Introduction

Since the late 1990 s, the expansion of social media technology has radically transformed human interconnectedness, creating new dynamics of communication and collaboration. However, healthy and responsible use of these technologies remains a major challenge, especially in populations where their use is particularly prominent, such as children and adolescents1,2. The latest international Health Behaviour in School-aged Children (HBSC) report indicates that 11% of adolescents use social media problematically. This prevalence represents a slight increase since 2018 (7%) and is higher among girls than boys3. In addition, the ESPAD Group warns that in Europe, almost half of adolescents (47%) perceive problems associated with their use4.

The negative consequences of social media abuse have sparked enormous interest among scientists and educators. Two main approaches stand out from the conceptual debate: the cognitive-behavioural paradigm of problematic use5,6 and the addiction components model7,8, the latter being the most widely used. The conceptualisation of this phenomenon as an addiction arises from an atheoretical perspective and within the framework of substance abuse9, without sufficient consensus10 for its inclusion in the main diagnostic manuals11,12. This conceptual disparity creates heterogeneity in research, hindering the replicability and comparison of results by not ensuring that different methodologies measure the same underlying construct13. Therefore, throughout this paper, the term ‘problematic use’ will be used according to the cognitive-behavioural approach, adapting Caplan’s model5 to social media, which allows us to cover all the nomenclatures used to refer to this phenomenon. Understanding problematic social media use (PSMU) as the excessive use of social media platforms that results in harmful consequences for a person’s personal, social, or professional functioning, experiencing negative psychosocial outcomes5,14.

These psychosocial consequences of PSMU are particularly relevant at a time when the mental health of children and adolescents is a global concern. It is estimated that 13% of adolescents between the ages of 10 and 19 suffer from some mental disorder, of which 40% are diagnosed with depression and anxiety disorders15. The relationship between social media use and the presence of depressive symptoms in adolescence is complex, and despite increased research in recent years, the results are inconclusive. Generally, they show weak associations between social media use and depressive symptoms, with greater relevance in girls16,17,18, who tend to experience more persistent depressive symptoms over time19. Nevertheless, in many cases, these associations lack clinical significance or are explained by sociodemographic or emotional variables20. Other authors did not find significant relationships with subjective well-being or mental health in general21,22. A key aspect is the direction of the effect, as depressive symptoms could predict greater use of social media, rather than the reverse23,24.

Nonetheless, problematic use has been suggested to be a variable with greater explanatory and predictive power than simple time of use25,26. Furthermore, gender has been identified as a moderating variable, with adolescent women and gender-diverse individuals showing greater vulnerability to the negative effects of PSMU, reflecting significant differences in terms of psychological well-being27,28. Other moderating variables on emotional well-being include other-oriented use29, self-esteem30, social comparison25, cybervictimisation25 and the emotional valence of social media experiences31. It is also suggested that increased activity by followers increases the frequency of use of social media, mediated by emotional regulation32. Furthermore, adolescence itself is a period of greater neurobiological vulnerability to the effects of digital technology use33. The negative correlation between social media use and overall life satisfaction is more pronounced in children aged 10 to 15 than in those over 16 years of age34.

The reviewed studies share common methodological limitations, including the lack of a unified conceptual framework, heterogeneity in measurement, insufficient longitudinal designs, and the use of statistical models that do not always adequately capture intra-individual changes26,35,36. Consequently, there is a need for further longitudinal studies with more robust designs that integrate more accurate measurements and employ analytical models focused on individual change. Such an approach would help clarify causal directions, detect risk profiles, and establish more effective interventions based on individual and contextual characteristics.

Therefore, the overall objective of this study is to analyse the influence of social media use on depressive symptoms in the adolescent population. Specifically, to determine the influence of PSMU, frequency, and number of followers on social media on depressive symptoms, as well as to determine the moderating effect of gender and age on the presence or absence of depressive symptoms. Another objective is to determine whether the presence of previous depressive symptoms enhances the negative effects of PSMU, following vulnerability-stress models that suggest that risk factors may have amplified effects in predisposed individuals37,38.

