Youth Screen Media Activity Patterns and Associations With Behavioral Developmental Measures and Resting-state Brain Functional Connectivity

Abstract

Objective

Screen media activity (SMA) consumes considerable time in youth’s lives, raising concerns about the effects it may have on youth development. Disentangling mixed associations between youth’s SMA and developmental measures should move beyond overall screen time and consider types and patterns of SMA. We aimed to identify reliable and generalizable SMA patterns among youth and examine their associations with behavioral developmental measures and developing brain functional connectivity.

Method

Three waves of the Adolescent Brain and Cognitive Development (ABCD) data were examined. The Lifespan Human Connectome Project in Development (HCP-D) was interrogated as an independent sample. ABCD participants included 11,878 children at baseline. HCP-D participants included 652 children and adolescents. Youth-reported SMA and behavioral developmental measures (neurocognitive performance, behavioral problems, psychotic-like experiences, impulsivity, and sensitivities to punishment/reward) were assessed with validated instruments. We identified SMA patterns in the ABCD baseline data using K-means clustering and sensitivity analyses. The generalizability and stability of the identified SMA patterns were examined in HCP-D data and ABCD follow-up waves, respectively. Relationships were examined between SMA patterns and behavioral and brain (resting-state brain functional connectivity [RSFC]) measures using linear-mixed-effect modelling with false-discovery-rate (FDR) correction.

Results

SMA data from 11,815 children (Mean_age = 119.0 months, SD_age = 7.5; 6,159 (52.1%) boys) were examined, and 3,151 (26.7%) demonstrated a video-centric higher-frequency SMA pattern and 8,666 (73.3%) demonstrated a lower-frequency pattern. The SMA patterns were validated in similarly-aged HCP-D youth. Compared to the lower-frequency-SMA-pattern group, the video-centric-higher-frequency-SMA-pattern group showed poorer neurocognitive performance (Beta=−0.12, 95%CI, [−0.08, −0.16], FDR-corrected p<.001), more total behavioral problems (Beta=0.13, 95%CI, [0.09, 0.18], FDR-corrected p<.001), and more psychotic-like experiences (Beta=0.31, 95%CI, [0.27, 0.36], FDR-corrected p<.001). The video-centric-higher-frequency-SMA-pattern group demonstrated higher impulsivity, more sensitivity to punishment/reward and altered RSFC among brain areas implicated previously in cognitive processes. Most of the associations persisted with age in the ABCD data, with more individuals (n=3,378, 30.4%) in the video-centric higher-frequency SMA group at one-year follow-up. A social-communication-centric SMA pattern was observed in HCP-D adolescents.

Conclusion

Video-centric SMA patterns are reliable and generalizable during late childhood. A higher-frequency-video-entertainment-SMA-pattern group showed altered RSFC and poorer developmental measures that persisted longitudinally. The findings suggest public health strategies aiming to decrease excessive time spent by children on video-entertainment-related SMA are needed. Further studies are needed to examine potential video-centric/social-centric SMA bifurcation to understand dynamic changes and trajectories of SMA patterns and related outcomes developmentally.

INTRODUCTION

Electronic screens have become ubiquitous recreational devices, and screen media activity (SMA) consumes considerable time in the lives of many children and adolescents.¹ This situation may have been exacerbated during the COVID-19 pandemic.² Late childhood and adolescence represent periods of psychosocial maturing and neurobiological development with increased sensitivity to appetitive and aversive external cues, like novel stimuli in SMA.³ Such stimuli may impede or promote developmental processes;⁴ thus, effects of SMA have been debated.⁵ Accumulating data link SMA to poor cognitive development,⁶ mental health problems,⁷ and brain structural differences in white-matter integrity and cortical morphometry.^8,9 However, seemingly inconsistent findings exist,¹⁰ generating discussion¹¹ and recommendations that disentangling observed associations requires consideration of SMA types and patterns.¹² Use-and-gratification theories have proposed that individuals are driven to use different types of screen media, leading to different SMA use patterns.¹³ Theories of problematic/addictive patterns of SMA use also consider certain patterns of SMA use in relation to specific personal characteristics (e.g. trait-like measures), which may be related to specific vulnerability and different psychopathological outcomes.^14,15

Although different SMA patterns have been differentially linked to psychosocial health measures, studies have tended to focus solely on restricted facets of SMA (e.g. social media platforms) or have mixed educational and recreational use.^16,17 Moreover, most studies have lacked out-of-sample replication and have not examined generalizability of identified SMA patterns.^16–18

