We recruited young-adult participants with lifetime-residential geographies from CHIMGEN (n=3306) and IMAGEN second follow-up (FU2) (n=561). Detailed inclusion and exclusion criteria are presented in Supplementary Tables 1 and 2. Demographics of samples used in statistical analyses and sample attrition are described in Supplementary Table 3 and Supplementary Figs.1-2. Demographic comparisons between the analysed sample and total sample are shown in Supplementary Table 4. Demographic variables showing significant differences between the analysed sample and excluded sample were adjusted during analyses (Online Methods and Supplementary Tables 5-6).
UrbanSat: a satellite-based measure of urbanicity
To develop a satellite-based measure of urbanicity, we selected information from nine types of satellite registrations relevant for detecting and characterising urban settlements, including night-time light (NL), normalized difference built-up index (NDBI), normalized difference water index (NDWI), normalized difference vegetation index (NDVI), and five measures derived from land cover mapping (Built-up%, cropland%, grassland%, forest% and water body%) (Online and Supplementary Methods, Supplementary Table 7). After imputing the nine annual satellite registrations for the participants with missing values using Bayesian data augmentation (Supplementary Table 8 and Online Methods), we carried out a ten-fold cross validation stratified by spatiotemporality to optimize the confirmatory factor analysis (CFA) models, and to predict annual UrbanSat scores of each participant from birth to the age of recruitment (Online Methods). UrbanSat was generated by the optimized CFA model consisting of NL, built-up%, cropland% and NDVI, which best captured variation of urban features while maximizing goodness-of-fit. UrbanSat in CHIMGEN (Supplementary Tables 9-10) and IMAGEN-FU2 (Supplementary Tables 11-12) had a Tucker-Lewis-Index (TLI) and comparative-ﬁt-index (CFI)>0.95, root mean-square-error of approximation (RMSEA)<0.06 and standard root mean-square-residual (SRMR)<0.08, indicating excellent model ﬁt (Online Methods). UrbanSat was robust across time and geographies, as validated by its correlations with ground-level population-density from GHSL-POP9 for China and Europe for the years 1990, 2000 and 2015 (Fig.2). Histograms of the distribution of UrbanSat score in each center are shown in Supplementary Fig.3. UrbanSat showed higher correlations with population-density in different residential categories (rural, town, city and overall), countries (Asia and Europe) and years (1990, 2000 and 2015) than any individual satellite-measures (Fig. 2).
Correlations of UrbanSat with brain structure
Voxel-wise multiple-regression of mean UrbanSat before age 18 with brain gray-matter-volume (GMV) was performed in CHIMGEN (n=2176). We controlled for age in all analyses, thus accounting for the older and wider age spread in CHIMGEN (age: 23.54±2.33 years) compared to IMAGEN (age: 18.89±0.66 years). We also controlled throughout for gender, education, site, body-mass-index (BMI), genetic population-stratification and socioeconomic status (SES) (Supplementary Tables 13-14). Total intracranial volume was controlled in all imaging analyses, except for the analyses of cortical thickness (CT) and surface area (SA), where mean CT and total SA were controlled, respectively. Parental history of mental illness was an exclusion criterion for CHIMGEN and controlled for in IMAGEN. Uncorrected statistical maps of the association of UrbanSat with brain GMV in CHIMGEN adjusting for confounding covariates under parametric testing and non-parametric permutation testing are shown in Fig.3a and Supplementary Fig. 4. We found negative correlation of UrbanSat with medial prefrontal cortex (mPFC) volume (peak MNI-coordinate: x=-7.5, y=30, z=45; 676 voxels; peak t-value=-6.42; Fig.3b) and a positive correlation with cerebellar volume (peak MNI-coordinate: x=10.5, y=-51, z=-18; 978 voxels; peak t-value=6.22; Fig.3b) (Parametric testing Pc<0.05, family-wise error (FWE) corrected for voxel numbers, imaging modalities and data categories; see Online Methods). We confirmed the results with non-parametric permutation-testing (TFCE-FWE, Pc<0.05, Supplementary Methods and Supplementary Fig.4). Potential imputation-bias was ruled-out by sensitivity-analyses in 1491 participants with complete satellite and neuroimaging data (Supplementary Fig.5 and Supplementary Results). Uncorrected and adjusted vertex-wise correlation maps of UrbanSat with whole-brain CT and SA are shown in Supplementary Fig.6 (n=2164). The mPFC-region of interest (ROI) from GMV-analyses was projected onto fsaverage surface of Freesurfer v5.3.0 (Supplementary Fig.7). UrbanSat was correlated with mean SA (rho=-0.07, P=8.12×10-4) but not mean CT (rho=-0.02, P=0.28) of the mPFC cluster (Fig.3b and Supplementary Table 15). Using voxel-wise multiple regression of individual satellite- measures with GMV, we found significant correlation with mPFC-GMV being driven by NL and built-up%, and correlation with cerebellar-GMV being driven by NL and cropland%. We found no correlation of NDVI with either mPFC- or cerebellar-GMV (Supplementary Fig.8). We observed similar results in GHSL ground-level population-density data, thus validating the relation of UrbanSat and GMV (Supplementary Fig.8).
In IMAGEN-FU2 (n=415), we replicated CHIMGEN findings. The uncorrected statistical correlation map of UrbanSat with brain GMV in CHIMGEN showed significant spatial correlation (r=0.40, P<0.001) with that of IMAGEN-FU2 (Supplementary Fig.9). UrbanSat was correlated with GMVs of the mPFC (rho=-0.20, P=4.49×10-5) and cerebellum (rho=0.11, P=0.03), and SA of the mPFC (rho=-0.15, P=3.01×10-3), but not CT of the mPFC (rho=-0.03, P=0.58) (Fig.3b and Supplementary Table 15). In voxel-wise analyses, we validated negative correlation of UrbanSat with mPFC volume and positive correlation with cerebellar volume (Supplementary Fig.10), both driven by NL and built-up% (Supplementary Fig.11).
To exclude possible scanner and site effects, we performed separate analyses for each acquisition-site of CHIMGEN and IMAGEN-FU2, carrying out a meta-analysis with an inverse variance-weighted random-effects model (Online and Supplementary Methods). UrbanSat remained significantly negatively correlated with mPFC-GMV and SA, and positively correlated with the cerebellar-GMV (Supplementary Fig.12 and Supplementary Table 16). Heterogeneity of effect sizes was from low to moderate for all regions (I2-range: 0.19%-41.35%) (Supplementary Table 16). Thus, the observed correlation between UrbanSat and brain structure is robust across geographies and socio-cultural conditions.
We applied distributed lag models (DLMs) to identify susceptibility-periods of lifetime UrbanSat on GMV and SA (Online Methods). We observed a negative association of UrbanSat with mPFC-GMV from age 4 to 15 (Fig.3c) and SA from age 5 to 7 (Fig.3c), indicating a susceptibility-period during childhood and adolescence, driven by NL, built-up%, cropland% and NDVI (Supplementary Fig.13). Correlation of UrbanSat with cerebellar GMV was significant from age 1 to 10 years, indicating a susceptibility-period during childhood (Fig.3c), driven by NL, built-up% and cropland% (Supplementary Fig.13).
To investigate the relation between UrbanSat and brain development, we used the longitudinal IMAGEN dataset to calculate volumetric (n=340) and SA/CT change-rate/year (n=325) between baseline (BL) at 14 and FU2 at 19 years (IMAGEN BL-FU2). Consistent with the susceptibility periods identified, UrbanSat was significantly correlated with brain volumetric development in the mPFC-ROI (rho=0.17, P=2.10×10-3), but not with the cerebellum-ROI (rho=0.02, P=0.70). This correlation was driven by mPFC SA changes (rho=0.24, P=1.31×10-5), not by CT changes (rho=-0.01, P=0.93) (Supplementary Table 15).
