Prenatal cannabis exposure, the brain, and psychopathology during early adolescence

Abstract

Prenatal cannabis exposure (PCE) is associated with mental health problems in early adolescence, but the possible neurobiological mechanisms remain unknown. In a large longitudinal sample of adolescents (ages 9–12, n=9,322–10,186), we find that PCE is associated with localized differences in gray and white matter of the frontal and parietal cortices, their associated white matter tracts, and with striatal resting state connectivity, even after accounting for potential pregnancy, familial, and child confounds. Variability in forceps minor and pars triangularis diffusion metrics partially longitudinally mediate PCE-ADHD associations. PCE-related differences in brain development may confer vulnerability to worse mental health in early adolescence.

Alongside increasingly permissive sociocultural attitudes and laws, cannabis use during pregnancy doubled between 2002 (3.4%) and 2017 (7%)¹, despite evidence of potential adverse consequences and discouragement from governmental health agencies (e.g., Surgeon General, Food and Drug Administration)^2,3 and professional organizations (American College of Obstetricians and Gynecologists)⁴. Accumulating studies link prenatal cannabis exposure (PCE) to adverse birth-outcomes, including low birth-weight and preterm birth, in both human samples and preclinical rodent models^5,6. Concerningly, there is now growing evidence that PCE in humans is associated with related behavioral outcomes during childhood, adolescence, and early adulthood, including increased psychopathology and reduced cognition^7–10 (see also^11–13).

Pre-clinical models have shown that it is biologically plausible that PCE may influence psychopathology-related traits. Cannabis constituents traverse the placenta where cannabinoid receptors are expressed^14,15. The placenta then synthesizes neurotransmitters and neurosteroids that in turn influence brain development^16,17, with evidence that placental gene expression influences risk for psychiatric traits, including schizophrenia¹⁸. Cannabis constituents further interface with the fetal endocannabinoid system, which critically contributes to neurodevelopment (e.g., axonal elongation and synaptic pruning)¹². PCE has, in turn, been observed in rodent models to disrupt neural development and functioning, including axonal morphology, glutamatergic neurotransmission, and pyramidal neuron excitability, and to result in impaired cognition and emotional behavior^20–22.

Evidence of variability in child brain metrics as a function of PCE provides support for preclinical findings in human studies²³. PCE has been observed to be associated with a variety of neuroimaging metrics in children and adolescents, including differences in brain volume, white matter integrity, and resting-state connectivity^24–29. However, it remains unclear whether prior findings represent associations that are independent of correlated risk factors for PCE, including family history of mental health and drug problems, and prenatal exposure to other drugs²⁴. Further, there is as yet only limited evaluation of the extent to which neuroimaging associations with PCE may in turn be associated with mental health outcomes. Identifying whether the association of PCE with mental health could be plausibly partially explained by differences in brain structure, function, and/or development is a crucial step towards establishing whether observed causal effects in pre-clinical models also reflect a portion of the biological mechanisms at play in humans who are prenatally exposed to cannabis. Such evidence will have important implications for our understanding of the safety of cannabis use during pregnancy and the mechanisms associated with its influence on child health.

Using data from the Adolescent Brain Cognitive Development^SM (ABCD^®) Study³⁰ (data release 5.0) of 11,875 children, we tested whether PCE before and after maternal knowledge of pregnancy is associated with multimodal brain metrics. These included resting-state fMRI (rs-fMRI) cortical and subcortical connectivity, volume, surface area, thickness, and sulcal depth, and measures from diffusion tensor imaging (DTI) and restriction spectrum imaging (RSI) models applied to diffusion weighted data from white matter tracts, cortical white matter, and gray matter. Notably, endocannabinoid receptors are not expressed in the fetus until 5 to 6 weeks’ gestation^31–33, which approximately corresponds to when, in this study, mothers learned they were pregnant (mean [SD], 6.9 [6.8] weeks). Thus, we hypothesized that the strength of associations with cannabis exposure would be stronger among children with PCE after maternal knowledge of pregnancy, though indirect effects through the placenta are also plausible during earlier pregnancy. We also tested whether associated brain metrics were, in turn, associated with psychopathology during early adolescence, and whether associations with psychopathology may be partially mediated by brain differences.

Results

Analyses included data from the first two waves of neuroimaging data collection (baseline visit and follow-up wave 2), at ages 9–10 and 11–12. This comprised 16,641 observations from 10,186 participants who had complete usable data for at least one neuroimaging modality (6,455 participants had two data points available, 3,731 had one), including 373 prenatally exposed to cannabis only prior to the mother’s knowledge of her pregnancy (pre-knowledge only; ns = 337–373 per modality) and 195 exposed both before and after knowledge of her pregnancy (pre- and post-knowledge; ns = 172–195 per modality; n=5 exposed only after knowledge were coded as pre- and post-knowledge; Table 1, Methods; Supplemental Data). Demographic differences included differences in pubertal status, parental education and income, and self-reported race (Table 1). Groups also differed along variables that reflect aspects of predispositional risk, including family history of mental health and substance problems, prenatal exposure to substances other than cannabis, and variables related to pregnancy risk (i.e., planned pregnancy, use of prenatal vitamins, mother’s age at birth, and gestational age when the mother learned of her pregnancy; Table 1). All of the above were included as covariates in primary or secondary regression models (Methods).

Table 1:

Participant demographics.

