Associations of cannabis use disorder with cognition, brain structure, and brain function in African Americans

Abstract Although previous studies have highlighted associations of cannabis use with cognition and brain morphometry, critical questions remain with regard to the association between cannabis use and brain structural and functional connectivity. In a cross‐sectional community sample of 205 African Americans (age 18–70) we tested for associations of cannabis use disorder (CUD, n = 57) with multi‐domain cognitive measures and structural, diffusion, and resting state brain‐imaging phenotypes. Post hoc model evidence was computed with Bayes factors (BF) and posterior probabilities of association (PPA) to account for multiple testing. General cognitive functioning, verbal intelligence, verbal memory, working memory, and motor speed were lower in the CUD group compared with non‐users (p < .011; 1.9 < BF < 3,217). CUD was associated with altered functional connectivity in a network comprising the motor‐hand region in the superior parietal gyri and the anterior insula (p < .04). These differences were not explained by alcohol, other drug use, or education. No associations with CUD were observed in cortical thickness, cortical surface area, subcortical or cerebellar volumes (0.12 < BF < 1.5), or graph‐theoretical metrics of resting state connectivity (PPA < 0.01). In a large sample collected irrespective of cannabis used to minimize recruitment bias, we confirm the literature on poorer cognitive functioning in CUD, and an absence of volumetric brain differences between CUD and non‐CUD. We did not find evidence for or against a disruption of structural connectivity, whereas we did find localized resting state functional dysconnectivity in CUD. There was sufficient proof, however, that organization of functional connectivity as determined via graph metrics does not differ between CUD and non‐user group.


| INTRODUCTION
Views on cannabis use are changing in the United States. After legalization of medical cannabis and more recent decriminalization/legalization of recreational use in many states, it is likely that more people will use cannabis (Pearlson, 2020). Despite a recent meta-analysis reporting no link between legalization of medical cannabis and increased use (Sarvet et al., 2018), cannabis use and cannabis use disorder (CUD) are increasing (Compton, Han, Jones, Blanco, & Hughes, 2016;Hasin et al., 2017;Wang, Davies, Halmo, Sass, & Mistry, 2018), whereas the perceived risk of cannabis use is decreasing (Compton et al., 2016;Wen, Hockenberry, & Druss, 2019). A critical issue is that the evidence base regarding the safety of cannabis use is incomplete. Specifically, more information is needed regarding potential risks of long-term cannabis use, particularly on brain function.
The importance of examining the effect of cannabis on brain phenotypes is underscored by evidence that chronic use is associated with lower global neuropsychological function and underperformance in tasks of verbal learning and memory, attention, and psychomotor function (Broyd, van Hell, Beale, Yücel, & Solowij, 2016;Crane, Schuster, Fusar-Poli, & Gonzalez, 2013;. Although this association may be bidirectional (Jackson et al., 2016;Meier et al., 2018) the cognitive effects of cannabis are particularly relevant for adolescents and young adults (Volkow, Baler, Compton, & Weiss, 2014). Furthermore, effects are more prominent within 72 hr of last use with limited evidence that effects persist after longer periods of abstinence (Scott, Slomiak, et al., 2018).
While inconsistencies in the literature could reflect a number of methodological differences, three issues appear to be limiting our understanding of the relationship between cannabis use and brain structure and function. First, many studies use small sample sizes with limited statistical power, which reduces the chance of detecting a true effect and decreases the likelihood that a statistically significant result reflects a true effect (Button et al., 2013). Second, the definition of cannabis use varies dramatically between studies, potentially limiting generalizability. Third, ascertainment strategies can result in subtle biases that can skew results, particularly for case-control studies where cases and controls are drawn from different populations (Delgado-Rodriguez & Llorca, 2004;Kopec & Esdaile, 1990). To overcome these limitations, we studied the association of CUD on cognitive and neuroimaging measures in a relatively large sample (n = 205) of individuals recruited without regard to their cannabis use status and used DSM criteria to define CUD. Additionally, the use of multimodal imaging and detailed cognitive assessment enables the simultaneous investigation of the effect of CUD on multiple brain and behavioral variables in the same individuals.
We hypothesized that CUD would be associated with abnormal structural and functional connectivity involving brain regions that are rich in cannabis receptors, including cerebellum, subcortical structures, cingulate, and frontal areas (Burns et al., 2007;Glass, Faull, & Dragunow, 1997). We also expected to replicate prior findings on poorer cognitive function in verbal tasks and negative findings regarding the associations between cannabis use and brain morphometry.

