Task and resting state fMRI modelling of brain-behavior relationships in developmental cohorts

Abstract

Functional magnetic resonance imaging (fMRI) data are often used to inform individual differences in cognitive, behavioral, and psychiatric phenotypes. These so-called “brain-behavior” association studies come in many flavors and are increasingly the focus of investigations utilizing large population neuroscience datasets. Still, many open questions surrounding the utility of task and resting state fMRI for modelling brain-behavior relationships remain, including the feasibility of conducting these investigations in developmental cohorts. With the growing availability of large neurodevelopmental datasets such as that provided by the Adolescent Brain Cognitive Development (ABCD) Study, we are now able to conduct well-powered analyses using large samples of longitudinal neuroimaging data collected from diverse populations of youth. Here we provide a high-level review of current controversies and challenges in this growing subfield of neuroscience, highlighting examples where task fMRI data and resting state fMRI data - either in isolation or combined - have yielded significant insights into brain-behavior associations. Challenges include issues related to measurement noise, appropriate estimation of effect sizes, and limits to generalizability due to insufficient diversity of samples. Innovative solutions involving advanced MRI data acquisition protocols, application of multivariate analysis methods, and more robust consideration of phenotypic complexity are reviewed. We propose that additional future directions for developmental cognitive neuroscience should include more reliable behavioral measures, multimodal neuroimaging brain-behavior studies, and greater consideration of environmental and other contextual influences on brain-behavior associations.

Controversies and challenges in modelling brain-behavior relationships

Brain-behavior association studies involve exploring relationships between a metric derived from fMRI data (e.g. magnitude of brain activation during a task, or functional connectivity of a region-of-interest in the resting state) and a metric derived from cognitive task performance or clinical phenotype (e.g. performance on an attention task, or severity of depression symptoms). The assumption underlying these analyses is that a high correlation between a brain metric and a phenotype gives us insight into the neurobiology underlying complex human cognition and behavior.

Early critiques of brain-behavior association studies focused on the “puzzlingly high” correlations often reported. Specifically, some analyses presented in early task fMRI studies could be considered circular in that they selected brain voxels for analysis that were already shown to be highly correlated with the cognitive/behavioral measure of interest (1). As reliability estimates for brain-derived metrics and cognitive/behavioral measures have upper limits, brain-behavior correlations that exceed those limits are likely to be spurious. Increasing the reliability of both neural and psychological measurements has been proposed as a means for optimizing the detection of between-person effects in neuroimaging studies of individual differences (2). Questionable assumptions about the nature of psychological phenomena under investigation (e.g. functional localization, one-to-one mapping of neural ensembles to psychological categories, and independence from contextual factors such as the rest of the brain or environment) can also contribute to failure to replicate brain-behavior relationships (3).

Large population neuroscience datasets are now being utilized to conduct well-powered analyses (4). A study analyzing nearly 50,000 MRI datasets from the Adolescent Brain Cognitive Development (ABCD), Human Connectome Project (HCP), and UK Biobank recently claimed that thousands of subjects are needed to arrive at reproducible brain-behavioral phenotype associations using univariate analytic approaches (5). The authors examined brain metrics including cortical thickness and resting state functional connectivity to estimate relationships with multiple behavioral phenotypes in the cognition and mental health domains and argue that small brain-wide association effects and population sampling variability can result in inflated and irreproducible brain-phenotype associations in sample sizes smaller than thousands of subjects. Another type of variability that can exacerbate these issues is that different researchers may select different brain derived features to relate to phenotypes. Further, variation in analytic pipelines can have large effects on results (6). A recent systematic evaluation of 768 fMRI data processing pipelines concluded that most of them failed to meet criteria for producing consistent results (7). There is indeed a great deal of dependence on user-defined criteria inherent to all analyses of this type (8). From data preprocessing to selection of neuroimaging-derived features to include in modelling brain-behavior relationships, there are multiple methodological choice points (Figure 1). This analytic flexibility (6) can make it difficult to determine which brain-behavior associations are robust.

Figure 1. fMRI data analysis workflows and methodological decision points.

Researchers must make methodological choices at every step of the fMRI data analysis workflow, from preprocessing (which involves the steps illustrated below, often combined with additional data denoising steps) to statistical analyses (which can include univariate and multivariate approaches) to selection of neuroimaging-derived features with which to model brain-behavior associations.

