Similarity in functional connectome architecture predicts teenage grit

Abstract

Grit is a personality trait that encapsulates the tendency to persevere and maintain consistent interest for long-term goals. While prior studies found that grit predicts positive behavioral outcomes, there is a paucity of work providing explanatory evidence from a neurodevelopmental perspective. Based on previous research suggesting the utility of the functional connectome (FC) as a developmental measure, we tested the idea that individual differences in grit might be, in part, rooted in brain development in adolescence and emerging adulthood (N = 64, 11–19 years of age). Our analysis showed that grit was associated with connectome stability across conditions and connectome similarity across individuals. Notably, inter-subject representational similarity analysis revealed that teenagers who were grittier shared similar FC architecture with each other, more so than those with lower grit. Our findings suggest that individuals with high levels of grit are more likely to exhibit a converging pattern of whole-brain functional connectivity, which may underpin subsequent beneficial behavioral outcomes.

Introduction

Grit, an intrapersonal character to persevere and have sustained passion for long-term goals amidst setbacks (Duckworth et al., 2007), has been suggested as one of the key predictors of an individual’s academic success and beyond. Grittier high school students tend to earn higher grade point averages (GPA) and are more likely to graduate; grittier teachers, nurses and salesmen are more likely to maintain their jobs; and grittier men are more likely to sustain marriage (Duckworth et al., 2007; Lee and Sohn, 2013; Eskreis-Winkler et al., 2014; Robertson-Kraft and Duckworth, 2014; Jeong et al., 2019). The beneficial effects of grit reach one’s psychological realm: a recent meta-analysis highlighted a positive association between grit and subjective well-being (ρ = 0.46) (Hou et al., 2022).

Although the effects of grit have been studied in various age groups, from elementary school children to the elderly (Kim and Lee, 2015; Zhang et al., 2022), adolescence is especially relevant because grit is shown to be malleable during this period (Park et al., 2018), and youth grit may yield long-lasting positive outcomes (Eskreis-Winkler et al., 2014; Jiang et al., 2019; Tang et al., 2019). As such, a recent meta-analytic study on the relations between grit and achievement shows that ∼38% of studies have focused on adolescence with larger effect sizes than childhood and adulthood (Lam and Zhou, 2022).

Thus far, the majority of the prior work has relied on self-reported questionnaires, while a few have measured actual behavioral outcomes (e.g. GPA, standardized achievement test scores and job turnover) (Eskreis-Winkler et al., 2014; Robertson-Kraft and Duckworth, 2014). Even fewer studies have investigated its neural correlates, leaving research on the neural underpinnings of grit still in its infancy. Prior studies using resting-state functional magnetic resonance imaging (fMRI) indicated the prefrontal cortex (PFC) as a key region (Wang et al., 2017, 2023) and its functional connections with the striatum (Myers et al., 2016). These results suggest that grit might be represented in cognitive and affective networks that closely align with the two subcomponents of grit: perseverance and passion. Supporting the functional involvement of the frontostriatal network, Wang et al. (2018) showed that adolescents’ grit was associated with the gray matter volumes of dorsolateral PFC and putamen, providing neuroanatomical evidence for the role of grit in persistence and self-regulation (Moriguchi and Hiraki, 2013). More recently, grit was proposed as a potential mediator for dorsolateral PFC functional connectivity and posttraumatic growth following the pandemic in young adults (Wang et al., 2023). In another study, interindividual variability of grit in children was correlated with the shape of the nucleus accumbens (Nemmi et al., 2016), one of the major components of the ventral striatum where its dopamine system has been robustly linked with effortful and instrumental choices for rewards (Salamone and Correa, 2012; Treadway et al., 2012).

Although these neuroimaging studies have offered a useful starting point in the search for neural markers of grit, more nuanced approaches are needed. The aforementioned fMRI studies were limited to resting-state conditions, which is an important caveat considering that movie-watching paradigms are now known to exhibit better reliability (Meer et al., 2020). These studies also mostly relied on a priori regions of interest (ROIs), which often leaves out the chance to observe integrative processes that may occur at the brain network level. In addition, as the functional architecture of the brain undergoes significant reorganizations during adolescence and early adulthood (Kelly et al., 2009; Power et al., 2010), a whole-brain network-level analysis of movie-watching fMRI data would be well suited for capturing the dynamic trajectory of brain development and individual differences in grit.

Meanwhile, the functional connectome (FC)—an intrinsic and relatively stable whole-brain map of functional connectivity patterns between pairs of different regions—has emerged as a useful neural feature in explaining and predicting behavioral variability (Finn et al., 2015; Horien et al., 2019). For instance, Kaufmann et al. (2017) demonstrated that FCs stabilize and individualize markedly in adolescence enabling the identification of individuals’ FC across runs, and delay in such distinctiveness was observed in those with clinical symptoms. Another study also noted that the uniqueness of FC profiles starts to exist as early as 12 years of age and had a negative relationship with mental disorder symptoms (Shan et al., 2022), placing adolescence in the foreground of the brain developmental period.