| Participants
Participants (aged 18-70 years) were recruited via flyers, advertisements in local newspapers, and Craigslist as the control group of an ongoing study into psychosis (Mathias et al., 2017). To maintain generalizability, participants were allowed to have common psychiatric disorders (except for psychosis) and rates of psychiatric disorders in our sample reflect national averages in this population (Tables 1 and S1a-d). All participants underwent formal diagnostic assessments using the Structured Clinical interview for DSM-IV (First, Spitzer, Gibbon, & Williams, 2002) (see Supporting Information). Exclusion criteria included a history of major medical disorders, severe head injury, MRI contraindication, IQ < 70, dementia, traces of drugs other than THC in urine, and drug intoxication during cognitive or MRI assessment.
Participants were diagnosed with CUD (n = 57) if they (a) met DSM-IV criteria for current or partially remitted CUD (n = 7 for abuse, n = 23 for current dependence, n = 11 for partial remission); or (b) met criteria for remitted CUD for >1 year but had positive THC urine test (n = 9) or mentioned they used socially (n = 6). Individuals who did not use were defined as non-CUD (n = 148). Non-CUD participants were excluded if they tested positive for THC (n = 13) or mentioned that they used occasionally (n = 18) or use was unknown (n = 1). Figure S1 provides an overview of final sample size after quality control for each modality. Since (supervised) abstinence of 3 to 28 days normalizes functional connectivity (Blanco-Hinojo et al., 2017;Jacobus et al., 2012), cannabinoid receptor density (D'Souza et al., 2016;Hirvonen et al., 2012), and cognitive performance (Scott, Slomiak, et al., 2018), individuals in full remission were classified as past-CUD and placed in a separate group. This group was too small (n = 22) for analyses and was excluded from the current study. The protocol was approved by institutional review boards at Hartford Hospital and Yale University. All participants provided written informed consent.

| Cognitive assessment
Under supervision, participants completed a cognitive test battery ("Charlie," https://github.com/sammosummo/Charlie, (Mathias et al., 2017)), which included fully computerized and computer-aided administration of tests listed in Subjects list as many words as possible in 60 s. In the first three trials, the words must begin with the letters, F, A, and S (verbal fluency). In the fourth trial, the words must be animals (semantic fluency).
Processing speed and executive function Trail making A and B Computerized Trail making: Subjects click on circles presented on the screen in a specified order. The order is either consecutive letters (A), or alternating numbers and letters (B). Output is total time taken to complete the trail.
Processing speed Digit symbol A key of digit symbol pairs is presented at the top of the screen and subjects indicate whether a target digit symbol pair presented at the center of the screen matches any pair from the key. Subjects complete as many trials as possible in 90 s.
Verbal memory CVLT: California verbal learning test Subjects hear an audio recording of 16 words and repeat out loud as many words as they can recall. This is repeated for five trials. Sum correct is the total number of correct recalls over all five trails. On the sixth trial (recall condition), subjects list as many words as they can recall without first hearing the audio recording. In the recognition condition, subjects recognize the 16 target words presented alongside 32 non-target words.
Working memory Span forward and backward Letter number sequencing Subjects hear sequences of letters or numbers that increase in length throughout the trials and repeat these sequences out loud. In the forward condition, subjects repeat the sequences in the order in which they heard them. In the backward condition, subjects repeat the sequences in reverse order. Same as above, but the sequences contain both letters and numbers and subjects repeat the letters in numerical order, followed by the letters in alphabetical order.
Facial memory Penn face memory test Subjects see images of faces and are asked to try to remember them. After they have seen all the faces, they perform a recognitionmemory task. Each trial comprises a face (either an old face or a new one), and subjects make old/new judgments. Direct and delayed correct responses were averaged.
Emotional memory Penn emotion recognition test Subjects see a color image of a face expressing an emotion (happy, sad, fearful, and angry) or with a neutral expression. Subjects make their responses by clicking on the words printed to the screen. There is no feedback and there are no practice trials.