This study aimed to identify SMA patterns in a large and diverse youth sample from the Adolescent Brain and Cognitive Development Study (ABCD).¹⁹ Lifespan Human Connectome Project in Development (HCP-D) data were used to examine generalizability of the identified patterns.²⁰ Based on prior data,^16,18 we hypothesized that SMA patterns differing by type and pattern (frequency) would be identified, and more frequent patterns would associate with adverse measures (cognitive, behavioral). In addition, given significant brain plasticity during adolescence, a need exists for examining relationships between SMA and brain development in adolescents living in the current digital environment.²¹ Studies have suggested SMA-related differences in brain morphological features,^8,9 yet often lack evidence from large samples that can help clarify relationships between SMA patterns and developing brain function. Thus, this study also aimed to explore such associations and focused on resting-state brain functional connectivity (RSFC). RSFC is the co-fluctuation of the blood-oxygen-level-dependent signals between two brain regions and may statistically predict cognitive functioning and developmental disorders.^22,23 More frequent SMA has been related to weaker RSFC between brain regions involved in language and cognitive control,²⁴ and larger screen/reading time ratios have been linked to stronger RSFC among the salience network, language regions, primary visual networks, and motor-related regions.²⁵ However, SMA-pattern-specific differences in RSFC have not been comprehensively investigated; such studies may advance understandings of neural substrates of SMA-related developmental outcomes. In exploratory analyses, we preliminarily hypothesized that more frequent SMA patterns (versus less frequent SMA patterns) would be associated with RSFC between brain regions implicated in visual (occipital cortex), language (temporal cortex) and cognitive-control (cingulate or frontal parietal regions) processes. Furthermore, as excessive/problematic SMA has been linked to aberrant resting-state brain functional activities involved in abnormal emotion regulation and dysregulation in rewarding-stimuli processing, we hypothesized that more frequent SMA patterns (versus less frequent SMA patterns) may be associated with RSFC among regions/networks involved in emotional and reward-responsiveness processes, including the salience network and basal ganglia (especially amygdala, putamen, caudate and nucleus accumbens).^4,26 Given associations between excessive SMA and attention deficits,²⁷ we further hypothesized that more frequent SMA patterns (versus less frequent SMA patterns) would be associated with RSFC in ventral and dorsal attention networks implicated in bottom-up, stimulus-driven attention and top-down, goal-directed attention, respectively.²⁸

To investigate, we first conducted unsupervised machine learning clustering to identify SMA patterns and examine their robustness, generalizability and stability using sensitivity analyses, independent sample validation, and re-identification in follow-up data, respectively. Second, we examined differences in behavioral developmental measures and RSFC across different SMA groups in cross-sectional analyses. Third, we examined differences between SMA groups in longitudinal analyses, hypothesizing that SMA patterns would persist. Finally, we explored SMA patterns in adolescents to identify potential predictors of SMA pattern transitions.

METHOD

Study Design and Participants

The ABCD is a multi-site, longitudinal neuroimaging study with 11,876 participants enrolled at baseline from 22 U.S. sites.^19,29 Data analyzed here are from the ABCD 4.0 release tabulated data, which includes the baseline (2016–2018, ages 9–10 years), one-year follow-up (2017–2019, ages 10–11 years), and two-year follow-up (2018–2020, ages 11–12 years) assessments. Participants with missing SMA data at baseline (n=61) were excluded, leaving N_{ABCD-baseline}=11,815.

The HCP-D is a cross-sectional neuroimaging project spanning ages 5–21 years at four U.S. sites, with embedded longitudinal cohorts around puberty.²⁰ The HCP-D and ABCD are distinct yet include overlapping measurements.²⁰ The HCP-D 2.0 release data including 652 participants from the cross-sectional visit one were analyzed here, in which N_HCP-D=624 (aged 8–21 years) with available SMA data were used.

An analytic flowchart with list-wise deletions for missing values is provided in Figure S1, available online. More information about ABCD and HCP-D are provided in Supplement 1, available online. This study followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guidelines.

Variables

Details regarding the demographics, developmental measures of interest and brain-imaging-derived variables included in analyses are provided in Supplement 2 (Table S1 and S2, available online). Brief descriptions are listed below.

Screen Media Activity

SMA was assessed using the 14-item self-reported ABCD Youth Screen Time Questionnaire (STQ). Youth were asked to indicate the time spent on six types of SMA (e.g., watching TV or movies, playing video games) during the weekday and weekend as well as the frequency of mature-rated games and movies.

Mental Health

Parent-reported youth behavioral problems were assessed using the Child Behavior Checklist (CBCL), and t-scores of the CBCL scales were used. Youth-reported prodromal psychotic symptoms (PPS; i.e., psychotic-like experiences) were assessed using the ABCD-modified Prodromal Questionnaire Brief Version. Higher scores reflect more problems or psychotic-like experiences, respectively.

Neurocognition and Trait-like Measures

The age-corrected standard scores of the neurocognitive assessment from the NIH Toolbox were used. Higher scores reflect better cognitive performance. Trait-like measures included impulsivity and Behavioral Inhibition/Activation System (BIS/BAS), which were respectively assessed using the ABCD modified UPPS-P Impulsive Behavior Scale and the ABCD modified BIS/BAS scale.