Recently, a novel methodological approach using FCs as features of brain development has been proposed (Vanderwal et al., 2021). According to this method, a nuanced brain–behavior relationship during development could be captured by correlating (i) connectomes across conditions (e.g. task-based or resting-state functional runs) within an individual (i.e. within-subject connectome stability) and (ii) connectomes across individuals in each condition (i.e. between-subject connectome similarity). To be more specific, within-subject connectome stability denotes how stable one’s FCs are across conditions. FC stability, therefore, estimates the degree of preservation of an individual’s functional connectivity patterns despite being in different states, as functional connections altogether form intrinsic and distinguishable profiles (Finn et al., 2015; Kaufmann et al., 2017). In contrast, between-subject connectome similarity indicates how similar one’s FC in a given condition is to that of others. This metric was suggested to capture the deviation of one’s functional connectivity patterns from the group when the subjects are believed to be in the same state (e.g. by showing an identical naturalistic stimuli) (Vanderwal et al., 2021). An important assumption is that connectome stability and similarity reflect two maturational processes occurring hand in hand across development for optimized neural processing. These two lines of transformative functional changes thus entail that brain development might not only involve FCs being more stable across conditions within an individual, but they might also represent normative or template-like functional connectivity patterns shared among individuals in each condition. Supporting this idea, greater connectome stability and similarity were associated with higher social skills in a sample of 6- to 21-year olds (Vanderwal et al., 2021). In a similar vein, a recent study demonstrated that within-subject stability and between-subject similarity to high scorers in cognitive tasks successfully predicted sustained attention and working memory abilities in adults, highlighting the potential utility of FC stability/similarity features in behavioral prediction models (Corriveau et al., 2022).

Considering these methodological advances in using whole-brain functional connectivity patterns, the current study aims to test whether individual differences in grit are reflected in the putative brain development measures. There are several pieces of evidence implicating that grit might be related to more mature functional brain systems encompassing cognitive and affective networks. Although grit is conceptually distinct from cognitive abilities, pursuing long-term goals while dismissing momentary distractions does require engagement of higher-order functions such as self-regulation (Duckworth and Gross, 2014; Wolters and Hussain, 2015) for which the PFC is known to be responsible (Casey et al., 2011; Casey and Caudle, 2013). Considering that the PFC—a region suggested to be representative of brain maturity (Dosenbach et al., 2010)—was highlighted in prior neuroimaging research on grit, it is possible that gritty individuals might benefit from a more mature functional architecture of the brain. Furthermore, recent studies reported positive associations between grit and cognitive reappraisal (Millonado Valdez and Daep Datu, 2021; Kalia et al., 2022), an effective emotion regulation strategy that involves PFC engagement across development (McRae et al., 2012). This suggests that being gritty might capitalize on balanced interaction and maturation of cognitive and affective control networks resulting in sustained and passionate dedication to long-term goals. Taken together, these results offer insight into the possible link between grit and brain development.

Drawing on this speculation, we adopted the concept of connectome stability and similarity to test the following hypotheses. First, we expected that gritty individuals’ FCs might be relatively consistent across runs (i.e. high stability) and similar with others in each condition (i.e. high similarity)—the two proxies as features of brain development. Second, we hypothesized networks that contribute to the relationship between grit and connectome stability, and similarity would be distributed across the brain, notably networks that support cognitive–affective functions. In addition, we sought to test the potential utility of FC generated from fMRI data that do not require grit-specific tasks (e.g. movie watching) in identifying gritty individuals, which in turn would be able to shed further light onto the neural underpinnings of grit.

Materials and methods

Participants

All data used in this study were drawn from the publicly available Healthy Brain Network (HBN) dataset (Alexander et al., 2017). The dataset houses multimodal brain imaging as well as behavioral phenotypic data from various communities across New York City. The recruitment of the study was largely based on the families with clinical concerns in their child, and the study was approved by the Chesapeake Institutional Review Board. For participants <18 years, written consent and written assent were acquired from the legal guardians and the participants. Written informed consent was obtained for those >18 years. The dataset can be downloaded at http://fcon_1000.projects.nitrc.org/indi/cmi_healthy_brain_network/.

Data from the HBN Releases 7 through 10 were used for the present study, as the Grit scale was added to the protocol from Release 7. Among participants whose chronological age was between 11 and 19 years, a total of 238 participants with grit scores were initially accessed. Participant data were collected in one of two sites located at the Citigroup Biomedical Imaging Center (CBIC) or the Rutgers University Brain Imaging Center (RUBIC). Individuals without T1-weighted images as well as all four functional runs were excluded (n = 23). We removed low-quality data after manually inspecting the pre-processed output (n = 12) and excluded participants with excessive head motion for every functional run [mean framewise displacement (FD) > 0.2 mm] to avoid confounding effects from movement in the scanner (n = 129). After excluding data with FC construction failures (n = 10), the remaining 64 participants were analyzed as our main sample [N = 64, 22 females; age range = 11.07–18.82 years (M = 14.78); n = 48 from the CBIC, see Supplementary Methods, Fig. S1 for sample characteristics].

