Training TikTok creators in mental health communication can benefit their audiences

Training TikTok creators in mental health communication can benefit their audiences | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Social Sciences - Article Training TikTok creators in mental health communication can benefit their audiences Yuning Liu, Matt Motta, Kenzie Brenna, Shahem McLaurin, Katharine Speer, and 3 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6787693/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract The rapid rise of social media presents new opportunities for implementing network interventions to improve health outcomes 1 , 2 . In the mental health domain, content creators can serve as influential community leaders 3 , 4 , yet it remains unclear whether creator-focused training impacts the mental health attitudes and behaviors of their audiences. This study evaluates the effectiveness of creator-focused interventions on the mental health knowledge and knowledge-based skill bases of their audiences using two experiments. First, an on-platform randomized controlled trial (RCT) provided evidence-based mental health communication toolkits and training to TikTok creators 5 . Analyzing 188,169 comments from 1,882 videos by 49 creators (March–May 2023), we found a 4% increase in foundational mental health knowledge construction among viewers—defined as the initial stage of learning through reflection on personal experiences or opinions. Second, an off-platform survey experiment exposed a nationally representative sample of 1,000 U.S. youth (aged 14–22) to pre- and post-training videos created by a lifestyle influencer. Participants who viewed the post-training video showed significant improvements in both perceived and objectively assessed emotional support skills. By combining the external validity of the on-platform RCT with the internal validity of the survey experiment, our study provides robust evidence that creator training toolkits can enhance mental health knowledge and competencies among audiences, supporting the promise of scalable, evidence-based communication strategies on social media. Scientific community and society/Social sciences/Interdisciplinary studies Scientific community and society/Social sciences/Communication Health sciences/Risk factors Figures Figure 1 Figure 2 Figure 3 Introduction Social media has created new opportunities for implementing network interventions to promote health outcomes 1 , 2 . Network intervention refers to a set of strategies that leverage social network structures to promote behavior change 6 . These strategies have been successfully applied in interventions targeting various health outcomes, including exercise, smoking control, and mental health 7 , 8 . Video-based platforms such as TikTok, Instagram, and YouTube host online social networks for mental health communication, where users seek information, share personal experiences, and access community support 3 , 9 , 10 . According to a national survey by the U.S. CDC, 58.5% of adults used the internet for health or medical information in 2022 11 . On TikTok, views of the hashtag #MentalHealth surged from 25.3 billion in 2022 to nearly 44 billion in 2023, with over 22.1 million posts in 2025 12,13 . Notably, 66% of all TikTok users—and 91% of those aged 18–29—report using the platform to access mental health information or advice 3 . Mental health content creators can serve as influential online community leaders in promoting mental health within online social networks. Content creators have been shown to influence user behavior in fields such as fashion and lifestyle 14 , 15 , as well as promote hygiene practices during COVID-19 16 . Mental health content creators reach broad audiences 5 on platforms such as TikTok. More importantly, their content has gained trust from audiences. For example, a recent poll found that 53% of U.S. adults aged 18–29 trust health information found on TikTok 3 , and young adults trust social media health content nearly as much as information from national news outlets 17 . Moreover, younger users are more likely than older adults to act on this information—for example, by consulting a doctor (19% for ages 18–29, compared to 14% for ages 30–49 and 5% for 50+), or seeking mental health treatment (26% for ages 18–29, versus 9% and 2%, respectively) 3 . A 2025 study across 16 countries found 33% of participants aged 18 to 34 made health decisions influenced by content creators without medical training 4 . To promote evidence-based online mental health communication via content creators, organizations have initiated creator training programs. For example, Harvard T.H. Chan School’s Center for Health Communication (CHC), the World Health Organization (WHO), and YouTube Health have each launched programs for mental-health creators 5 , 18 . These programs aim to improve the quality of mental health communication on social media by providing creators with resources such as expert-vetted mental health information and training in evidence-based content practices. Meanwhile, organizations