Main

Adolescents around the world have experienced a decline in their mental health over the past decade1. Recent UK data suggests that one in six 7–16-year olds and one in four 17–19-year olds have a probable mental health condition, a clear rise from the one in nine and one in ten recorded in 2017, respectively2. As 48% of those with a mental health condition first experience relevant symptoms before the age of 18 years3, this increased mental health burden will negatively impact society and the economy, as well as adolescent and adult life4. Many have raised concerns that this trend has been caused, at least in part, by increased adolescent social media use, which has revolutionized how adolescents live, learn and interact: 93% of 12–17-year olds now report having a social media profile5.

To address these concerns, academic investigation of social media use and adolescent mental health has increased substantially in recent years6. Research teams have recruited adolescent populations in schools, universities or as part of broader community-based samples to identify cross-sectional and longitudinal links between increased smartphone or social media use and scores on questionnaires of depression7,8, anxiety9, disordered eating10 and other mental health symptoms10,11,12. These studies have primarily found small positive associations. Some researchers have used these to argue that there exists a causal link between social media use and mental health declines (that is, “screen time, perhaps especially social media, may have larger effects on adolescent girls’ mental health than on boys’ and that is indeed what we found, with social media significantly correlated with depressive symptoms […]”13 p. 13). Such arguments, in turn, have been used to call for restrictive policy regulations to limit smartphone and social media use in adolescent age groups14.

However, many researchers have also questioned the strength of the current evidence base and highlighted that existing studies do not support the idea that there is a causal relationship linking social media use to mental health. Indeed, the literature provides many conflicting results15. Researchers have not only debated about a lack of longitudinal or causal evidence16, but have also disagreed about what effect sizes matter17,18,19 and how to deal with the substantial individual differences present20,21, which have been linked to factors such as age22, gender23,24 and ethnicity25.

Across these topics of debate, however, researchers have largely overlooked how the kind of instruments used to measure mental health, as well as the populations being studied, limits their ability to draw meaningful inferences about the relationship of social media with adolescent mental health in the first place. So far, most studies have examined school- or community-based adolescent samples15,16,26, relating scores on mental health questionnaires (for example, the Hospital Anxiety and Depression Scale9,27) to time spent on social media. The rationale for doing so is that questionnaires capturing continuous clinical symptoms are informative when reasoning about social media use in relation to the whole spectrum of mental health, across types of severity and clinical presentation. However, this approach is not a suitable surrogate for studying links between social media and mental health in adolescents with versus without mental health conditions, for two main reasons. First, it reduces the complexity of clinical presentations to the tail end of variation in selected mental health symptoms among mostly healthy individuals. Second, it ignores the potentially important differences between those who endorse symptoms on a questionnaire and those who reach diagnostic criteria in standard clinical classifications. For example, an adolescent can score very highly on a questionnaire measuring depressive symptoms, but not meet the criteria for a diagnosis if the queried symptoms have only been present for a short time or if they are better explained by a different condition or situation.

To address this issue, select studies have moved beyond such an approach, dichotomizing symptom severity by applying cut-off scores to mental health questionnaires to reflect the presence or absence of a mental health diagnosis23. However, dichotomization does not solve many of the problems highlighted above and is known to have low sensitivity when predicting clinical diagnoses. Indeed, those with a mental health condition can score below the threshold on some scales28,29. In other words, although researchers would like to demonstrate the presence or absence of specific links between mental health conditions and social media use, the measures of psychopathology they employ might not be appropriate for these goals.

Importantly, the assumption that patterns of social media use found in non-clinical or community samples will generalize to those with mental health conditions has not yet been systematically tested. To our knowledge, only a few studies—most qualitative—have documented different social media use experiences in clinical adolescent populations, including those fulfilling stringent diagnostic criteria for a clinical condition, attending mental health services or being hospitalized for suicidal ideation and suicide attempts30,31,32,33. Adolescents in these studies reported both positive and negative social media experiences, such as enhanced social connection and trouble downregulating their use. Broadly speaking, these experiences aligned with established risk and protective factors previously linked to mental health in offline spaces and suggest there is no clear-cut positive or negative association between mental health and social media use. The studies also raise the idea that vulnerable youth might experience heightened emotional responses to social media use. However, this has not been directly assessed due to the lack of non-clinical comparison groups. Such comparisons are therefore necessary to identify differences in social media use between adolescents with and without mental health conditions.

Owing to the lack of research among young people with mental health conditions26, the important question of whether social media use varies across different types of conditions also remains unaddressed. For example, an adolescent with an internalizing condition (for example, generalized anxiety disorder or depressive disorder) might use and feel impacted by social media differently than an adolescent with an externalizing condition (for example, attention deficit hyperactivity disorder or conduct disorder). This is because, despite both groups presenting mental health symptoms at clinical levels, their experiences of psychopathology can be qualitatively different. Internalizing conditions involve negative emotionality towards the self, expressed through ruminative thought patterns, worries and social withdrawal34. On the contrary, externalizing conditions involve negative emotionality towards others, expressed through impulsivity, risk taking and disinhibition35. Studies that assess mental health with select questionnaires cannot comprehensively account for and investigate such clinical diversity. This is a substantial shortcoming, given the need for research to understand how social media use relates to the growing number of adolescents experiencing mental health symptoms at clinical levels.

This Registered Report provides critical data and evidence-based insights into how social media and mental health are related across adolescent populations who meet and do not meet diagnostic criteria for a wide range of mental health conditions. Given the cross-sectional nature of the data and planned analyses, the study results do not provide causal evidence. Hence, all reported coefficients indicate associations, with the possibility of bidirectional relationships and third variables affecting social media use, mental health or the relationship between the two. We analysed the nationally representative Mental Health of Children and Young People (MHCYP) study36, a cross-sectional survey carried out by National Health Service (NHS) Digital in 2017 that collected data from over 3,000 adolescents (11–19 years old) in England. In place of completing self-report measures of mental health, the participants in this study underwent multi-informant standardized diagnostic assessments evaluated by professional clinical raters for different mental health conditions (Table 1). We note that, in the stage 1 report, the terms ‘clinical and non-clinical population’ were used, but in stage 2 this was changed to ‘adolescents with and without mental health conditions’ to clarify that participants in this study were not recruited or diagnosed by a clinic, but instead underwent the mental health assessment as part of the MHCYP study.

Table 1 Summary of our categorization of mental health conditions into internalizing and externalizing

Further, to gain a comprehensive understanding of how social media use differs across adolescents with and without a mental health condition, we examined both quantitative and qualitative dimensions of social media use. Measuring only time spent provides a crude and simplistic estimate of social media use, conflating distinct analytical levels and missing a rich range of psychological factors such as appraisal and motivations that might vary as a function of mental health37. Researchers have therefore called for quantitative time-based measures of social media use to be complemented by more qualitative engagement-based measures capturing adolescents’ social media activities and their appraisal of them16,38,39,40,41,42. Such practice is, however, still relatively rare. In this Registered Report, we included both types of measures, namely, time spent on social media and dimensions of social media engagement that could incur mental health risks (that is, online social comparison, monitoring and impact of online feedback and lack of control over time spent online) or benefits (that is, online friendship, as well as opportunities for honest self-disclosure and authentic self-presentation). By complementing quantitative and qualitative dimensions of social media use, this work provides a more solid foundation for mechanistic research aimed at informing future targeted interventions, clinical practice and policy actions benefitting adolescent mental health.