Variable	None (n=9,618)	Pre-only (n=373)	Pre- and Post-knowledge (n=195)	Comparing all three groups		Comparing PCE groups
				Χ2/F	p	Χ2/F	p
Age (weeks)*	128.75 (14.06)	128.43 (14.32)	127.29 (13.79)	1.77	0.17	1.14	0.29
Sex (F)	4608 (47.88)	181 (48.53)	108 (55.38)	4.35	0.11	2.65	0.10
Pubertal stage*	1.77 (0.89)	2.06 (0.96)	2.13 (0.98)	33.81	2.31×10 −15	0.31	0.58
Mother’s education (yrs)*	15.38 (2.48)	14.32 (2.24)	13.88 (1.98)	17.20	4.44×10 −8	5.68	0.02
Household income, $
≤ 49,999	2781 (28.89)	209 (56.03)	125 (64.1)	230.03	1.12×10 −50	2.82	0.09
50,000 – 74,999	1335 (13.87)	49 (13.14)	28 (14.36)	0.20	0.90	0.04	0.83
75,000 – 99,999	1383 (14.37)	35 (9.38)	25 (12.82)	7.64	0.02	1.48	0.22
100,000 – 199,999	2974 (30.9)	60 (16.09)	15 (7.69)	84.40	4.70×10 −19	6.68	0.01
≥ 200,000	1152 (11.97)	20 (5.36)	2 (1.03)	36.85	9.93×10 −9	5.15	0.02
Race/ethnicity
White	7414 (77.03)	221 (59.25)	121 (62.05)	83.93	5.94×10 −19	0.50	0.48
Black	1801 (18.71)	150 (40.21)	84 (43.08)	170.37	1.01×10 −37	0.10	0.75
Hispanic	1936 (20.11)	84 (22.52)	25 (12.82)	7.80	0.02	6.21	0.01
Native American	296 (3.08)	15 (4.02)	12 (6.15)	6.82	0.03	0.78	0.38
Pacific Islander	54 (0.56)	5 (1.34)	1 (0.51)	3.75	0.15	0.21	0.65
Asian	613 (6.37)	11 (2.95)	5 (2.56)	11.72	0.003	0.00	1.00
Other	616 (6.4)	32 (8.58)	8 (4.1)	4.63	0.10	3.03	0.08
Other Prenatal Exposures
Alcohol before knowledge	2259 (23.47)	240 (64.34)	111 (56.92)	417.27	2.46×10 −91	2.24	0.13
Alcohol after knowledge	200 (2.08)	12 (3.22)	47 (24.1)	375.10	3.53×10 −82	59.67	1.12×10 −14
Tobacco before knowledge	983 (10.21)	206 (55.23)	129 (66.15)	1146.50	1.10×10 −249	6.07	0.01
Tobacco after knowledge	341 (3.54)	52 (13.94)	89 (45.64)	824.80	7.87×10 −180	70.08	5.68×10 −17
Other sub. before knowledge	38 (0.39)	11 (2.95)	23 (11.79)	381.91	1.17×10 −83	16.94	3.85×10 −5
Other sub. after knowledge	16 (0.17)	2 (0.54)	19 (9.74)	485.05	4.72×10 −106	28.73	8.32×10 −8
Family history of psychopathology
Depression	691 (7.18)	51 (13.67)	36 (18.46)	54.54	1.43×10 −12	2.25	0.13
Mania	36 (0.37)	5 (1.34)	7 (3.59)	48.41	3.07×10 −11	2.27	0.13
Conduct	84 (0.87)	22 (5.9)	33 (16.92)	425.24	4.58×10 −93	16.06	6.14×10 −5
Nerves	227 (2.36)	22 (5.9)	12 (6.15)	28.32	7.08×10 −07	0.00	1.00
Alcohol	120 (1.25)	19 (5.09)	31 (15.9)	277.86	4.61×10 −61	16.71	4.35×10 −5
Drugs	105 (1.09)	22 (5.9)	42 (21.54)	532.79	2.03×10 −116	30.93	2.68×10 −8
Pregnancy-related
Prenatal vitamins	8210 (96.27)	327 (96.46)	127 (90.71)	11.68	0.003	5.96	0.01
Planned pregnancy	5594 (65.6)	90 (26.55)	37 (26.43)	298.94	1.22×10 −65	0.00	1.00
Week learned pregnancy*	6.44 (1.67)	7.52 (1.66)	7.61 (1.7)	94.76	2.68×10 −22	0.40	0.53
Birth weight (oz)*	112.31 (23.4)	112.7 (22.78)	108.06 (22.9)	4.47	0.01	0.66	0.42
Mother age at birth (yrs)	29.85 (5.98)	25.63 (5.67)	24.64 (5.47)	208.85	3.99×10 −90	3.71	0.06

Continuous variables (*) were compared with an F-test, values reflect the mean and SD for each group. Binary variables were compared with a chi-squared test, values reflect the N and group percentage. Comparisons were made between all three groups, as well as between the two PCE groups.

Linear mixed-effect regression models (Methods) revealed three brain metrics significantly associated with PCE after false discovery rate (FDR) multiple-test correction for all tests (n_{brain variables}=2,907; Figure 1; Supplemental Data): 1) restricted normalized directional diffusion of the forceps minor, 2) transverse diffusivity of the right pars triangularis, and 3) rs-fMRI connectivity of the auditory network and left putamen. An additional 14 brain metrics were significantly associated with PCE after multiple-test correction within modality (Figure 1; Figure 2; Supplemental Data). These broadly included diffusion weighted metrics from gray and white matter, and white matter tracts, in the frontal and parietal lobes. Findings indicate that the association of PCE with brain metrics is relatively localized; no global measure (Methods) survived correction (smallest p=0.015 uncorrected). While several associations were, as hypothesized, driven by stronger effects in the pre- and post-knowledge group (Figure 2; e.g., transverse and mean diffusivity of the cortical gray matter in the right pars triangularis), there were also associations driven by stronger effects in the pre-knowledge group (Figure 2; e.g., hindered and restricted normalized directional diffusion in the cortical white matter of the inferior frontal cortex). Notably, regions where effects were driven by the pre-knowledge group did not show associations with psychopathology in subsequent analyses (see below). All findings were robust to additional post-hoc tests, including: 1) inclusion of pregnancy-related variables with high missingness (due to non-report); 2) restricting data to only the baseline visit; 3) adjustment for polygenic risk for cannabis use disorder³⁴; 4) restricting covariates to only essential variables (Methods; Supplemental Data).

