).

Comparison between bilateral subcortical results for alcohol use disorder (AUD), cannabis use disorder (CUD), depression (MDD), psychotic disorder (SCZ), bipolar disorder (BPD) and attention-deficit/hyperactivity disorder (ADHD). Error bars represent 95% confidence intervals. Significant volumetric variations when compared to age-, sex- and disorder-matched controls was observed when the confidence intervals did not overlap with non-effect line at 0 and survived false discovery rate correction. While significant reductions are observed in the ADHD for the putamen, amygdala and the caudate, none of these results remained in ADHD adult-specific analyses

Analysis of cortical thickness by AUD case–control comparisons shows FDR-corrected significant bilateral differences in the caudal anterior cingulate, fusiform, inferior temporal, parahippocampal, posterior cingulate, superior frontal and temporal pole. Table 2 presents effect sizes and CIs for cortical thickness in each ROI for AUD. Table 3 presents effect sizes and CIs for cortical thickness for CUD in each ROI. None of the CUD-control comparisons on cortical thickness were shown to survive FDR correction for multiple testing across all ROIs, but marginal, FDR-corrected effects were revealed for the medial orbitofrontal cortex (pFDR < .1), caudal middle frontal (pFDR = .1), precentral gyrus (pFDR = .1) and insula (pFDR = .1).

TABLE 2. Full meta-analytic results for volume and thickness of each bilateral structure for the alcohol use disorder versus controls comparison controlling for age, sex and intracranial volume (for subcortical regions only) ROI ES SE 95% CI.LB 95% CI.UB I2 p pFDR Controls Cases Thalamus −0.2272 0.0976 −0.4184 −0.0359 28.84 .0199 .0279 359 345 Caudate −0.0382 0.094 −0.2225 0.1461 28.12 .6844 .6844 361 437 Putamen −0.2656 0.0966 −0.455 −0.0762 30.53 .006 .0105 361 436 Pallidum