Resting-state Brain Functional Connections

The resting-state functional MRI (rs-fMRI) scanning protocol, parameter and data preprocessing pipeline have been described elsewhere.^29,30 RSFC involved functional connections among 13 cortical networks and 19 subcortical regions from the tabulated data, which included 91 cortical (78 between- and 13 within-network RSFC) and 247 cortical to subcortical region RSFC (see Figure S2, available online). We followed the recommended inclusion criterion for rs-fMRI quality control in ABCD 4.0 release.

Covariates

Age, sex/gender, race/ethnicity, handedness, parental marital status, parental highest education, family income, household size, and household structure for children were included as covariates in liner-mixed-effect (LME) modeling analyses. The child’s handedness and head motion during rs-fMRI scanning (mean framewise displacement) were additionally included in RSFC analyses (see details in Supplement 3, available online.)

Statistical Analysis

To identify SMA patterns, K-means clustering based on squared Euclidean distance was applied to the 14 items in the STQ. Several cluster evaluation criteria were adopted to determine the optimal cluster number, including the Davies-Bouldin Index, Calinski-Harabasz Index and silhouette coefficient; see details in Supplement 4, available online. This clustering procedure was also applied to the ABCD one-year follow-up wave to test the stability of and transitions regarding identified SMA patterns.

Given the nested structure of the ABCD data, three-level LME models were used to test differences in developmental measures (twenty-one regression models for neurocognitive and trait-like measures, twenty-two for mental health related measures, 338 for RSFCs) between identified SMA patterns, where the individual SMA pattern label was dummy coded as an independent variable. Site and family nested within site were treated as random intercepts, and LME models were weighted by population weights to control for specific sources of selection bias and prevent possible biased estimation (see details in Supplement 5, available online).³¹

Longitudinal analyses were performed using an LME-implemented three-level conditional growth model, in which time was coded as a wave. Random slopes and intercepts at site and nested individuals were estimated.³²

All tests were two-sided. Benjamin-Hochberg false discovery rate (FDR) correction was used to adjust for multiple comparisons, both in cross-sectional and longitudinal analyses (see details in Supplement 6, available online).

Sensitivity Analysis

To examine the robustness and generalizability of the identified SMA patterns, we conducted a battery of sensitivity analyses. First, for internal validation, we randomly selected a set of subsamples from ABCD data with various sampling proportions (60%, 70%, 80%, and 90%) and re-clustered individuals in every subsample. This procedure was repeated 500 times across all sampling proportions to prevent potential bias in random selection. Second, for generalizability examination, we applied the same clustering procedure on HCP-D data, where we hypothesized finding SMA patterns with high similarity. Squared Euclidean Distance (SQD) was used to quantitatively measure similarities between cluster centroids, with smaller SQD values reflecting greater similarity; see details in Supplement 4, available online. To avoid possible methodological bias, we also applied hierarchical clustering and latent profile analysis for identifying SMA patterns.

Given that individuals with missing values were excluded from LME analyses, reassignment analyses based on randomly selected subsamples described above were used to examine influences of list-wise deletion. Reassignment accuracy was used to evaluate robustness of findings. It was hypothesized that individuals’ SMA pattern labels would not dramatically change after excluding parts of other individuals, which would reflect high reassignment accuracy.

RESULTS

Identification and Validation of SMA Patterns

In the ABCD baseline wave, two SMA patterns were identified among 11,815 youth aged 9–10 years, and all cluster evaluations consistently indicated it was the optimal clustering solution (Figure 1A and Figure S3A, available online). Internal validation testing suggested that SMA patterns could be reliably replicated in randomly selected subsamples (cluster centroids were very similar or even identical), and all cluster evaluations worsened with increasing cluster number (Figure 1B and Figure S4, available online). The average reassignment accuracy across all sampling proportions reached 99.83% [0.12%] (mean [SD], Figure S5, available online). Thus, the SMA patterns and individuals assigned to them remained almost unchanged after excluding large numbers of individuals. In addition, hierarchical clustering and latent profile analyses also provided similar results (Figure S6 and Table S3, available online).

Figure 1. Identification and Validation of the Screen Media Activity Patterns.

A) The internal validation of the ABCD baseline wave at sampling proportion of 80%. The solid line indicates the cluster centroids (subgroup average of item scores) identified in the full sample of ABCD baseline wave (N=11,815). The dotted lines indicate the cluster centroids identified in the randomly selected subsample (n=11,815*80%=9452, sampling without replacement). B) The cluster evaluation criteria values for the internal validation of the ABCD baseline wave. Error bars indicate the standard deviations. C) Comparison of the identified SMA patterns between ABCD baseline wave and age-separated HCP-D (age demarcation point: 13 years old). Lines indicate the cluster centroids in the full sample without random selection. N_{ABCD-baseline}=11,815, N_HCPD-younger=227, N_HCPD-older=397. D) Double y-axis scatter plot for the HCP-D Individual Squared Euclidean Distance. Every point indicates a participant from the HCP-D. Red points (left y-axis) indicate the SQD between participants’ STQ data and ABCD Subgroup 1. Blue points (right y-axis) indicate the SQD between participants’ STQ data and ABCD Subgroup 2. Smaller SQD values indicate higher similarity. The gray dotted line represents the age demarcation point.