Phenotypic data

Phenotypic assessments used in this study were obtained on the first and third visits of the HBN schedule. As a main behavioral variable of interest, the 12-item Grit scale was used to measure perseverance of effort (e.g. ‘I have overcome setbacks to conquer an important challenge.’) and consistency of interest (e.g. ‘New ideas and projects sometimes distract me from previous ones.’) for long-term goals on a 5-point scale (Duckworth et al., 2007). Items for consistency of interest were reverse-coded and the average scores for the two subscales were calculated. The total grit score was averaged after adding up all the items resulting in a range from one (not at all gritty) to five (extremely gritty). For completeness, Grit-S scores were additionally calculated by omitting four items (items 1, 3, 10, 11), as the Grit-S scale was previously known to be more efficient with stronger psychometric properties (Duckworth and Quinn, 2009) despite the cost of lower content validity (Duckworth et al., 2021). We note that, in the present sample, reliability metrics between the two versions of the scale showed comparable results, with slightly better performance for the Grit-O scale (Grit-O, Cronbach’s α = 0.73; Grit-S, α = 0.70).

To show that grit is related to positive behavioral variables and to conceptually distinguish from cognitive abilities such as IQ or attention, we utilized several related phenotypic assessments available in the dataset. First, the Screen for Child Anxiety Related Disorders (SCARED) (Birmaher et al., 1999) and Mood and Feelings Questionnaire (MFQ) (Messer et al., 1995) were used as self-reported behavioral measures for anxiety and depression. Self-reported version of the Columbia Impairment Scale (CIS) was used as a measure for global impairment across domains including school performance, interpersonal relationship, psychopathology and use of leisure time (Bird et al., 1993). Parent ratings on attention-deficit and hyperactivity levels assessed by the Strengths and Weaknesses of Attention-Deficit/Hyperactivity Disorder Symptoms and Normal Behavior Scale (SWAN) were used (Swanson et al., 2001).

Two additional behavioral phenotypes were used as control variables in our main analyses. Based on the previous research demonstrating a link between FC stability and similarity measures and social skills (Vanderwal et al., 2021), we sought to ensure that the observed grit–brain association from our analyses was not attributable to social skills. To this end, the Social Communication Questionnaire (SCQ), a parent measure with 40 yes-or-no items often used as a screening instrument for autism spectrum disorder (Chandler et al., 2007), was used. The higher score in this assessment indicates impaired social communication skills including a child’s body movements and language or gesture usage. Among those who completed the Grit scale, all but one had SCQ scores and were thus analyzed (n = 63). We also used a composite score of full-scale intelligence quotient (FSIQ) from the Wechsler Intelligence Scale for Children (WISC), Fifth Edition (Wechsler, 2014) to ensure that the observed effect was not derived by general intellectual abilities. A subsample (n = 41) in the range of 6–17 years of age provided both grit and IQ scores.

Image acquisition

Magnetic resonance imaging (MRI) data were collected using a Siemens 3T Tim Trio at the RUBIC and a Siemens 3T Prisma at the CBIC on the second visit schedule. Identical scan parameters were used for both centers: 3D anatomical T1-weighted MPRAGE image [224 slices, repetition time (TR) = 2500 ms, echo time (TE) = 3.15 ms, flip angle (FA) = 8°, slice thickness = 0.8 mm], functional runs with echo-planar imaging sequences (TR = 800 ms; TE = 30 ms; FA = 31°; slice thickness = 2.4 mm; field of view = 204 mm; multi-band acceleration factor = 6; voxel size = 2.4 mm isotropic).

The four functional runs were administered in the following order: two consecutive resting-state runs of 5.1 min each (Rest1 and Rest2; 375 TRs each) and two movie-watching runs of 10 min [a clip from ‘Despicable Me’ (MovieDM); 750 TRs] and 3.47 min [a short animated movie ‘The Present’ (MovieTP); 250 TRs] length. In resting-state runs, participants were asked to keep their eyes open and focus on the fixation cross on the screen. The latter two naturalistic stimuli were played with sound. In the MovieDM condition, a clip of a DVD version of the movie ‘Despicable Me’ was extracted where the main character unwillingly reads the three kids a bedtime story. A full-length, short animation ‘The Present’, story of a boy getting a puppy as a surprise gift from his parents, was shown in MovieTP. We present full-scan-length results as our primary main analysis. For completeness, results from a volume-matched analysis leaving only the first 250 volumes in the three functional scans (Rest1, Rest2 and MovieDM) to fit the shortest scan length of MovieTP (250 TRs) are reported separately (for the results from the last 250 volumes, see Supplementary Results).

Image pre-processing

All fMRI data were pre-processed using fMRIPrep 21.0.2 (Esteban et al., 2019). This includes blood-oxygen-level-dependent (BOLD) signal reference image estimation, head-motion estimation, slice-time correction, co-registration, resampling onto standard space based on the developmental sample from 4.5 to 18.5 years of age (MNI PediatricAsym:cohort-1) (Fonov et al., 2011) and confounds estimation. To increase signal-to-noise ratio, the data were smoothed with a 4 mm full-width at half-maximum Gaussian kernel using 3dmerge in AFNI (Cox, 1996; Cox and Hyde, 1997). We applied relatively small spatial smoothing since a kernel with less than twice the voxel size was recommended for mitigating overestimation in correlation-based functional activation mappings (Liu et al., 2017). The first four dummy volumes were discarded and confound regression was performed with 27 regressors (24 head-motion parameters: three translational and three rotational deviations and their squares, six temporal derivatives and their squares, and three mean tissue signals: global, cerebrospinal fluid and white matter). Global signal regression was adopted to control for the effects of motion on functional connectivity (Power et al., 2015) since head motion is often a major concern in the developmental sample. Finally, we set our high-pass filter cut-off as 0.01 Hz and low-pass filter cut-off as 0.10 Hz to remove the physiological noise in the data.