Spatial memory
Corsi SCAP: Spatial capacity delayed response test Subjects observe the sequence of blocks in which circles appear, and then repeat the sequence back in order. The task starts with a small number of blocks and gradually increases in length up to nine blocks. Three different conditions (clicking the block where circles appeared (order irrelevant), clicking in order (similar to the original Corsi), clicking blocks in order where circles appeared, then click in order the blocks were crosses appeared) were averaged to obtain one score. On each trial, the subject sees a study array comprising three to five yellow circles in random positions on the screen. The study array is removed and, after a delay, is replaced by a single green circle (the probe). The subject indicates whether the probe has the same spatial location as one of the original circles. In this version of the SCAP, there are 14 three-, 14 four-and 14 five-item trials.

| Neuroimaging
Imaging was conducted on a Siemens Skyra 3 T scanner at the Olin Tract-Based Spatial Statistics (TBSS) (Smith et al., 2006). Average FA was calculated for the whole skeleton and 20 white matter tracts based on the John's Hopkins University white matter atlas (Hua et al., 2008). Two motion estimates calculated during the DTI distortion correction were used as covariates. Final group sizes were 110 non-CUD, 42 CUD.
Preprocessing of resting-state data was based on HCP pipelines . Framewise displacement (FD) and root-meansquare change in blood-oxygen-level-dependent (BOLD) signal from volume to volume (DVARS) was calculated and timepoints with FD > 0.5 mm were removed. Runs with >20% high motion time points were discarded. Most subjects (67%) had all 4 runs (in both groups).
Final group sizes were 92 non-CUD, 39 CUD. See Supporting Information for details on preprocessing.
The cortex was parcellated according to the Gordon atlas (Gordon et al., 2016), subcortical regions were segmented with the Harvard-Oxford atlas (Desikan et al., 2006), and cerebellum with FSL's probabilistic atlas (Diedrichsen, Balsters, Flavell, Cussans, & Ramnani, 2009). Total number of brain parcels was 382. Global signal regression was applied (based on all voxels) and pair-wise correlations between all 382 regions were Fisher r-to-z transformed. The correlation matrix was checked for outliers. First, edge-wise outliers (>4 SD) were detected. Second, for each subject the number of outlier-edges was calculated. Based on the histogram of the number of outlieredges per subject, five subjects (two CUD, three non-CUD) where defined as outliers (>1,000 outlier-edges). These subjects did not deviate in their mean connectivity. Analyses were repeated without these subjects.