Based on the evidence presented, we propose that problematic social media use will positively predict the presence of depressive symptoms (Hypothesis 1). Likewise, greater depressive symptoms will be recorded in individuals with higher daily frequency (Hypothesis 2) and with a higher number of followers on social media (Hypothesis 3). We expect both age and gender to have a moderating effect. Younger age (Hypothesis 4) and being female (Hypothesis 5) will increase the scores of depressive symptoms. Finally, we expect that people with previous depressive symptoms will experience an increase in their symptoms (Hypothesis 6).

Results

Descriptives

In general terms, women reported higher levels of depressive symptoms than men at both T1 and T2, although these differences can be considered moderate (see Table 1). The most notable difference is observed in the daily frequency of social media use (SMU), which is considerably higher in women than in men. Regarding problematic social media use (PSMU), this sample shows less marked differences between genders, with slightly higher values in women. Regarding the number of followers on social media (SMF), the results reflect similar values in both groups, with a slight advantage for women.

Table 1 Descriptive statistics of study variables.
Full size table

Although the statistics show good stability overall, it is important to note that the estimates for the gender category ‘Other’ should be interpreted with caution due to the small sample size (n = 36), which leads to greater variability and lower statistical precision.

The bivariate correlations between the main variables of the study are shown in Table 2 using Spearman’s correlations. Both the PSMU and the SMU show a positive and significant relationship with depressive symptoms, especially at T2. Likewise, the SMF also correlates positively with depressive symptoms, although the magnitude of this association is smaller. Finally, age showed statistically significant, albeit weak, correlations with the other variables, indicating a moderate role. These preliminary results support the use of quantile regression models.

Table 2 Spearman correlations between the study variables.
Full size table

Base model

The base model obtained is presented in Table 3. PSMU showed a positive and significant association in all quantiles, indicating that greater problematic use is related to an increase in depressive symptoms at T2. Similarly, the presence of depressive symptoms at T1 is significantly associated with an increase in symptoms at T2. On the other hand, SMF and SMU did not show a significant association with depressive symptoms, suggesting that the main effects of these factors are not relevant to depressive symptoms. The baseline model showed adequate fit across quantiles, with pseudo-R² values of 0.172 for τ = 0.25, 0.208 for τ = 0.50, and 0.220 for τ = 0.75.

Table 3 Quantile regression model for depressive symptomatology at time 2 (Base model).
Full size table

Concerning gender, girls consistently presented higher levels of depressive symptoms compared to boys in all three quantiles, especially in the second and third (p =.002 and p <.001, respectively), suggesting a systematic gender gap in depressive symptoms. On the other hand, no main effects of age on depressive symptoms were observed at T2.

The marginal effects of the base model showed that previous depressive symptoms were the strongest predictor (β = 0.478–0.568), followed by a moderate effect of problematic social media use (β = 0.161–0.217) and gender for girls (β = 0.075–0.147). The main effects of the other variables were smaller (β < 0.05), suggesting the need to examine interaction effects (see Fig. 1).

Fig. 1
figure 1

Marginal effects of baseline quantile regression model for depressive symptoms.

Full size image

Given that theoretically oriented interactions did not show solid statistical evidence in confirmatory analyses, an alternative model based on previous exploratory analyses was explored. Systematic comparison of eight candidate models (Supplementary Material 1) revealed that the model with SMU × Age, SMF × Gender, and SMF × SMU interactions showed the best fit (AIC = −3394.31 vs. −3385.29 for the base model, ΔAIC = −9.02 points) and robust statistical evidence for the included interactions.

