Probing the digital exposome: associations of social media use patterns with youth mental health

The study addresses urgent public health concerns regarding how adolescent social media use relates to mental health. With nearly universal smartphone ownership among U.S. teens and increasing screentime, questions remain about whether and how specific social media behaviors contribute to psychopathology beyond general screentime and non-digital adversities. Prior research has shown small or mixed associations between social media use and mental health, with heterogeneity across studies. This study uses a data-driven exposome-wide association approach in a large, diverse ABCD cohort (age ~12) to test whether specific facets of social media exposure (e.g., app types, time patterns, cyberbullying, problematic/addictive use, parental monitoring) are associated with overall youth mental health, beyond general non-social screentime and established non-digital adversities. Hypotheses: (1) specific social media exposures would show stronger associations with worse mental health than general screentime and independent of non-digital adversity; (2) disparities would be evident in exposure levels and potentially in effects across sex, race/ethnicity, and sexual and gender minority (SGM) identity.

The paper synthesizes evidence that adolescent mental health problems have risen, with disparities among youth of color and SGM youth. Smartphone ownership exceeds 95% among U.S. teens and daily use is pervasive. Prior work, including ABCD analyses and meta-analyses, typically finds small associations between screentime/social media use and internalizing, externalizing, and depressive symptoms, with substantial heterogeneity and unclear directionality. Some studies suggest associations may be due to displacement of in-person activities or social comparison, while longitudinal evidence is mixed. CDC survey data links higher daily screentime with greater suicidal ideation, but it is unclear if screentime per se or specific digital exposures (e.g., cyberbullying) drive risk. The exposome framework has been used to model cumulative environmental risk for physical and mental health, motivating a comprehensive, multiexposure approach to the digital environment.

Design and data source: Cross-sectional analyses using ABCD Study Data Release 4.0, assessing social media and screentime at the 2-year follow-up (2019–2020). Participants: N=10,147 adolescents (mean age ~12). The full ABCD cohort initially recruited 11,876 children aged 9–10 across 21 U.S. sites. Dataset split: Using ABCD Reproducible Matched Samples (ABCD_3165), the sample was split into matched training (n=5,082) and testing (n=5,065) subsamples, matched on age, sex, ethnicity, grade, parent education, family income, and relatedness across sites. Outcomes: Primary outcome was adolescent self-reported psychopathology via the Brief Problem Monitor (BPM) Total Problems T-score; sensitivity outcomes included BPM Internalizing T-score and parent-reported CBCL Total Problems T-score. A higher-severity outcome was lifetime suicide attempt (self-reported via KSADS-5). Digital exposure variables: Youth and parent surveys provided 52 candidate digital/social media exposures; after removing low-endorsement items (<1%) and highly collinear variables (|r|>0.9), 41 exposures remained across five domains: (1) screentime (weekday/weekend minutes/hours), (2) parental monitoring (including suspected or self-reported secret accounts), (3) apps used (e.g., TikTok, Instagram, Snapchat account/use), (4) problematic/addictive patterns (e.g., perceived compulsive use, interference with schoolwork), and (5) peer interactions (e.g., cyberbullying victimization, feeling connected on phone). Non-social screentime (e.g., for schoolwork, non-social purposes) was computed from 13 items as a covariate. Data processing: Missing exposure data were imputed separately in training/testing via missForest (RandomForest). Extreme outliers on continuous variables were removed (e.g., followers). Mental health outcomes were not imputed. Exposome-wide association study (ExWAS): In the training sample, each of the 41 exposures was entered into a separate linear mixed-effects model (lme4::lmer) predicting BPM Total T-score, with random intercepts for family nested within site. Fixed covariates: age, sex, race (binary indicators for Black and White), and Hispanic ethnicity. False discovery rate correction was applied across 41 tests; exposures with positive coefficients were considered risk, negative coefficients protective. Digital exposomic risk score construction: For each participant in the testing sample, an aggregate exposomic risk score was computed as the weighted sum of all significant exposures, using their coefficients from the ExWAS LMEs. Higher scores reflect greater cumulative digital risk burden. Validation and specificity analyses: In the testing sample, successive LME models predicted BPM Total T-scores with fixed covariates and random effects (family nested within site). Model-1: demographics only; Model-2: + non-social screentime; Model-3: + non-digital childhood adversity exposome (from prior work); Model-4: + digital exposomic risk score. Variance explained (Nakagawa marginal R²) was compared across models. Suicide attempt analyses: Logistic regression models in the testing sample examined associations between the digital exposomic score and lifetime suicide attempt history (KSADS-5), covarying demographics, non-social screentime, and then additionally non-digital childhood adversity. Disparities analyses: Differential exposure was tested by comparing exposomic risk scores across sex, race/ethnicity (non-Hispanic White, non-Hispanic Black, Hispanic), and SGM identity using nonparametric tests: Kruskal–Wallis with Dunn’s post hoc tests (Holm-adjusted) for three-group comparisons and Mann–Whitney tests (Glass rank-biserial r) for two-group comparisons. Differential effects were tested via interaction terms between exposomic risk and sex, race/ethnicity, and SGM identity in the main LME models; significant interactions were probed with stratified analyses. Sensitivity analyses: Re-ran ExWAS with list-wise deletion instead of imputation; repeated validation without outlier removal; excluded youths without smartphones or without social media use; examined alternative outcomes (BPM Internalizing, CBCL Total); conducted E-value analysis with BPM Total T-score dichotomized (top 10% vs. others) to assess robustness to unmeasured confounding. Random effects and covariates: All LMEs included random effects for families nested within site; fixed covariates included age, sex, race, Hispanic ethnicity, household income (ordinal 1–10), and parent education (highest grade/degree).