Fig. 1: Association of PCE with brain metrics.

Plot shows the uncorrected P value (log scale; x axis) for the association between brain metrics and PCE from mixed-effect regressions. Each brain region is a separate point; each imaging modality is in a distinct color. The vertical dashed line is placed at P < 0.05 (two-sided) FDR-corrected (PFDR) for all 2,907 tests. Regions (n = 3) FDR significant for all tests are labeled; points with a dark black border are significant at PFDR < 0.05 for all comparisons within a given metric type (for example, corrected only for all comparisons of cortical thickness or only for all comparisons of fractional anisotropy of white matter tracts, and so on). Additional statistical information is provided in Supplementary Table 2. sMRI, structural MRI; fiber, white matter fiber tracts; WM, cortical white matter; GM, cortical and subcortical gray matter.

Fig. 2: Significant associations of PCE with brain metrics, by exposure group.

Standardized regression β effect sizes and 95% confidence intervals (CIs) from mixed-effect regressions assessing the association of PCE before maternal knowledge of pregnancy or pre- and post-maternal knowledge of pregnancy compared with no exposure. a, Transverse diffusivity (DTI) of cortical gray matter. b, Mean diffusivity (DTI) of cortical gray matter. c, Hindered directional diffusion (RSI) of cortical white matter. d, Restricted directional diffusion (RSI) of cortical white matter. e, Restricted isotropic diffusion (RSI) of cortical white matter. f, Restricted total diffusion (RSI) of cortical white matter. g, Restricted directional diffusion (RSI) of white matter tracts. h, Restricted total diffusion (RSI) of white matter tracts. i, Average cortical-subcortical connectivity (rs-fMRI). Nonsignificant outcomes are not shown. *Significant after FDR correction for all 2,907 tests (two-sided; a, P = 1.65 × 10⁻⁵; h, P = 4.1 × 10⁻⁵; i, P = 5.1 × 10⁻⁵). The corresponding brain regions are shown on the right of each panel; each metric is shown in a different color. Additional statistical information is provided in Supplementary Table 3.

To evaluate whether the 17 significant PCE-brain associations may plausibly contribute to behavioral variability, we estimated the association of the 17 regions with 13 measures of mental health that have previously been associated with PCE in this sample (Figure S1)^24,35,36, including measures reflecting psychotic-like experiences, aggressive behavior, attention problems, and social problems (Methods). Seven associations survived multiple-test correction for all 221 tests, of which five remained significant when controlling for all PCE-related covariates used in primary analyses above; an additional eleven survived multiple-test correction within measure-type, of which six remained significant when controlling for additional PCE-related covariates (Figure 3A; Supplemental Data; Methods). Significant associations were all externalizing-related behaviors (i.e., attention and conduct problems, ADHD, and rule-breaking behavior) and were restricted to four metrics (Figure 3A) - including two which survived correction for all tests in the above analyses of associations with PCE - the right pars triangularis (transverse and mean diffusivity of the cortical gray matter) and the forceps minor (restricted directional and total diffusion). The directions of these associations are consistent with these brain regions mediating the effect of PCE on psychopathology, including that the association of PCE with these metrics is driven by the pre- and post-knowledge group. Longitudinal mediation analysis (Methods) of these 11 brain-behavior associations revealed 8 indirect associations that were robust to multiple testing correction, of which 7 remained significant when controlling for additional PCE-related covariates (Figure 3B; Supplemental Data; Methods), wherein brain metrics partially mediated the association between PCE and attention and ADHD problems at a subsequent wave. Significant associations included two regions which survived FDR-correction for all tests in primary PCE analyses: 1) restricted normalized directional diffusion of the forceps minor (mediating effects on ADHD problems), 2) transverse diffusivity of the right pars triangularis (mediating effects on ADHD and attention problems). Effects were uniformly small, indicating that 1.5–2% of the association of PCE with attention-related psychopathology may be plausibly mediated by these brain metrics.

Fig. 3: Association of brain metrics with psychopathology in early adolescence.

a, Standardized regression β effect sizes and 95% CIs from mixed-effect regressions assessing the association of significant (PFDR < 0.05) brain metrics with log-transformed psychopathology scales across the full sample (N = 10,584). b, Standardized mediation effect and 95% parametric CIs from longitudinal mediation analyses testing whether significant brain metrics mediate the association of PCE with psychopathology (N = 7,990). Analyses were run first collapsing across PCE groups (that is, any exposure), and post hoc analyses computed effects for each group separately. c, Mediation path diagrams for significant mediations (PFDR < 0.05) with regions that survived correction for all comparisons in primary analyses (Fig. 2; restricted total diffusion of the forceps minor and transverse diffusivity of the right pars triangularis gray matter). Values reflect standardized effects and 95% CIs. Additional statistical information is provided in Supplementary Table 4. c, total effect; c’, direct effect; DSM-5, Diagnostic and Statistical Manual of Mental Disorders 5th Edition ⁶⁶; Prob, problems.