According to their SMA patterns, the two clusters were respectively named as a higher-frequency SMA subgroup (hereafter refer to subgroup 1, n=3,151) and lower-frequency SMA subgroup (subgroup 2, n=8,664). Compared to subgroup 2, subgroup 1 spent significantly more time on video-related entertainment (i.e., watching TV, videos and playing games, Table S4, available online), regardless of weekday and weekend. However, both groups showed little time on social communications (i.e., texting, social networking and video chatting), with means or medians of these items close to zero in both subgroups (Table S4, available online).

The ABCD demographics are listed in Table 1 (see HCP-D demographics in Table S5, available online). The generalized LME with logistic regression illustrated that subgroup 1 was characterized by higher proportions of boys, Black and mixed/other race individuals, and specific familial factors, such as lower parental education, lower family income and dwelling in non-single households (Table S6, available online).

Table 1.

Baseline Demographics for the identified Screen Media Activity (SMA) patterns in the Adolescent Brain and Cognitive Development (ABCD)

	Higher-frequency SMA patterna	Lower-frequency SMA patterna	Nb	t/χ2	p	effect sizec
Demographics	subgroup 1 N=3151 (26.67%)	subgroup 2 N=8664 (73.33%)
Age (months)	119.16 ± 7.47	118.92 ± 7.5	11815	−1.55	0.12	−0.03
Gender			11815	305.09	<0.001	0.16
girls	1089 (34.56%)	4567 (52.71%)
boys	2062 (65.44%)	4097 (47.29%)
Handedness			11815	6.99	0.03	0.02
Right	2452 (77.82%)	6929 (79.97%)
Left	236 (7.49%)	609 (7.03%)
Mixed	463 (14.69%)	1126 (13.00%)
Relationship in family			11815	20.39	<0.001	0.04
single	2196 (69.69%)	5659 (65.32%)
sibling	424 (13.46%)	1376 (15.88%)
twin/triplet	531 (16.85%)	1629 (18.80%)
Youth’s Education Leveld			11814	0.93	0.35	0.02
Race			11694	605.03	<0.001	0.23
White	1601 (51.30%)	5886 (68.66%)
Black	909 (29.13%)	953 (11.12%)
Asian	33 (1.06%)	240 (2.80%)
Mixed/other	578 (18.52%)	1494 (17.43%)
Ethnicity			11815	5.29	0.02	−0.02
Hispanic	681 (21.61%)	1706 (19.69%)
Non-Hispanic	2470 (78.39%)	6958 (80.31%)
Parents’ Marital Status			11723	371.69	<0.001	−0.18
Single	1223 (39.31%)	1857 (21.56%)
Married or living with partner	1888 (60.69%)	6755 (78.44%)
Parents’ Highest Education			11801	375.35	<0.001	−0.18
high school or less	1268 (40.36%)	1940 (22.40%)
college education	1874 (59.64%)	6719 (77.60%)
Parents’ Employment Status			10871	429.11	<0.001	0.20
Married, 2 in LFe	1077 (37.80%)	4272 (53.25%)
Married, 1 in LFe	476 (16.71%)	1739 (21.68%)
Married, 0 in LFe	81 (2.84%)	170 (2.12%)
Single, in LFe	794 (27.87%)	1277 (15.92%)
Single, Not in LFe	421 (14.78%)	564 (7.03%)
Family Sized			11542	5.06	<0.001	0.11
Household Structure			11734	43.06	<0.001	0.06
Single household	2666 (85.59%)	7750 (89.92%)
Another household	449 (14.41%)	869 (10.08%)
Country of Birth			11630	0.26	0.61	−0.004
Other	50 (1.61%)	126 (1.48%)
USA	3057 (98.39%)	8397 (98.52%)
Family Incomed			10803	25.25	<0.001	0.55

^amean(SD) for continuous variables, No. (Percentage) for categorical variables.

^bthe number of individuals with available data, may not be equal to 11817 due to missing data. The denominator of percentages for categories is their corresponding N.

^cHedges’ g for continuous variables (two sample t-test), Cramer’s V or Cramer’s phi for categorical variables (chi-square test).

^dThese variables were rank-coded to ordinal continuous variables, see details in Supplement 2, available online.

^eLF: labor force.

