Brain network coupling associated with cognitive performance varies as a function of a child’s environment in the ABCD study

Prior research indicates that lower resting-state functional coupling between two brain networks, lateral frontoparietal network (LFPN) and default mode network (DMN), relates to cognitive test performance, for children and adults. However, most of the research that led to this conclusion has been conducted with non-representative samples of individuals from higher-income backgrounds, and so further studies including participants from a broader range of socioeconomic backgrounds are required. Here, in a pre-registered study, we analyzed resting-state fMRI from 6839 children ages 9–10 years from the ABCD dataset. For children from households defined as being above poverty (family of 4 with income > $25,000, or family of 5+ with income > $35,000), we replicated prior findings; that is, we found that better performance on cognitive tests correlated with weaker LFPN-DMN coupling. For children from households defined as being in poverty, the direction of association was reversed, on average: better performance was instead directionally related to stronger LFPN-DMN connectivity, though there was considerable variability. Among children in households below poverty, the direction of this association was predicted in part by features of their environments, such as school type and parent-reported neighborhood safety. These results highlight the importance of including representative samples in studies of child cognitive development.


Supplementary Note 1. Additional demographic and environmental information.
Supplementary Table 1. Wider environmental information. Variables included in the ridge regression predicting cognitive test scores. All except income were used in primary models; additional tests confirmed that income did not add predictive power above and beyond these variables. P-values without correction obtained from two-sided t-tests, calculated using the tableone package in R.

Supplementary Note 2. Identification of environmental variables
In order to identify environmental variables to include in our ridge regression, we began by identifying all measures which were collected for all families at the baseline timepoint of the ABCD study which may characterize children's environments. This included those relating to demographics, neighborhood, school, parenting, and culture. We did not include items more directly related to the child's behavior, like screen time or substance use, nor items more directly related to family members' health and wellbeing.
In general, we aimed to include each of these measures. However, there were several exceptions, as we also wanted to limit the absolute number of measures: 1. When there were measures that were likely to be measuring the same construct, we chose to retain only those variables which previous literature could theoretically link to children's test performance (e.g., "census: median home value" might be better captured by the family's precise combined income and by other neighborhood measures such as "census: income disparity and census: percentage of families below poverty"). 2. Similarly, when the same survey measure of the environment was administered to both parent and child, we chose the child's response over the parent's. 3. When there were multiple variables that could be subsumed under a single summary measure, and we had more reason to believe that this summary was meaningful as opposed to each separate measure, we used the summary measure (e.g., both "parent and parent partner highest level of education" were recoded to indicate "combined highest year of education," as in previous work).
Once we gathered our list of environmental variables, we pre-registered these prior to running any analyses. The purpose of the pre-registration was to ensure that we thought carefully about each variable we selected ahead of time and did not alter the list on the basis of our results. A full table listing each and our use is included below.

Supplementary Note 3. Scatterplots relating resting state metrics and cognitive test performance
For ease of viewing, Figure 2 in the main text displays trend lines of our primary models without the data points. Data points underlying Figure 2 are plotted in Supplementary  Figure 1, below. This figure illustrates the extent of individual variability in the relation, and the sheer number of participants.
Supplementary Figure 1. Scatterplots with data points for relations between resting state network metrics and cognitive test score residuals, after accounting for fixed effects of age and motion and a random effect for study site, for children living above poverty (dark blue) and below poverty (light blue). Trend lines are presented as mean values +/-95% confidence intervals for a linear model, using the geom_smooth function in ggplot. Networks functionally defined using the Gordon parcellation scheme; lateral frontoparietal network (LFPN) shown in yellow, default mode network (DMN) shown in red; figures adapted from 1 and reprinted with permission from the authors.
We also repeated analyses using the NIH Toolbox Fluid Cognition composite, which includes two tests of working memory ( connectivity and NIH-TB fluid abilities composite score residuals, for children living above poverty (dark blue) and below poverty (light blue). Models include fixed effects for age and motion and a random effect for study site. Data are presented as mean values +/-95% confidence intervals for a linear model, calculated and displayed using the geom_smooth function in ggplot.