We used existing literature on adolescents’ mental health in relation to both online and offline contexts (Table 2) to guide our hypotheses and analyses of the data along three lines of enquiry. Specifically, we evaluated whether social media use differs in adolescents with versus without a mental health condition (Question 1), with an internalizing or externalizing condition versus without a condition (Question 2) and with an internalizing versus externalizing condition (Question 3).

Table 2 Review of key social media and mental health literature used to formulate the hypotheses of our study

First, we expected adolescents with any mental health condition to report engaging with social media differently than those without a condition. For instance, gathering information about peers is particularly important during adolescence when young people develop a sense of personal and social identity43. However, high levels of upward social comparison (that is, comparisons with those believed to be of higher status than the self) have been associated with poorer mental health44,45,46,47,48. Previous work suggests that most social comparisons made on social media sites are upward rather than downward49, possibly because individuals tend to portray themselves in an ideal manner online50,51. Social media could further amplify these processes as platforms offer continuous and more concrete opportunities for comparing oneself with others, such as browsing profiles without initiating social interaction52. Indeed, engagement in self-denigrating online social comparisons was a common theme raised by adolescent psychiatric inpatients during qualitative interviews31. We therefore expected adolescents with a mental health condition to engage in more online social comparison than those without a condition (hypothesis H1.1b). Similarly (see Table 2 for a detailed overview of the literature in support of each hypothesis), we expected them to spend more time on social media8,15 (H1.1a), report more lack of control over time spent online31,53 (H1.1c), while also monitoring54,55 (H1.1 d) and feeling more impacted by online feedback56,57 (H1.1e).

In contrast, we expected adolescents with any condition to engage less with social media in ways that might be protective for their mental health. For instance, we hypothesized that adolescents with a mental health condition are less happy with the number of friends they have online than adolescents without a condition (H1.2f). This is because social connections protect against long-term adverse physical and emotional outcomes, particularly during adolescence48, and young people with mental health conditions often report difficulties with peers, having few friends and wanting to be alone58,59,60. Similarly, we expected adolescents with a mental health condition to engage in less honest online self-disclosure61,62 (H1.2g) and authentic online self-presentation63,64 (H1.2h) compared with adolescents without a condition (Table 2).

Second, we predicted social media use to vary between adolescents with different symptomatology, focusing specifically on how the social media use of adolescents with an internalizing or externalizing condition (Table 1) differed from those without a condition.

We know from the mental health literature that internalizing conditions involve negative self-views, rumination, worries and social withdrawal34,65. Notably, these symptoms could be relevant to how young people present themselves and engage with others on social media31,32,66. For instance, adolescents with internalizing conditions are more likely to notice discrepancies between their ideal and actual selves67 and may compensate for these discrepancies via impression management offline. We expect this process to also occur online, given the multiple affordances social media platforms offer to curate one’s image (for example, deleting old posts and editing new posts). Hence, we hypothesized that adolescents with internalizing conditions would be less likely to engage in authentic self-presentation on social media than adolescents without a mental health condition68,69,70 (H2.2h). Further (Table 2), we hypothesized that adolescents with internalizing conditions would spend more time on social media8,15,71 (H2.1a), engage in more online social comparison30 (H2.1b) and online feedback monitoring54 (H2.1d), feel more impacted by online feedback56 (H2.1e), be less happy about online friends (H2.2f) and engage in less honest self-disclosure62 (H2.2g) than those without a condition. In contrast, we did not expect a difference in whether they perceive a lack of control over the time they spend online (H2.0c)53.

Compared with internalizing conditions, externalizing conditions involve impulsivity, low self-monitoring and risk taking34,65. These symptoms could be reflected in how social media is used by these groups. Hence, we hypothesized that adolescents with externalizing conditions would be more likely than adolescents without a condition to spend more time on social media72 (H2.3a) and perceive they lack control over the time they spend online53 (H2.3c). Further, we expected them to also be more dissatisfied with their number of online friends73 (H2.4f). In contrast, we did not expect differences in the other dimensions of social media use (H2.0b,d,e,g,h; Table 2), since they primarily relate to symptoms of internalizing rather than externalizing conditions.

Our third question examined whether adolescents with internalizing conditions use social media differently than those with externalizing conditions. Specifically, we expected adolescents with internalizing conditions to report engaging in more online social comparison30 (H3.1b) and online feedback monitoring55 (H3.1d), feeling more impacted by online feedback56 (H3.1e), engaging in less honest self-disclosure61 (H3.2g) and less authentic self-presentation on social media69 (H3.2h) than those with externalizing conditions. In contrast, we expected adolescents with externalizing conditions to report having less control over the time they spend online53,73 compared to adolescents with internalizing conditions (H3.2c). Further, since adolescents with internalizing conditions tend to be unhappy with their social status74 and adolescents with externalizing conditions tend to have trouble making and keeping friends in both online and offline contexts73,75, we expect both groups to be similarly dissatisfied with their number of online friends (H3.0f). We also hypothesized that both groups would not differ in the amount of time they spend on social media (H3.0a), given that both adolescents with internalizing and externalizing symptoms have been reported to spend more time on social media than adolescents without these symptoms33,76.

Altogether, this study comprehensively maps and compares different dimensions of social media use in adolescents with and without mental health conditions. Hence, it will lay the foundation for future mechanistic and translational research studying which specific social media dimensions relate to mental health in different adolescent groups. This will be a crucial first step to inform translational research and clinical practice, as well as the design of targeted interventions and policies to improve children’s and adolescents’ mental health.

Results

Sample description

Our final sample included 3,340 young people (Table 3) aged 11–19 (mean 14.77, s.d. 2.48) years. The sample was 50% male and 50% female, 16% of participants had at least one mental health condition (N = 519), 8% had an internalizing condition (N = 282) and 3% had an externalizing condition (N = 104). Descriptive statistics split by social media items, sex and diagnostic groups are reported in Supplementary Tables 58 and Supplementary Figs. 35). Further, an overview of our hypotheses and results is provided in Supplementary Table 9.