- ExWAS results (training sample): Of 41 digital/social media exposures tested, 35 were significantly associated with higher BPM Total T-scores after FDR correction. Notable risk factors with larger effects included: cyberbullying victimization, having secret social media accounts unknown to parents, and addictive/compulsive patterns of social media use; weekday video game screentime also showed significant associations. Having a private (friends-only) vs. public account showed a protective association. - Validation and variance explained (testing sample): Sequential LME models showed: Demographics explained 1.14% of variance (Model-1). Adding non-social screentime increased explained variance to 6.18% (b=1.38, 95% CI [1.19–1.56], t=14.46, p<0.001; Model-2). Adding non-digital childhood adversity increased variance to 9.07% (b=1.31, 95% CI [1.07–1.56], t=10.65, p<0.001; Model-3). Adding the digital exposomic risk score further increased variance explained to 15.61% (estimate=1.78 per z-score, 95% CI [1.58–1.98], t=17.41, p<0.001; Model-4). - Suicide attempts: Higher digital exposomic scores were associated with higher odds of lifetime suicide attempt (OR=1.76, 95% CI [1.39–2.23], z=4.69, p<0.001) when adjusting for demographics and non-social screentime. This association remained significant and strengthened when additionally adjusting for non-digital childhood adversity (OR=2.76, 95% CI [1.88–4.05], z=5.18, p<0.001). Non-social screentime was not significantly associated with suicide attempts (OR=1.10, 95% CI [0.82–1.47], z=0.63, p=0.53). - Disparities in exposure: Digital exposomic risk scores were highest among non-Hispanic Black youth (median=0.38), intermediate among Hispanic youth (median=0.01), and lowest among non-Hispanic White youth (median=−0.44); Kruskal–Wallis χ²=480.90, p<0.001; all pairwise Holm-adjusted p<0.001. No sex differences (p=0.17). SGM youth had higher scores than non-SGM peers (rank-biserial r=0.33, p<0.001; medians 0.42 vs −0.26). - Differential effects: A significant interaction indicated a weaker association between digital exposomic risk and BPM scores among Black youth compared to others (interaction estimate=−0.12, t=−3.37, p=0.001). No significant interactions by sex or SGM identity. - Sensitivity analyses: Findings were robust to outlier handling, imputation vs. list-wise deletion, restricting to youths with smartphones or social media use, and using alternative outcomes (BPM Internalizing; CBCL Total). E-value analysis suggested an unmeasured confounder would need ~3.3-fold associations with both the exposome score and high BPM to explain away effects beyond measured covariates.

The study demonstrates that specific digital exposures, particularly within social media contexts, are associated with adolescent mental health burden above and beyond general non-social screentime and non-digital adversity. By leveraging an exposome framework, the authors show that aggregating multiple correlated digital risk factors into a weighted score captures substantial variance in psychopathology and relates to suicide attempt risk, suggesting that the type and quality of digital interactions, not total screentime per se, are critical. Identified high-impact exposures (cyberbullying, secret accounts, problematic/addictive use) align with established risk mechanisms, supporting the validity of the approach. The results also delineate disparities in exposure: Black and SGM youth carry greater digital risk burdens; however, Black youth show weaker exposure–outcome coupling, potentially reflecting buffering by supportive online communities or differing salience of offline risks. Clinically and for public health, these findings argue for interventions targeting modifiable digital exposures (e.g., reducing cyberbullying, promoting privacy and parental monitoring practices, addressing problematic use) and for tailoring strategies to subgroups with higher exposure burdens. The exposomic approach provides a scalable tool for risk stratification and hypothesis generation to inform prevention and policy.

This work introduces an exposome-wide, data-driven method to quantify the cumulative burden of digital/social media exposures on youth mental health. The derived digital exposomic risk score robustly associates with general psychopathology and suicide attempt history, independent of demographics, non-social screentime, and non-digital adversity. The study highlights specific modifiable digital risks (cyberbullying, secret accounts, problematic use) and documents exposure disparities among Black and SGM youth. Future research should employ longitudinal ABCD waves and other cohorts to clarify directionality and causality; incorporate richer, platform- and behavior-specific measures (active vs. passive use, public vs. private accounts, weekday vs. weekend patterns); integrate multimodal digital phenotyping (passive sensing, wearables) and objective usage logs; and examine resilience-promoting and positive online activities. Interventions should focus on mitigating harmful digital exposures while acknowledging potential benefits of online communities, particularly for minoritized youth.

- Cross-sectional design limits causal inference and directionality. - Digital exposure assessments were not specifically designed for comprehensive social media interrogation in ABCD, limiting scope and depth; reliance on self- and parent-report may introduce bias. - Residual collinearity among exposures is possible despite removing highly collinear variables; exposomic scores do not fully model correlated structures. - Generalizability should be examined in independent samples and evolving digital ecosystems; platform usage trends (e.g., TikTok’s rise) may shift associations over time. - Potential unmeasured confounding remains, though E-value analysis indicates effects are robust to moderate confounding. - Lack of detailed measures of minority stress, especially for SGM youth, and limited assessment of positive or resilience-promoting online experiences.