Multivariate approaches are increasingly used in brain-behavior association studies. These types of prediction algorithms can produce highly replicable results with samples as small as 100 subjects (9), highlighting that sample size estimation depends on the analytic method being employed and the effect size of the brain-behavior relationship in question (10). Some have proposed that within-subject designs provide another means for maximizing statistical power (11). Reproducible brain-behavior associations might also be more readily observed when targeted fMRI tasks are used in studies of specific clinical populations (12).

Here we selectively review studies that have used fMRI data to probe developmental brain-behavior associations. As we are not aiming to present a comprehensive review of brain-behavior associations for every phenotype, we focus on a few key domains of particular interest for developmental cognitive neuroscience - executive function, reward processing, and social cognition. We discuss methodological and theoretical issues, pointing out strengths and weaknesses of various approaches. In doing so, we highlight some of the major open challenges and ongoing controversies in the field. We conclude by suggesting potential future directions for increasing the reliability and reproducibility of brain-behavior investigations in developmental cohorts.

Task fMRI brain-behavior relationships during development

Research on child and adolescent development often focuses on the maturation of cognitive control and reward-related processes. The interplay between the two, specifically the notion that maturation of cognitive control processes is delayed relative to reward and motivation-related processes, forms the basis of dual-systems theories of adolescent development (13–15). Although there is evidence for age-related improvements in prefrontally-mediated cognitive control (16–19), variation in the patterns of age-related activation increases and decreases and evidence that motivational systems (social and affective factors) may drive the magnitude of inhibitory control observed combine to undermine a simple model of linear improvements related to frontal cortex maturity (20). Findings of age-related changes in reward processes are also mixed (19,21,22). Compared with adults, adolescents show both reduced and increased activity in the ventral striatum during reward anticipation (21). Discrepant findings may reflect a mixture of task paradigms, small samples, and cross-sectional designs. Exploiting the large, longitudinal sample of the IMAGEN study (23), Cao and colleagues reported brain activation during reward anticipation to show a mix of activation decreases in subcortical areas and activation increases in frontal regions (24).

Beyond these topics, the literature includes investigations of reasoning (25), working memory (26), error signaling (27), episodic memory (28), emotional regulation (29), emotional processing (30), and vicarious reward processes (31), to name a few. Task fMRI has utility for studying theories of neurodevelopment such as dual-systems theories, as they typically implicate specific psychological processes that can be engaged by well-designed tasks. Relatedly, task-based data typically yields a performance measure enabling individual differences and developmental trajectories in brain-behavior associations to be quantified. Numerous task designs exploit the availability of real-time behavioral measures to adaptively adjust task design or difficulty to accentuate performance abilities (32).

The changing relationships across development between brain and behavior on cognitively demanding tasks creates complexity in interpretation. Obtaining valid, sensitive, and reliable measures of a participant’s abilities and their performance during scanning is critical to understanding the neurobiological underpinnings of normative development and individual differences (33). Convergence from tasks completed outside of the scanner is also important. Ideally, the reliability of in-scanner task performance would be confirmed with more extensive out-of-scanner cognitive assessments, thereby demonstrating the generalizability of any observed brain-behavior associations.

Head motion is a challenge to the validity of neuroimaging data and especially problematic for developmental studies, as age tends to be inversely related to motion. Task fMRI measures are less sensitive to these effects, as they rely on aggregations of trial-related activity and subtractions between conditions rather than the inter-regional correlations which can be artifactually inflated by motion (34). However, task connectivity analyses are a relatively underutilized data resource. Many of the sophisticated analytic methods that have been developed in the last two decades to characterize resting state functional connectivity can be applied to task data. For example, Cao and colleagues showed increases between ages 14 and 19 in functional connectivity between the ventral striatum and frontal, temporal, parietal and occipital regions (24). In addition, graph theory metrics showed higher inter-regional connectivity and topological efficiency at age 19. Poon and colleagues reported that the strength of prefrontal-limbic connectivity when adolescents viewed negative emotional images was predictive of emotional regulation difficulties a year later, and these difficulties mediated associations with future internalizing and externalizing symptoms two years later (35). Exploiting both task and resting state data, Lee et. al. showed that the ability to reconfigure the reward network (i.e., the extent to which it changes during task compared to rest) increased with age (36).