For the generalizability examination, clustering results in a precisely age-matched HCP-D subsample (n=116, 9–10 years old) successfully replicated the ABCD-identified SMA patterns, which showed highly similar SMA subgroups (Figure S7, available online). While in the total HCP-D sample (N=624, 8–22 years old), clustering results cannot stably replicate these SMA patterns (Figure S8, available online), which may be explained by the broader age range in HCP-D and potential confounding effects of age. Thus, we speculated the identified SMA patterns may generalize to an age-similar subsample. Then, to determine to which age group the identified SMA patterns could generalize, we calculated SQD values between the HCP-D individual STQ data and ABCD-identified SMA patterns. Interestingly, we found a noticeable cliff in the SQD plot for all HCP-D participants, which precisely separated individuals at 13 years of age (Figure 1D). We next applied K-means separately on the HCP-D younger subsample (n=227, age ≤ 13 years) and HCP-D older subsample (n=397, age ≥ 13 years). As expected, the SMA patterns identified in the ABCD baseline wave were reliably replicated in the HCP-D younger subsample (Figure 1C, Figure S3C, and Figure S9A; demographics in Table S7, available online). To prevent possible influences from subjective separation, we directly applied k-means on the total HCP-D sample with the target cluster number being set as four (k=4), which theoretically should be equivalent to the 2(young versus old)-by-2(target cluster number, k=2) clustering, and we obtained similar results (Figure S10, available online).

Cross-sectional Results

The individual SMA pattern label was dummy-coded as a fixed-effect regressor in LME models (subgroup 2 as reference). Unadjusted LME models showed larger standardized betas. To avoid exaggerating effect sizes, we mainly reported results from adjusted LME models and provided effect size comparisons in eResults. After FDR correction, statistically significant SMA subgroup differences survived in neurocognitive measures (Figure 2A and Table S8, available online). Subgroup 1 showed poorer performance on multiple neurocognitive assessments and the Cognition Total Composite Score (Standardized Beta=−0.12, 95%CI, [−0.08, −0.16], FDR-corrected p<.001), except for processing speed and Flanker task. For trait-like measures, subgroup 1 was associated with higher impulsivity (UPPS-P sum score: Standardize Beta=0.36, 95%CI, [0.32, 0.41], FDR-corrected p<.001), BIS sum score (Standardized Beta=0.16, 95%CI, [0.11, 0.20], FDR-corrected p<.001) and BAS sum score (Standardized Beta=0.34, 95%CI, [0.29, 0.38], FDR-corrected p<.001).

Figure 2. Cross-sectional Differences in Neurodevelopmental Measures between Screen Media Activity Subgroups.

A, B) Differences in neurocognition, trait-like measures and mental health between the identified SMA subgroups in the ABCD baseline wave. Lines with square points indicate the subgroup average (z-scores, subtracted the mean of total sample and divided by its standard deviation). Shadowed areas indicate the standard errors of mean. * FDR-corrected p<.05; ** FDR-corrected p<.01; *** FDR-corrected p<.001; Text colored by light gray indicates FDR-corrected p>.05. See abbreviations in eMethods (Table S1, available online). C, D) Differences in RSFC between the identified SMA subgroups in the ABCD baseline wave. AuN: Auditory Network; CON: Cingulao-Opercular Network; CPN: Cingulo-Parietal Network; DMN: Default Mode Network; DAN: Dorsal Attention Network; FPN: Fronto-Parietal Network; none: None Network; RsTN: Retrosplenial Temporal Network; SmHN: Sensory/Motor Hand Network; SmMN: Sensory/Motor Mouth Network; VAN: Ventral Attention Network; ViN: Visual Network; Ventral DC: ventral diencephalon, which is a name given by Freesurfer for a group of mid-brain structures including the hypothalamus, substantia nigra, and ventral tegmental area; L: left hemisphere; R: right hemisphere. RSFCs with FDR-corrected p<.05 were drawn and different colors indicate different RSFCs; The positive network represents RSFC positively associated (subgroup 1 > subgroup 2) with the binary SMA subgroup variable (high-frequency SMA subgroup = 1, low-frequency SMA subgroup = 0) and vice versa for the negative network. Line width indicates the absolute of standardized beta; the thicker the line, the larger the standardized beta. The circular graphs do not show the within-network RSFC (self-looped connections); please see the matrix plot in Figure S11, available online, for this information.

As compared with subgroup 2, subgroup 1 demonstrated worse mental health (Figure 2B, Table S9, available online), including higher scores on total behavioral problems (CBCL TotProb: Standardized Beta=0.13, 95%CI, [0.09, 0.17], FDR-corrected p<.001) and more psychotic-like experiences (PPS severity score: Standardized Beta=0.31 ,95%CI, [0.26, 0.35], FDR-corrected p<.001).

In addition, we also found 38 RSFC features distinguishing subgroups 1 and 2 (Figure S11 and Figure S12, available online). These RSFC features were separated into positive (subgroup 1 > subgroup 2, positively associated with the dummy-coded binary SMA subgroup categorical variable, Figure 2C) and negative (subgroup 1 < subgroup 2, negatively associated with the SMA subgroup categorical variable, Figure 2D) networks. The positive network included six cortical network RSFCs and ten cortical to subcortical RSFCs, while the negative network included six cortical network RSFCs and sixteen cortical to subcortical RSFCs. The RSFC between the salience network and left ventral diencephalon (ventral DC) had the largest effect size (standardized beta=0.13, 95% CI, [0.08, 0.18], FDR-corrected p<.001). The nodes that were most frequently involved in the connections that showed a group difference were the dorsal attention network (DAN) and the default mode network (DMN; Figure S12.C, available online).