FC construction

Fully pre-processed fMRI images were then transformed into FCs using 3dNetCorr in AFNI (Taylor and Saad, 2013). By providing the Shen atlas parcellation with 268 ROI masks (Shen et al., 2013) to the time-series data of each functional run as input, we generated a 268 × 268 whole-brain functional connectivity matrix for each individual. The correlation coefficients from each pair of brain regions were Fisher z-transformed. As a result, four FCs (Rest1, Rest2, MovieDM and MovieTP) were created for each participant.

Within-subject connectome stability

The overall workflow of this study is presented in Figure 1. We calculated FC features based on the approach proposed by Vanderwal et al. (2021) to facilitate comparisons with previous work. Every whole-brain FC matrix was left with an off-diagonal lower triangular part and was vectorized. This left us with 35778 connections (i.e. edges) between two different brain regions (i.e. nodes). In order to calculate FC stability within individuals, the functional connectivity value of each edge was correlated across two conditions. Such edge-wise correlation resulted in three within-subject connectome stability measures: cross-rest (Rest1 and Rest2), cross-state (MovieDM and Rest1) and cross-movie (MovieDM and MovieTP) stability (Figure 1A). The mean stability for each individual was calculated by averaging the three stability measures.

Fig. 1.

A summary of the analysis steps. (A) Stability and similarity measures were created after vectorizing FCs of four conditions. Within-subject stability was calculated by correlating every functional connection in two different conditions for each subject (e.g. S_i). Between-subject similarity in a given condition was the average of all edge-wise functional connectivity correlations of one subject and all others. This procedure resulted in a total of six FC features (three stability and three similarity measures) for each subject. (B) IS-RSA adopting the AnnaK model was conducted as a confirmatory analysis by comparing similarity matrices of brain features (i.e. between-subject FC) and behavior variables (i.e. grit). According to the AnnaK model, high scorers in grit would share similar brain features more so than those with lower scorers.

Between-subject connectome similarity

FC similarity across individuals in a given condition (i.e. MovieDM, MovieTP and Rest1) was computed by averaging edge-wise correlation coefficients between one participant’s FC and everyone else’s. In other words, FC in each run was correlated with FCs of all others excluding oneself. This resulted in three between-subject connectome similarity measures: Rest1, MovieDM and MovieTP FC similarity (Figure 1A). The three similarity measures were averaged to calculate mean similarity.

Linking FC stability and similarity with behavior

Linear regression analysis was conducted to test whether the six FC measures (three stability measures and three similarity measures) predicted behavioral variables of interest (i.e. grit, social skills and IQ). Participants’ age, sex, scan site and head motion for the corresponding functional runs (e.g. average mean FD of MovieDM and MovieTP for cross-movie stability) were entered as covariates. A leave-one-out cross-validation (LOOCV) on the regression model was performed along with non-parametric permutation tests (1000 iterations) by correlating actual and predicted grit scores to validate the findings. To investigate FC stability and similarity at the network level, we correlated grit and FC measures computed from predefined 10 canonical networks (Finn et al., 2015; Noble et al., 2017) separately.

Inter-subject representational similarity analysis

In addition to the main analyses using FC stability and similarity, inter-subject representational similarity analysis (IS-RSA) (Finn et al., 2020) was conducted to confirm the relationship between FC similarity and grit. Specifically, via IS-RSA, we adopted (i) a non-parametric Mantel test where permutation-based null distributions of correlation values are used to determine significance (Mantel, 1967) and (ii) a computational lesion approach whereby the relative contributions of canonical networks are quantified by creating a ‘virtual lesion’ that knocks out specified edges per iteration. As IS-RSA allows flexible modeling of brain–behavior similarities, we assumed that gritty individuals may share more similar FCs (i.e. express normative functional connectivity patterns while in the same fMRI condition) with other gritty individuals, while those who are not may exhibit heterogenous FCs that are less similar to everyone else.

The idea of the IS-RSA framework is to compare the brain similarity matrix and behavioral similarity matrix (Figure 1B). To this end, the pairwise Pearson correlations of FC between one and every other subject were used to create a 64 × 64 inter-subject brain similarity matrix. For grit similarity, individuals were ranked by grit scores from low (e.g. Ranks 1 and 2) to high (e.g. Ranks 63 and 64). The mean of the two individuals’ rank was regarded as a behavioral similarity metric for the pair (e.g. behavioral similarity for subjects Ranks 63 and 64 is 63.5, Ranks 1 and 2 is 1.5) resulting in a 64 × 64 inter-subject behavioral similarity matrix. This approach, called ‘the Anna Karenina (AnnaK)’ model, assumes that high scorers on a given behavior show higher brain feature similarity with each other, whereas low scorers exhibit more variability (Finn et al., 2020) (Figure 1B). Then, we computed inter-subject representational similarity using Spearman’s correlation between brain and behavior similarity matrices. The corresponding P-value was derived from the Mantel test with 5000 permutation test iterations. The same procedure was conducted for other behavioral phenotypes that were correlated with grit as control analyses.