like the Mental Health Storytelling Coalition and CHC maintain publicly available resource libraries for mental health content creators 19 . Empirical creator-engaged studies have recently shown that training programs can effectively change social media content. A large-scale randomized controlled trial conducted on TikTok 5 demonstrated that providing creators with toolkits containing evidence-based mental health talking points increases the prevalence of such content in the videos creators produced. Creators also reported increasing motivation to incorporate evidence-based research into their videos after attending an in-person training 20 . Moreover, the creator training program significantly boosted views of trained creators' content 5 , demonstrating the effectiveness of promoting evidence-based content to social media users through a network-based strategy of "influencing the influencers." However, it remains unclear whether the effects of creator training influence mental health attitudes and behaviors of the content consumers – which is the primary goal of network interventions. According to health behavior change theories 21 , 22 , exposure to evidence-based mental health information can shift beliefs and perceptions, improve health literacy, and potentially prompt protective behaviors. Yet, no previous studies have empirically examined these training programs' effects on viewer-level mental health outcomes. Our research aims to address this gap. We examine the potential of training content creators to improve mental health communication within online social networks through two randomized controlled trials (RCTs). The first is an on-platform RCT on TikTok that evaluates how training creators in evidence-based content influences mental health knowledge construction—the process of understanding concepts, phenomena, and situations—as evidenced by user comments left on each video. The second is an off-platform, nationally representative survey-based RCT of U.S. youth aged 14–22, assessing how training creators in how to offer emotional support to friends affects the audience’s ability to provide such support. In so doing, our approach combines the external validity benefits associated with on-platform experiments (i.e., the fact that video interventions were embedded in “real life” social media interactions) with the internal validity associated with survey-based experiments (i.e., standardization of the content to which people were exposed; ability to measure attitudinal and behavioral data from even those participants who may be reluctant to publicly comment on a video). Our first study, an on-platform RCT, tests the hypothesis that increases in the supply of evidence-based mental health communication content on TikTok, driven by creator training programs 5 , can enhance mental health knowledge construction in video comments. Knowledge construction is the consequence of informal learning among social media users—a non-didactic, socially collaborative process driven by individual interests—across various domains, including climate change 23 , education 24 , geoscience communication 25 , and mental health 26 . Users’ comments offer real-time insights into their online communication and activities 27 , as well as their processes of socialization and idea sharing 12 , 28 . For example, a study of comments on TikTok mental health videos found that 66% offered validation or support of peers and 56% discussed other mental health issues 12 . An overview of the on-platform RCT is presented in Fig. 1 . We assess mental health knowledge construction in TikTok video comments, by using the Interaction Analysis Model (IAM). IAM isa widely applied framework that conceptualizes knowledge construction as a cascading process of increasing complexity. Specifically, we examine whether evidence-based mental health content enhances both the foundational mental health knowledge construction—the initial stage of reflection on personal experiences or opinions—and the advanced stages of mental health knowledge construction, which are characterized by expressions of agreement or disagreement, clarification questions, reinterpretation of knowledge, or its application to other contexts. We apply Large Language Model (LLM)-based content analysis to identify knowledge construction stages in > 180,000 user comments. Based on the experimental assignment protocol embedded in our previous RCT research 5 , we requested all comments (n = 188,169) from all videos (n = 1,882) created by the 49 mental health content creators enrolled in the RCT in March, April, and May 2023 using TikTok’s Research Tools API 29 (Figure A1). We fine-tuned BERT-based LLMs (Table A1) to predict the study outcomes (Table A2)—foundational and advanced mental health knowledge construction—using 4,152 human-annotated comments and augmented texts (Table A3) (see Methods: Outcome Measurement). Our second study, an off-platform survey experiment, addresses the selection bias inherent in the on-platform RCT, where individuals who choose to comment on videos