More than 1 in 8 adults and adolescents worldwide live with a mental disorder, with anxiety and depression being the most commonly reported mental health symptoms [1]. Yet, a significant lack of access to mental health services has left more than 70% of those with mental disorders untreated [2,3]. The limited access to mental health services stems from widespread stigma surrounding mental health conditions, a lack of certified counselors and therapists [2,3], and a high employee turnover rate exacerbated by the global pandemic [4,5]. This global mental health crisis not only harms people’s physical health [6], personal wellness [7], and interpersonal relationships [8], but can also result in economic losses amounting to US $1 trillion [9].

Online mental health interventions delivered through social-media-based programs can potentially alleviate the mental health crisis. Traditional in-person mental health services often face limitations in accessibility, affordability, and reach. In comparison, social-media-based interventions can be delivered at scale [10], with a low cost [11], and require no extensive training [12]. These features facilitate people’s access to mental health resources and alleviate the burdens of health care workers. Furthermore, social-media-based interventions can also offer real-time peer support, access to resources (ie, psychoeducation), and therapeutic interactions in a familiar and engaging environment, creating a safe and supportive environment, especially for individuals who may feel stigmatized or isolated.

The growing popularity of social media in recent years presents an opportunity to critically evaluate its potential in promoting mental well-being [13]. However, while plenty of meta-analyses on social media interventions have focused on physical health-related outcomes such as sexual health and substance use [14], weight-related behavior [15], obesity [16], diet and exercise [17], there is a severe lack of research focusing on psychological health outcomes [18]. The pressing need to explore cost-effective interventions for mental health problems, especially with the current shortage of counselors [19] and the growing national crisis in mental health [20], makes our meta-analysis timely and necessary. In light of the common criticism of “garbage in, garbage out” on review work [21], this review only includes rigorously designed randomized controlled trials (RCTs) to provide high-quality evidence on what works. The results of this meta-analysis will give us an overview of the ways we can use social media to benefit our mental health.

Prior meta-analyses have examined extensively online [22,23], digital [24-27], eHealth [28], computer therapy [29], or internet-based [30-32] mental health interventions. However, up till the time of writing this paper, there are only two meta-analyses of social-media-based mental health interventions for patients with cancer [33,34], one scoping review probing into social-media-based mental health interventions for children [35], and a systematic review focusing on social networking sites in mental health interventions for young people [36]. To the best of the authors’ knowledge, no meta-analysis so far has focused on rigorously designed social-media-based mental health interventions for general populations. Therefore, we are motivated to conduct this meta-analysis.

Since there is no meta-analysis of social-media-based mental health interventions as a reference, we selected meaningful moderators based on past online and digital mental health interventions, a broader category that includes social-media-based interventions. We included the following 7 moderators.

In this meta-analysis, we intend to answer two research questions.

This meta-analysis was preregistered at Open Science Framework [49]. In practice, there were 3 deviations (population scope, intervention channel of focus, and outcome type) from our preregistered protocol. While the protocol focused on adults, we observed substantial variation in the age groups studied during the screening process. We also recognized that adolescents’ social media use patterns and platforms might differ from those of adults [50]. Thus, we decided to include studies involving all age ranges but conduct a moderator analysis using age groups.

The second change involves changing from including both internet-based and social-media-based interventions to including only social-media-based interventions. This decision was made to reduce heterogeneity in intervention delivery channels and allow for more in-depth analysis. Broader categories like mobile- or internet-based interventions include diverse formats such as digital platforms, chatbots, mHealth, and telehealth, which would have introduced significant variability. While numerous meta-analyses exist on general online or mobile mental health interventions, very few focus specifically on social media. To the best of our knowledge, the only existing meta-analyses on this topic target patients with cancer [33,34]. We identified this gap in the literature and adjusted our scope accordingly to contribute new and timely insights.