| Organization of functional connectivity
The organizational structure of resting-state functional connectivity Graph-theoretical metrics were estimated using the Brain Connectivity Toolbox (BCT, www.brain-connectivity-toolbox.net). We focused on metrics of (a) global connectivity (global efficiency); (b) regional connectivity (degree, strength, and clustering coefficient); and (c) modular organization (participation coefficient, modularity).
Because graph metrics are sensitive to the density of the graph (i.e., the fraction of present connections to all possible connections), we selected the same number of edges for each participant, namely Motor speed Orientation test Subject sees either a blue square (during the first 10 trials) or a blue square and a red circle (during the last 10 trials) positioned randomly on the screen. The task is to click on the blue square as quickly as possible. After each trial the blue square becomes smaller and the red square becomes larger. It is similar to the mouse practice task from Gur et al., 2001. Output is total time taken to complete the test.
the top 1 to 40% (increments of 1%) of the strongest positive or negative edges. This upper bound was based on the density of positive versus negative edges in each participant, thus ensuring that the top 40% of positive edges included the same number of positive only edges for each participant (and likewise for top 40% negative edges).
Graphs based on positive and negative correlations were analyzed separately for global efficiency, clustering coefficient, and strength.
Modularity and participation coefficient were computed on the union of positive-and negative-correlation based graphs (thus, graph density for these was twice that of other metrics). A Louvain approach with negative asymmetry was used to estimate modularity (Rubinov & Sporns, 2011).
Graph-theoretical metrics were computed for each density threshold (1-40%), thus forming a curve as a function of density. The difference between the groups was computed as the area between their respective curves. Permutations (n = 50,000) of group membership were used to assess significance. When regional output was created (clustering coefficient, strength, degree, participation coefficient), p-values were corrected for multiple testing (382 regions) with an FDR-correction. Effect sizes and Bayes factors were computed over the mean over all densities.

| Statistical analyses
Covariates common to all analyses were age, age 2 , sex, and their interactions. Cognitive, morphometry, and DTI data were visually checked for normality (density and Q-Q plots) and transformed if necessary (only for cognitive data, with either data e ; data −1 ; log[data], depending on the distribution). In the structural data, outliers (>4 SD) were removed as they likely reflect segmentation errors. Differences between groups were assessed via t tests, and multiple regression (glm function) in R (R Core Team, 2017). The threshold for statistical significance was set at p < .05 after FDR correction. Bayes factors (BF) (Jarosz & Wiley, 2014;Jeffreys, 1961) were computed on imputed data with BayesFactor package in R (Morey & Rouder, 2018) against a null-interval of (−0.1, 0.1) (Morey & Rouder, 2011). BF > 3 is considered substantial, BF > 10 is strong evidence for the alternative, and BF < 0.3 is substantial evidence for the null (Jeffreys, 1961 Post hoc analyses examined the relationship of (log-transformed) duration of cannabis use and age of onset of cannabis use on measures that significantly differed between groups. We tested for the influence of confounding variables by regressing out (in separate analyses) the effects of number of alcoholic drinks/month, lifetime diagnosis of alcohol use disorder (AUD), lifetime diagnosis of substance use disorder other than cannabis or nicotine (SUD), and years of education approximated as an ordinal variable (see Supporting Information), and social disability status as a proxy for social economic status (although this did not differ between groups, p = .292). Because correlations between nicotine use and brain structure (Liao, Tang, Liu, Chen, & Hao, 2012) and function (Filbey et al., 2018) have been reported, we also regressed out the effects of nicotine dependence (Fagerström test) for any significant brain metrics.

| Cognition
Seven tests showed FDR-corrected group differences in the domains of general cognitive ability, verbal intelligence, verbal memory, working memory and motor speed (Cohen's d = −0.39 to −0.80, p = <.011, BF > =1.9, Figure 1, Table 3). As cognitive test-scores are often correlated ( Figure S2), we performed a multiple regression analysis including each of the test scores (excluding g; variance inflation factors <2.67), to determine if a single measure was differentially associated with CUD status. Only the total sum of correctly recalled words significantly differed between the CUD and non-CUD group (p = .0006) in this analysis.

| Brain morphometry
No statistically significant neuroanatomical differences were observed in subcortical volumes, cerebellar volume, cortical thickness or cortical surface-area (Cohen's d = −0.40 to 0.37; p > .026 uncorrected, BF < 1.5; Figure 1, Tables S2 and S3). BF was <0.3 for the majority of the regions, including many frontal and subcortical regions, providing support for the null hypothesis (Jeffreys, 1961). Although findings were inconclusive (0.3 < BF < 1.5) for 22 regions, these regions did not differ between the groups when correcting for the number of tests (PPA < 0.14 when BF < 1.5 and prior probability <0.1).