are an unrepresentative subset of the audience. To overcome this, we conducted a pre-registered 30 , survey-based RCT with a demographically representative sample (N = 1,000) of U.S. youth aged 14–22. This study examines whether exposure to videos created using evidence-based mental health communication toolkits influences perceived and objective emotional support competencies; i.e., the process by which people offer guidance and understanding to peers experiencing hardship. We partnered with a popular social media content creator to produce a short, TikTok-style video about emotionally supporting friends using their conventional media production strategies (control, prior to engaging with any interventional materials). The creator then completed training analogous to that featured in our on-platform experiment that provided guidance on how to create evidence-based emotional support information for young people. After the session, the creator made a second video on the same topic (treatment), incorporating as they saw fit the evidence-based content while keeping the format as similar as possible to the first video. Respondents were randomly assigned to view either the treatment or control video and then asked questions about (1) their perceived efficacy in providing emotional support (perceived support) and (2) hypothetical vignette scenarios that put respondents in a position to offer emotional support, pursuant with best practice recommendations (objective support). The training on emotional support was designed based on key competencies identified in prior research 31 – 33 . These competencies include: (a) validating emotions by acknowledging what the person is feeling and why it is understandable, (b) active listening through asking questions, paraphrasing, and avoiding unsolicited problem-solving, and (c) supporting peers' emotions by refraining from minimizing, challenging, or dismissing their concerns. Correspondingly, we expected that treated respondents would be more likely to both express perceived competencies in providing emotional support to friends experiencing mental health challenges and exhibit greater competencies in providing such support, in line with evidence-based best practices. Results Study 1 – On-platform RCT: Effect of creator training on mental health knowledge construction Table 1 Effect of toolkits-only and sessions + toolkits interventions assignment on mental health knowledge construction in video comments Model result Sample size (N) % changes in comments with knowledge construction due to treatment SE P value Comments Videos Creators Toolkits-Only group (ref = control) Foundational + 4.15% 1.85 0.025 123323 1207 29 Advanced -0.86% 1.29 0.502 Sessions + Toolkits group (ref = control) Foundational + 2.29% 1.86 0.219 77910 1171 35 Advanced + 0.64% 1.46 0.662 Note: Parameter estimates are standardized and from three-level multilevel linear probability models (LPM), with comments nested in videos that are created by creators engaged in the study (sessions + toolkits group N = 20, toolkits-only group N = 14, control group N = 15). Multilevel LPMs were separately fitted for sessions + toolkits (compared to the control) and toolkits-only (compared to the control). The study outcomes are foundational mental health knowledge construction—the initial stage involving personal reflection—and advanced mental health knowledge construction—the stages with deeper cognitive engagement and higher-order thinking through agreeing, disagreeing, clarifying, reconceptualization, and application—in mental health measured following Interactive Analysis Model (IAM) by LLM-assisted content analysis. Foundational and advanced knowledge construction are binary variables. The models identify the effect of change in the outcomes attributable to exposure to our study’s interventions by interacting dichotomous fixed effect indicators of treatment group assignment (sessions + toolkits vs. toolkits-only, with the control serving as a reference group), with the fixed effect indicator of whether or not each video was produced pre or post intervention. The pre- and post-intervention periods were defined based on the timing of the study intervention. For the toolkits-only group, the pre-intervention period spanned from March 1, 2023, at 12:00 AM ET to May 2, 2023, at 9:30 AM ET, while the post-intervention period covered May 2, 2023, at 9:30 AM ET to May 31, 2023, at 11:59 PM ET. For the sessions + toolkits group, the pre-intervention period lasted from March 1, 2023, at 12:00 AM ET to April 7, 2023, at 11:59 AM ET, and the post-intervention period extended from April 8, 2023, at 12:00 AM ET to May 31, 2023, at 11:59 PM ET. The estimated prevalence of knowledge construction outcomes in treatment and control groups at the pre and post period are presented in Table A4. Table 2 Effect of creator training toolkits (the toolkits-only intervention) on foundational mental health knowledge construction in video comments by video and creator characteristics Model result Level of analysis Sample size (N) % changes in comments with