Table 3 Descriptive information by mental health condition

Any mental health condition versus no condition

We first compared adolescents with versus without mental health conditions, irrespective of condition type (Fig. 1, H1). In line with our hypothesis, we found that adolescents with any mental health condition reported spending more time on social media than adolescents without a condition (H1.1a; g = 0.46 (90% confidence interval (CI) 0.38 to 0.54); NHST (null hypothesis significance testing): 𝛽 = 0.87, s.e.m. of 0.08, t = 10.48, P = 2.745 × 10−25; EQV (equivalence testing): t(633.8) = 1.21, P = 0.916). Specifically, differences were positive, statistically significant and non-equivalent. That is, we could not reject differences as large or larger than our preregistered equivalence bounds (smallest effect size of interest (SESOI), g = 0.4), indicating potentially meaningful differences between adolescents with and without a condition. On the contrary, the results did not support our hypothesis for online social comparison (H1.1b; g = 0.30 (90% CI 0.22 to 0.39); NHST: 𝛽 = 0.42, s.e.m. of 0.07, t = 6.52, P = 7.973 × 10−11; EQV: t(641.08) = −1.89, P = 0.026), lack of control over time spent online (H1.1c; g = 0.27 (90% CI 0.19 to 0.35); NHST: 𝛽 = 0.39, s.e.m. of 0.07, t = 5.52, P = 3.614 × 10−8; EQV: t (674.42) = −2.68, P = 0.002) and impact of online feedback on mood (H1.1e; g = 0.29 (90% CI 0.21 to 0.38); NHST: 𝛽 = 0.38, s.e.m. of 0.06, t = 6.36, P = 2.269 × 10−10; EQV: t(608.98) = −2.06, P = 0.016). For these dimensions of social media engagement, effect sizes were positive, statistically significant but equivalent (that is, they fell within the preregistered SESOIs). This indicates a difference that is too small to be theoretically meaningful between adolescents with and without a mental health condition. Last, for monitoring of online feedback, we found differences that were not statistically significant and also equivalent, indicating no statistically nor theoretically meaningful differences between adolescents with versus without a condition (H1.1d; g = 0.08 (90% CI −0.01 to 0.15) NHST: 𝛽 = 0.11, s.e.m. of 0.07, t = 1.55, P = 0.121; EQV: t (653.24) = −6.47, P = 0.000).

Fig. 1: Differences in social media use for the three group comparisons.
figure 1

Top: hypothesis 1 (H1, any mental health condition versus no condition). Middle: hypothesis 2 (H2, internalizing/externalizing versus no condition). Bottom: hypothesis 3 (H3, internalizing versus externalizing condition). Data are presented as mean differences based on Hedges’s g effect size (g) and its corresponding 90% CI. The shaded area indicates the SESOI (g = 0.4, corresponding to d = 0.4). If the 90% CI lies completely within the SESOI, we concluded equivalence, and therefore no meaningful differences. Bolded effect sizes reflect comparisons that supported our hypotheses, while faded effect sizes reflect comparisons that did not support our hypotheses. The starting sample size includes social media users (N = 519 for adolescents with any mental health condition, N = 282 for internalizing conditions, N = 104 for externalizing conditions and N = 2,821 for no mental health condition). However, the exact sample sizes for each comparison are reported in Supplementary Tables 5 and 6 and differ for each dimension of social media use, given that we planned to analyse those separately.

Source data.

We further hypothesized that adolescents with any mental health condition would score lower than adolescents without a condition on dimensions of social media use that could incur mental health benefits, namely, happiness about the number of online friendships (H1.2f), honest self-disclosure (H1.2g) and authentic self-presentation (H1.2h). In line with our hypothesis, we found lower happiness about the number of online friendships (H1.2f; g = −0.37 (90% CI −0.45 to −0.29); NHST: 𝛽 = −0.33, s.e.m. of 0.04, t = −8.23, P = 2.660 × 10−16; EQV: t(590.4) = 0.56, P = 0.277), for which the effect size was negative, statistically significant and non-equivalent (that is, large enough to be potentially meaningful). In contrast, we did not find differences in honest self-disclosure (H1.2g; g = −0.30 (90% CI −0.38 to −0.22); NHST: 𝛽 = −0.39, s.e.m. of 0.06, t = −6.37, P = 2.213 × 10−10; EQV: t(629.39) = 1.931, P = 0.028) and authentic self-presentation (H1.2h; g = −0.19 (90% CI −0.28 to −0.11); NHST: 𝛽 = −0.24, s.e.m. of 0.06, t = −3.98, P = 7.085 × 10−5; EQV: t(624.53) = 4.04, P = 0.000). In these cases, effect sizes were negative and statistically significant but equivalent, suggesting that differences between those with and without a mental health condition were too small to be theoretically meaningful.

Internalizing/externalizing conditions versus no condition

Our second question concerned the extent to which adolescents with internalizing or externalizing conditions differed in their social media use from adolescents without a condition (Fig. 1, H2).

Internalizing versus no condition

Our hypotheses were grounded in the mental health literature, which suggests that anxiety and depressive disorders are characterized by negative self-views, rumination, worries and social withdrawal. We expected these symptoms to be mirrored in adolescents’ online experiences. The results supported our hypotheses for time spent on social media (H2.1a; g = 0.62 (90% CI 0.51 to 0.73); NHST: 𝛽 = 1.12, s.e.m. of 0.11, t = 10.32, P = 1.609 × 10−25; EQV: t(317.39) = 3.21, P = 0.999), online social comparison (H2.1b; g = 0.54 (90% CI 0.43 to 0.65); NHST: 𝛽 = 0.76, s.e.m. of 0.08, t = 9.12, P = 1.304 × 10−19; EQV: t(318.57) = 2.11, P = 0.994) and the impact of online feedback (H2.1e; g = 0.38 (90% CI 0.27 to 0.49); NHST: 𝛽 = 0.51, s.e.m. of 0.08, t = 6.61, P = 4.494 × 10−11; EQV: t(306.67) = −0.27, P = 0.385), where we found positive, statistically significant, and non-equivalent effect sizes, suggesting potentially meaningful differences between adolescents with internalizing versus no condition. In contrast, the results did not support our hypothesis for monitoring of online feedback, where differences were not statistically significant and were also too small to be considered meaningful (H2.1d; g = 0.13 (90% CI 0.03 to 0.25); NHST: 𝛽 = 0.20, s.e.m. of 0.09, t = 2.13, P = 0.033; EQV: t(324.47) = −4.13, P = 0.000).

For those with internalizing conditions, we also hypothesized decreased levels of happiness about the number of online friendships (H2.2f), honest self-disclosure (H2.2g) and authentic self-presentation (H2.2h). Our hypotheses were confirmed for happiness about the number of online friendships (H2.2f; g = −0.45 (90% CI −0.55 to −0.35); NHST: 𝛽 = −0.40, s.e.m. of 0.05, t = −7.91, P = 3.49 × 10−15; EQV: t(304.45) = −0.69, P = 0.776) and honest self-disclosure (H2.2g; g = −0.31 (90% CI −0.42 to −0.20); NHST: 𝛽 = −0.41, s.e.m. of 0.08, t = −5.16, P = 2.670 × 10−7; EQV: t(314.2) = 1.32, P = 0.088), where we found negative, statistically significant and potentially meaningful differences. In other words, those with internalizing conditions scored lower than adolescents with no condition. In contrast, we did not find support for meaningful differences in authentic self-presentation (H2.2h; g = −0.19 (90% CI −0.30 to −0.08); NHST: 𝛽 = −0.25, s.e.m. of 0.08, t = −3.16, P = 0.002; EQV: t(310.64) = 3.08, P = 0.500 × 10−4), where the effect size was statistically significant but equivalent, and therefore too small to be considered meaningful. Last, we expected no differences in lack of control over time spent online for adolescents with internalizing versus no condition (H2.0c). The results did not support our hypothesis, showing positive, statistically significant and potentially meaningful differences (g = 0.43 (90% CI 0.33 to 0.55); NHST: 𝛽 = 0.60, s.e.m. of 0.09, t = 6.74, P = 1.91 × 10−11; EQV: t(336.39) = 0.489, P = 0.689).