Supplementary Note 5. Relations between LFPN-LFPN connectivity and cognitive test performance, separated by test
Among the children in poverty, the direction of association between LFPN-LFPN connectivity and each cognitive test were inconsistent, matrix reasoning:

Supplementary Note 6. Bootstrapped distribution of LFPN-DMN connectivity ~ test performance parameter estimates
Our first test was designed to probe how frequently the parameter estimate observed in the children in poverty would be expected to be observed in a larger sample of children living above poverty. In order to derive an estimate for observed parameter estimates in a population of higher-income children, we randomly sampled 500 data points from the children living above poverty, with replacement. For these 500 data points, we fit our primary linear mixed effects model to the data, predicting children's cognitive test scores, and calculated the average parameter estimate for LFPN-DMN connectivity. We repeated this process 999 times, generating a distribution of parameter estimates likely within the larger population of higher-income children from which our participants were drawn.
Next, we compared these bootstrapped parameter estimates to the parameter estimate observed for children in poverty in our sample. If the brain-behavior relation does not differ systematically as a function of poverty status-in other words, if the observed relation between LFPN-DMN connectivity and cognitive test scores for the children in poverty would be likely to be observed in a larger, population-level sample of children above poverty-the parameter estimate for children in poverty should fall within the 95% confidence interval of the bootstrapped parameter estimates.
Thus, we estimated the expected distribution of LFPN-DMN coefficients for the prediction of cognitive test scores among the higher-income children in the dataset. The results of this analysis confirmed that the observed estimate for children in poverty fell outside of the 95% CI, and was higher than 987 out of 999 bootstrapped samples, p = 0.013.
Repeating this bootstrapping procedure for children living below poverty revealed a similar effect. Bootstrapped coefficients ranged from -2.87 to 7.66, with a mean of 2.26, 95% CI [2.16, 2.35]. The observed estimate for children above poverty fell outside of the 95% CI, and was lower than 990 out of 999 bootstrapped estimates, p = 0.010. The bootstrapped distributions from the two samples are plotted side by side in Supplementary Figure 3. Supplementary Figure 3. Bootstrapped distributions for lateral frontoparietal-default mode network (LFPN-DMN) connectivity parameter estimates in the models predicting cognitive test performance, for children above (dark blue) and below (light blue) poverty. Estimates calculated from models run on 500 data points drawn with replacement from each sample separately, repeated 999 times. Bootstrapped coefficients for children above poverty ranged from -6.47 to 3.74, with a mean of -1.41, 95% CI [-1.50, -1.31], mirroring our observed parameter estimate for the higher-income group. The observed estimate for children in poverty fell outside of the 95% CI, and was higher than 987 out of 999 bootstrapped samples, p = 0.013. P-value caculated based on the number of times the estimate was higher for the children in poverty than for the bootstrapped distribution, divided by the number of bootstrapped observations plus one.

Supplementary Note 7. Permutation testing of LFPN-DMN connectivity ~ test performance parameter estimates
To further confirm the dissociation with LFPN-DMN connectivity and test performance for children living above or below poverty, we performed a permutation procedure. This procedure examined the extent to which the model parameters fit in the higher-income children alone could explain the data in the children in poverty.
Briefly, we used the model parameters generated from the higher-income children to predict test performance in the children in poverty, and calculated the mean difference (observed values minus predicted values). We next randomly permuted the labels of each group, such that assignment into the higher-versus lower-income group was now arbitrary. We repeated the process above, now fitting model parameters to our arbitrary higher-income group and using them to predict cognitive test performance in our arbitrary lower-income group. Again, we calculated the mean error between observed and predicted values. This permutation and prediction procedure was repeated 999 times, generating a distribution of mean differences when the distinction between the two groups was arbitrary. If the model parameters generated from our actual higher-income group could reasonably be applied to our actual lower-income group, we would expect that the mean error would fall within the 95% confidence interval of our distribution of permuted mean errors.
The results of this permutation procedure revealed that the model parameters fit on the children above poverty over-estimated the performance of children below poverty, on average (mean difference between observed and predicted test scores = -1.23). To contextualize whether this difference in prediction was larger than what would be expected by chance, we compared this to the distribution of 999 randomly group permuted labels, such that assignment into the higher-versus lower-income group was now arbitrary. The mean difference between the actual groups fell above all differences in the 999 permutations (range = -0.22 -0.22), suggesting this difference is larger than would be expected by chance.

Supplementary Note 8. Relations between LFPN-DMN connectivity and cognitive test performance, for children with low thresholds of motion
Head motion, which is known to influence functional connectivity estimates (Power et al., 2015), differed significantly as a function of poverty status (see Table 1) and was correlated with cognitive test performance (B = -1.91, SE = 0.16, p < 0.001). Our reported analyses use a stringent motion exclusion criteria, in which participants were retained only if they had at least 12.5 minutes of data with low head motion (FD < 0.2 mm). Additionally, there was stringent motion correction in the analysis pipeline, as reported in the main text.
Still, we repeated these analyses with only those children who met a highly stringent motion criterion of less than or equal to 0.2 mm of average framewise displacement (N = 4444; 589 below poverty). Specifically, we fit a linear mixed effects model with site as a repeated measure, testing the relation between cognitive test scores and LFPN-DMN connectivity, controlling for age and head motion. In this subsample of participants who met our threshold for low motion (N = 4444; 589 below poverty), results were consistent, if not stronger. Specifically, for children living below poverty, the main effect of LFPN-DMN connectivity on test scores was positive and significant, B = 4.92, SE = 1.92, t (583) = 2.57;  2 (1) = 6.61, p = 0.010. Children living above poverty, in contrast, showed a negative main effect of LFPN-DMN connectivity, B = -1.27, SE = 0.62, t (3844) = -2.039;  2 (1) = 4.15, p = 0.041. The interaction between poverty status and LFPN-DMN connectivity was significant,  2 (1) = 11.93, p = 0.001 (consistent with the interaction effect in the full sample,  2 (1) = 8.99, p = 0.003). Thus, results were consistent-and seem to be even stronger-in this subsample of lowmotion children.