Discussion

These findings align with the hypothesis that the association of PCE with psychopathology may be partially attributable to effects of PCE on brain development. Our results build upon prior work using small sample sizes and a restricted subset of measures and regions^24–28 to show that associations of PCE with brain metrics are relatively localized. In particular, we find robust associations with lower transverse and mean diffusivity of the right pars triangularis cortical gray matter and greater restricted directional and total diffusion of the forceps minor. Further, we demonstrate that these brain metrics are in turn associated with externalizing traits, particularly attention problems, consistent with the interpretation that differences in brain development partially mediate PCE-adolescent psychopathology associations.

Significant associations with PCE in our study included DWI (Diffusion Weighted Imaging) measures of cortical gray and white matter in the frontal cortex (pars triangularis and orbitalis, lateral orbitofrontal) and parietal lobe (inferior frontal lobe and post central gyrus), DWI measures of white matter tracts that connect the frontal cortex (forceps minor) and the frontal and parietal cortices (superior longitudinal fasciculus), and resting state connectivity of the putamen and auditory network. Cannabis involvement in adults and adolescents (e.g., use, problematic use) has been linked, albeit inconsistently, to variability in other metrics within many of the same regions observed here (e.g., forceps minor and superior longitudinal fasciculus fractional anisotropy, pars triangularis thickness, lateral orbitofrontal volume)^37–39. Acute cannabinoid administration has also been observed to disrupt connectivity between the cortex and the striatum⁴⁰. As such, the present results support the hypothesis that the putative effects of PCE on brain development may partially intersect with the effects of self-administered cannabis use.

While the present analyses included many DWI metrics, multiple-test correction was applied both across all measures and independently within each measure-type. As such, it is notable that the majority of findings were in measures derived from DWI scans. These findings may reflect differential gray matter cellular organization (e.g., neurite density) and white matter tract integrity of these regions, echoing pre-clinical observations that PCE alters axonal development and morphology²⁰. However, these data are also consistent with several alternative mechanisms. Physiological processes that increase isotropic diffusion (i.e., inflammation) would also be expected to result in the observed increased mean and transverse cortical diffusivity, and decreased restricted directional and total diffusion of white matter bundles^41,42. Thus, our findings could instead reflect relatively reduced neuroinflammation in PCE participants that may begin in the intrauterine environment^43,44, consistent with evidence of the effects of cannabinoids on peripheral inflammation and placental transcription of inflammation-related genes^44,45. Alternatively, cortical mean diffusivity decreases over adolescence and restricted directional diffusion increases over adolescence⁴⁶, and thus these results may indicate accelerated neurodevelopment with PCE, consistent with prior findings that chronic cannabis use is associated with accelerated aging⁴⁷, possibly due to chronic cannabis use co-occurring with exposure to post-combustion hydrocarbons⁴⁸.

Significant brain metrics were in turn associated with psychopathology in early adolescence, including externalizing-related measures (i.e., attention and ADHD problems, conduct problems, and rule breaking behavior), social problems, and self-report of psychotic-like experiences (Supplemental Data). The strength of the observed associations with psychopathology were small (Figure 3A), consistent with prior brain-wide neuroimaging studies, which have observed similar magnitude effects across a variety of traits and metrics⁴⁹. Associations with social problems and self-report of psychotic-like experiences did not survive correction for confounding variables reflecting predispositional risk (i.e., family history of mental health and substance problems, prenatal exposure to other drugs). The remaining externalizing measures all showed both cross-sectional and longitudinal associations with diffusivity metrics of the forceps minor and right pars triangularis. These findings align with prior meta-analyses, which have linked ADHD to variability in other metrics in these same regions (forceps minor fractional anisotropy and pars triangularis surface area)^50,51. Notably, metrics in both regions have also been linked to cognitive control^52,53, which is known to be impaired in externalizing disorders, including ADHD⁵⁴. The contributing role of cognitive control in the emergence of externalizing psychopathology in PCE is an avenue that future studies may wish to explore.

As hypothesized, associations with diffusivity metrics of the forceps minor and right pars triangularis were driven by participants with exposure both pre- and post-knowledge of pregnancy. This suggests that the outcome of PCE may depend on the developmental timing of exposure, and that improved measurement of the timing, mode, quantity, and frequency of exposure could yield larger effects. Further, associations between these brain metrics and psychopathology were consistent with a mediational effect. Indirect associations between prenatal exposure and psychopathology via brain metrics were significant, and notably the observed cross-sectional associations between brain metrics and psychopathology remained significant longitudinally. This highlights the potential informativeness of cross sectional work and suggests that these associations may be relatively stable over time. This result indicates that it is statistically plausible that variability in these brain metrics partially longitudinally mediate the association between PCE and future adolescent externalizing behaviors.

We were able to account for many known familial, pregnancy-related, and child-related confounding variables, but the role of unmeasured confounders cannot be discounted. Post-hoc analyses used a polygenic risk score for cannabis use disorder³⁴ as an additional control, but it should be noted that the risk score only captures a portion of genetic risk, and these control analyses were limited to only participants with genomically-confirmed European ancestry. Prenatal cannabis exposure may thus act as a proxy for predispositional risk factors that were only partially accounted for by model covariates (i.e., family history), including risk for substance use, parental mental health, personality, and brain structure, as well as unmeasured environmental risk factors (e.g., maternal stress during pregnancy, exposure to a permissive home environment). Thus, while the present findings provide some of the strongest evidence to date that the brain mediates effects of PCE on mental health during early adolescence, we cannot rule out the alternative explanation that the present results reflect the effects of unmeasured confounders.