To check if there was a particular functional network contributing to the correlation between FC similarity and grit, the 10 canonical networks (Finn et al., 2015; Noble et al., 2017) were computationally lesioned. In this approach, in an iteration of the analysis, all nodes from a given network were excluded from the whole-brain FC to simulate a virtual lesion. If a network made a unique contribution to the prediction, we would expect a marked decrease in the inter-subject representational similarity recalculated without the said network.

Results

Correlations between grit and other behavioral measures

Correlations between grit and developmentally relevant behavioral measures are presented in Table 1. Grit had a significant negative relationship with anxiety (r = −0.30, P < 0.05) and marginally significant relationship with depression (r = −0.24, P = 0.06). Grit was also inversely correlated with general impairment (r = −0.42, P < 0.01), indicating that gritty teenagers tend to be less anxious and less susceptible to global impairment. In terms of social skills, there was a significant negative association (r = −0.27, P < 0.05), such that gritty teenagers tended to possess higher social communication skills. Neither attention-deficit scores nor IQ scores were correlated with grit (r = −0.20, P = 0.12; r = −0.02, P = 0.92, respectively).

Table 1.

Descriptive statistics and correlations of study variables (N = 64)

	M (s.d.)	1	2	3	4	5	6	7
1. Grit	3.07 (0.53)
2. SCARED	22.92 (15.29)	−0.30*
3. MFQ (n = 61)	13.23 (12.61)	−0.24	0.75**
4. CIS	11.16 (7.90)	−0.42**	0.49**	0.57**
5. SWAN	0.32 (0.87)	−0.20	−0.04	−0.06	0.24
6. SCQ (n = 63)	8.05 (4.28)	−0.27*	0.04	0.15	0.23	0.44**
7. WISC FSIQ (n = 41)	104.71 (17.85)	−0.02	0.16	−0.09	0.11	−0.24	−0.36*	−

Note: In the case of a subsample, the number of participants is noted.

^* P < 0.05,

^** P < 0.01.

FC stability and similarity

As a summary measure of FC stability and similarity, we present the results based on mean stability and mean similarity. The group average of mean stability was higher than mean similarity (full-volume, mean stability = 0.45, mean similarity = 0.34; volume-matched, mean stability = 0.25, mean similarity = 0.19). Importantly, mean stability and mean similarity showed high-positive correlation even after controlling for age, sex, head motion and scan site (r = 0.64, P < 0.001), and this relationship remained in the volume-matched analysis (r = 0.49, P < 0.001). That is, individuals who showed highly stable functional connectivity patterns across conditions were more likely to resemble others’ FCs, replicating previous work by Vanderwal et al. (2021) and supporting the proposition that connectome stability and similarity may serve as concurrent brain development proxies. Stability and similarity measures were not correlated with age (Supplementary Results, Table S1), also replicating previous findings (Vanderwal et al., 2021). Independent samples t-tests revealed that none of the FC measures (both full-volume and volume-matched analyses) showed statistically significant sex differences.

FC features and grit

Partial correlation values from linear regression are reported in Table 2. We present the main findings using grit scores derived from the Grit-O scale (see Supplementary Results, Table S3 for the results using the Grit-S scale). Cross-movie stability and MovieTP similarity significantly predicted grit after controlling age, sex, scan site and head motion (cross-movie stability, r = 0.34, P < 0.008; MovieTP similarity, r = 0.37, P = 0.003; Bonferroni corrected, P < 0.008). When the time window was set to the first 250 volumes, cross-movie stability was no longer a significant predictor, r = 0.17, P = 0.21 (for the last 250 volumes, see Supplementary Results). LOOCV and the following permutation tests confirmed the results such that MovieTP similarity alone significantly predicted grit, though at a marginal level (P = 0.045) (see Supplementary Results, Table S4 for additional validation results). We also note that the significant positive relationship was largely driven by one of the two subscales of grit: perseverance of effort (cross-movie stability, perseverance of effort, r = 0.30, P < 0.05, consistency of interest, r = 0.02, P = 0.87; MovieTP similarity, perseverance of effort, r = 0.36, P < 0.01, consistency of interest, r = 0.22, P = 0.08). As the FC similarity in MovieTP was the only robust predictor when holding the scan-length constant, we chose to focus on MovieTP similarity in subsequent analyses.

Table 2.