Last, we changed from including both positive and negative mental health outcomes to including only negative mental health outcomes. Our review found that study designs varied greatly depending on whether the intervention aimed to reduce negative mental health issues or enhance positive outcomes (eg, well-being, life satisfaction, and happiness), resulting in high heterogeneity. We also realized that enhancing positive mental health outcomes was not the same as reducing negative outcomes. For instance, the interventions that can effectively increase happiness are very different from interventions that can effectively reduce anxiety. To reduce heterogeneity, we chose to include only studies targeting negative mental health outcomes, such as depression, anxiety, stress, negative affect, and psychological distress.

To ensure comprehensive literature coverage, the first author (QZ) conducted a combined search strategy including database search, hand searching, and backward citation tracking. Using a predefined set of keywords on social media, intervention method, and mental health issues (Table S1 in Multimedia Appendix 1), the first author searched for literature in 7 databases, including the Education Resources Information Center, PsychINFO, Scopus, PsychArticles, Communication and Mass Media Complete, PubMed, and Proquest (Table S2 in Multimedia Appendix 1). Supplementary to the database search, targeted hand searching was performed through Paperfetcher [51] across selected reputable journals specific to the field (Table S3 in Multimedia Appendix 1). Finally, forward citation tracking and backward citation chasing were performed using CitationChaser on relevant systematic reviews and meta-analyses (Table S4 in Multimedia Appendix 1 [14,15,29,36,52-67]). We completed all literature searches and updates by April 2025, yielding a total of 11,658 studies, and subsequently imported them into Covidence for screening. Covidence was chosen for its functionality in facilitating full-text review and availability of software licenses through the authors’ affiliated institutions [68].

We used the population, intervention, comparison, and outcome framework to help refine inclusion criteria. In this meta-analysis, we focused on general populations across different age groups and investigated mental health interventions delivered through social media platforms with either CAU or waitlist control groups. The outcomes of interest are negative mental health issues, such as depression, anxiety, stress, etc.

Eligible studies must meet the following twelve criteria to be included in this meta-analysis: (1) Studies must be RCTs. We excluded quasi-experimental study designs because we only want to include rigorous causal inference study designs. (2) Studies must have at least 30 participants per experimental condition at the baseline measurement. This is because extremely small samples can inflate ESs [69]. (3) Interventions must be largely delivered through social media platforms (eg, Facebook, Instagram, WhatsApp, and WeChat). We excluded social media abstinence interventions. (4) Based on the What Works Clearinghouse standards [70] for high-quality evidence, the difference between conditions at baseline on mental health measures must be less than 0.25 SDs. (5) Based on What Works Clearinghouse standards [70] for high-quality evidence, differential attrition between treatments and control groups must be less than 15%. (6) Intervention or instruction should be delivered by nonresearchers. This is because treatments delivered by researchers might be more challenging to sustain or replicate in the real world. For example, Guo et al [71] were excluded because the intervention was delivered by the authors. (7) Outcomes of interest measurements must include quantitative measures of negative mental health issues, such as depression, anxiety, stress, psychological distress, etc. The authors must either directly provide the ESs or offer enough statistics for us to compute Hedges g. For instance, Wang et al [72] were excluded because of insufficient data. (8) Full text must be available on the Internet and written in English. This is because we need to screen the full text to assess studies’ eligibility for inclusion, and our researchers need to double-code each study in English. (9) Articles must be published on or after January 1, 2005. This is because social media only started to be prevalently used after 2005 [73]. (10) Studies must be primary studies instead of secondary analyses. This prevents duplicated datasets from being included in this meta-analysis multiple times. (11) We excluded studies that use one-item measures because one-item measures are often unreliable [74]. (12) We excluded interventions that only have a single session because the extremely short program duration might inflate ESs.

Covidence was used for screening. All authors participated in the title and abstract screening as well as the full-text review. Each study was independently and double-blindly screened by at least 2 authors. Each eligible study was coded by 2 reviewers using Google Spreadsheets. The items extracted from each included paper can be found in Table S5 in Multimedia Appendix 1. We resolved any conflicts through weekly group discussions and reached a consensus with 100% agreement.