Longitudinal Results

We next tested whether differences persisted developmentally. Longitudinal analyses revealed statistically significant fixed effects of time and baseline SMA subgroup in trait-like measures and mental health problems (Figure 3A). The interaction term (Time×baseline SMA subgroup) was also significant except for CBCL TotProb (Table S10, available online). Subsequent results from simple effect analyses can be found in Table S11, available online. In addition, sixteen RSFCs showed significant fixed effects of SMA subgroups, of which the simple effect analyses showed four RSFCs kept the SMA subgroup differences at the two-year follow-up wave (Figure 3D, Figure S13 and Figure S14, available online). There were no significant interaction effects for RSFCs surviving FDR correction.

Figure 3. Stability of Screen Media Activity Patterns and Longitudinal Differences between Identified Subgroups.

A) Longitudinal results from all available behavioral measures data. *** p<.001 (Simple effect analysis results, corrected by Bonferroni method); n.s.: non-significant (p>.05). All variables were converted to z-scores by subtracting mean and dividing its standard deviation within each time point. Error bar indicates the standard error of mean. B) Comparison of the identified SMA patterns between ABCD baseline and one-year follow-up waves. C) The transition matrix of individual SMA pattern identified in baseline and one-year follow-up waves with available STQ data (the denominator of these percentage is N=11,125). McNemar’s Chi-square test was applied on the black box, where every cell indicates the number of individuals. For example, the cell at the first row with the second column (within-cell text: 1042, 9.38%) represents they were assigned into SMA subgroup 1 at baseline but into SMA subgroup 2 at one-year follow-up. The third row and the third column are the summation along the vertical (one-year follow-up) and horizontal (baseline) directions, respectively. D) The left panel shows four statistically significant RSFCs in a simple effects analysis for LMEs (Bonferroni Correction). The red lines indicate positive RSFCs (subgroup 1 > subgroup 2, positively associated with the binary SMA subgroup label), and blue lines indicate negative RSFCs (subgroup 1 < subgroup 2). Z-values indicate statistics (z.ratio from emmeans R package) from simple effects analyses, which compare RSFCs between two SMA subgroup at two-year follow-up wave. The right panel shows the details about the simple effects analyses and their results. The Y-axis indicates the estimated marginal means of RSFC strength. Error bar indicates the 95% CI for the estimated marginal means. See Figure S14, available online, for the descriptive means of RSFC comparisons.

* p <.05; ** p<.01; *** p<.001.

Regarding transitions and stability of SMA patterns, highly similar SMA patterns were identified in the ABCD one-year follow-up wave (N=11,125, Figure 3B, Figure S3B and S15, see similarity matrix in Figure S9B, available online), which also showed high reassignment accuracy (99.80% [0.13%], mean[SD], Figure S5, available online). However, it is noted that the cluster centroids after one-year were generally larger than at baseline except for watching TV or movies, which may be explained by overall increased SMA time (Table S12, available online). The week-average screen time significantly increased by 0.73 hours (paired t-test, 95% CI, [0.67, 0.79], Hedges’ g=0.22, p<.001). Furthermore, the proportion of SMA subgroups was significantly different between baseline and one-year follow-up (McNemar’s Chi-square test: Chi-square=82.81, p<.001). More individuals were assigned into the higher-frequency SMA group (i.e. subgroup 1, Figure 3C). An exploratory analysis with step-wise multinomial logistic regression found youth’s baseline age, gender, race, ethnicity, BAS, impulsivity, cognition total composite, total behavioral problems, PPS severity score sand RSFC between the salience network and left ventral DC statistically predicted individual SMA pattern transitions (see details in Supplement 7 and Table S13 and S14, available online). In addition, beyond SMA group differences, within-group associations between week-average screen time and developmental measures are provided in Figure S16–S18, available online, with different association strengths observed among SMA subgroups.

DISCUSSION

In the baseline wave of the ABCD cohort, two stable SMA patterns were identified which were respectively characterized by more frequent screen time involving television, videos and games (subgroup 1) and overall less frequent screen time (subgroup 2). Identified SMA patterns were validated in an independent sample from the HCP-D cohort. As hypothesized, youth with higher-frequency SMA showed poorer neurocognitive functioning, more severe mental health problems and higher trait-like measures of impulsivity and other constructs in comparison to the lower-frequency screen-using youth.