While the ABCD study is among the largest longitudinal studies of neurobiology and adolescent behavior that assessed prenatal exposures, limitations include the relatively small sample of prenatal cannabis-exposed offspring. It should be noted that the measure of prenatal cannabis exposure used here reflects a retrospective report of behavior that occurred approximately 10 years earlier, which may have resulted in biased reporting and misclassification. Further, there are limited or no data on potency, mode, frequency, timing, or quantity of cannabis exposure in this data set. Analyses rely on parent-report for mental health outcomes, which may be biased⁵⁵. Further, some data suggest that externally observable behavior, such as externalizing, is more likely to be captured by parent-report as opposed to potentially unobservable internalizing behavior^56,57. Mediation analyses show that a mediating effect of brain metrics is statistically plausible, but we cannot rule out reciprocal effects (i.e., psychopathology leads to changes in brain metrics) or unmeasured causal variables that drive changes in both brain metrics and psychopathology, among other explanations⁵⁸.

Conclusions

Limitations notwithstanding, our study provides evidence that PCE is associated with differences in brain development that may partially mediate associations between PCE and increased psychopathology during early adolescence. Future work incorporating improved measurement of prenatal cannabis exposure and recruiting familial samples with discordant exposures⁵⁹ may yield larger effects and more insight into the causal mechanisms underlying these associations.

Methods

MRI Acquisition and Processing:

Casey et al., 2018³⁰ provide an in-depth description of the ABCD Study^® imaging acquisition protocol and parameters and Hagler et al., 2019⁶⁰ provide an in-depth description of the ABCD Study^® image processing and analysis methods⁴⁶. The present analyses used tabulated neuroimaging data provided as part of the 5.0 data release, the acquisition and processing streams are briefly described below.

Structural MRI:

1 mm isotropic T1-weighted structural magnetic resonance images (MRI) were acquired on 3 T (Siemens, Phillips and GE) MRI scanners using either a 32-channel head or 64-channel head-and-neck coil. Scan protocols were carefully harmonized across the three MRI vendor platforms to reduce scanner-caused variability. MRI data were processed with the Multi-Modal Processing Stream software package that includes FreeSurfer 5.3. Besides a modified intensity normalization process used by the ABCD processing pipeline, the standard FreeSurfer cortical and subcortical reconstruction pipeline was run to generate structural measures including volume, cortical thickness, cortical surface area, and cortical sulcal depth. Here we use measures from the Desikan cortical atlas and the Freesurfer Aseg subcortical atlas. A description of the quality-control measures conducted on the processed data is provided in Hagler et al. MRI analyses included only participants whose structural MRI reconstructions passed QC tests. Global variables used as covariates in analyses included intracranial volume, mean thickness, total surface area, and mean sulcal depth.

Diffusion MRI:

Diffusion MRI (dMRI) data were acquired in the axial plane at 1.7 mm isotropic resolution with multiband acceleration factor 3. Diffusion-weighted images were collected with: seven b= 0 s/mm² frames and 96 non-collinear gradient directions, 6 directions at b= 500 s/mm², 15 directions at b= 1000 s/mm², 15 directions at b= 2000 s/mm², and 60 directions at b= 3000 s/mm². 3D T2-weighted fast spin echo with variable flip angle scans were acquired at 1 mm isotropic resolution with no multiband acceleration. Scanning protocols were harmonized across sites.

Data were corrected for eddy current distortion. Images were rigid-body-registered to the corresponding volume synthesized from a robust tensor fit. Dark slices caused by abrupt head motion were replaced with values synthesized from the robust tensor fit, and the diffusion gradient matrix was adjusted for head rotation. Spatial and intensity distortions caused by B0 field inhomogeneity were corrected and gradient nonlinearity distortions were corrected for each frame. Data were registered to T1w structural images and dMRI data were then resampled to 1.7 mm isotropic resolution. Major white matter tracts were labeled using AtlasTrack. The diffusion tensor imaging (DTI) model was used to calculate standard measures related to microstructural tissue properties, including fractional anisotropy and mean, longitudinal (or axial), and transverse (or radial) diffusivity (MD, LD, and TD).

A Restriction Spectrum Imaging (RSI) model was also fit. This linear estimation approach allows for mixtures of “restricted” and “hindered” diffusion within individual voxels. RSI was used to model two volume fractions, representing intracellular (restricted) and extracellular (hindered) diffusion, with separate fiber orientation density (FOD) functions, modeled as fourth order spherical harmonic functions, allowing for multiple diffusion orientations within a single voxel. Measures derived from the RSI model fit include: restricted normalized isotropic, restricted normalized directional, restricted normalized total, hindered normalized isotropic, hindered normalized directional, hindered normalized total, and free normalized isotropic. Normalized isotropic and hindered normalized total reflect varying contributions of intracellular and extracellular spaces to isotropic diffusion-related signal decreases in a given voxel. Restricted normalized directional (a proxy for oriented myelin organization) and hindered normalized directional reflect oriented diffusion--diffusion that is greater in one orientation than others. Restricted normalized directional is similar to FA, except that it is unaffected by crossing fibers⁴⁶. Restricted normalized total and hindered normalized total reflect the overall contribution to diffusion signals of intracellular and extracellular spaces.

Mean DTI and RSI measures were calculated for white matter fiber tract ROIs created with AtlasTrack and for ROIs derived from FreeSurfer’s automated subcortical segmentation. DTI and RSI measures were also sampled onto the FreeSurfer-derived cortical surface mesh to make maps of diffusion properties for cortical gray matter and white matter adjacent to the cortex. Here we use measures of these properties taken from the Desikan cortical atlas. dMRI analyses included only participants whose structural MRI and dMRI reconstructions passed QC tests. Global variables used as covariates in analyses included the mean of all regions/tracts for a given metric.