| Diffusion imaging
Mean FA of several bundles was lower the CUD group (Cohen's d < −0.41, p < .05 uncorrected; Figure 1;  Figure S4). This was also the only edge that had sufficient evidence for a group difference (PPA > 0.95).

| Graph-theoretical metrics
There were no group differences in global or regional metrics. There were no group differences in mean FC between resting state modules; mean FC strength within module or mean FC strength between each module and all other modules; or mean FC over the entire matrix. See Figure 1 for effect sizes, Table S5 for an overview of marginal findings at p < .005 uncorrected. Although BF > 6, PPA < 0.01 indicates strong evidence that none of the metrics differed between the groups.

| Global signal regression
To test for bias related to the global signal, we reran the FC analyses on data without global signal regression (GSR). There were no group differences in the NBS approach. Preparing the data for the graph  F I G U R E 2 Affected network in CUD at F ≥ 15, based on intensity (a) and extent (b). Red/thick edge, stronger functional connectivity in CUD; Blue/thin edge, weaker functional connectivity in CUD. SM, somatomotor cortex; Attn, attention both positive and negative metrics in the main GSR analyses). There were several group differences at p < .005 uncorrected (Table S6).
However, as in the GSR analyses, none of the findings survived FDR correction and PPA suggested an absence of a group difference (PPA < 0.20 for 24 of 28 variables, the remaining 4 were inconclusive PPA < 0.33). Findings did not overlap with results done on GSR data with the same density range (Table S7).

| Duration of use
Significant results mentioned above were tested for correlation with duration of cannabis use and age of first cannabis use (in CUD group only). None of the cognitive or resting state variables correlated with these metrics, although evidence is inconclusive (0.2 < BF <1.8; see Supporting Information).

| Cross modality correlations
Cognitive and neuroimaging measures found to differ between the CUD and non-CUD group were correlated to examine potential common effects across cognitive and brain measures. None of these correlations survived FDR correction and BFs were inconclusive at best (BF < 9.8, PPA < 0.15; see Supporting Information).

| Confounding variables
Analyses showing significant results were repeated after regressing out the effects of confounding variables to determine if results were influenced by comorbidity.
Group differences in cognition remained significant after adjusting for alcoholic drinks/month and AUD, but four tests did not reach FDR significance after covarying for SUD (Table 3). When correcting for education, only verbal learning reached (FDR-corrected) T A B L E 4 Mean (SD) functional connectivity and effect sizes of edges in the NBS component (intensity, F ≥ 15) and FDR significant edges significance. Correcting for social disability status did not change the results.
The NBS-intensity components were roughly the same when alcoholic drinks/month, AUD, SUD, education, social disability, or DVARS were added as covariates. The connections between the SPG and right insula were the most stable. The NBS-intensity component did not reach significance when corrected for nicotine dependence, and showed a different network when excluding five outliers. The NBS-extent component remained significant only after correction for SUD and social disability. Edges found at FDR significance were stable after controlling for above mentioned confounding variables. See Table 4.