Characterization of the identified SMA patterns

The identified SMA patterns mainly differed in video-related entertainment instead of social communication, and this pattern was validated in the HCP-D younger subsample. In the older subsample, we found another social-communication-centric SMA pattern (green line in Figure 1C). Although these cross-sectional analyses cannot prove the video-entertainment/social-communication bifurcation is a consequence of development, the possibility of change in SMA with development has been suggested,³³ and the 13-year-old demarcation point has also been suggested in a recent meta-analysis.⁷ In addition, 13 years has been highlighted in theoretical frameworks as a developmental milestone.³⁴ At this age, individuals begin to establish autonomy and peer identification, where electronic screens may become a useful tool for social communications. For children younger than 13 years, self-reported time on video-entertainment-related items in the higher-frequency SMA pattern was approximately above 2 hours per typical day, which resonates with the Canadian 24-Hour Movement Guidelines for Children and Youth that recommends 2 hours or less for recreational screen time per day.⁵ Although the video-centric SMA pattern accounts for a minority of the sample (26.16% at ABCD baseline, 30.03% at ABCD 1-year follow-up), the potential overuse of video-related entertainment may contribute to differences in developmental outcomes for many youths.

Cross-sectional difference between SMA patterns

Differences in behavioral measures observed between SMA subgroups are generally consistent with previous studies investigating associations with screen time, such as poorer cognitive abilities,³⁵ higher impulsivity and BIS/BAS³⁶ and more behavioral problems.⁷ However, there are some seemingly inconsistent findings. On the one hand, we found significant differences in working memory and card-sorting, which were not reported as being associated with screen time in previous studies.^37,38 Two cognitive tasks (Flanker and processing speed) and two CBCL syndrome scales (anxious/depressed and withdrawn/depressed) did not show significant differences between SMA subgroups. On the other hand, larger effect sizes were observed than in a recent study utilizing ABCD data.³⁹ These seemingly inconsistent findings should be interpreted cautiously because the current results mainly indicated group-level differences, which, in a methodological sense, differ from association analyses. Moreover, the ages of participants varied in the previous studies, possibly explaining seemingly mixed results.

Regarding RSFC, it should be noted that statistical significance was sensitive to covariate adjustment where the unadjusted model showed more complex results (Figure S11 and S12, available online). Nonetheless, both the adjusted and unadjusted models revealed concordant findings. For example, the sign of standardized beta remained the same, and some RSFC with small effect sizes were not significant in the adjusted model, upon which we focus below.

Before discussing differences in RSFC strength, it should be noted that functional connection strength in rs-fMRI represents the magnitude of co-fluctuations between blood-oxygen-level-dependent signals in two brain regions, which is calculated by a correlation of time series instead of the actual cellular connections. As in our hypothesis, in the higher-frequency SMA subgroup, greater RSFC strength (hereafter referred to hyper-connection and vice versa for hypo-connection) was observed between visual and auditory networks. Paulus et al. have reported on greater brain structural maturation in visual areas in association with more-frequent SMA.⁹ The greater connections possibly suggest that frequent exposure to video-centric SMA may alter audio-visual systems leading to stronger functional integration among these regions.⁴⁰ Of more possible clinical concern, the DMN was found to be hyper-connected with the visual and dorsal attention networks in the higher-frequency SMA subgroup, which may be respectively related to low language functioning and difficulties in switching between internal and external stimuli.⁴¹ The greater connections between the salience network and ventral DC has been proposed to be involved in supersensitive reward responsiveness and motivational salience, which may potentially serve as neural substrates for frequent SMA.⁴² In the negative network, the ventral and dorsal attention networks were hypo-connected to a set of subcortical structures. The cingulo-opercular network showed weaker within-network connections and hypo-connections to subcortical regions. As the cingulo-opercular network has been implicated in sustained top-down, goal-directed control processes,²⁸ the weaker functional connection within it may reflect difficulties in executive control, working memory and task maintance.^25,43 The functional connection between the cingulo-opercular network and subcortical regions, especially the amygdala, has been associated with affective disorders (e.g., anxiety)²² and may moderate associations between interoceptive sensibility and bodily symptoms.⁴⁴ Similarly, the hypo-connections among the ventral/dorsal attention networks and these subcortical regions may reflect abnormal attention resource allocation and attention processing.²⁵ However, inconsistent with our hypothesis, we did not find significant difference in RSFC between frontal parietal network and visual or language network (the ventral attention network has been previously implicated in language functions).²⁸

Longitudinal findings and SMA transitions

Using available data in the ABCD 4.0 release, we examined SMA patterns at one-year follow-up and found similarities with baseline results. The transition matrix (Figure 3C) showed more individuals were assigned into subgroup 1 (one-year follow-up: 30.03%, baseline: 26.16%) and 13.53% of individuals changed from lower-frequency to higher-frequency SMA patterns, suggesting that with increasing age, more youth engaged in higher-frequency SMA. In longitudinal analyses, similar results have been reported for mental health problems over time,⁴⁵ and simple effects further revealed group differences in BAS, impulsivity and mental health remained significant, suggesting that relationships may persist over time. However, BIS scores were not significantly different at two-year follow-up, where the baseline SMA subgroup 1 decreased faster than subgroup 2 (Table S11, available online). Although, in the longitudinal findings, we found 16 RSFCs with significant fixed effects of SMA subgroups (overlap ratio with cross-sectional findings is 36.84% (14/38)), only four survived in simple effect analyses that compared the two SMA subgroups at the two-year follow-up wave, which included hyper-connections between the salience network and left ventral DC and hypo-connections among ventral attention and cingulo-pareital networks and left caudate and between the DMN and right amygdala (Figure 3D). The SMA subgroup differences in the four RSFCs have been detected in the baseline wave and persisted over a two-year period. These findings are inconsistent with our hypothesis regarding longitudinal effect of RSFCs and may be explained by high intra-individual variations in RSFC in conjunction with the brain development.⁴⁶ These longitudinal results suggest that the higher-frequency video-centric SMA pattern should be investigated further and may require public health interventions.