To measure the relation between age of migration and brain structure, we split CHIMGEN participants into who migrated to the city before age 14 years (n=229, mean-age at migration=8.24±4.86 years), after age 14 (n=1385, mean-age at migration=17.17±2.68 years), and life-long city-dwellers (n=562) (Fig.3d). We found that participants born in the city or early migrants showed smaller mPFC-GMV (P=0.040) and SA (P=7.28×10-9) as well as greater cerebellar-GMV (P=5.00×10-5) than those with later exposure (Fig.3e and Supplementary Table 17).
No correlation of UrbanSat with white-matter microstructure
Using tract-based spatial statistics (TBSS) analysis of diffusion-tensor imaging (DTI) data, we did not find significant correlation of UrbanSat with brain fractional anisotropy (FA) in either CHIMGEN or IMAGEN-FU2 (TFCE-FWE Pc <0.05).
Correlations of UrbanSat with resting-state functional network connectivity
Using group-independent-component-analysis (GICA) of estimated 30 independent components (Supplementary Methods), we identified 17 resting-state networks (RSNs) related to cognitive and sensory-motor processes13 in both CHIMGEN and IMAGEN (n=2156) (Supplementary Fig.14). For each RSN, we tested the relation between mean UrbanSat and within-network functional connectivity (WNFC). A voxel-wise multiple-regression analysis controlling for all confounders revealed a negative correlation of UrbanSat with WNFC in the mPFC of the anterior default-mode-network (aDMN) (peak MNI-coordinate: x=-3.5, y=69, z=0; 142 voxels; peak t-value=-6.96), and positive correlations in the cerebellar vermis of the cerebellar-network (CN) (peak MNI-coordinate: x=3, y=-72, z=-9; 122 voxels; peak t-value =7.21), in the left lingual gyrus (LG) of the medial visual-network (mVN) (peak MNI-coordinate: x=-12, y=-90, z=3.5; 143 voxels; peak t-value=6.97) and in the left LG of the lateral visual-network (lVN) (peak MNI-coordinate: x=-24, y=-81, z=-12; 141 voxels; peak t-value=6.97) (FWE Pc<0.05, additionally corrected for 17 RSNs, Online Methods) (Fig. 4a). Voxel-based correlations of individual satellite-measures with WNFCs of each RSN are shown in Supplementary Fig.15. The correlations of UrbanSat with WNFCs in CHIMGEN were replicated in ROI-based analyses in IMAGEN-FU2 (n=351) (aDMN: rho=-0.18, P=7.20×10-4; CN: rho=0.21, P=1.41×10-4; mVN: rho=0.24, P=1.05×10-5; lVN: rho=0.19, P=3.97×10-4) (Fig.4b and Supplementary Table 15). Only the aDMN and CN results were replicated in voxel-wise analyses in IMAGEN-FU2 (Supplementary Fig.16).
In 136 between-network functional connectivity (BNFC), UrbanSat was correlated with 49 BNFCs in CHIMGEN (Pc<0.05, 10,000 permutations, see Online Methods) (Fig.4e), four of which were replicated in IMAGEN-FU2 (Fig.4f). These four BNFCs (aDMN-CN, aDMN-ECN, aDMN-rFPN and rFPN-lFPN) connect five brain functional-networks (aDMN; CN; executive-control-network (ECN); rFPN/lFPN, right or left frontoparietal-network (FPN)), implicated in self-referential thoughts14 and executive control15.