Adjusted model with interactions

The final model showed a superior fit and robust evidence for the interactions included, especially SMU × Age, significant in all three quantiles, and SMF × Gender, significant in the first and second quantiles. Compared to the baseline model (see Table 3), the final model with interactions (see Table 4) showed improved fit across all quantiles. The pseudo-R² increased from 0.172 to 0.178 at τ = 0.25 (improvement of 3.4%), from 0.208 to 0.214 at τ = 0.50 (improvement of 2.9%), and from 0.220 to 0.226 at τ = 0.75 (improvement of 2.9%). The largest improvement was observed at the lower quantile (τ = 0.25), consistent with the finding that interaction effects were most prominent at this distributional level. Quantile regression analyses showed that in the adjusted model, PSMU and SMU maintain a positive and significant main effect in all quantiles, while SMF does not show significant associations with depressive symptoms (see Table 4). Likewise, being female rather than male only increases depressive symptom scores in the third quantile, being significant in the first two in interaction with the number of followers. Overall, the results indicate that the final model offers a more detailed representation of the relationships between the predictor and dependent variables, albeit with moderate increases in pseudo-R². In particular, the interaction between SMU and age remains a robust effect in all three quantiles, reinforcing its importance in predicting depressive symptoms.

Table 4 Final quantile regression model with key interactions for predicting depressive symptomatology at time 2.
Full size table

Firstly, the main effects of the technological variables show that PSMU retains a positive association with depressive symptoms across all quantiles. This confirms that more severe and problematic use of social media is consistently related to higher levels of depressive symptoms.

On the other hand, SMU presents a more complex pattern. Although its direct effect remains in all quantiles, the results indicate that this effect is moderated by age. Figure 1 shows how this interaction manifests itself differently depending on the level of depressive symptoms (quantiles τ = 0.25, 0.50 and 0.75) and the age of the adolescents. At age 13, more intensive use of social media is associated with an increase in depressive symptoms at all three levels of the distribution. However, as age increases, this relationship is reversed: at age 16, increased frequency of use is linked to a slight decrease in symptoms, especially at the lower and middle levels of the distribution. This pattern suggests that frequency of use may have different implications depending on the stage of adolescent development, and that the age between 14 and 15 could be a particularly sensitive stage.

This visual interpretation is supported by the statistical results: the SMU × Age interaction coefficient is negative and significant in all three quantiles, indicating that age modulates the effect of daily social media use on depressive symptoms at T2 (see Table 4; Fig. 2).

Fig. 2
figure 2

Relationship between daily social media use and depressive symptoms at different quantiles, moderated by age.

Full size image

The results indicate gender differences in depressive symptoms at low and medium SMF levels. The SMF × Female interaction is significant in both the lower quantile (τ = 0.25, p =.022) and the median (τ = 0.50, p =.045), indicating that the impact of this variable varies according to gender. Figure 3 clearly illustrates this pattern: while in girls a higher number of followers is associated with a progressive increase in depressive symptoms, in boys the opposite trend is observed. This crossover suggests that, especially at the lower and middle levels of symptomatology, digital social visibility may have different emotional implications depending on gender. For girls, accumulating followers could be linked to greater pressures or expectations, while for boys it could have a neutral or even slightly protective effect.

Fig. 3
figure 3

Relationship between social media followers and depressive symptoms at different quantiles, moderated by gender.

Full size image

Finally, the model shows an interaction effect between the number of followers on social media and daily frequency of use (SMF × SMU) that is significant at the 0.25 quantile (p =.009) but not at the 0.50 (p =.163) or 0.75 quantiles (p =.340). As shown in Fig. 4, this interaction displays a crossover pattern: among adolescents with high daily use, more followers associate with slightly lower predicted symptoms, while among those with low daily use, more followers associate with slightly higher symptoms. However, several factors warrant caution in interpreting this pattern. First, the interaction appears only at the lowest quantile (adolescents with minimal baseline symptoms), not at moderate or higher symptom levels. Second, post-hoc tests revealed that interaction coefficients did not differ significantly across quantiles (p >.15), providing limited evidence for true differential effects. Third, the effect size is small, and the pattern lacks clear theoretical precedent in the literature. One speculative explanation is that for light users, a large follower count may create social pressure without developed engagement skills, while for frequent users, it may provide social validation. However, this interpretation remains tentative and requires independent replication before drawing firm conclusions.