General discussion

First, although the higher-frequency video-centric SMA subgroup was not assessed to have a problematic/addictive SMA use pattern per se, the differences in the developmental correlates that characterized this SMA subgroup were consistent with the Interaction of Person-Affect-Cognition-Execution (I-PACE) model for specific internet-use disorders.^14,15 As a dynamic view, developing toward unhealthy and problematic screen use or specific internet-use disorder may be conditioned by frequent SMA for seeking gratification, which may be considered a risk factor.¹⁵ As for trait-like measures, the higher-frequent SMA subgroup showed relatively high sensitivity to punishment and reward (BIS/BAS),⁴⁷ which, combined with high impulsivity, may be considered as predisposing temperamental features in the I-PACE model.¹⁵ Second, the higher-frequency SMA subgroup exhibited more severe mental health problems, especially with attention-deficit/hyperactivity disorder (ADHD) having the largest effect size, which is similar to the psychopathological features noted in the I-PACE model and data linking the polygenetic risk of ADHD to excessive SMA use among youth longitudinally.²⁷ However, the current study did not examine the relationship between the higher-frequency video-centric SMA subgroup and internet-use disorder. Thus, the extent to which these findings link to specific internet-use problems and related functional impairment warrants additional study.

Several limitations of the present study should be addressed. First, some results came from cross-sectional assessments could not allow drawing of causal inferences. Further, the longitudinal analyses are preliminary, which predominantly investigate group differences between SMA subgroups. Given that individual SMA patterns may change throughout development, the underlying mechanisms and potential video-social bifurcations remain to be investigated further in more nuanced longitudinal investigations. Second, our findings are limited by the retrospective self-report nature of SMA. Although a pilot study from the ABCD Novel Technologies Workgroup has revealed there is a moderate correlation between self-report and objectively measured SMA (r=0.49), it also found participants used screens more than they reported.⁴⁸ Further studies utilizing some advanced approaches (e.g. passive-sensing and experience-sampling methods) may help examine the identified SMA patterns using more objective measurements, thus reducing recall bias. Third, the neuroimaging-derived RSFC investigated in this study is relatively coarse-grained and only captures large-scale network-level co-fluctuations, rather than the direct physiological links. Some significant RSFCs had negative connection strengths (Figure 3D, Figure S14, available online), of which the psychological and clinical meanings are challenging to interpret.⁴⁶ Further studies are needed to examine these data in a more fine-grained way (e.g., calculating connectivity based on an elaborated brain functional parcellation from the preprocessed brain images). Fourth, the higher-frequency SMA pattern may not reflect a single factor with respect to all measures of well-being, and future studies are needed to disentangle multiple factors linking SMA and health measures among youth longitudinally. Such information may help inform the updating of screen use guidelines and related interventions for youth in the contemporary digital technology environment. Our study also possesses several strengths. First, to our knowledge, this is the first study combining two independent samples (the ABCD and HCP-D) to identify SMA patterns among youth and examine generalizability. We also conducted a battery of sensitivity analyses to augment the robustness and reliability of the identified SMA patterns. Second, we systematically compared many developmental measures between the two SMA subgroups and investigated the underlying brain-behavior associations while controlling for potential confounders. These findings provide a more holistic view of SMA patterns and their relationship with cognition, trait-like measures, mental health and RSFC.

In this study, we combined two large cohort datasets and reliably identified two higher- and lower-frequency video-centric SMA patterns generalizable to youth under the age of 13 years. The cross-sectional findings indicated the higher-frequency SMA pattern was associated with adverse developmental measures, and the longitudinal analyses suggested persistence over a year-long period. These findings suggest the importance of considering SMA patterns when investigating effects of using digital devices and a possible target group for public health interventions. Further studies are required to understand the trajectories of SMA patterns, investigate their formation mechanisms and identify further implications for developing brains.

Diversity & Inclusion Statement:

We worked to ensure sex and gender balance in the recruitment of human participants. We worked to ensure race, ethnic, and/or other types of diversity in the recruitment of human participants. We worked to ensure that the study questionnaires were prepared in an inclusive way. We actively worked to promote sex and gender balance in our author group. We actively worked to promote inclusion of historically underrepresented racial and/or ethnic groups in science in our author group. While citing references scientifically relevant for this work, we also actively worked to promote sex and gender balance in our reference list. The author list of this paper includes contributors from the location and/or community where the research was conducted who participated in the data collection, design, analysis, and/or interpretation of the work.