Supplementary Note 9. Relations between LFPN-DMN connectivity and cognitive test performance, controlling for number of usable frames
A related concern is that our finding was driven by group differences in the number of usable frames of resting state data. Indeed, resting state metrics become more stable with more data 2 . In our data, the number of frames participants contributed after outliers were excluded ranged from 376-2170. We also found that LFPN-DMN connectivity was related to participants' number of usable frames, even when controlling for mean framewise displacement,  2 (1) = 21.23, p < 0.001. However, frames of usable data no longer contributed to model fit when considering participants with relatively more usable frames (top 75% of usable frames, >759:  2 (1) = 2.03, p = 0.154; top 50% of usable frames, >1005:  2 (1) = 1.34, p = 0.247; top 25% of usable frames, >1199:  2 (1) = 1.36, p = 0.244).
To address whether scan length affected our results, we first reran our primary model testing the interaction between LFPN-DMN and poverty status in predicting cognitive test scores, with the additional covariate of number of usable frames after outliers were removed. The interaction between LFPN-DMN and poverty status remained significant, B = 3.14, SE = 1.06, t (6825) = 2.97,  2 (1) = 8.81, p = 0.003, and the number of usable frames did not contribute to model fit above and beyond framewise displacement,  2 (1) = 1.91, p = 0.167. (Framewise displacement continued to contribute significantly to model fit,  2 (1) = 20.43, p < 0.001.) Moreover, the interactive effect remained when restricting analyses to only those participants in the top 75 th , 50 th and 25 th percentiles of usable frames (see Supplementary Figure 4 below for results and associated Ns).

Supplementary Note 10. Relations between LFPN-DMN connectivity and age
Given prior evidence that the LFPN and DMN become less correlated during childhood, we asked whether there was an effect of age in the current study. Indeed, even within this very restricted age range, LFPN-DMN connectivity was lower among two-thirds of the children in poverty predicted the held-out sample of children in poverty at above chance levels in more than 95% of iterations (in a cross-validation framework, this means R 2 > 0). The mean of this distribution was 0.023, 95% CI [0.021, 0.024]. Thus, while our split and model was on the high end of possible model fits, the model did consistently perform above chance.

Supplementary Note 12. Deviations from pre-registration
Both pre-registrations were written before knowing which data from the ABCD study would be available for analysis, which led to some necessary changes upon receipt of the data. In our first pre-registration, we planned to examine the relation between reasoning performance and specific node-to-node connectivity; however, only summary network measures had been released when we conducted our investigation. We also planned to look at test scores longitudinally, but found that only the first timepoint of cognitive assessments had been completed. Thus, we focused our analyses on one of our two primary planned questions. In addition, we planned to run simple linear regressions; these did not take into account the nested structure of the data, which we ultimately addressed in a data-driven fashion using linear mixed effects models, as described in the analysis section of the main text. The nested structure of the data also made our planned cross-validation approach less feasible, and we therefore did not cross-validate this first set of analyses. Finally, we planned to define our poverty threshold based on the Supplemental Poverty Threshold for each study site; however, due to privacy issues with de-identifying study site, we were only able to use a coarser threshold averaging across study sites. In our second pre-registration, we listed three environmental variables that were not collected at the baseline visit: self-reported discrimination, negative life events, and positive life events. Finally, we made the decision to include ethnicity separate from race, as it was collected, to retain maximal information. Otherwise, all analyses were performed as planned.
Several previously unspecified decisions were also made in the analysis process. First, we chose to use raw, rather than age-standardized, cognitive test scores. The rationale was that for using raw scores was that (1) the age range within our sample is relatively tight, (2) brain-behavior relations are of interest within the sample, not in relation to test norms based on a different sample of children, and (3) brain imaging data we are using aren't age normalized. Second, several factor levels within the environmental variables had a very low incidence in the whole sample, with less than 15 participants total (school setting: cyber school; school setting: vocational/tech school; race/ethnicity: Native Hawaiian); these were grouped into the "other" designation for the given factor, to allow for successful cross-validation when the sample was split further. Third, we made the decision to impute missing data from the environmental variables to preserve sample size.