Fig. 4
figure 4

Relationship between social media followers and depressive symptoms at τ = 0.25, moderated by daily social media use.

Full size image

Discussion

This study aimed to analyse the influence of problematic social media use on depressive symptoms in the adolescent population. Specifically, it sought to determine the influence of PSMU, time spent on social media, and number of followers on depressive symptoms, as well as to determine the moderating effect of gender. To this end, a sample of secondary school students was followed up and asked to complete a questionnaire on social media use on two separate occasions, one year apart. Quantile regression models were performed, including three significant interactions.

The final model confirmed that PSMU and being female rather than male increase levels of depressive symptoms, thus verifying hypotheses 1 and 5. The results are in line with those found in previous studies, which support the idea that problematic social media use increases depressive symptoms38,39, as well as the development of other mental health problems25,40. Gender stands out as a significant variable, with girls showing higher scores on depressive symptoms41,42. Hypotheses 2 to 4 are not confirmed in terms of simple effect, but secondary effects do appear through interactions.

About the frequency of social media use and the number of followers, the results are more heterogeneous and complex, showing significant effects in interaction with other variables or among themselves on depressive symptoms. Firstly, the interaction between frequency of use and age suggests that the effect of SMU diminishes with age. These results could be because at younger ages, the ability to regulate intensive use of social media is lower, leading to an increase in depressive symptoms. These findings only partially confirm the hypothesis 2 which predicted a positive effect of daily use (SMU) on depressive symptoms, and are in line with studies that support the idea of early adolescence as a vulnerable stage in the development of mental health problems43,44 due to the maturational gap between the limbic system and the prefrontal cortex45,46 and the dopaminergic hypersensitivity47. The results on the effect of age support the debate on the minimum age for access to social media without parental consent. Although the decision to set the minimum age for social media use at 16 is debatable, the findings suggest that regulation in this area is advisable, as well as an evaluation of its effectiveness. In this regard, the European Union’s General Data Protection Regulation sets the minimum age at which a minor can give consent for the processing of personal data on social media at 16 years48. In the United States, some states, such as Utah and Arkansas, have passed laws restricting access to networks for minors under the age of 18 without parental consent and limiting night-time use. Australia has also passed new legislation that will prohibit children under the age of 16 from accessing social media. These examples show an international trend towards tightening access conditions and towards the joint responsibility of technology platforms in protecting minors. The entry into force of these measures provides an opportunity to assess whether the ban is an effective form of protection or whether, on the contrary, it is a low-impact strategy compared to other educational measures.

The interaction between the number of followers and gender has shown a differential effect between girls and boys. This finding reinforces the importance of considering the mechanisms of social validation and aesthetic pressure that particularly affect adolescent girls, which have been documented in studies of self-representation on social media49, aesthetic pressure50 and appearance validation and social compensation51. Nevertheless, there is a lack of consensus on its role as a moderating variable between social media use and the development of mental health problems52. Hypothesis 3, which anticipated a direct effect of the number of followers (SMF) on symptoms, is not confirmed in the main models. However, significant effects emerge in interaction with gender and frequency of use, suggesting that the variable becomes relevant only when combined with other factors.

Thirdly, a significant interaction has been observed between frequency and number of followers on social media. Specifically, people who use social media less frequently and have many followers may experience greater pressure for validation compared to those who have fewer followers or use social media more frequently. This interaction suggests that the effects of exposure and social visibility on social media are intensified when both factors are present and cannot be fully understood in isolation when considering patterns of social interaction and validation53,54. This specific risk profile—low frequency of use coupled with high visibility—can give rise to several practical challenges. For example, adolescents in this group may be more susceptible to negative comments, greater anxiety about maintaining their online reputation, or greater vulnerability to social comparison55. These risks highlight the complex emotional landscape created by social media use and underscore the need for specific interventions to help young people navigate these dynamics by addressing the different mechanisms that mediate and moderate these interactions, such as the quality of time spent or the quality of relationships with followers56.