Following the call for open science, we shared all procedures, coding spreadsheets, and analysis code on GitHub [75] to ensure accessibility and future reproducibility.

We used the R statistical software’s (R Foundation for Statistical Computing) metafor package for analysis [76]. This meta-analysis used weighted mean ESs and meta-analytic tests such as Q statistics. Weights were assigned to each study based on inverse variance [77] and were adjusted according to Hedges’ [78] recommendations. We used a random-effects model in meta-regression due to the presence of a range of ESs dependent on various factors [79]. For each primary study, we calculated standardized mean differences in Hedges’ g [80] through R’s metafor package’s function called escalc [76].

The moderator analysis included 7 sets of moderators. We grand-mean-centered all moderators and covariates to aid interpretation. Mean ESs were derived from the meta-regression model, which accounted for potential moderators and covariates. To assess publication bias, we used selection modeling instead of traditional methods (eg, funnel plot, Egger’s regression, fail-safe N) due to their limitations [81]. For instance, funnel plots are subject to meta-analysts’ interpretations, which are often misled by the plot shapes [81]. The fail-safe N technique is often criticized for the arbitrary choice of zero, which ignores heterogeneity in primary studies [82,83]. Selection modeling involved a weight function model implemented using the weightr package [84]. In adherence to the principles of open science, the complete description of the intervention characteristics is available together with the data and R code in GitHub [75].

As for risk of bias analysis, we applied the JBI Critical Appraisal Checklist for RCTs (Textbox 1 presents the 13 criteria). Since we applied stringent inclusion criteria, all the included studies already met the JBI Checklist’s criteria 1, 3, 7, 9, 10, 11, 12, and 13. Therefore, we only coded the criteria 2 (Was allocation to treatment groups concealed?), 4 (Were participants blind to treatment assignment?), 5 (Were those delivering treatment blind to treatment assignment?), 6 (Were outcomes assessors blind to treatment assignment?), and 8 (Was follow-up complete, and if not, were differences between groups in terms of their follow-up adequately described and analyzed?) to examine risk of bias. Following the double-blinded coding procedure, 2 independent reviewers conducted the quality assessment. A third reviewer was consulted to resolve the disagreement. In the online data repository [75], we listed the initials of the authors responsible for each study.

The search strategy retrieved 11,658 published studies. Figure 1 presents the PRISMA (Preferred Reporting Items for Systematic reviews and Meta-Analyses) screening process (Checklist 1). A total of 4019 records were removed as duplicates. After deduplication, 7639 studies were screened by title and abstract, and 761 potentially eligible studies were reviewed in full text. During the full-text review stage, the top 5 reasons for exclusion included: wrong study design (n=345), not social media-based (n=112), absence of a negative mental health outcome (n=77), fewer than 30 participants per condition (n=73), and social media abstinence interventions (n=35). Finally, 17 studies were included in this meta-analysis. Table 1 presents the main characteristics of the included studies.

Applying the weight-function model, we found that the mean effect estimates were upwardly adjusted, implying that statistically nonsignificant effects were less likely to be reported than significant results (g is 0.98 when the cut-off point for P value is between .025 and .5 [ie, P<.01]; g is 0.02 when the cut-off point for P value is between .5 and 1 [ie, P=.28]).

Table 5 presents the risk of bias analysis. Among the 17 included studies, 6 studies concealed allocation to treatment groups, 3 studies did not, and 8 studies were unclear. Four studies blinded participants to treatment assignment, 7 studies did not, and 6 studies were unclear. Only 2 studies blinded those delivering the treatment, 7 studies did not, and 8 studies were unclear. Four studies blinded outcomes assessors, 1 study did not, and 12 were unclear. Six studies completed follow-ups, while 11 did not clearly report this information. Overall, the risk of bias across the included studies can be considered low, as the mean appraisal score was relatively high (9.29 out of 13).