Externalizing versus no condition

Externalizing conditions are characterized by impulsivity, low self-monitoring, and risk taking. We expected such symptoms to be reflected in how social media is used by this group. The results supported our hypotheses for time spent on social media (H2.3a; g = 0.31 (90% CI 0.13 to 0.48); NHST: 𝛽 = 0.58, s.e.m. of 0.17, t = 3.40, P = 0.001; EQV: t(105.73) = −0.83, P = 0.232), where we found positive differences that were statistically significant and large enough to be theoretically meaningful. The results did not support our hypotheses for the lack of control over time spent online (H2.3c; g = 0.19 (90% CI 0.01 to 0.37); NHST: 𝛽 = 0.27, s.e.m. of 0.15, t = 1.83, P = 0.068; EQV: t(100.39) = −2.01, P = 0.017) and happiness about the number of online friendships (H2.4f; g = −0.16 (90% CI −0.32 to 0.02); NHST: 𝛽 = −0.12, s.e.m. of 0.09, t = −1.39, P = 0.165; EQV: t(96.65) = 2.34, P = 0.023), where differences were not statistically significant and were too small to be meaningful.

Last, we expected no differences between adolescents with externalizing and no conditions in online social comparison (H2.0b), monitoring of online feedback (H2.0d), feeling impacted by online feedback (H2.0e), honest online self-disclosure (H2.0g) and authentic self-presentation (H2.0h). The results supported our hypotheses for online social comparison (H2.0b; g = −0.10 (90% CI −0.28 to 0.07); NHST: 𝛽 = −0.13, s.e.m. of 0.14, t = −0.97, P = 0.331; EQV: t(104.81) = 2.94, P = 0.000), monitoring of feedback (H2.0d; g = 0.09 (90% CI −0.08 to 0.27), NHST: 𝛽 = 0.13, s.e.m. of 0.16, t = 0.82, P = 0.410; EQV: t(97.07) = −2.92, P = 0.004), honest self-disclosure (H2.0g; g = −0.21 (90% CI −0.38 to −0.03), NHST: 𝛽 = −0.26, s.e.m. of 0.13, t = −2.01, P = 0.045; EQV: t(98.82) = 1.79, P = 0.040), and authentic self-presentation (H2.0h; g = −0.09 (90% CI −0.25 to 0.08), NHST: 𝛽 = −0.11, s.e.m. of 0.13, t = −0.85, P = 0.395; EQV: t(98.29) = 2.96, P = 0.003), as these effect sizes were not statistically significant and were also too small to be considered meaningful. In contrast, the results did not support our hypothesis for the impact of feedback on mood (H2.0e; g = 0.27 (90% CI 0.10 to 0.45), NHST: 𝛽 = 0.34, s.e.m. of 0.13, t = 2.71, P = 0.007; EQV: t(96.78) = −1.14, P = 0.120), where we found positive, significant and potentially meaningful differences.

Internalizing versus externalizing conditions

Our third question focused only on adolescents with a mental health condition and specifically examined whether those with an internalizing condition use social media differently than those with an externalizing condition (Fig. 1, H3).

The results supported our hypotheses for online social comparison (H3.1b; g = 0.64 (90% CI 0.45 to 0.85); NHST: 𝛽 = 0.89, s.e.m. of 0.16, t = 5.75, P = 0.000, EQV: t(203.72) = 2.187, P = 0.993), where we found positive differences that were statistically significant and large enough to be theoretically meaningful in adolescents with internalizing compared with externalizing conditions. In contrast, the results did not support our hypotheses for the monitoring of online feedback (H3.1d; g = 0.05 (90% CI −0.15 to 0.24); NHST: 𝛽 = 0.07, s.e.m. of 0.18, t = 0.40, P = 0.689, EQV: t(157.17) = −2.924, P = 0.002) and impact of online feedback on mood (H3.1e; g = 0.12 (90% CI −0.07 to 0.32); NHST: 𝛽 = 0.16, s.e.m. of 0.14, t = 1.12, P = 0.261, EQV: t(171.21) = −2.421, P = 0.008), where differences were not statistically significant and were too small to be meaningful.

We further hypothesized that adolescents with internalizing conditions would score lower than adolescents with externalizing conditions in lack of control over time spent online (H3.2c), honest self-disclosure (H3.2g) and authentic self-presentation (H3.2h). However, we found neither significant nor meaningful differences across these dimensions (H3.2g; g = −0.11 (90% CI −0.30 to 0.09); NHST: 𝛽 = −0.15, s.e.m. of 0.15, t = −0.99, P = 0.323, EQV: t(170.73) = 2.47, P = 0.004; H3.2h; g = −0.11 (90% CI −0.29 to 0.08); NHST: 𝛽 = −0.14, s.e.m. of 0.15, t = −0.93, P = 0.351, EQV: t(185.78) = 2.51, P = 0.005). Further, the results were inconclusive for lack of control over time spent online (H3.2c; g = 0.24 (90% CI 0.04 to 0.43); NHST: 𝛽 = 0.33, s.e.m. of 0.17, t = 1.98, P = 0.048, EQV: t(156.91) = −1.37, P = 0.080), where we did not find statistically significant differences nor we could reject meaningfully large effect sizes.

Last, we hypothesized that adolescents with internalizing conditions would not differ from adolescents with externalizing conditions in time spent on social media (H3.0a) and happiness in the number of online friendships (H3.0f). The results did not confirm our hypotheses. For time spent on social media, we found positive differences (internalizing higher than externalizing) that were statistically significant and potentially meaningful (H3.0a; g = 0.27 (90% CI 0.07 to 0.47); NHST: 𝛽 = 0.54, s.e.m. of 0.191, t = 2.76, P = 0.006, EQV: t(171.33) = −1.15, P = 0.1125). Further, we found that adolescents with internalizing conditions reported lower happiness about the number of online friends than adolescents with externalizing conditions. In this case, differences were negative, statistically significant and potentially meaningful (H3.0f; g = −0.32 (90% CI −0.51 to −0.14); NHST: 𝛽 = −0.29, s.e.m. of 0.10, t = −2.93, P = 0.003, EQV: t(206.5) = 0.707, P = 0.223).

Exploratory and sensitivity analyses

To extend our findings, we conducted four sets of sensitivity analyses. First, we included adolescents with between-group comorbidities in question 2, such as any internalizing condition with a comorbid externalizing condition, or vice versa, and compared them with those without a condition (Supplementary Tables 15 and 16). Second, we examined the association between mental health severity, conceptualized as the number of conditions (irrespective of diagnostic type), and social media use (Supplementary Table 17). Third, we focused on adolescents with specific conditions and compared their social media use with that of adolescents without a condition. In this case, we only tested conditions for which we were sufficiently powered, namely, major depressive disorder (N ≈ 86; Supplementary Table 18), generalized anxiety disorder (N ≈ 75; Supplementary Table 19) and social anxiety disorder (N ≈ 70; Supplementary Table 20). Last, we tested our hypotheses for time spent on social media separately for school days and weekends, rather than using a composite score (Supplementary Table 21).

Overall, the results were largely in line with our primary findings with a few exceptions. Namely, we found that adolescents with externalizing and between-group comorbidity (Supplementary Table 16) reported less honest self-disclosure than those without a condition. For the sensitivity analysis on mental health severity, we found that the number of conditions, irrespective of type, was associated with time spent on social media, social comparison, monitoring of feedback and impact of feedback on mood (Supplementary Table 17). Last, we did not find differences for time spent on weekdays versus weekends/holidays (Supplementary Table 21). We note that these sensitivity analyses were exploratory and conducted on relatively small sample sizes, which limits the robustness of these findings.