Task fMRI data offer several options for characterizing brain connectivity (37). For example, connectome-based predictive modeling (CPM) (38) creates edges (pairwise correlations, typically between the timeseries data of regions-of-interest). Feature selection and cross-validation approaches identify the edges that are related to phenotypes of interest, and have shown notable success in explaining individual differences among adolescents in alcohol use based on a response inhibition task’s connectivity (39) and in cannabis use based on a reward task’s connectivity (40). Alternative approaches focus on connectivity among task timeseries data that have had task-related signals removed (thereby approximating the task-free data typically preferred in resting state analyses) or that isolate connectivity specific to the task-induced signal (41,42). Zhao et. al. show that task-related connectivity may be superior for capturing individual differences that are either engaged by a task or that are similar in their cognitive demands (41). Other approaches such as psychophysiological interaction (PPI) (43) or beta-series correlation (44) reveal how functional connectivity differs across task conditions, affording further insights into brain-behavior associations (45).

Chief among concerns with task fMRI may be reliability (46,47). Many task designs yield inadequate signal (e.g., too few trials) to properly characterize a participant’s abilities and their underlying brain signature. The attentiveness which the resting state fMRI community has shown to the importance of scan duration has not been matched within the task fMRI community. Tasks are often chosen for their ability to yield robust group-activation maps, which may make them especially unsuitable for capturing individual differences (48,49) or individual-level developmental trajectories. To inform the investigation of brain-behavior associations, tasks typically yield individual differences in performance, and while this is a strength, it can create analytic challenges. Inadequate specification of behavioral differences in the MRI timeseries analyses (e.g., the inclusion of regressors to capture trial-specific response times) can lead to imprecise and biased activation measures (50). Timeseries analyses of the MRI signal need to be suitably flexible to accommodate intra- and inter-individual differences, including developmental differences in hemodynamic responses (e.g., utilizing a range of basis functions or empirically derived hemodynamic response functions).

Tasks are often selected with no more than face validity as probes of psychological processes, and their utility in exposing theoretically useful brain-behavior associations is limited by how successful they are at isolating the process of interest. Computational modeling approaches to task data offer an important path forward. Computational models that clearly specify and parameterize the psychological processes involved in task performance can yield more mechanistic insights into brain-behavior associations (51,52). For example, Weigard and colleagues applied a drift diffusion model to performance on the working memory EN-Back task of the ABCD Study (53,54). An efficiency of evidence accumulation (EEA) factor was used to distinguish performance that was related to stimulus familiarity (e.g., responses to lure stimuli that were previously seen but were not targets) from performance related to goal-relevant stimulus features (i.e., familiar stimuli that were targets). The EEA parameter was highly correlated across levels of the task, showed robust developmental differences, high test-retest reliability, and was associated with clinical measures.

Finally, multivariate approaches also hold substantial promise. By incorporating broadly distributed activation measures, multivariate analyses often show improved brain-behavior associations and better reliability than their univariate counterparts (55). Based on over 9,000 participants performing the EN-Back working memory task from the ABCD Study, Baranger and colleagues (56) constructed a neural signature of working memory. This multivariate classifier sought to optimally discriminate activation during the 2-back and 0-back conditions and derived a single score per participant indicating how well each individual’s activations differed between the two conditions. This neural signature was more reliable and better correlated with EN-Back task performance, out-of-scanner cognitive measures, and psychopathology symptom scores compared with standard measures of regional brain activation. Multivariate approaches can also accommodate the rich phenotypic characterization of participants that is often available in large neurodevelopmental datasets. Utilizing canonical correlation analysis (CCA), which seeks to identify associations between sets of multiple variables, Kohler and colleagues showed that a latent dimension of reward that incorporated negative urgency and delay discounting behaviors could be linked to covariation between striatal, thalamic, and anterior cingulate responses during reward anticipation (57).

Resting state fMRI brain-behavior relationships during development

As rsfMRI data acquisition is relatively straightforward - participants are asked to do nothing for a relatively short period of time - it is increasingly used to collect neuroimaging data from pediatric and otherwise difficult-to-scan populations (58). rsfMRI allows researchers to sidestep the problem of complex relationships between changes in task fMRI brain activation related to performance differences versus those related to brain maturation (59). Over the past decades, the availability of large rsfMRI datasets (Table 1) has enabled brain-behavior association research in pediatric neuroimaging to flourish.

Table 1.