Since the 70% threshold for defining a majority-female sample is somewhat arbitrary, we conducted a sensitivity analysis using a 50% threshold instead. The result remained robust, indicating that studies with a majority-female participants were still more effective than those with a majority-male participants.

This meta-analysis found that rigorous social-media-based RCTs were effective in reducing mental health issues. In particular, social-media-based interventions are more effective when the participants are mainly female, when the programs are human-guided, social-oriented, and when control groups are CAU.

Figure 2 shows that the number of rigorously designed social-media-based mental health RCTs peaked from 2020 to 2024, reflecting the impact of the COVID-19 pandemic as well as the increasingly rigorous methods. After the pandemic, the number started dropping but remained higher than the prepandemic average. The trend in this research area highlights the need for review work to understand the overall picture of these interventions.

In descriptive results, interestingly, all studies conducted in Muslim countries (Iran, the Kingdom of Saudi Arabia, Jordan, and Indonesia) recruited 100% female participants. This might be attributed to gender roles and cultural expectations. The online channels might be more appealing to this population since Muslim women prefer to be seen by a female health care provider [104,105], which would not be a barrier in social-media-based programs. Research has shown that Muslim women are not only more vulnerable to mental health issues, but they are also more likely to experience mental health stigma and a lack of social support [106]. It is worth investigating whether the anonymity features on social media encourage more help-seeking behaviors among Muslim women and whether the researchers in the Middle East deliberately designed interventions targeting women.

The weighted-average ES showed that social-media-based interventions were effective at reducing depression, anxiety, stress, negative affects, and psychological distress. This reinforces previous meta-analyses conducted by Siew et al [33] and Wang et al [34]. Although they found a slightly larger ES (Siew et al [33]: 0.41 and Wang et al [34]: 1.1 and 1.37) compared with our 0.32, one reason could be their focus on patients with cancer only. Moreover, compared with general online and internet-based mental health interventions that found higher ESs, this social-media-based meta-analysis showed lower ESs. For instance, Christ et al [32] focused on CBT interventions only and reported ESs of 0.51 on depression and 0.44 on anxiety. Another example is Alrashdi et al [23] who found ESs of 0.41 on depression, 0.40 on stress, and 0.45 on anxiety. Since social-media-based mental health interventions are relatively new channels compared with digital and online interventions, the developing designs and implementation fidelity might affect the overall ESs.

Our moderator analysis showed that social-media-based interventions are more effective when the participants are more than 70% female. This is probably because women tend to obtain more social support from social media platforms [107] and engage more in an interactive and reflective community [47]. According to the tend-and-befriend theory, females give more social support than males when they experience stress, whereas men often respond to stress with the fight-or-flight mindset [108]. Since females give more social support, they might receive more social support due to the norm of reciprocity. Furthermore, females generally have higher rates of help-seeking and engagement with mental health resources than males [47].

In future studies, researchers could investigate more into how these social-media-based programs work differently for different sexes. It is possible that the type of social support provided through social media is more appealing and therapeutic for women [109]. However, this means that future interventions should seek answers for what works best for the majority of male participants. Readers should note that, among the 17 studies, only one study is of a majority-male sample (more than 70% male).

We found that programs that primarily provide social interaction are more effective than programs that provide exercises, information, and tasks. In mental health therapy, therapeutic relationships are important [45,46]. This moderator might also confound with the sex moderator since previous studies found that women prefer communicative channels while men prefer solution-focused or skills-training programs [47]. Since most programs have more than 50% female participants, the sex factor might have interacted with the program orientation.