FC features and their correlations with grit using full and truncated TRs (first 250 volumes)

	Within-subject FC stability			Between-subject FC similarity
	Cross-rest	Cross-state	Cross-movie	Rest1	MovieDM	MovieTP
Full volume
Grit	0.03 (0.8266)	−0.09 (0.5036)	0.34 (0.0076)	0.04 (0.7353)	0.18 (0.1785)	0.37 (0.0032)
Perseverance of effort	0.01 (0.9569)	0.06 (0.6725)	0.30 (0.0211)	−0.04 (0.7825)	0.03 (0.8195)	0.36 (0.0045)
Consistency of interest	0.03 (0.7933)	−0.17 (0.1949)	0.23 (0.0729)	0.09 (0.4776)	0.22 (0.0895)	0.23 (0.0812)
Volume-matched
Grit	−0.09 (0.5133)	0.04 (0.7823)	0.16 (0.2115)	0.19 (0.1498)	0.21 (0.1133)
Perseverance of effort	−0.11 (0.4219)	0.16 (0.2321)	0.25 (0.0545)	0.08 (0.5257)	0.13 (0.3108)
Consistency of interest	−0.03 (0.8011)	−0.08 (0.5502)	0.02 (0.8513)	0.20 (0.1345)	0.18 (0.1671)

Notes: Partial correlation coefficients (r) between grit (total score and its subscales) and stability/similarity measures, controlling for age, sex, scan site and head motion, are reported. Corresponding P-values are in parentheses, and significant correlations are in bold (Bonferroni corrected, P < 0.008).

In the MovieTP condition where FC similarity successfully predicted grit, none of the 10 networks was significantly correlated with the total grit score by themselves (Supplementary Results, Table S2), implicating that simultaneously considering all of the networks offered better predictive performance than any single canonical network.

Inter-subject representational similarity analysis

Based on the findings showing that grittier individuals shared more similar FCs with others in the MovieTP condition, IS-RSA was conducted as a confirmatory analysis to formally examine the shared geometries between FC and grit (Figure 2). The correlation between MovieTP FC similarity and grit similarity matrices was significant (r = 0.28, P < 0.01) (Figure 2A), confirming the FC similarity–grit association from the previous analysis. We then ran the same analysis using the subscales of grit and found that perseverance of effort and consistency of interest both showed similar results (perseverance of effort, r = 0.22, P < 0.05; consistency of interest, r = 0.21, P < 0.05). As the present IS-RSA adopted the AnnaK model, this implies that gritty individuals shared normative functional connectivity patterns with other gritty individuals, while those who are not exhibited heterogenous FCs that are less similar to everyone else (Figure 2B).

Fig. 2.

Results of IS-RSA. (A) MovieTP FC similarity (left) and grit score similarity (right) matrices were compared (5000 permutation test iterations). The composite score of grit yielded a higher r-value than its subcomponents. (B) A visual summary of FC similarity sorted by grit scores (from low to high). Gritty individuals in the bottom right section showed higher FC similarity among themselves. (C) Results of the computational network lesion approach. A larger drop in inter-subject representational similarity (r) of each network lesion compared to that of the whole-brain result (dashed line) is interpreted as a higher contribution of a network in IS-RSA. Single network lesions did not show a significant drop-off in representational similarity (Bonferroni corrected, P < 0.005).

Results of the computational lesion approach suggest that, once again, the entirety of whole-brain networks, rather than any given single network, contribute more to the observed grit–brain relationship (Figure 2C). There was no significant decline in the correlation for any of the network lesions (Medial frontal, r = 0.28, P = 0.003; Fronto-parietal, r = 0.28, P = 0.002; Default mode, r = 0.28, P = 0.003; Motor, r = 0.28, P = 0.002; Visual I, r = 0.27, P = 0.003; Visual II, r = 0.28, P = 0.003; Visual association, r = 0.28, P = 0.002; Limbic, r = 0.28, P = 0.004; Basal ganglia, r = 0.28, P = 0.003; Cerebellum, r = 0.28, P = 0.004; Bonferroni corrected, P < 0.005). This finding may imply that predictive brain systems of grit do not rely on a single network, but rather they are distributed across the brain.

Control analyses

We performed a set of control analyses to validate the main finding that grit is related to neurodevelopmental features (Figure 3). First, considering that most of the participants received clinical diagnoses at the point of data collection (Figure 3A), it is possible that gritty individuals may have been different from less gritty individuals in terms of the prevalence of clinical conditions. Thus, a chi-squared test was carried out to test whether there was a significant association between grit and clinical diagnoses. We split participants into half according to the median grit score (3.00) resulting in a high-grit group (n = 31) and a low-grit group (n = 33). The number of participants with diagnoses (i.e. attention-deficit hyperactivity disorder, anxiety disorders, major depressive disorders, cognitive abilities-related disorders and autism spectrum disorders) was not significantly related to grit scores [χ² (6, N = 64) = 7.98, P = 0.24]. Additionally, we provide group-averaged brain maps of functional connectivity patterns for high and low grit, which suggests that both groups had similar nodes exhibiting the strongest connections in the MovieTP condition (Supplementary Results, Fig. S2).

Fig. 3.

A summary of control analyses. (A) Stacked bar graph of the participants’ clinical conditions, grouped by grit scores (median-split). Abbreviations: ADHD, attention-deficit hyperactivity disorder; AD, anxiety disorder; MDD, major depressive disorder; Cog, cognitive-related disorder; ASD, autism spectrum disorder. (B) Partial correlation coefficients (r) between control variables (mean FD, SCQ and WISC FSIQ) and FC features in full-volume and volume-matched analyses (Bonferroni corrected, P < 0.008). (C) Heatmaps where FC similarities are sorted by behavior scores indexing anxiety, global impairment and social skills, respectively. Unlike grit, all three heatmaps do not show any discernible pattern. (D) A heatmap depicting that gritty individuals share FC similarity with others, based on IS-RSA results from a subsample of CBIC participants.