Normative developmental neuroimaging data collection and sharing initiatives

	Sample size	Age Range (years)	Data Collected
Philadelphia Neurodevelopmental Cohort (PNC)	1,445	8–21	Neuroimaging: sMRI, DWI, ASL, tfMRI, rsfMRI Other: Kiddie-SADS; Penn CNB
Adolescent Brain Cognitive Development (ABCD)	11, 878	9–10 at baseline, follow up every two years	Neuroimaging: sMRI, DWI, tfMRI, rsfMRI Other: Kiddie-SADS and other clinical measures, NIH toolbox, physical, cultural, biological measures
Human Connectome Project (HCP) Lifespan Children	1,350	5–21	Neuroimaging: sMRI, multishell DWI, tfMRI, rsfMRI Other: extensive battery of social, behavioral, and neurocognitive measures
Enhancing Neuro Imaging Genetics through Meta Analysis (IMAGEN)	> 2,000	14, follow up at 16, 19, 22	Neuroimaging: sMRI, DWI, tfMRI, rsfMRI Other: extensive battery of social, behavioral, and neurocognitive measures
Nathan Kline Institute-Rockland Sample 1	1,500	6–85	Neuroimaging: sMRI, DWI, tfMRI, rsfMRI Other: extensive battery of social, behavioral, and neurocognitive measures
Pediatric Imaging, Neurocognition, and Genetics (PING)	> 1,000	3–20	Neuroimaging: sMRI, DWI, rsfMRI Other: extensive battery of social, behavioral, and neurocognitive measures

sMRI = structural MRI; DWI = Diffusion Weighted Imaging; ASL = Arterial Spin Labeling; tfMRI = task fMRI; rsfMRI = resting state fMRI; Kiddie-SADS = Kiddie-Schedule for Affective Disorders and Schizophrenia; Penn CNB = Computerized Neurocognitive Battery

The idea that neurodevelopment is characterized by processes of both integration and segregation across large-scale functional brain networks has been around since the early days of rsfMRI (60). These maturational processes are thought to be related to the weakening of short-range functional connections and strengthening of long-range functional connections in the brain (61). Understanding the impact of these neurodevelopmental changes on cognition and behavior has been a major focus of brain-behavior association research.

The last few decades of pediatric rsfMRI research have revealed that atypical functioning of the salience network (62) (with key nodes in anterior cingulate and anterior insular cortices) and the default network (63) (with key nodes in posterior cingulate and medial prefrontal cortices) is a transdiagnostic feature of many neurodevelopmental disorders (64) (Figure 2). Brain-behavior associations studies are starting to provide insights into why this might be the case. Across two pediatric neuroimaging datasets collected in China (n = 202) and the United States (n= 2,186), a study found that salience/ventral attention network functional connectivity in children relates to broad measures of intellectual functioning and cognitive ability (65). This association is an example of a robust finding that was replicated in an independent sample. Another cross-sectional investigation in the Philadelphia Neurodevelopmental Cohort (PNC) sample spanning ages 8–22 (n = 754) found that in both resting state and working memory task fMRI data, age-related increases in eigenvector centrality could be observed in the cingulo-opercular network. This centrality measure, which reflects the number and strength of connections of a given node to the rest of the network, was positively associated with working memory task performance (66). Another study of 6–17-year-olds (n=119) found that resting state functional connectivity features derived from the whole brain can be used to predict valence bias - individual differences in propensity toward positive or negative appraisals of ambiguity. Interestingly, that study also found that the cingulo-opercular network was a significant contributor to this prediction (67). Note that there is significant spatial overlap between what some researchers have called the ‘cingulo-opercular network’ and others have called the ‘salience/ventral attention network’ (68). This suggests that networks involving the anterior insula and anterior cingulate cortex (69,70), regardless of what they are named (71,72), are critically important in typical and atypical development.

Figure 2. Anatomy of core large-scale functional brain networks.

Resting state functional connectivity is often used to parcellate the brain into distinct functional networks. One of the most widely-used functional atlases is the Yeo parcellation (106) that divides the cortex using a clustering approach to identify functionally coupled brain regions (107).