We found that programs guided by nurses, clinicians, coaches, or therapists are more effective than self-guided programs. This finding aligns with previous systematic reviews that guided programs, regardless of qualifications, are more effective than nonguided programs [110,111]. Support from others can enhance motivation, reduce attrition, and create therapeutic relationships for improved counseling effectiveness [112]. Human supporters can also hold users accountable to engage more in the designed programs by providing timely help, which eventually leads to more effectiveness of the program [113]. This shows that although social-media-based programs are less costly to scale up, incorporating certain human elements is essential in maintaining the programs’ value and effects. Another thing to note is that most social-oriented programs are human-guided (86%) rather than self-guided. It is possible that these 2 moderators’ effectiveness affects each other.

This finding is unexpected and contradicts our hypothesis. While the conventional assumption is that participants assigned to the waitlist control will not engage in anything, this might not hold true for this topic. Most studies recruited people who voluntarily signed up through social media advertisements. Participants assigned to waitlist control groups might actively seek other online treatments since telehealth therapy is much easier to access in the postpandemic period. They might also seek in-person one-on-one therapy, which further confounds this comparison.

In terms of age, we did not find significant differences between adolescents and adults, although some previous reviews [35] argued that younger populations might be more fluent in social media and internet literacy. Therefore, they might reap larger benefits. The meta-regression results did not support this argument.

There are several limitations worth readers’ attention when interpreting results. First, during the moderator analysis, there are a few subgroups of moderators that have small sample sizes. For example, we only found one ES on psychological distress, 2 ESs on negative affect, and 1 study on late adulthood that met our criteria. The small sample sizes in these categories not only revealed a lack of intervention studies on these specific moderators but also reminded us to interpret these results with caution.

Moreover, the statistical power of this review was limited by the small sample size available. Since we only included rigorously designed RCTs, and only 17 studies met our inclusion criteria, our statistical power is limited, which is shown in the low degrees of freedom in the meta-regression model.

Finally, although there is a low risk of bias and the sensitivity analysis suggested that the model results were robust, the selection bias analysis suggested that the mean effect estimates were likely upwardly adjusted. Readers should be aware of this selection bias.

Based on this review, future researchers investigating social-media-based mental health interventions can focus on designing more rigorous RCTs. Researchers should focus on developing and implementing robust methodologies and exploring new social-media intervention models to effectively address mental health challenges.

We found that social-oriented programs are more effective than task-oriented programs. Future researchers can investigate the interaction between program types and other factors (ie, personality, sex, level of comfort with technology, and different mental health issues). It is possible that certain subgroups would prefer certain program orientations. Understanding this would help us design more targeted and effective programs for different users.

We found that programs with active and CAU controls are more effective than programs with waitlist control. Future studies can design 3-arm trials to assess the comparison more directly.

Last but not least, we did not collect enough information on racial distribution, which could be a valuable moderator to investigate the equity issue in the design and application of high-quality interventions. Although we planned to code for racial distribution, most studies did not report race. Even among studies that did report relevant information, they might have reported nationality instead of race, or used different ways to categorize racial groups. As most studies do not report on race, we encourage future studies to report the racial distribution of participants or consider the racial information as a variable in future RCTs of social media.

Social-media-based mental health interventions provide an important modality to deliver treatment with various benefits, including accessibility, cost-effectiveness, scalability, and acceptability. Despite intense debates surrounding the effects of social media use on mental health, a more proactive approach is to explore how we can harness these platforms to design interventions that enhance mental health outcomes. As former US President Barack Obama stated [114],

This meta-analysis synthesized the best evidence on this topic and found that, overall, high-quality social-media-based RCTs were effective in reducing depression, anxiety, stress, negative affect, and psychological distress. Given the benefits of scalability and cost-effectiveness of social-media-based approaches, mental health services should consider integrating online interventions into routine practice. These interventions are especially beneficial for individuals facing barriers to traditional therapy, such as those in geographically isolated areas or those unable to afford in-person care. Policymakers could consider collaborating with social media platforms to perhaps integrate these interventions into public health systems and promote equitable access to mental health care through social media platforms.