| DISCUSSION
We investigated the association between CUD and measures of cognition, brain structure, and brain function in a large sample selected without regard to cannabis use status. Consistent with prior studies, CUD individuals showed poorer performance on measures of verbal intelligence, verbal memory, working memory, and motor speed. We did not find conclusive evidence for lower FA in the CUD group.
Functional connectivity (FC) was altered in a component that included lower FC between the bilateral superior parietal gyrus (SPG) and right insula. In contrast, there was substantial evidence that neither neuroanatomical measures nor did graph-theoretical metrics of FC differed between groups. Results were not related to alcohol drinking behavior, AUD, or SUD, although nicotine dependence explained a portion of the variance in FC.
Similarly, cognitive impairments (Scott, Slomiak, et al., 2018;Tait, Mackinnon, &Christensen, 2011) andFC (Blanco-Hinojo et al., 2017) are recovered after abstinence, suggesting a causative role. However, longitudinal twin studies show that individuals at risk for substance abuse could have cognitive vulnerabilities before onset of cannabis use (Jackson et al., 2016;Meier et al., 2018;Ross et al., 2019), although there is also evidence from a longitudinal study for cognitive effects after persistent use (Meier et al., 2012).
When correcting for education, many of the cognitive and FA differences between CUD and non-CUD disappeared, except for verbal memory and FC. Verbal memory is one of the most robustly found domains to be affected in relation to cannabis use (Broyd et al., 2016;Crane et al., 2013). Possibly, these metrics may be associated with CUD through a different pathway.
The absence of significant morphometric differences associated with CUD is consistent with previous large studies on recreational cannabis use (Orr et al., 2016;Thayer et al., 2017), with similar effect sizes Weiland et al., 2015). Our findings add to the literature that morphometric differences are absent even when individuals who meet CUD diagnostic criteria are compared with non-users from the same population. However, a few recent large studies did find morphometric differences in association with cannabis use. ;   Although the literature is consistent in reporting lower structural connectivity in participants with CUD (Arnone et al., 2008;Ashtari et al., 2009;Becker et al., 2015;Epstein & Kumra, 2015;Gruber et al., 2014;Jacobus et al., 2009;Jakabek et al., 2016;Manza et al., 2019;Murray et al., 2017;Orr et al., 2016;Shollenbarger et al., 2015;Zalesky et al., 2012), we did not find conclusive evidence for lower FA in the CUD group (0.3 < BF < 3). Functional connectivity studies, however, are more inconsistent. Studies using a whole brain approach to investigate CUD-related resting state FC (Cheng et al., 2014;Filbey et al., 2018;C. Orr et al., 2013;Thijssen et al., 2017) have reported that cannabis users have: higher fractional amplitude of low-frequency fluctuations in right superior frontal gyrus, SPG, semilunar node of the cerebellum, and weaker interhemispheric connectivity in the medial frontal cortex and pyramid of the cerebellum (C. Orr et al., 2013); stronger connectivity between frontal and precentral gyri, and between frontal and cingulate gyri (Cheng et al., 2014); weaker connectivity in the salience network and posterior cingulate gyrus (Filbey et al., 2018). The largest study (130 CUD, 47 controls) found no cannabis-use associated differences in intra-network connectivity but did report a negative association between duration of cannabis use and connectivity of the executive control network with the auditory and a sensorimotor network in cannabis users (Thijssen et al., 2017). We did not find differences in internetwork connectivity of data processed with global signal regression (GSR); in non-GSR data there could be (evidence was inconclusive) stronger inter-network FC between the cingulo-parietal (CP) and auditory network, and between the CP and somato-motor network in CUD. These inconsistent results, emphasized by the difference between results with and without GSR, illustrate the need for more standardized methodological approaches and resist drawing generalized conclusions.
We report weaker FC between bilateral motor-hand regions in the SPG and the bilateral insula in the NBS-intensity and FDR approach, even after correcting for several confounders. In contrast, the NBS-extent finding of stronger FC between nodes in the occipital cortex reached only marginal significance and did not survive correction for confounding variables. Although without GSR there were no group differences in NBS networks, the global signal (as inferred from mean FC over the entire matrix) was higher in the CUD group. Our results suggest that after removal of this signal, the CUD group still had different FC between specific regions. It should be stressed that the multiple comparison correction of NBS is at the component-level rather than the edge-level, meaning that the null hypothesis is not rejected for any specific edge, only for the component as a whole.