Discussion

In this Registered Report, we analysed differences in social media use between adolescents with and without mental health conditions in a UK sample of over 3,000 participants. Overall, we found significant and meaningful differences across both quantitative (time spent) and qualitative (for example, online social comparison and happiness about the number of online friends) dimensions of social media use.

Interestingly, the largest difference in social media use between those with and without mental health conditions was in the time spent on social media, with the former reporting higher usage. It is important to note that this measure was self-reported, which is known to have only a moderate correlation with objective measures such as sensing data77. This raises the question of whether those with mental health conditions perceive that they spend more time on social media or whether they actually do so. Further, we observed that adolescents with a mental health condition reported lower satisfaction with the number of their online friends. In offline contexts, social connections serve as a protective factor against long-term adverse physical and emotional outcomes, especially during adolescence. Our findings therefore suggest that the difficulties with peer relationships experienced by youth clinical groups offline may also be reflected in their online interactions. For the other dimensions of social media use examined, our results followed the hypothesized direction and were statistically significant (with the exception of monitoring of online feedback). However, the differences were not large enough to be considered meaningful. This might be explained by the relatively high threshold set as our SESOI, which was a moderate effect size grounded in literature on sleep and physical exercise, both established markers of psychopathology.

We next compared social media use between adolescents with a specific mental health condition (internalizing or externalizing) versus no condition. In this case, the results largely supported our hypotheses, whereby adolescents with internalizing conditions demonstrated higher time spent on social media, increased social comparison, greater impact of social media feedback on mood, lower satisfaction with the number of online friends and lower honest self-disclosure compared to those without a mental health condition. Unexpectedly, we also found that adolescents with internalizing conditions reported a higher lack of control over their time spent online, an engagement dimension that we had instead hypothesized would be more pronounced in those with externalizing conditions. For adolescents with externalizing conditions, the only meaningful difference when compared with those with no condition was increased time spent online, with no notable differences across other dimensions of social media use. These results might be explained by the fact that the dimensions of social media engagement used in this study were largely framed around internal experiences (that is, they enquired about one’s emotions and thoughts) that could be more effective indicators of internalizing rather than externalizing conditions.

Finally, by limiting our focus to only adolescents with mental health conditions, we compared those with internalizing to those with externalizing conditions. We found that adolescents with internalizing conditions engaged in higher online social comparison. This finding aligns with existing research indicating that adolescents with depressive and anxious symptomatology tend to unfavourably compare themselves with others on social media. Given that social media platforms provide continuous and concrete opportunities for social comparison—such as browsing others’ profiles without initiating social interaction52—social comparison may represent a critical mechanism associated with internalizing symptoms on social media. Further, adolescents with internalizing conditions were less happy about the number of their online friends. This may be because their tendency towards negative self-evaluation and social comparison leads them to make negative evaluations of their social status78. In contrast, adolescents with externalizing conditions might be less focused on social comparison and more on immediate social interactions, resulting in greater satisfaction with their online friendships.

Altogether, this study has three key strengths. First, the complex survey design, which utilized random probability sampling, produced a nationally representative UK sample. Second, all participants in the sample underwent a standardized multi-informant assessment by professional clinical raters. This method provided comprehensive information about participants’ mental health without only relying on self-reported questionnaires, self-diagnoses or mental health service use, measures that only capture a subset of individuals with mental health conditions79,80. Last, we examined both qualitative and quantitative aspects of social media use, offering insights beyond time spent into relevant engagement dimensions such as online social comparison and the impact of feedback on mood.

We also acknowledge some limitations of our study. First, regarding study design, we analysed raw associations from cross-sectional data. Therefore, no causal or directional inference can be drawn from these findings, including whether the onset of mental health conditions affects the examined dimensions of social media use or vice versa. Further, given the relatively small sample size in our externalizing group (N = 104) and the relatively large number of individuals with between-group comorbidities (N = 57), our findings concerning externalizing conditions should be interpreted with caution. Specifically, our question 2 and 3 results exclude adolescents with a relatively common clinical profile of comorbid externalizing and anxiety disorders. Additional work on how young people with this clinical profile use and experience social media is needed.

Further, although the sample was nationally representative, we cannot determine the extent to which our findings apply outside the UK. Given the socio-cultural factors influencing social media use and mental health conditions, replicating these results in other regions across both the Global North and Global South is crucial if generalizations are to be made25. We also acknowledge that the data were collected in 2017. While the examined dimensions of social media use remain relevant, the rapid evolution of platforms and user behaviours presents a potential limitation when applying our findings to current trends. Last, we acknowledge limitations regarding our measures. Specifically, we relied on self-reported social media data. Self-reports capture participants’ perceptions, such as online social comparison, that cannot be objectively or reliably measured otherwise. However, they often fail to accurately reflect actual patterns in usage, particularly when estimating time spent on social media77.

Future research should aim to replicate and expand on these findings in three key areas. First, studies using experimental and longitudinal designs are essential to clarify the temporal and causal dynamics linking various social media patterns to mental health conditions. Doing so would allow us to disentangle the within-person variation and directional relationship between social media use and mental health symptom presentation, onset or recovery. Second, regarding social media use, future studies could investigate other self-reported and clinically relevant aspects of engagement, such as time displacement from offline activities81, the similarity between perceived self on social media and offline82, as well as goal-directed social media use. Additionally, studies could assess differences in objective social media use, such as time spent on various apps, posting behaviours, active messaging and content exposure. Third, research involving adolescents with intellectual and learning disabilities is necessary to identify differences in this specific clinical group, which was not included in this study.

The results have implications for clinical practice. Specifically, we find key aspects of social media engagement that could inform the creation of guidelines for patient consultations and early intervention strategies. For example, this could include psychoeducation and cognitive-behavioural reappraisal techniques specifically aimed at online social comparison or the impact of social media feedback (for example, ‘likes’) on mood for adolescents with internalizing conditions.

Over the past years, there has been increasing concern that social media is negatively impacting young people’s mental health, but very little research has compared social media use in those with and without mental health conditions26. In one of the first studies of its kind, we find that young people with mental health conditions report engaging with social media in different ways from those without a condition. This highlights aspects of social media use that might present an increased risk to this already vulnerable group and provides a window for future research to ensure that the digital world is safe for all children regardless of mental health status.

Methods

Ethics information

The MHCYP 2017 survey was reviewed and approved by the West London and GTAC Research Ethics Committee (reference: 16/LO/0155) and the Health Research Authority Confidentiality Advisory Group (reference: 16/CAG/0016) in 2016. Both parents and children provided consent to take part in data collection and were compensated with a £10 voucher for their time. Parents of children under 16 years were interviewed first and permission was sought to interview their child afterwards; the child then provided assent. Conversely, 17–19-year olds were directly asked for their consent, with permission subsequently sought for their parents to be interviewed. Access to the data was granted to the research team by NHS Digital (DARS-NIC-424336-T7K7T-v0.6 r).