Next, due to the potential confounding effects of head motion during scans, we examined the correlation between head motion and the six measures of FC stability and similarity, where head motion was calculated by averaging the mean FD in the corresponding conditions. FC measures were not significantly correlated with head motion in either the full-volume or volume-matched analyses (Figure 3B), which may be attributed to a conservative head-motion threshold applied in this study.

To test whether the observed effects of the main results were specific to grit, the six FC measures were correlated with social skills and IQ while controlling for the effects of age, sex, scan site and head motion. Social communication skills had overall negative correlations (i.e. higher score in SCQ is regarded as lower social communication skills), but none of them survived the correction for multiple comparisons. Results of 41 participants who had both grit and IQ scores also showed that the six brain features were not predictive of IQ (Figure 3B). In addition, when IS-RSA was conducted using other phenotypic data that were correlated with grit, including anxiety, global impairment and social communication skills, there were no significant associations found for any of these measures with MovieTP similarity (SCARED, r = −0.04, P = 0.65; CIS, r = −0.13, P = 0.17; SCQ, r = −0.08, P = 0.40) (Figure 3C). These results provide supporting evidence for a unique association between MovieTP similarity and grit when tested with the AnnaK model.

Finally, even though the neuroimaging data collected in the HBN dataset followed identical scan parameters, MRI scanners from the two sites might yield different results. To address this possibility, an additional IS-RSA was performed in which the data were restricted to the 48 participants from the CBIC. Results still showed that gritty individuals do share similar FCs with one another (r = 0.33, P < 0.01) (Figure 3D), consistent with the results from IS-RSA with the full sample (Figure 2B).

Discussion

The purpose of the current study was to investigate whether brain development measures derived from FCs capture individual differences in teenage grit. Here, we showed that grit was significantly related to movie-related within-subject FC stability and between-subject FC similarity. When matched for scan duration (i.e. number of data points), FC similarity in the MovieTP condition was the sole predictor of grit, meaning that gritty teenagers showed more similar whole-brain FCs with others. IS-RSA with the AnnaK framework revealed that gritty individuals shared FC organizations among themselves in contrast to those with lower levels of grit. Such results were not observed in other behavioral measures correlated with grit such as anxiety, global impairment and social skills. With respect to network contribution, we demonstrated that the observed association between FC similarity and grit was not explained by any given single functional network.

As expected, the two components of grit—an individual’s tendency to persevere and remain passionate—were related to several behavioral measures that are important in developmental contexts. We found that higher grit was related to lower anxiety, general impairment and social communication skills, illustrating that grit may be associated with better mental health and functioning across domains in daily life. Also of importance is that grit was not correlated with IQ, suggesting that grit is not simply reflecting general intellectual ability and is consistent with the proposition that grit is better conceptualized as a non-cognitive trait (Duckworth et al., 2007).

Our analyses provided preliminary evidence that FC features might be associated with grit. Notably, these features (i.e. cross-movie stability and MovieTP similarity) were both movie-related, supporting the growing usage of naturalistic paradigms in developmental research due to their suitability in individual differences in fMRI research and utility in reducing head motion (Vanderwal et al., 2019; Frew et al., 2022). Although we cannot pinpoint which particular feature of MovieTP was crucial for its association with grit, there are several possibilities worth considering. First, we expected that the context mattered more than the length of the scans as previous studies showed that longer scan duration does not necessarily yield better prediction (Sanchez-Alonso et al., 2021; Vanderwal et al., 2021). Instead, as Meer et al. (2020) demonstrated that subjective engagement in movie-watching sessions was related to the representations of functional dynamics, it might be possible that the emotionally engaging presentation and narrative of MovieTP ended up drawing more attention from the participants, compared to MovieDM which was an edited clip extracted from a feature-length movie. The latter might have required previous context for it to be as engaging and comprehensive as the former, which may have influenced its association with grit. We add that individual differences in attention deficit likely do not explain this interpretation due to the weak correlation found between grit and SWAN scores (Table 1). Another possibility to note is the fact that MovieTP was always presented at the tail end of the four fMRI conditions. Considering that the entire MRI session lasted over an hour (64.7 min), simply staying engaged toward the end could be a burdensome task, especially in a developmental sample. This may have inadvertently induced a demanding situation where trait-like grit could be manifested during the MovieTP condition. Although these interpretations are speculative and require more systematic examination in future studies, it is possible that some combination of the movie stimuli per se and the grit-invoking scanning environment might have contributed to the present findings.

Our analyses evinced the utility of whole-brain FCs such that no particular functional network was solely associated with grit, a finding that was corroborated by the computational lesion method in IS-RSA. Of relevance, recent studies demonstrated that FC showed greater reliability than individual edges, which implies that it may not be the simple sum of its components (Noble et al., 2017; Pannunzi et al., 2017). While it is still a possibility that specific grit-related functional networks could be revealed when using methods such as fine-scale FC (Busch et al., 2022), our findings highlighting the shared whole-brain neural representations among gritty teenagers encourage future research to examine functional systems across the brain.