Resting-state fMRI:

Twenty minutes of resting-state functional MRI (rs-fMRI) data were collected across four 5-minute scans with eyes open and passive viewing of a cross hair. The FIRMM real-time head motion monitoring system was implemented for motion detection in resting state fMRI scans at sites using Siemens scanners⁶¹. FIRMM allows scanner operators to adjust the scanning paradigm based on a participant’s degree of head motion (i.e., the worse the motion, the less usable data and greater the need for more data to be acquired). rs-fMRI data were acquired with: a matrix of 90 × 90, 60 slices, FOV size = 216 × 216, voxel size = 2.4 × 2.4 × 2.4 mm³, TR = 800 ms, TE = 30 ms, FA = 52°, and a multiband acceleration factor of 6. Head motion was further corrected by registering each frame to the first and images were corrected for distortions due to gradient nonlinearities. To correct for between-scan motion, each scan was resampled with cubic interpolation into alignment with a reference scan that is chosen as the one nearest to the middle of the set of fMRI scans for a given participant. Further processing included the removal of initial frames, normalization, regression of motion and mean signal time courses, and temporal filtering. Time points with a framewise displacement greater than 0.2 mm were excluded, as were periods with fewer than five contiguous, sub-threshold time points. Preprocessed time courses were sampled onto the cortical surface for each individual subject and average time courses were calculated for cortical surface-based ROIs using a functionally-defined parcellation based on resting-state functional connectivity patterns from the Gordon atlas. Average time courses were also calculated for subcortical ROIs. The correlation between each pair of ROIs were averaged within or between networks to provide summary measures of network correlation strength. rs-fMRI analyses included only participants whose data passed QC tests. Analyses of resting-state data did not include a global variable as a covariate.

Stability:

The 2-year longitudinal stability of individual regions and metrics (ICC) ranged from 0.01–0.97 and is reported in Supplemental Figure 2 and in the Supplement Data. Notably, rs-fMRI and cortical DWI metrics were the least stable (median ICC = 0.14 – 0.47) while traditional structural indices (i.e., cortical thickness and surface area, volumes) and DWI metrics of white matter tracts were the most (median ICC = 0.5 – 0.93). Analyses did not include task-based measures as prior work examining the stability of task-based measures in this sample has found that it is quite low (average stability = 0.072)⁶².

Measures of psychopathology

CBCL:

The Child Behavior Checklist (CBCL)⁶³ is a 113-item questionnaire on which caregivers rated items representing specific problems in the past six months. The CBCL was completed annually. Subscales include aggressive behavior, anxious/depressed, attention problems, rule-breaking behavior, somatic complaints, social problems, stress problems, thought problems, and withdrawn/depressed. There is a broadband scale for internalizing problems, which sums the anxious/depressed, withdrawn-depressed, and somatic complaints scores, and a broadband scale for externalizing problems, which sums rule-breaking and aggressive behavior. The total problems score is the sum of the scores of all the problem items. The CBCL also includes a set of DSM-oriented scales for depressive problems, affective problems, anxiety problems, somatic problems, ADHD problems, oppositional defiant problems, conduct problems, obsessive-compulsive problems, and sluggish cognitive tempo. Analyses using the CBCL included 12 scales that were previously found to be significantly associated with PCE in this sample (Supplemental Figure 1)³⁵: Total problems, externalizing factor, rule-breaking behavior, aggressive behavior, social problems, thought problems, attention problems, sluggish cognitive tempo, stress problems, obsessive compulsive problems, ADHD problems, and conduct problems. Analyses examined the raw CBCL scores. Analyses included both higher-order factors (i.e., total problems and externalizing), as well as lower-order scales, as prior work has identified both shared and distinct associations between brain metrics and psychopathology⁶⁴.

Psychotic-like Experiences:

Children completed the 21-item Prodromal Questionnaire-Brief Child Version (PQ- BC)⁶⁵ in which they are first asked to respond (yes/no) to whether they experienced a thought/feeling/experience (e.g., do familiar surroundings sometimes seem strange, confusing, threatening, or unreal to you?) before reporting on whether it was distressing and if so, the extent that it bothered them. From these data, Total (i.e., the sum of endorsed items) was used as a measure of psychotic-like experiences.

Additional variables

Child Substance Initiation:

Neuroimaging study visits were excluded from analyses once participants reported substance use initiation. Alcohol, nicotine, and cannabis use initiation variables were derived based on endorsement of substance use at any yearly substance use interview or mid-year substance phone interview included in ABCD release 5.0. Specifically, alcohol initiation was defined as endorsing a full drink containing alcohol, outside of the context of religious ceremonies; nicotine initiation was defined as “more than a puff” in any form including tobacco cigarettes or cigars, e-cigarettes, hookah or pipes, or use of smokeless tobacco or chew, or nicotine patches, outside of the context of religious ceremonies; and cannabis initiation was defined as “more than a puff” in any form including smoking or vaping flower, oils, or concentrates, smoking blunts, or consuming edibles or tinctures, but not including synthetic cannabis or cannabis-infused alcoholic drinks.

Covariates were coded as in prior studies^24,35,36. In brief:

Pubertal status:

Parents and children both completed a 5-item scale on the child’s pubertal development⁶⁶, combined to a summary score. The parent rating was used as the primary measure. The child rating was used if the parent rating was unavailable.

Child Race:

Parents/caregivers selected from 26 categories. Dichotomous groups were formed for the most prevalent categories of race (i.e., White, Black, Asian, Pacific Islander, Native American) with remaining participants being assigned to Other. All variables were dummy coded as non-mutually exclusive dichotomous variables; as such, participants could be coded within more than one category.

Child Ethnicity:

Parents/caregivers reported whether they consider the child to be hispanic/latinx.

Maternal Education:

Maternal education was recoded such that 12th grade, HS grad, and GED =12 years; some college and associate’s degree = 14 years; Bachelor’s degree = 16 years; Master’s degree = 18 years; Professional and Doctoral degrees = 20 years.