The results indicate that problematic social media use can be considered a risk factor for mental health, specifically for the onset of depressive symptoms. This finding reinforces the importance of early interventions during childhood and adolescence, aimed at responsible use of social media57, in line with the recommendations of various institutions, such as the American Psychological Association58. This study is not without limitations, mainly related to methodological issues. Although the selection of educational centres was random, only those centres that gave their consent participated in the study. Another limitation could derive from not having differentiated between types of use: active versus passive. The measures used are based on self-reporting, which exposes them to social desirability bias, recall bias, and underestimation or overestimation of actual network usage time. In particular, the variable ‘daily usage time’ was assessed using a single ad hoc question, without objective recording (e.g., use of digital data or screen monitoring applications), which may affect the accuracy of the results. Nor were the types of use (active vs. passive) differentiated, despite evidence that they have different impacts on depressive symptoms59. The use of quantile regression provides a differentiated view by symptom levels, but the pseudo-R² obtained is moderate, indicating that other variables not considered could explain a significant part of the variance.

Finally, this study gives rise to different areas of opportunity for further research. The role of problematic social media use as a predictor of the development of other mental health or psychological problems, such as fear of missing out (FOMO), self-esteem, or social comparison. As well as delving into other variables related to social media, such as the type of social media used. Conducting more longitudinal studies is a key aspect in this area of knowledge, selecting a larger sample with a wider age range to refine the most vulnerable developmental stages for the development of mental health problems in interaction with social media use. As for the applicability of the conclusions obtained, beyond general preventive measures, our findings on the interaction with age have a crucial practical implication. The results suggest that waiting beyond mid-adolescence to implement digital literacy and healthy technology use programmes may be insufficient, as habits of high daily social media use may become entrenched. Therefore, intervening at early stages, such as pre-adolescence or early adolescence, would provide minors with the necessary tools before their exposure to and autonomy in the use of social media intensify. This would enable the development of much more efficient preventive measures in educational and family settings, thus anticipating periods of greater vulnerability and promoting greater awareness of the current uses of social media by young people.

Method

Procedure and participants

The study was approved by the educational authority, school managing board, and parents or legal guardians. Participants were recruited from randomly selected secondary schools in the Valencian Community (Spain). Data were collected at two time points one year apart: Time 1 (T1) from October 2022 to January 2023, and Time 2 (T2) from October 2023 to January 2024.

Sample refinement and matching

Initial participation included 5,248 students at T1 and 4,448 at T2 across 39 and 31 schools, respectively. Data quality control involved four sequential exclusion criteria: (1) elimination of centres providing fewer than five cases, (2) removal of cases with incomplete or inconsistent responses, (3) exclusion of cases with response times outside the 15–90-minute range, and (4) age restriction to 13–15 years at T1 and 13–17 years at T2. This process resulted in refined samples of 4,556 participants at T1 (39 schools) and 3,887 at T2 (31 schools).

Longitudinal matching was restricted to the 29 schools that participated in both time points, comprising 3,834 eligible students at T1 and 3,538 at T2. Anonymous individual codes were used for matching, yielding a final matched sample of 2,121 students (retention rate = 55.3%). This retention rate is consistent with longitudinal studies examining social media use and mental health in adolescent populations, which commonly face substantial attrition challenges due to school transfers, absences, and the complexities of anonymous tracking systems24,38,60,61. To address potential selection bias from differential attrition, comprehensive analyses confirmed that the matched sample remained representative of the initial population across key demographic and psychological variables (see Attrition Analysis).

The mean age was 13.79 years (SD = 0.74) at T1 and 14.73 years (SD = 0.76) at T2. The sample comprised 1,111 girls, 974 boys, and 36 participants selecting “Other” gender.