Design

The MHCYP study is one of a series of national surveys on the mental health of children and young people in England administered in 1999, 2004, 2017, 2021 and 2022. In this Registered Report, we analysed the 2017 wave collected between January and October 2017: the most recent wave to be made available to researchers as well as the first wave to collect comprehensive data on adolescents’ social media use and to include 17–19-year olds. We only analysed data from adolescents who reported being social media users aged 11–19 years, a total of 3,340 participants (50% male and 50% female) out of the full sample of 9,117. The survey was collected using a stratified probability sample of children and young people living in England who were registered with a general practitioner36. Data were collected via face-to-face interviews with adolescents and their parents. At the same time, if the family agreed, questionnaires were mailed to teachers (for the available data and key demographics see MHCYP 201736,83).

Measures

Time spent on social media

The study measured time spent on social media using two questionnaire items: “When you use social media sites or apps, how much time do you spend using them on a typical school day?” (SMTimeSpentS) and “When you use social media sites or apps how much time in total do you spend using them on a typical weekend or holiday day?” (SMTimeSpentW). Participants answered both questions on a nine-point Likert scale: 1, less than 30 min; 2, more than 30 min but less than an hour; 3, 1–2 h; 4, 2−3 h; 5, 3–4 h; 6, 4–5 h; 7, 5–6 h; 8, 6–7 h; 9, more than 7 h. A single variable reflecting average social media hours was created from these two variables (SMTimeSpent; see Supplementary Table 1 for more details). To do so, we first calculated the mean time in hours for each response. For example, if a participant responded “one to two hours”, we recoded this as 1.5 h. Participants that responded “more than seven hours” were recoded to 7.5 h, while participants that responded “less than 30 min” were recorded to 15 min (that is, 0.25 h). Weekday hours were then multiplied by 5, and weekend hours were multiplied by 2, and the products were summed and divided by 7 to establish a daily mean social media use variable, measured in hours. The SMTimeSpent variable was coded as continuous84.

In the survey, the questions regarding time spent on social media for both weekdays (SMTimeSpentS) and weekends (SMTimeSpentW) were only asked of adolescents who responded that they use social media sites daily or on most days on a previous questionnaire item (SMFreqofUse; 1, daily or most days). Hence, participants that reported a lower frequency of social media use (that is, SMFreqofUse; 2, a few times a week; 3, once a week; 4, a few times a month; 5, once a month; 6, less often than once a month) were not asked these questions. To handle the resulting missing data in the SMTimeSpent variable, we coded any adolescents who responded to the SMFreqofUse question that they use social media between “a few times a week” and “once a week” to 45 min (0.75 h) and adolescents that responded “a few times a month” and “less often than once a month” to 15 min (0.25 h) on the SMTimeSpent question (for more information about our approach to missing data, see Supplementary Table 2).

Social media engagement

We analysed seven qualitative dimensions of social media engagement measured with questionnaire items tapping into experiences associated with both risks and benefits to adolescent mental health. All measures (summarized with related literature in Table 2) were developed in consultation with a young person advisory group (see the Supplementary Methods for more information), where children and adolescents defined the dimensions of social media use most relevant to them. The measures related to mental health risks encompassed online social comparison (“I compare myself to others on social media”), lack of control over time spent online (“I spend more time on social media than I mean to”), monitoring of online feedback (“I monitor the amount of likes, comments and shares I get on social media”) and the impact of online feedback (“The amount of likes, comments and shares I get on social media has an impact on my mood”). On the contrary, the measures indicative of mental health benefits included online friendship (“I am happy with the number of friends I have on social media”), honest self-disclosure (“I can be honest with people on social media sites and apps about how I am feeling) and authentic self-presentation (“My social media profile is a true reflection of myself”).

Participants responded to these measures using a five-point Likert scale (1, disagree a lot; 2, disagree a little; 3, neither agree nor disagree; 4, agree a little; 5, agree a lot). We omitted the “Don’t know” responses and coded 1–5 responses as continuous, given research suggesting that five-point continuous classifications perform as well as or occasionally better than categorical classifications84. We performed a sensitivity analysis to test whether examining social media use on weekdays versus weekends/holidays separately, rather than as a weighted average, changed the main results.

Mental health conditions

Face-to-face interviewers completed the Development and Wellbeing Assessment (DAWBA)85 with parents and adolescents aged 11 years or over to establish mental health conditions. During the interview, participants were first led through the 25-item Strengths and Difficulties Questionnaire86 (Supplementary Table 3). Second, the interviewer administered the DAWBA, a diagnostic tool shown to have good validity87 and reliability88. The DAWBA uses structured and semi-structured questions to assess the presence and severity of symptoms for a wide range of DSM-5 or ICD-10 mental health disorders (Table 1). Each module starts with a few screening items, which, if answered negatively (indicative of the lack of symptoms), allow the interview to proceed to the next module with no loss of accuracy87 (for example, Supplementary Table 4). There is one exception: if a participant scores highly on the Strengths and Difficulties Questionnaire, the interviewer is directed to ask in-depth internalizing disorder DAWBA modules, even if participants screened negative on the initial questions.

After the initial screening items, the subsequent structured items in the DAWBA modules relate directly to diagnostic criteria in DSM-5 and ICD-10. They are close-ended questions about specific mental health symptoms (for example, “In the last 4 weeks, have there been times when you have been very sad, miserable, unhappy or tearful?”). If a participant responds positively to these structured items, they are subsequently asked open-ended questions about these problems (for example, “Please describe your mood—sadness or irritability—and your level of interest in things”). During the assessment, interviewers transcribe open-ended responses verbatim and are also able to add personal comments beneath each response.

A team of clinical raters assessed the DAWBA’s structured and qualitative information from all informants to decide whether an adolescent showed evidence of a DSM-5 or ICD-10 mental health condition that would warrant clinical treatment (Supplementary Figs. 1 and 2). To determine the presence of a condition, clinical raters (1) checked that the answers to structured comments were understood by the participants accurately; (2) interpreted any conflicts between child, parent and teacher responses and decided which assessment to prioritize; and (3) identified clinically impairing disorders that did not perfectly fit current operationalized diagnostic criteria or “not otherwise specified diagnosis” such as “other anxiety disorder”.

Overall, we coded information on mental health conditions into two separate variables: (1) a binary variable indicating the presence of any mental health condition (diagnosis or no diagnosis) and (2) a categorical variable for the type of condition (internalizing diagnosis, externalizing diagnosis or no diagnosis). We subdivided conditions identified via the diagnostic assessment into internalizing and externalizing using existing classifications of psychiatric diagnoses34,65 (Table 1). This distinction draws from transdiagnostic research showing that different conditions (for example, anxiety, depression and eating disorders) are often comorbid, share underlying core symptoms89 and can therefore be grouped to reflect clinical presentations with more validity34. Further, when we had enough power, we ran additional exploratory analyses examining responses to all social media questions (a–g) separately for individual conditions (specifically, major depressive disorder, generalized anxiety disorder and social anxiety disorder).

Comorbidity data were coded into two separate variables: (1) a binary variable indicating within-group comorbidity (any two diagnoses of either internalizing or externalizing: yes or no, for descriptive purposes only) and (2) a binary variable indicating internalizing–externalizing between-group comorbidity (any comorbid internalizing and externalizing diagnoses: yes or no). Individuals who showed between-group comorbidity were removed before the analysis of questions 2 and 3, given our goal to compare social media use between these groups. To increase the clinical utility of this work, we ran a sensitivity analysis for question 2, including people with between-group comorbidity.