Another study using PNC rsfMRI data (n=553) showed that those with higher reading achievement had stronger lateralization of frontoparietal networks, a brain-behavior relationship that was mediated by cognitive control abilities. No significant interaction with age was observed, suggesting stability of this specific brain-behavior relationship across adolescent development (73). A few studies have begun to test whether brain-behavior associations derived within a particular age cohort generalize to individuals outside that age range. Using rsfMRI and structural MRI metrics derived from the CamCAN dataset (n=550) divided into young, middle, and older age-groups, Yu and colleagues constructed prediction models for behavioral outcomes within specific age groups. Using predictive models formed from data collected from younger participants, behavioral outcome predictions were most accurate in the younger subjects, with lower accuracy for the middle-aged and older groups. When predictive models were constructed using data collected from older participants, they did not observe significant differences in predictive accuracy when applied across age groups (74). This suggests an asymmetry whereby brain-behavior associations in older individuals may generalize to younger individuals, but not the other way around. Another study of 4–55 year old participants (n=468) used CPM to predict working memory capacity from rsfMRI data and showed that different functional connectivity features predict working memory in different age groups, with the worst overall performance in the younger groups (75). Finally, a study using rsfMRI data provided by the Enhanced Nathan Kline Institute - Rockland Sample (age 6–85, n =724) found that the association between brain signal variability and executive function differs as function of participant age. The association between brain signal variability in the executive control network and neuropsychological assessments of executive function was positive in adolescence, but negative in older adulthood (76). Taken together, these studies suggest that brain-behavior associations can be age-specific.

The availability of the ABCD dataset has inspired multiple recent developmental brain-behavior association studies. Using a “cognition composite” score derived from the NIH toolbox cognition battery, Marek and colleagues found an association between default mode network-dorsal attention network (DMN-DAN) resting state functional connectivity and higher general cognitive ability (77). Chang and colleagues similarly found that increased DMN-DAN anticorrelation was related to lower intraindividual variability in task performance and reduced attention problems (78) (see also (79)). Other ABCD results suggest that cognitive flexibility performance is associated with DMN and frontoparietal network functional connectivity (80).

Multivariate methods have yielded significant insights beyond those achieved using univariate approaches. CCA applied to multiple brain metrics and phenotypic variables in the ABCD dataset revealed replicable associations between neighborhood environment, parental characteristics, family life quality, physical health, cognition, and psychopathology and brain measures (morphometric, myelination, white matter integrity, and resting state functional connectivity) (81). Thapaliya and colleagues used rsfMRI-derived functional network connectivity matrices to predict intelligence in the ABCD dataset (82).

One of the most well-studied brain-behavior topics is the development of psychopathology. One study by Royer and colleagues found that a general psychopathology factor explained the most covariance among multimodal neuroimaging features, while internalizing, externalizing, and neurodevelopmental dimensions were associated with distinct brain morphology and functional connectivity profiles (83). Another by Xiao and colleagues computed a latent functional connectome pattern that could predict the transition of diagnosis across disorders at longitudinal follow-up (84). Yet another focused specifically on internalizing symptoms, finding that around 10% of the variance in these scores could be explained by resting state functional connectivity in DMN, DAN, and cingulo-parietal networks (85). Taken together, these results suggest a myriad of forms of brain-behavior associations in the domains of cognition and psychopathology that can be meaningfully derived from rsfMRI.

Summary and Future Directions

Which is better, task or rest?

For the purposes of examining brain-behavior associations, which is better: task or rest? Whole-brain network architecture computed from resting state and task fMRI data is highly similar (86), although task-rest differences in network topology have been noted in studies examining graph metrics (87). With limited resources, researchers are often faced with the decision of whether to include more task or more rsfMRI in their neuroimaging protocols. The answer to this depends largely on the questions of interest. In the context of maximizing brain-behavior prediction accuracy, a recent comprehensive examination determined that 30 minute scans are the most cost-effective, with task fMRI enabling better prediction than rsfMRI over the same scan duration (88). This aligns with earlier work demonstrating that predictive models generated from task fMRI data outperform models built from rsfMRI data when explaining variance in individual phenotypes such as fluid intelligence (89).

A substantial strength of rsfMRI is that it affords great flexibility in the selection of regions and networks for analyses, including post hoc investigations and is thus suitable for data-pooling and secondary analyses. In contrast, task fMRI requires a commitment to probing a specific function. While a task fMRI paradigm will be optimal for elucidating the neurobiology underlying specific cognitive processes (and this may generalize to related processes), rsfMRI may have greater utility for identifying brain-behavior associations in a broader range of phenotypes. Other designs such as movie watching may strike a balance by engaging a wide range of psychological processes suitable for numerous post hoc enquiries (90).