Instruments and variables

The independent variables were daily frequency of social media use, number of followers, problematic social media use, age, and gender. The dependent variable in this study was depressive symptoms. The instruments used to assess each of these variables were as follows:

Depressive symptoms (CESD). This was assessed using the CES-D questionnaire62,63. It has seven items with four alternative Likert-type responses (1 = Never/4 = Always). The reliability of this questionnaire was Cronbach’s α = 0.83.

Problematic social media use (PSMU). This was assessed using the social media subscale of the MULTICAGE-TIC questionnaire64, which evaluates problems associated with the use and abuse of social media. This subscale consists of four dichotomous response items (YES/NO), with scores ranging from 0 to 4 points. Regarding the consistency of this questionnaire, it obtained a Cronbach’s α of 0.93.

Frequency of social media use (SMU). This was assessed using an ad hoc question: ‘On a normal day, how many hours on average do you spend on social media?’ This was answered on a scale ranging from 0 to more than 10 h.

Number of followers on social media (SMF). This was assessed with an ad hoc question: ‘How many friends or followers do you have on your two favourite social media platforms?’ This was answered using a scale ranging from 0 to more than 1,000 followers for each social media platform.

The questionnaire used can be found in the Supplementary Material 2.

Statistical analysis

Descriptive statistics for all key variables, including measures of central tendency, dispersion, and data quality, were calculated by gender and for the total sample (see Table 1).

Attrition analysis

To assess potential selection bias due to longitudinal attrition, we compared baseline characteristics between participants successfully retained in the longitudinal sample (n = 2,121) and those lost to follow-up from the eligible sample of 29 schools participating at both time points (n = 1,713). This analysis evaluates the representativeness of the longitudinal cohort prior to any analytical exclusions.

Comparisons were conducted using independent-samples t-tests for continuous variables and chi-square tests for categorical variables. Effect sizes were calculated using Cohen’s d and Cramer’s V respectively, with values of 0.2, 0.5, and 0.8 representing small, medium, and large effects65.

Six baseline (T2) variables were examined: age, gender, depressive symptoms (CESD), problematic social media use (PSMU), social media use frequency (SMU), and follower count (SMF). Results revealed statistically significant differences in five variables: retained participants were slightly younger (Mretained = 13.79, SD = 0.74 vs. Mlost = 14.07, SD = 0.86; t = −10.63, p <.001, d = −0.35), more likely to be female (χ² = 23.11, df = 2, p <.001, V = 0.08), had slightly lower baseline depressive symptoms (Mretained = 2.06, SD = 0.67 vs. Mlost = 2.15, SD = 0.70; t = −3.88, p <.001, d = −0.13), reported lower social media use frequency (Mretained = 3.45, SD = 2.46 vs. Mlost = 3.92, SD = 2.73; t = −5.41, p <.001, d = −0.18), and had fewer followers (Mretained = 720, SD = 793 vs. Mlost = 881, SD = 929; t = −5.67, p <.001, d = −0.19). No significant difference was found for problematic social media use (p =.252, d = −0.04).

Importantly, all significant effect sizes were small to trivial (|d| ≤ 0.35, V < 0.1), indicating that differences are not practically meaningful66,67. The largest differences (age and gender) are controlled as covariates in all regression models, statistically adjusting for these baseline imbalances. The minimal effect sizes for outcome and predictor variables (CESD, SMU, SMF) further suggest that longitudinal attrition introduces negligible bias and does not substantially compromise the validity or generalizability of findings68.

Missing data analysis and treatment

Due to the very small number of cases who selected the “Other” gender option (n = 36), these cases were excluded from the analysis because of low representativeness and minimal statistical power in regression models, resulting in a refined sample of 2,085 adolescents.