Household Income:

Due to low endorsement of the first five of 10 household income levels, this variable was recoded such that the first five categories were assigned a value of one (i.e., <$50,000). The subsequent categories used were coded as two ($50,000-$74,999), three ($75,000-$99,999), four ($100,000-$199,999), and five ($200,000 or more), respectively.

Prenatal exposure to other substances:

Child prenatal exposure to alcohol, tobacco, cocaine/crack, heroin/morphine, and oxycontin both before and after maternal knowledge of pregnancy were coded as separate dichotomous variables (cocaine/crack, heroin/morphine, and oxycontin were collapsed into a single variable due to low endorsement) based upon parent/caregiver retrospective report.

Additional covariates reported by parent/caregiver:

Birth weight, family history of psychopathology (first-degree relative - depression, mania, conduct problems, nerves, alcohol problems, and drug problems), unplanned pregnancy, maternal age at birth, use of prenatal vitamins, and gestational age when mother learned of pregnancy.

MRI covariates: scanner manufacturer and mean in-scanner motion during diffusion-weighted and resting state scans (not available for structural scans).

Genetics

Genotyping, Quality Control, and Imputation: Saliva samples were genotyped on the Smokescreen array by the Rutgers University Cell and DNA Repository (now incorporated with other companies as Sampled; https://sampled.com/). The Rapid Imputation and COmputational PIpeLIne for Genome-Wide Association Studies (RICOPILI)⁶⁷ was used to perform quality control (QC) on the 11,099 individuals with available ABCD Study phase 3.0 genotypic data, using RICOPILI’s default parameters (genotyping call rate >98%, inbreeding coefficient (F) < ±0.2, sex checks). The 10,585 individuals who passed these QC checks were aligned with broad self-identified racial groups using the ABCD Study parent survey. Of the 6,787 parents/caregivers indicating that their child’s race was only “white,” 5,561 of those individuals did not endorse any Hispanic ethnicity/origin. Next, variants were filtered to exclude those with missingness >2% and HWE p-value <1e-06. Using data from unrelated individuals (pi-hat ≤ 0.2) and an LD pruned set of common (MAF>0.05) and non-palindromic SNPs (and excluding MHC and chromosome eight inversion region), principal components analysis (PCA) was performed in RICOPILI using EIGENSTRAT⁶⁸ to confirm the genetic ancestry of these individuals. This was done by merging the ABCD Study data with the 1000 Genomes reference panel, computing the mean and standard deviations for the 1000 Genomes ancestry populations for the top 6 PCs, and establishing that the previously identified group of 5,561 participants were genetically similar to the 1000 Genomes European panel by assessing whether they fell within 3 SDs of the mean for the top 6 PCs within this 1000 Genomes European reference population. After another round of QC on this subset, 5,556 European-ancestry individuals were retained. The European ancestry subset was then imputed to the TOPMed imputation reference panel16. Imputation dosages were converted to best-guess hard-called genotypes, and only SNPs with Rsq > 0.8, MAF > 0.01, missingness < 0.1, and HWE p-values >1E-06 were kept for PRS analyses.

Polygenic score:

Polygenic scores (PGS) for cannabis use disorder (CUD) were generated using summary statistics from a recent CUD GWAS³⁴. PGS were generated using PRS-CS⁶⁹ (phi=default, iterations = 10,000, burn-in = 5,000), which uses bayesian regression and continuous shrinkage priors to create PGS that are competitive with state of the art methods.

Statistical analyses

All outcome variables were winsorized (+/− 3SDs) prior to analyses to reduce the influence of outlier observations, and measures of psychopathology were log-transformed due to high skew. Post-hoc analyses found that all results remained FDR-significant without these transformations and results were robust to the removal of observations with extreme values. All variables were z-scored prior to analyses so that effect sizes would be comparable across imaging modalities. Linear mixed-effect models were fit in R (v 4.1.5)⁷⁰ using the ‘lme4’ package (v 1.1–34)⁷¹. We adopted a mass-univariate approach in these analyses as there are many fewer participants with PCE than not. Alternative approaches, such as machine learning, are difficult to implement in this context, as it becomes challenging to achieve greater than chance performance using only a subset of the data for model training or testing⁷². We also adopt a mass-univariate approach, as opposed to implementing a multivariate approach to combine imaging metrics (e.g., principal components analysis), as such an approach would assume that the effects of PCE are widely shared across the brain, which may not necessarily be the case.

Linear mixed effect models testing the association of PCE with brain metrics included random intercepts for participant ID, site, and family ID. Fixed effects included familial relationship (i.e., twin, sibling), measures that capture aspects of development (i.e., age, age², pubertal status), sex, variables associated with risk for prenatal cannabis exposure and mental health problems (i.e., parental education and income, race, family history (i.e., first-degree relative) of drug, alcohol, or other mental health problems (n=6)), potential confounding exposures (i.e., prenatal exposure to alcohol, tobacco, or other drugs), as well as variables that influence imaging metrics, including scanner manufacturer and in-scanner motion (not available for structural scans), and the corresponding neuroimaging global variable (e.g., intracranial volume, mean fractional anisotropy of all white matter fibers, etc., except when testing said global metric as the outcome). Models were fit with and without the two PCE variables, and a log-likelihood ratio test was used to compare the two models. A significant p-value indicates that the addition of both PCE variables improves model fit, but does not give information on what drives the effect. We then examined the standardized regression effect size and associated confidence interval for the two PCE variables.