Analysis plan

We conducted all statistical analyses in R version 4.3.1 (R Core Team, 2021), testing the association between time- and engagement-based measures of social media use and mental health conditions using equivalence tests and linear regression models90,91. To control for the type I error rate across multiple tests, we set a corrected alpha level of 0.0125, accounting for the false discovery rate across our four tested hypotheses for any given social media item92. Our analytical approach was based on regressions rather than analysis of variance, as the former allow for more diverse predictors, unbalanced groups and inclusion of covariates in potential exploratory analyses93. Below, we describe the statistical analyses we used to test our questions and hypotheses (see the Supplementary Information for more details). Further, the analysis code is available on OSF94.

Questions and hypotheses

Question 1: investigate whether adolescents with any mental health condition use social media differently than those without a condition

To address question 1, we tested the association between social media use and mental health conditions. Specifically, we estimated linear regression models with mental health condition as a binary predictor (two levels: diagnosis versus no diagnosis) and social media use as a continuous outcome. The no diagnosis group was set as the reference level for these analyses. Below, we report our null and directional hypotheses, first for dimensions of social media use expected to reflect mental health risks and second for dimensions expected to reflect mental health benefits in adolescents with versus without a condition.

  • H0(1.0): adolescents with any mental health condition will not differ from adolescents without a condition in (a) time spent on social media, (b) online social comparison, (c) lacking control over time spent online, (d) monitoring of online feedback, (e) feeling impacted by online feedback, (f) happiness about the number of online friendships, (g) honest online self-disclosure and (h) authentic self-presentation online.

  • H1(1.1): adolescents with any mental health condition will score higher than adolescents without a condition in (a) time on social media, (b) online social comparison, (c) lacking control over time spent online, (d) monitoring of online feedback and (e) feeling impacted by online feedback.

  • H1(1.2): adolescents with any mental health condition will score lower than adolescents without a condition in (f) happiness about the number of online friendships, (g) honest online self-disclosure and (h) authentic self-presentation online.

  • To examine whether social media use varies with mental health severity, we conducted sensitivity analyses to test for a linear effect of the number of diagnoses on the social media responses using linear regression models.

Question 2: investigate whether adolescents with an internalizing or externalizing condition use social media differently than those without a condition

After assessing differences in social media use in adolescents with versus without any mental health condition, we examined whether adolescents with internalizing or externalizing conditions use social media differently than adolescents without a condition. Hence, we conducted linear regression models with diagnostic category as a categorical predictor (three levels: internalizing diagnosis, externalizing diagnosis and no diagnosis) and social media use as a continuous outcome. To test our hypotheses, we examined comparisons between two levels of the diagnostic category variable, with no diagnosis set as the reference level. For hypotheses 2.0i, 2.1 and 2.2, we reported regression coefficients for internalizing versus no diagnosis, while for hypotheses 2.0e, 2.3 and 2.4 we reported coefficients for externalizing versus no diagnosis. The null hypotheses marked by ‘e’ or ‘i’ indicate that, for the considered comparison, the null hypothesis was our primary hypothesis. Hence, for those dimensions of social media use, we expected no difference between adolescents with internalizing or externalizing conditions and those without a condition. We used ‘e’ to indicate primary null hypotheses related to externalizing versus no condition and ‘i’ to indicate our primary null hypotheses related to internalising condition.

  • H0(2.0): adolescents with internalizing or externalizing conditions will not differ from adolescents without a condition in (a) time on social media, (b) online social comparisone, (c) lacking control over time spent onlinei, (d) monitoring of online feedbacke, (e) feeling impacted by online feedbacke, (f) happiness about the number of online friendships, (g) honest online self-disclosuree and (h) authentic self-presentation onlinee.

  • H1(2.1): adolescents with internalizing condition will score higher than adolescents without a condition in (a) time on social media, (b) online social comparison, (d) monitoring of online feedback and (e) feeling impacted by online feedback.

  • H1(2.2): adolescents with internalizing conditions will score lower than adolescents without a condition in (f) happiness about the number of online friendships, (g) honest online self-disclosure and (h) authentic self-presentation online.

  • H1(2.3): adolescents with externalizing conditions will score higher than adolescents without a condition in (a) time spent on social media and (c) lack of control over time spent online.

  • H1(2.4): adolescents with externalizing conditions will score lower than adolescents without a condition in (f) happiness about the number of online friendships.

Question 3: investigate whether adolescents with an internalizing mental health condition use social media differently than those with an externalizing condition

To address question 3, we examined how adolescents with internalizing conditions differed in social media engagement compared to adolescents with externalizing conditions. To this aim, we compared the internalizing and externalizing levels of the diagnostic category variable described in question 2, with internalizing as the reference level. Also, in this case, the null hypotheses marked by ‘c’ indicate our primary hypotheses. Hence, for those dimensions of social media use, we expected no difference between adolescents with internalizing and externalizing conditions.

  • H0 (3.0): adolescents with internalizing conditions will not differ from adolescents with externalizing conditions in (a) time on social mediac, (b) online social comparison, (c) lacking control over time spent online, (d) monitoring of online feedback, (e) feeling impacted by online feedback, (f) happiness about the number of online friendshipsc, (g) online self-disclosure and (h) authentic self-presentation online.

  • H1(3.1): adolescents with internalizing conditions will score higher than adolescents with externalizing conditions in (b) online social comparison, (d) monitoring of online feedback and (e) feeling impacted by online feedback.

  • H1(3.2): adolescents with internalizing conditions will score lower than adolescents with externalizing conditions in (c) lack of control over time spent online, (g) online self-disclosure and (h) authentic self-presentation online.

For all regression analyses, we treated the eight dimensions of social media use, including both time- and engagement-based measures, as separate outcomes predicted by information on mental health conditions as the only regressor. While it is common in research to use statistical control to remove confounding effects from a regression coefficient, appropriate control variables should be identified only after justifying a causal structure that includes the outcome, predictors and all theorized confounders. When the selected control variables are inappropriate or remain unjustified, controlling can result in biased regression estimates95. Further, recent literature warns against controlling for demographic factors such as sex without thought and instead prompts researchers to interrogate how this variable intersects with the predictors and outcomes under investigation96. In the present work, treating sex or age as a confounding variable would mean ignoring the possibility that there are meaningful sex or age differences in the examined relationships. As our goal is to investigate the overall association between social media use and mental health, we provided a descriptive account of the age and sex of adolescents included in each tested model rather than control for these demographics. For example, in question 1, age and sex of adolescents with any condition versus without a condition are reported for descriptive purposes (Table 3).

Equivalence testing

Given that null hypothesis significance testing does not allow a comprehensive interpretation of statistically non-significant results97, each model was complemented by equivalence tests97 using the bootstrapped two one-sided tests (TOST) with the ‘boot_T_TOST’ function from the TOSTER package91 to quantify support for the null hypothesis (H0(1–3)). This involves assessing whether the 90% CIs for the effect size lie inside prespecified equivalence bounds that indicate the SESOI. If the CIs lie inside the equivalence bounds, the effect size is interpreted as negligible. On the contrary, if the CIs spread outside the equivalence bounds, the effect size is considered meaningful in size. In this case, the 90% CIs are used rather than 95% because the effect size is tested against two equivalence bounds separately (that is, the upper and lower bound), reflecting (1 – 2α) × 100% (see the Supplementary Information for more details).