Prior to conducting the main analyses, the pattern and mechanism of missing data were systematically examined. Missing data analysis revealed that 6.04% of cases had incomplete information across study variables. Little’s MCAR test indicated that data were not missing completely at random (χ² = 233.20, p <.001), necessitating further investigation of the missing data mechanism. To distinguish between Missing At Random (MAR) and Missing Not At Random (MNAR) mechanisms, logistic regression analysis was conducted with data completeness as the dependent variable and baseline demographic and psychological variables as predictors. Results indicated that only the number of social media followers at baseline significantly predicted data completeness (OR = 1.001, 95% CI [1.000, 1.002], p =.048), representing a minimal effect size. Key demographic variables (age: p =.881; gender: p =.859) and psychological variables (baseline depressive symptoms: p =.198; social media use frequency: p =.508; appearance-related social media consciousness: p =.145) did not significantly predict missingness patterns. These findings support a MAR mechanism, where missingness depends primarily on observed variables rather than unobserved data. Given that the only significant predictor of missingness (social media followers) was included as a covariate in all analytical models, listwise deletion was considered methodologically appropriate. Under MAR conditions, listwise deletion provides unbiased parameter estimates when variables that predict missingness are controlled for in the analytical model68. The final analytical sample comprised 1,959 adolescents, representing a retention rate of 93.96%.

Variable operationalization

The primary predictor variables, averages of T1 and T2 measurements, were calculated to capture stable patterns in digital behaviour and reduce temporal measurement error. Test-retest correlations between time points demonstrated moderate to high temporal stability for social media use frequency (SMU; r =.51) and social media followers (SMF; r =.66), empirically supporting the use of averaged scores. For problematic social media use (PSMU), the correlation was lower (r =.26), reflecting greater temporal variability. However, averaged scores were maintained for methodological consistency, as literature suggests that while social media use problems may fluctuate temporally, they reflect stable underlying individual patterns69. All composite scores were calculated only for cases with complete data on component variables.

Analytical approach

Prior to hypothesis testing, Spearman rank-order correlations were computed to examine bivariate associations among all study variables, providing initial insights into the strength and direction of relationships between social media use patterns and depressive symptoms.

Model selection was informed by diagnostic evaluation of ordinary least squares (OLS) regression assumptions. Tests revealed violations of key parametric requirements: non-normal residual distribution (Shapiro-Wilk: p <.001), heteroscedasticity (Breusch-Pagan: p <.001), and nonlinear residual patterns. These violations would compromise the validity of OLS estimates.

Quantile regression was therefore employed as the primary analytical method. Unlike conventional linear regression, this technique provides robust estimates without requiring normality or homoscedasticity, even when these parametric assumptions are violated70,71. Additionally, quantile regression enables examination of differential predictor effects across the entire conditional distribution of the dependent variable70. In the context of depressive symptoms, this provides richer information about how social media-related risk factors may have heterogeneous effects depending on symptom severity, allowing identification of specific vulnerability patterns across different population segments71. Quantile regression models were estimated at three quantiles (τ = 0.25, 0.50, 0.75), representing low, moderate, and high levels of depressive symptoms at T2, respectively. Standard errors were computed using the Normal Inverse Density (NID) method to ensure stable coefficient estimates and address potential issues with non-positive definite Hessian matrices encountered in preliminary analyses. Given the multifactorial nature of adolescent depression and heterogeneous social media usage patterns, potential interaction effects explaining differential vulnerabilities were systematically explored. Initially, theoretically motivated interactions between social media variables and key individual characteristics (gender, age, baseline depressive symptoms) were evaluated. However, these interactions showed insufficient statistical evidence across quantiles. Subsequently, interaction effects were systematically explored based on vulnerability-stress theoretical frameworks, suggesting that social media risk factors may have differential effects across individual characteristics37, and empirical evidence indicating that age and gender moderate social media effects on adolescent mental health28,33. Specific interactions examined included SMU × Age (reflecting developmental vulnerability periods), SMF × Gender (based on documented gender differences in social media effects), and SMF × SMU (exploring synergistic effects between different usage dimensions consistent with cognitive-behavioural models of problematic use5. Model selection was guided by pseudo-R² values using the Koenker and Machado approach72, with the final interaction model demonstrating superior fit compared to the baseline main-effects model.

All data coding and analyses were performed using R version 4.4.273. Data is provided within the Supplementary Material 3.