False discovery rate (FDR) correction for multiple comparisons was applied across all metrics. In order to balance the risk for false negative results that comes with correcting for so many tests, FDR correction was additionally applied separately within each measure-type (e.g., structural volume only, white matter tract fractional anisotropy only, etc.). Post-hoc analyses confirmed that associations remained nominally significant with: 1) the inclusion of additional pregnancy-related covariates with higher missingness (i.e., mother’s age at the birth, birth weight, prenatal vitamins, planned pregnancy, when mother learned of pregnancy), 2) when analyses were restricted only to the baseline wave of collection, and 3) when including only a more limited set of covariates (random intercepts for participant ID, site, and family ID, and fixed effects for age, age², sex, pubertal status, familial relationship (i.e., twin, sibling), scanner manufacturer, in-scanner motion, and the corresponding neuroimaging global variable). Further post-hoc analyses 4) confirmed that the direction of association remained consistent when restricted to participants with genomically-confirmed European ancestry (N=4,962), which additionally included a polygenic score for cannabis use disorder and 10 principal components reflecting genomic ancestry as fixed effects.

Linear mixed effect models testing the association of brain metrics with psychopathology included random intercepts for participant ID, site, and family ID, and fixed effects for age, age², sex, pubertal status, familial relationship, scanner manufacturer, in-scanner motion, and the corresponding neuroimaging global variable. Models examined linear associations between brain metrics with psychopathology, as there is limited evidence for non-linear associations⁵¹. Post-hoc polynomial regressions (including 2nd and 3rd order polynomials) for the association between brain metrics and CBCL scores found that polynomials did not significantly improve model-fit after FDR-correction.

False discovery rate (FDR) correction for multiple comparisons was again applied within each measure-type. Follow-up analyses included the full set of covariates used in primary PCE analyses (see above), to confirm that associations remained nominally significant with these covariates.

Mediation analyses were performed in R with the ‘mediation’ package (v 4.5.0)⁵⁸. Analyses tested whether brain metrics at baseline mediated the association of PCE with psychopathology at the 1-year follow-up, and whether brain metrics at the 2-year follow-up mediated the association of PCE with psychopathology at the 3-year follow-up. As the ‘mediation’ package only accepts one random intercept, which needed to be a random intercept for Participant ID, study-site was modeled as 20 dummy-coded fixed effects, instead of a random effect. Further, only one individual from each family was included (chosen at random, unless there was at least one family member who was part of a PCE group), so that a random intercept for Family ID was not needed. As the package does not allow for categorical treatment variables with more than two levels in a mixed effect model, PCE was coded as a binary variable indicating any prenatal exposure. This was deemed reasonable as post-hoc tests in the significant brain metrics indicated that effects were directionally consistent between the two PCE groups (Figure 2). Post-hoc analyses were run for each PCE group separately to derive a separate estimate for each group. The mediation effect size reflects the expected effect if the association between PCE and psychopathology were fully mediated by the brain metric. Parametric confidence intervals were computed via Monte Carlo simulation of model parameters (10,000 simulations), followed by a mediation analysis for each simulation. The resulting confidence interval is equivalent to a quasi-Bayesian approximation of the posterior distribution. FDR correction for multiple tests was applied across all analyses. Post-hoc analyses confirmed that associations remained nominally significant with the inclusion of additional pregnancy-related covariates with higher missingness (i.e., mother’s age at the birth, birth weight, prenatal vitamins, planned pregnancy, when mother learned of pregnancy). Further post-hoc analyses 2) confirmed that the direction of association remained consistent when restricted to participants with genomically-confirmed European ancestry (N=4,962), which additionally included a polygenic score for cannabis use disorder and 10 principal components reflecting genomic ancestry as fixed effects.

Supplementary Material

Sup_Figs_1_2

Figure S1. Association of prenatal cannabis exposure with psychopathology in early adolescence. After Baranger et al. (2022)³⁵, reflecting results from a multilevel model that used data from the baseline visit, 1-year, and 2-year follow-ups.

Figure S2. 2-year stability of neuroimaging metrics. The stability (ICC; i.e., variance explained by a random-intercept for Subject ID) of imaging metrics in unrelated participants (Ns = 4950–6113) with data at both waves (full statistics are provided in the Supplemental Data). Density plots represent the distribution of ICCs across all regions for a given metric. Analyses included a random intercept for site and fixed-effects for age, age², pubertal stage, sex, scanner model, motion, and the corresponding global metric (where applicable)

Data availability

Data used in the preparation of this article were obtained from the Adolescent Brain Cognitive Development^SM (ABCD) Study (https://abcdstudy.org), held in the NIMH Data Archive (NDA). This is a multisite, longitudinal study designed to recruit more than 10,000 children aged 9–10 and follow them over 10 years into early adulthood. The ABCD Study^® is supported by the National Institutes of Health and additional federal partners under award numbers U01DA041048, U01DA050989, U01DA051016, U01DA041022, U01DA051018, U01DA051037, U01DA050987, U01DA041174, U01DA041106, U01DA041117, U01DA041028, U01DA041134, U01DA050988, U01DA051039, U01DA041156, U01DA041025, U01DA041120, U01DA051038, U01DA041148, U01DA041093, U01DA041089, U24DA041123, U24DA041147. A full list of supporters is available at https://abcdstudy.org/federal-partners.html. A listing of participating sites and a complete listing of the study investigators can be found at https://abcdstudy.org/consortium_members/. ABCD consortium investigators designed and implemented the study and/or provided data but did not necessarily participate in the analysis or writing of this report. This manuscript reflects the views of the authors and may not reflect the opinions or views of the NIH or ABCD consortium investigators. The ABCD data repository grows and changes over time. The ABCD data used in this report came from http://dx.doi.org/10.15154/8873-zj65. DOIs can be found at https://nda.nih.gov/abcd/abcd-annual-releases.html. Dataset identifier: http://dx.doi.org/10.15154/dxx6-fk12. Analysis code is available at https://github.com/WashU-BG/PCE_MRI.