Social media use and mental health during the COVID-19 pandemic in young adults: a meta-analysis of 14 cross-sectional studies

Studies were included which met the following criteria: (1) use of the English language; (2) conducted after March 11, 2020 (date the WHO declared a pandemic) and published by December 20, 2020; (3) collected data using a validated tool of mental health symptoms (e.g., Patient Health Questionnaire: PHQ9, Generalized Anxiety Disorder-7 items: GAD-7); (4) full texts available; (5) measured time spent on social media platform in either continuous or categorical variable; (5) provided their results in OR, β, and/or Pearson’s r, and (6) studies measured mental health symptoms such as anxiety and depression.

Studies with the following characteristics were excluded: (1) Studies examined traditional social media (e.g., television and radio); (2) case reports, letters, comments, and narrative reviews without quantitative results, and (3) studies using a language other than English.

Studies investigating the association between time spent on social media and mental health outcomes (e.g., anxiety and depression) were summarised in Supplementary Material 1. The pooled effect size of this meta-analysis was mainly presented in an odds ratio (Fig. 2).

The search strategy principles were as follows: (1) “Social media” or individual names of social media in the title, keyword and abstract results; (2) Terms referring to mental health with COVID-19 specified in the title (e.g. depression, anxiety or blue).

A systematic literature search of the PubMed, Embase, and Cochrane Library databases was performed to identify studies. Publication date restrictions are from March 2020 to December 20, 2020. The search terms for a systematic search were as following: (1) (“COVID-19“ OR “corona“) AND (“mental health” OR depress* OR anxiety) AND (“social media” OR “Instagram” OR “Facebook” OR “twitter”) for PubMed, (2) (“coronavirus disease 2019’/exp/mj) AND (“mental health“/exp/mj OR “depression“/exp OR “anxiety“/exp) AND (“social media”/exp./mj OR “Facebook”/exp. OR “twitter”/exp. OR “Instagram“/exp) for Embase; (3) (“COVID-19″ OR “corona”) AND (“mental health“ OR depress* OR “anxiety”) AND (“social media“ OR ‘Instagram” OR “Facebook” OR “twitter”) for Cochrane Library.

Articles were first screened by reviewing titles, followed by a full-text review. Every selection stage involved three independent researchers (two medical doctors [SJJ and YRL] and one graduate student from the Epidemiology Department [YJJ]). Every article was independently evaluated by two researchers (YJJ and YRL) in first hand, and a third researcher (SJJ) mediated the final selection in case of differences in opinion.

Study data were extracted by two independent researchers (YRL and YJJ). A single author first extracted the information and a second author checked for accuracy. The extracted information is as follows: country of study, participant group sampled, age group of sample, date of data collection, mental health measures, effect size information, social media use time, and whether the adjustment was made for each analysis (see Supplementary Material 1). Studies were subdivided into categories according to the summary estimate of effect sizes (odds ratio [OR], beta estimate from multiple linear regression [β], and correlation coefficient [Pearson’s r]).

The final studies after retrieval measured the amount of time spent on social media, which was either categorical or continuous variables (see Supplementary Material 1). It was measured based on the response to an item in the questionnaire: “How often were you exposed to social media? [categorical]” and “How long (in hours) were you exposed to social media? [continuous].” The measurement of exposure was expressed in different wordings as follows: “Less” vs. “Frequently,” “Less” vs. “Often”, “less than 1 hour” vs. “2 hours or more,” or “less than 3 hours” vs. “3 hours or more.” To calculate the overall effect, these individually measured exposure levels were operationally redefined (e.g., “Less” and “Few” were considered the same as “less than 2 hours;” “less than 1 hour,” “Frequently,” and “Often” were treated the same as “2 hours or more” and “3 hours or more”).

The outcomes of included studies were “anxiety”, and “depression”. Anxiety was ascertained by using GAD-7 (cut-off: 10+), DASS-21, and PHQ-9, while depression was measured using PHQ-9 (cut-off: 10+), WHO-5 (cut-off: 13+), and GHQ-28 (cut-off: 24+). Anxiety and depression measured by using screening tools with cut-offs presented results in odds ratios (see Supplementary Material 1).

All statistical analyses and visualisations were performed with the “meta,” “metaphor,” and “dmeter” package of R version 3.6.3 (https://​cran.​r-project.​org/​), using a random-effect model [13‐15]. The effect measures were odds ratio, regression coefficient, and Pearson’s r, which calculated the association between the increase in social media use time and anxiety and depressive symptoms. In each study, the association with the mental health level of the social media frequent use group (compared to the low frequency group) was calculated as the odds ratio, and the association with the increase in the mental health level per hour increase was calculated as the regression coefficient (β) and Pearson’s r. Statistics used for calculating pooled effects (e.g., odds ratio, regression coefficient, and Pearson’s r) were utilized as its adjusted value with covariates from each study, not the unadjusted crude values.