Unlike grit, age was unrelated to any connectome stability and similarity features in our sample of teenagers. Such lack of age effects may appear to undermine the utility of these measures as proxies for brain development (e.g. 19-year olds are expected to have more mature functional architecture than 11-year olds). We note, however, that several studies utilizing larger study samples with a wider age range did show a substantial increase in brain development measures such as connectome distinctiveness (Kaufmann et al., 2017; 797 individuals aged 8–22 years) and algorithm-based estimation of functional maturity (Dosenbach et al., 2010; 238 individuals aged 7–30 years). These findings suggest that the trajectory of age-related changes in connectome stability and similarity features may become more readily visible when considering brain development across childhood, adolescence and adulthood.

Our study is not without caveats that could be addressed in future work. First, as the sample size was modest, future studies with larger samples are necessary to validate and generalize these findings. We do note that, to the extent of our knowledge, the HBN dataset is the only publicly available developmental neuroimaging dataset that includes the Grit scale as a part of its protocol. Using the dataset, our analyses did begin with 238 subjects but >54% of the initial data had to be discarded due to excessive head motion during scanning, highlighting the challenges in procuring fMRI data with sufficient quality in developmental samples. Additionally, we focused on teenagers ranging from early adolescence to early adulthood based on prior work suggesting that this is an important period for academic or career achievements, interpersonal relationships, mood and conduct disorders (Kenny et al., 2013; Lee et al., 2014; Negru-Subtirica and Pop, 2016), coupled with dynamic neural changes encouraging the need to progress in interventions or predictions (Rosenberg et al., 2018). Future research with longitudinal neuroimaging data including other age groups would fill the gap by investigating whether FC features can be meaningful predictors of grit and related positive outcomes from different developmental stages. Another limitation is the use of behavioral assessments. Although we controlled for several correlated measurements in our study, the unique predictive validity of grit could be more rigorously tested if other related psychological constructs, such as self-control and conscientiousness, had been available. We acknowledge the ongoing debates about the incremental validity of grit compared to these measures (Muenks et al., 2017; Schmidt et al., 2018; Vazsonyi et al., 2019; Wang et al., 2023). Specifically, some researchers argue that grit adds little to individual’s outcomes beyond self-control and conscientiousness (Rimfeld et al., 2016; Vazsonyi et al., 2019), while others present evidence to the contrary (Suzuki et al., 2015; Tedesqui and Young, 2018; Hong and Lee, 2019; Jiang et al., 2020). Therefore, we encourage future studies to replicate our findings by explicitly delineating grit and its related psychological constructs. Additionally, there are still other avenues to investigate brain–behavior relationships. For instance, BOLD signal variability per se can be utilized as an alternative metric for stability, based on recent work demonstrating a link between functional connectivity patterns and inter-regional signal dynamics (Baracchini et al., 2021). It is also possible that alternative metrics such as fractional amplitude of low-frequency fluctuations (Zou et al., 2008), regional homogeneity (Zang et al., 2004) or functional connectivity density (Tomasi and Volkow, 2010) might yield meaningful results, as evidenced by a previous resting-state fMRI study of grit (Wang et al., 2017). Another suggestion is to use experimental tasks during fMRI scans that better capture an individual’s objective grittiness to enhance predictive power. Although the present results suggest the possibility of grit-induced situations (e.g. order of scans and scan duration) being a conducive factor that capture individual differences in self-reported grit, utilizing appropriate tasks and corresponding brain states could help circumvent the potential biases embedded within subjective behavioral assessments.

In conclusion, we provide initial evidence supporting the relationship between grit and neurodevelopmental features in adolescence and emerging adulthood. Grit was correlated with positive behavioral phenotypes such as lower anxiety, lesser general impairment and better social skills. We showed that within-subject FC stability and between-subject FC similarity in movie-related scans were associated with grit. Notably, FC similarity while watching a complete animation film was higher among gritty individuals but not with others. Taken together, these results highlight potential common neural representations of grit that are distributed across the whole brain and reveal that gritty teenagers share similar FC architecture.

Supplementary data

Supplementary data are available at SCAN online.

Data availability

The HBN dataset (http://fcon_1000.projects.nitrc.org/indi/cmi_healthy_brain_network/) is publicly available. Codes for FC stability and similarity analysis and IS-RSA are also uploaded and can be accessed via doi:10.5281/zenodo.7986646.

Author contributions

Sujin Park: Conceptualization, Methodology, Formal analysis, Writing - Original Draft. Daeun Park: Supervision, Writing - Reviewing and Editing, Funding acquisition. M. Justin Kim: Conceptualization, Supervision, Writing - Reviewing and Editing, Project administration, Funding acquisition.

Funding

This research was supported by the National Research Foundation of Korea (NRF-2022R1A2C1091871 and NRF-2021R1F1A1045988). This research was also supported by the Sungkyunkwan University and the BK21 FOUR (Graduate School Innovation) funded by the Ministry of Education (Korea) and the National Research Foundation of Korea.

Conflict of interest

The authors declared that they had no conflict of interest with respect to their authorship or the publication of this article.