The correlations of UrbanSat with WNFCs and BNFCs were stable in a meta-analysis of all CHIMGEN and IMAGEN sites (Supplementary Fig.12 and Supplementary Table 16). Brain localization (Fig.4b and Fig.4f) and susceptibility- periods (Fig.4d and Fig.4h) of WNFCs and BNFCs in CHIMGEN and IMAGEN were consistent with those observed for brain structure (Supplementary Fig.15-17), except for non-significant during adolescence. The WNFCs and BNFCs changes between 14 and 19 years in IMAGEN were correlated with UrbanSat (Fig.4b and Fig.4f and Supplementary Table 15 and Supplementary Results). In CHIMGEN, WNFCs and BNFCs were correlated with age of migration to the city (Fig. 4c and Fig.4g and Supplementary Results).
Correlations of UrbanSat with behavior
We investigated whether UrbanSat is related to measures of cognition and mental- health, i.e. depression and anxiety. In CHIMGEN (n=2148), the social-cognition measure ‘perspective-taking’, perceiving a situation from an alternative point of view16, was positively correlated with UrbanSat (reaction-time for perspective-taking: rho=-0.14, Pc<0.05, Bonferroni-corrected for data-categories and 21 behavioral assessments, see Online Methods) and replicated in IMAGEN-FU2 (rho=0.14, Pc<0.05) (Table 1). A negative correlation between UrbanSat and reaction-time for perspective-taking performance was observed from 12 to 22 years in CHIMGEN (Fig.5a).
UrbanSat was correlated with depression-symptoms assessed by Beck-Depression-Inventory (BDI) in CHIMGEN (n=2170) (rho=0.14, Pc<0.05) (Table 1) with a susceptibility-period from 3 to 12 years (Fig. 5a). As BDI was not available in IMAGEN, we validated this association using an instrument measuring core features of depression, the Ruminating-Scale-Questionnaire (RSQ) (rho=0.14, Pc<0.05) (Table 1 and Supplementary Methods).
In CHIMGEN and IMAGEN, increased NL and built-up%, decreased NDVI and cropland% were significantly correlated with enhanced perspective-taking performance and increased depression-symptoms (Table 1). The susceptibility-periods for individual satellite-measures were similar to UrbanSat in CHIMGEN (Supplementary Fig.18). Although most correlations of UrbanSat with brain and behaviour were consistent between males and females, some correlations, especially with brain development in IMAGEN BL-FU2, were sex-specific (Supplementary Tables 19-20).
Multiple mediation in UrbanSat-brain-behavior
We applied multiple-mediation analysis to investigate if the significant brain imaging measures mediate correlations of UrbanSat with perspective-taking and depression-symptoms in CHIMGEN and IMAGEN-FU2 (Online Methods). In CHIMGEN, 19.32% of the correlation between UrbanSat and reaction-time for perspective-taking was mediated by brain, namely mPFC-GMV (2.55%), the cerebellar-GMV (2.94%), WNFCs in aDMN (2.89%) and CN (3.21%), as well as by the BNFCs of the aDMN-CN (2.27%), aDMN-ECN (4.54%) and aDMN-rFPN (4.14%) (Fig. 5b). Mediation was replicated in IMAGEN-FU2, with the association of UrbanSat with perspective-taking being mediated by the mPFC-GMV (1.47%), WNFCs in the aDMN (1.96%) and CN (0.89%) as well as BNFCs of the aDMN-CN (1.45%) and aDMN-rFPN (1.22%) (Fig.5b). There was no mediation of the cerebellar GMV in IMAGEN-FU2 (Supplementary Table 21).
In CHIMGEN, 20.32% of the correlation between UrbanSat and BDI was mediated by brain, namely mPFC-GMV (4.81%) and SA (1.80%), cerebellar-GMV (6.88%), WNFCs in aDMN (2.45%) and mVN (2.18%), BNFC of aDMN-ECN (4.04%) (Fig.5c). In IMAGEN-FU2 the correlation between UrbanSat and rumination was mediated by mPFC GMV (1.93%), WNFC in aDMN (0.96%) and BNFC of the aDMN-ECN (1.62%) (Fig.5c), but not by the cerebellar-GMV (Supplementary Table 21).