To obtain a longer fMRI timeseries, some studies have concatenated task and rsfMRI data (91). Overall, the field stands to benefit from collection of both task and rsfMRI data going forward. These modalities can potentially provide complementary information about brain function when analyzed separately, while at the same time could be meaningfully combined in some analytic approaches where longer timeseries are required to increase statistical power.

A roadmap for the future

Improve reliability and validity

Although much of the recent discussion surrounding the reproducibility of brain-behavior associations has focused on sample size, measurement reliability is also critically important (Table 2). A recent study demonstrated the impact of low reliability of phenotypes using UK Biobank data. The authors argue that increases in sample size from hundreds to thousands of participants can only improve brain-behavior predictions when applied to highly reliable data (92). An important consideration for any future data collection efforts will be to make every effort to only include the most reliable behavioral measures. For existing large datasets available to the research community, pre-registration of secondary analyses and the use of cross-validation or other replication schemes can greatly help in discovering reproducible effects. Improved brain-behavior associations may result from creating more valid measures of cognitive abilities through, for example, utilization of out-of-scanner task batteries and computational models of task abilities which may yield measures of more elementary cognitive processes that are more closely related to brain function. Additionally, ongoing efforts to extract the most meaningful signals from the MRI data and timeseries analyses remain a high priority.

Table 2.

Summary of ongoing controversies and potential future directions for the field

Ongoing Controversies	Potential Future Directions
Measurement noise in assessing phenotypes	Improve reliability of cognitive and behavioral measures
Statistical power and estimation of effect sizes	Employ multivariate analysis methods
Limits to generalizability	Collect data across the globe to improve diversity of samples
Impacts of methodological choices	Employ multiverse data analysis strategies to evaluate multiple pipelines
Impact of environmental factors on brain-behavior associations	Incorporate measures that constitute the “urbanome”

Improve diversity of samples

Insufficient diversity of samples can also hinder progress in brain-behavior prediction. For example, when rsfMRI data collected from only white participants is used to generate predictive models, prediction of phenotypes for Black participants is weaker (93). Understanding the full extent of neural variation associated with sociodemographic factors will be necessary to prevent this type of generalization failure (94). Large population neuroscience datasets of neurodevelopmental cohorts are now being created and released by countries outside the United States such as China (95). Leveraging these datasets sampling the global population will be an important future direction for combined and stratified brain-behavior association studies.

Incorporate multimodal neuroimaging

Studies of adults demonstrate that integrating across structural and functional neuroimaging modalities may boost prediction accuracy of cognitive abilities (96,97). Less progress has been made with respect to multimodal developmental neuroimaging. As pediatric structural neuroimaging datasets are more prone to motion artifacts that can lead to systematic biases in measurements of cortical thickness and surface area (98), manual quality control interventions may be required to ensure detection of meaningful brain-behavior associations when combining neuroimaging modalities.

Improve evaluation of impacts of methodological choices

The impact of different data preprocessing pipelines can dramatically alter resting state functional connectivity results in ways that can impact brain-wide association studies (99). Ongoing methodological work evaluating the impact of brain parcellations on prediction of behavior suggests that different types of parcellations might perform better for adult versus developmental datasets (100). Along these lines, individual-specific parcellations have been shown to slightly outperform group-level parcellations for prediction of behavior from resting state functional connectivity (101). There is clearly the need for the methodological advances in brain-behavior prediction that have been gained from studies of adult cohorts to be extended and applied to developmental cohorts in future studies. One potential path forward is the “multiverse” strategy, whereby multiple data analysis pipelines are evaluated on a dataset simultaneously (102). While this approach can be computationally intensive, it offers a means for evaluating consistency and robustness of results.

Improve incorporation of environmental factors

The influence of environmental context on neurodevelopment is well documented. The ABCD dataset includes a range of environmental variables that constitute the “urbanome”, encompassing important factors such as traffic patterns, air quality, and social interactions that occur in urban settings (103). A recent ABCD study found that in children from households with income less than $25,000, better performance on cognitive tests was related to stronger functional connectivity between the DMN and frontoparietal network, and the direction of this association was predicted by environment features such as school type and parent-reported neighborhood safety (104). Another ABCD finding linked children’s environment with functional network topography and longitudinal changes in cognition (105). These results highlight the importance of considering how environmental context can influence brain-behavior associations. Careful consideration of the complex mediating relationships that environmental and other contextual variables play will be an important future direction for increasing the potential policy impacts of developmental brain-behavior association studies.