We established the equivalence bounds by identifying a theoretically meaningful SESOI (Cohen’s d = 0.4; refer to the Supplementary Methods for a detailed explanation). After an extensive scoping exercise to identify a suitable theoretical foundation, we determined that the most relevant benchmark for our research questions are everyday behaviours linked to mental health, such as sleep and physical activity. These behaviours, much like social media use, are a regular part of daily routines but, unlike social media use, they are well-established markers of psychopathology, based on both theory98,99,100 and matched empirical evidence101,102. Consequently, if the actual effect size of social media use is comparable to that observed for behaviours such as sleep and physical activity, we can confidently conclude that social media use also represents an everyday behaviour that exhibits meaningful group-level differences between clinical and non-clinical populations. The interpretation and analysis plan are presented separately for each hypothesis in the Supplementary Information.

Overall, we inferred support for the null hypotheses (that is, no meaningful difference in social media use between groups) if the 90% CI for this association lies within the equivalence bounds. Of note, while a theoretical SESOI based on everyday behaviours and their link to mental health served as our primary effect size of interest to allow for a clear confirmatory approach, we also identified secondary SESOIs based on the effect sizes that are practically and clinically meaningful. These secondary SESOIs play an important role in supporting the interpretation of our findings (as detailed in Supplementary Methods), ensuring the applicability of our study to both academic and practical domains.

Overall, we inferred support for the alternative hypotheses if (1) the coefficient for the association of social media use and our grouping variable for mental health diagnosis was significant, (2) the association followed the hypothesised direction and (3) the CIs for the association did not fall within the equivalence bounds. The interpretation and analysis plan are presented separately for each hypothesis in the Supplementary Information. Further, we detail the planned exploratory analyses in the Supplementary Methods.

Given that not all assumptions of linear models were met (Supplementary Table 9), we complemented our preregistered parametric analyses with exploratory nonparametric tests for all hypotheses (Supplementary Tables 10 and 11). We implemented the Brunner Munzel test based on the ‘brunner_munzel’ function for NHST and the ‘simple_htest’ function for equivalence testing from the TOSTER package in R91. This test, also known as the generalized Wilcoxon test, is a nonparametric test of stochastic equality between two samples that tests against the null hypothesis that for randomly selected values X (that is, social media score in the adolescents without a condition) and Y (that is, social media score in adolescents with a condition), the probability of X being greater than Y is equal to the probability of Y being greater than X. We documented all discrepancies (20% overall) between the parametric and nonparametric test results in Supplementary Tables 1214.

Sampling plan

Inclusion criteria and sample size

We included all participants in the MHCYP 2017 survey aged between 11 and 19 years at the time of assessment who have answered any of the social media use questions with anything other than “don’t know”. Only individuals with diagnostic information, including the absence of a diagnosis, were included in the MHCYP dataset. Since we examined each social media use question separately, we included participants who answered at least one of the seven social media engagement questions or the question about time spent on social media. No specific documentation was available regarding the response rates for the social media questions. Given that the questionnaires were administered through interviews with professionals, we expected minimal missing data for the social media responses103. Hence, using a conservative estimate, we assumed a 5% missingness for our power calculations.

To address question 1, all adolescents with and without a condition were included in our analysis. Hence, in question 1, we included (1) participants with conditions other than internalizing or externalizing disorders (for example, autism spectrum disorder, tic disorder and psychotic disorders; Table 2); (2) participants with within-group comorbidity (multiple internalizing or externalizing diagnoses; for example, comorbid depressive and social anxiety disorder); and (c) participants with between-group comorbidity (both internalizing and externalizing conditions; for example, comorbid depressive and attention deficit hyperactivity disorder).

In contrast, for questions 2 and 3, we included (1) participants with internalizing or externalizing conditions only and (2) participants with within-group comorbidity (multiple internalizing or externalizing diagnoses). Hence, participants with other conditions or between-group comorbidity were excluded from the main analysis, given our goal to compare social media use in adolescents with externalizing and internalizing conditions. However, we ran sensitivity analyses to test the impact of including participants with between-group comorbidities in question 2.

On the basis of the summary demographics83, after accounting for potential missing data, we estimated the sample of complete cases to be around N = 3,854 (accounting for 5% missingness in the sample of N = 4,057, 11–19-year olds), approximately 15% of whom received at least one mental health diagnosis (N = 577). Available documentation83 suggests that approximately 19.2% of individuals with internalizing conditions have at least one comorbid externalizing condition, and approximately 28% of individuals with externalizing diagnoses have at least one comorbid internalizing diagnosis. This suggests approximately N = 370 with internalizing-only diagnoses and N = 199 with externalizing-only diagnoses.

Power calculations

We calculated the power by setting the SESOI (d = 0.4, refer to Supplementary Methods for a detailed explanation), the alpha level to 0.05 and the estimated sample size (estimated N = 3,854 based on existing MHCYP documentation). For the equivalence tests, power was calculated using the TOSTER package in R91,104. For the regression models, power was determined using the pwr package105 in R. The code for these calculations is available on the OSF94.

H0(1). We calculated the power for equivalence testing to detect a SESOI of d = 0.4 given at least N ≥ 577 individuals with a condition and N ≥ 3,277 individuals without a condition. The results indicate 100% power to reject the presence of effects that are larger than d = 0.4.

H1(1). We calculated the power to detect a statistical effect of condition on social media responses using linear regression. The results indicate 100% power to detect the SESOI (d = 0.4) with at least N ≥ 577 individuals with a condition and N ≥ 3,277 individuals without a condition.

H0(2). We calculated the power for equivalence testing to detect a SESOI of d = 0.4 given at least N ≥ 370 individuals with an internalizing only condition and N ≥ 3,277 with no condition. The results indicate 100% power to reject the presence of effects that are larger than d = 0.4. We calculated the power for equivalence testing to detect a SESOI of d = 0.4 given at least N > 199 individuals with an externalizing only condition and N > 3,277 with no condition. The results indicate 100% power to reject the presence of effects that are larger than d = 0.4.

H1(2). We calculated the power to detect a statistical effect of internalizing condition type (internalizing versus no condition) on social media responses using linear regression. The results indicate 100% power to detect the SESOI (d = 0.4) with at least N ≥ 370 individuals with an internalizing-only condition and N ≥ 3,277 with no condition. We calculated power to detect an effect of externalizing diagnosis type (that is, externalizing versus no condition) on social media responses using linear regression. The results indicate 100% power to detect the SESOI (d = 0.4) with at least N ≥ 199 individuals with an externalizing-only condition and N ≥ 3,277 with no condition.

H0(3). We calculated the power for equivalence testing to detect a SESOI of d = 0.4 given at least N ≥ 370 individuals with internalizing only and N ≥ 199 with externalizing only conditions. The results indicate 96% power to reject the presence of effects that are larger than d = 0.4.

H1(3). We calculated the power to detect a statistical effect of internalizing-only versus externalizing-only condition type on social media responses. The results indicate 99.8% power to the SESOI (d = 0.4) with at least N ≥ 370 individuals with internalizing only and N ≥ 199 with externalizing only conditions.

In addition to our a priori power calculations, we conducted power sensitivity analyses with the final sample size105. That is, we determined the smallest effect size that is observable with 95% power, given the corrected alpha of 0.0125 and our final sample size (Supplementary Table 22).

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.