The edges between the right SPG and right insula which reach FDR significance are not subject to this limitation. However, these findings were attenuated when controlling for nicotine dependence. A study on nicotine addiction reported that weaker FC between the insula (seed region) and pre-and postcentral gyri was related to higher chances of relapse (Addicott, Sweitzer, Froeliger, Rose, & McClernon, 2015). Although it is likely that these connections are not unique to nicotine addiction but substance abuse related (c.f. the involvement of the insula in drug craving [Naqvi, Gaznick, Tranel, & Bechara, 2014]), our results might be influenced by the higher prevalence of nicotine smokers in the CUD group. Taking that into account in combination with the marginal significance, our FC results should be treated with caution and replication is needed.
We did not find conclusive evidence for the presence or absence of a correlation between age of onset and duration of use and cognition or FC. The literature is also inconclusive on this association as some studies report that earlier onset and longer use have a more detrimental effect on cognition and the brain, whereas other studies report an absence of such association. For a comprehensive overview we refer the reader to (Nader & Sanchez, 2018). Discrepancies in the literature could be due to the definition of "cannabis users," the fidelity of self-report (Prince, Conner, & Pearson, 2018), or that earlyonset users may have a higher intensity and frequency of use compared with late-onset users (Gruber, Sagar, Dahlgren, Racine, & Lukas, 2012). Additionally, THC content in cannabis preparations has increased from 3 to 12% between 198012% between and 201412% between (ElSohly et al., 2000ElSohly et al., 2016). Although users may self-titrate their desired level of THC consumption (Bidwell et al., 2018), this makes the cumulative THC intake difficult to estimate and compare. Last, specific timing of high frequency and/or high potent cannabis use may have a differential effect on outcomes. Frequency and potency of use could differ while age of onset is the same; more frequent use during adolescence versus (early) adulthood is likely to have more detrimental effects on cognition and the brain (Volkow et al., 2014, but see Meier et al., 2019) given that adolescence is a vulnerable period of educational opportunities and brain maturation. Future research should attempt to take these dynamics of cannabis use into account.
This study has several limitations. First, causality cannot be inferred from cross-sectional studies. Second, we lack detailed information such as days since last use and intensity of recent use (with e.g., timeline follow-back, hair analyses). As such, it is not clear whether the findings reported here are due to current THC-bioavailability, long-term changes, or premorbid vulnerability to CUD.
However, a standardized DSM diagnosis can be seen as a more reliable phenotype than (long-term) frequency of use or cumulative lifetime use (Prince et al., 2018). It is also important to realize that CUD captures those people who not just use, but where use leads to problems, which is potentially a different phenotype regardless of amount of use. Third, results of the resting state analysis could be specific to the parcellation atlas, as low resolution parcellations can lead to unstable estimations. However, individual differences in graph metrics were largely conserved in parcellations with more than 250 regions , suggesting our findings on a 333-region cortical parcellation should be stable. Fourth, although our participants were not acutely intoxicated, cognitive functioning and structural/functional connectivity might improve after complete cessation of cannabis use for an extended period of time (Blanco-Hinojo et al., 2017;D'Souza et al., 2016;Hirvonen et al., 2012;Scott, Slomiak, et al., 2018;Tait et al., 2011). Yet, the required duration of abstinence may depend on the frequency and intensity of use. THC is extremely fat-soluble and can remain in fatty tissues (including brain) for extensive periods of time (many weeks), especially after chronic use. Last, our sample comprised African Americans, which could limit the generalizability of our findings. However, African Americans are an underrepresented group in psychiatry research while having a higher prevalence of CUD than Caucasians (Wu, Zhu, & Swartz, 2016) and we mostly confirm earlier findings.
In conclusion, in a large sample that was collected irrespective of cannabis use to minimize recruitment bias, we confirm the literature on poorer cognitive functioning in CUD, and an absence of volumetric brain differences between CUD and non-CUD. We did not find evidence for or against a disruption of structural connectivity. We find disrupted FC in specific regions, although there was sufficient proof that organization of FC as determined via graph metrics does not differ between CUD and non-user group.

ACKNOWLEDGMENTS
The authors would like to thank Dr. Alecia Dager for helpful comments and discussion on earlier versions of this manuscript. This research was supported by National Institute of Mental Health R01 MH106324 to DG and National Institute on Drug Abuse 1R01DA038807 to GP.

CONFLICT OF INTERESTS
The authors report no conflicts of interest.

DATA AVAILABILITY STATEMENT
The data that support the findings of this study are available from the corresponding author upon reasonable request.