The pooled effect sizes, Cochrane’s Q, and I² to assess heterogeneity were calculated. The pooled effect sizes, CIs, and prediction intervals were calculated by estimating the pooled effect and CIs using the Hartung-Knapp-Sidik-Jonkman method, which is known as the one of the most conservative methods [16]. The degree of heterogeneity was categorised as low, moderate, or high with threshold values of 25, 50, and 75%, respectively [17]. Possible causes of heterogeneity among study results were explored by statistical methods such as influential analysis, the Baujat plot, leave-one-out analysis, and Graphic Display of Heterogeneity analysis [18]. In addition, publication bias was assessed using funnel plots, Egger’s tests, and the trim-and-fill method [19].

Quality assessment was conducted by two independent researchers, a psychiatrist (SHK) and an epidemiologist (YRL), using the Risk of Bias Assessment Tool for Nonrandomized Studies (RoBANS), which can assess cross-sectional studies [20]. RoBANS has been validated with moderate reliability and good validity. RoBANS applies to cross-sectional studies and comprises six items: participant selection, confounding, exposure measurement, blinding of outcome assessments, missing outcomes, and selective reporting of outcomes. Each item is measured as having a “high risk of bias,” “low risk of bias,” or “uncertain.” For example, based on “participant selection,” each researcher marked an article as having a “high risk of bias” if, for example, the patient definitions of depression were generated by self-reported data. In cross-sectional studies, misclassification cases due to an unreliable self-contained questionnaire for categorizing depressive patients were rated as “high risk.” For the qualitative assessment, studies with two or more “high risk of bias” grades were then classified as “low quality”. The study was rated as “high quality” only if the evaluation of both raters was congruent. For sensitivity analysis, additional analysis including only “high quality” studies was conducted and it compared with the pooled estimates of overall results (see Table 1).

The preferred reporting items for systematic reviews and meta-analyses (PRISMA) guidelines 2020 were followed for this study. No ethical approval and patient consent are required since this study data is based on published literature. This meta-analysis review was registered with PROSPERO (https://​www.​crd.​york.​ac.​uk/​PROSPERO/​, registration No CRD42021260223, 15 June 2021).

Total of 346 studies were selected from the database search (288 from PubMed, 34 from Embase, and 24 from the Cochrane Library). After removing 19 duplicate publications, 327 studies were included for the title and full-text review (see Fig. 1). Non-original studies and those conducted with irrelevant subjects (n = 218) were excluded. Another 95 studies were excluded finally due to inconsistent study estimates. As summarised in Supplementary material 1 and 8, 13 papers studied anxiety as an outcome (6 studies in odds ratio, 3 in regression coefficient, 4 in Pearson’s r), and a total of 9 papers studied depression as an outcome (6 studies in odds ratio, 3 in regression coefficient). Each of the final distinct 14 studies (after excluding duplicate studies) measured multiple mental health outcome variables (i.e., anxiety and depression), and pooled effect sizes were calculated for each outcome. Six studies that dealt with anxiety symptoms and six with depression (Supplementary Material 1–1-1, 1–2-1) reported ORs and their 95% confidence intervals (CIs) (n = 9579 and n = 13,241 for anxiety and depressive symptoms, respectively). Three studies each on anxiety and depression (Supplementary Material 1–1-2, 1–2-2) reported their findings in β (n = 2376 and n = 2574 for anxiety and depression, respectively). All included studies were cross-sectional studies. The pooled effect size was presented in odds ratio.

Table 1 shows the result of the meta-analysis about the relationship between time spent on social media and mental health outcomes (i.e., anxiety and depression) of the selected cross-sectional studies. The increase in the time spent using social media platforms were associated with anxiety symptoms in overall studies (pooled OR = 1.55, 95% CI: 1.30–1.85, prediction intervals: [1.08–2.23]), and the heterogeneity between studies was mild (I² = 26.77%) (see Fig. 2). The three cross-sectional studies (presented in β) were insignificant (β = 0.05, 95% CI: − 0.32–0.15; a unit increment of each screening tool score per hour) with relatively high inter-study heterogeneity (I² = 76.07%). The overall estimate of the four cross-sectional studies (Pearson’s r) was 0.18 (95% CI: 0.10–0.27) with high inter-study heterogeneity (I² = 73.04%). The increase in social media use time was also associated with depressive symptoms (pooled OR = 1.43, 95% CI: 1.30–1.85, prediction intervals: [0.82–2.49]), and the heterogeneity between studies was moderate (I² = 67.16%) (see Fig. 2).

As result of quality assessment analysis, pooled effect size of studies classified as “high quality” was presented in Table 1. The results were similar to the overall outcome (anxiety: OR = 1.45, 95% CI: 1.21–1.96; depression: OR = 1.42, 95% CI: 0.69–2.90). High-quality studies had low inter-study heterogeneity (anxiety: I² = 0.00%; depression: I² = 0.00%). The kappa statistic (inter-rater agreement) was 33.3%, indicating fair agreement.

Publication bias was assessed by funnel plot analysis and Egger’s test (Supplementary Material 4–1). Funnel-plot analyses revealed symmetrical results (Supplementary Material 4–2). In addition, all results of the Egger test were statistically insignificant, indicating improbable publication bias. After applying the trim-and-fill method, the funnel plot revealed no asymmetry (Supplementary Material 5), indicating no significant publication bias.