outcome* due to training toolkits SE P value Comments Videos Creators Video featured themes in training toolkits Video Yes, featured + 8.59% 3.84 0.026 30030 296 24 No, not featured + 2.14% 2.14 0.318 93293 911 28 Video comment engagement (comment-to-view ratio) Video ≥0.97% (median) + 5.56% 2.46 0.024 81322 652 27 <0.97% + 0.59% 2.84 0.837 42001 555 26 Video like engagement (like-to-view ratio) Video ≥0.02% (median) + 6.46% 2.61 0.014 42202 603 26 <0.02% -1.24% 3.10 0.688 81121 604 28 Video share engagement (share-to-view ratio) Video ≥0.18% (median) + 1.18% 2.25 0.601 93555 640 26 <0.18% + 7.86% 3.15 0.013 29768 567 28 Creator engagement level** Comment ≥1.20% (median) + 10.03% 3.56 0.005 68541 557 15 <1.20% + 2.65% 2.36 0.261 54782 650 14 Creator follower count Comment ≥500,000 + 3.87% 3.04 0.204 82774 570 9 <500,000 + 4.75% 2.81 0.092 40549 637 20 Creator is licensed health professional Comment Licensed + 4.70% 2.42 0.053 50481 696 19 Not licensed + 4.29% 3.22 0.184 72842 511 10 * Outcome: foundational mental health knowledge construction ** Creator engagement level is calculated by: [(comment + like)/view] using data in Mar 2023-May 2023. Note Parameter estimates are standardized and derived from three-level multilevel linear probability models (LPMs), with comments nested within videos, which are nested within creators (toolkits-only group N = 14, control group N = 15). Separate LPMs were fitted for comment subsets based on video or creator characteristics. As indicated in the “level of analysis” column, video-level subgroup analyses include all comments from videos in a given subgroup (e.g., videos featuring toolkit-related themes), with one LPM per subgroup. The number of videos in each subgroup defined by a single characteristic (e.g., whether the video features a toolkit-related theme or not) sums to the total of 1,207 videos across the toolkits-only and control groups. Similarly, creator-level subgroup analyses include comments on videos by creators in each subgroup (e.g., ≥ 500,000 followers), with one LPM per subgroup. The number of creators in each subgroup defined by a single characteristic (e.g., follower count ≥ 500,000 vs. <500,000) sums to the total of 29 across the toolkits-only and control groups. The study outcome is the foundational mental health knowledge construction—a binary variable presenting the initial stage involving personal reflection in mental health measured following Interactive Analysis Model (IAM) by LLM-assisted content analysis. The models identify the effect of change in the outcomes attributable to exposure to our study’s interventions by interacting dichotomous fixed effect indicators of treatment group assignment (toolkits-only, with the Control serving as a reference group), with the fixed effect indicator of whether or not each video was produced pre or post intervention. The pre- and post-intervention periods were defined based on the timing of the study intervention. For the toolkits-only group, the pre-intervention period spanned from March 1, 2023, at 12:00 AM ET to May 2, 2023, at 9:30 AM ET, while the post-intervention period covered May 2, 2023, at 9:30 AM ET to May 31, 2023, at 11:59 PM ET. The estimated prevalence of knowledge construction outcomes in treatment and control groups at the pre and post period for different subgroups are presented in Table A5. With the LLM-assisted content analysis results on 188,169 comments, we first explored the pre- vs. post-treatment change in mental health knowledge construction. We assess mental health knowledge construction in video comments across content creators assigned to a “Toolkits-Only” training condition (N = 14 creators, N = 692 videos) that exposed TikTok creators to a series of asynchronous digital training toolkits, a “Sessions Plus Toolkits” condition (N = 20 creators, N = 656 videos) which exposed creators to both digital toolkits and a series of synchronous, virtual training sessions (see: Methods–Creator training program); all compared to a randomly-assigned control group (N = 15 creators, N = 515 videos). Figure A2 presents the estimated prevalence of foundational mental health knowledge construction in video comments across intervention groups over the study period, allowing us to examine the parallel trends assumption and assess the distribution of the raw data. Shown in Table 1 (Note that two-tailed p values are presented throughout), we find that exposure to videos by creators in the toolkits-only intervention significantly increased the likelihood that comments featured foundational mental health knowledge construction—the initial stage of reflection on personal experiences or opinions—by 4.15% (P = 0.025), relative to the control group; following exposure to our interventional materials. In the toolkits-only group, the prevalence of foundational mental health knowledge construction increased from 32.86% in the pre period to 35.02% in the post period, while in the control group, it decreased from 31.57–29.58% over the same period (Table A4). Additionally, neither the toolkits-only nor sessions + toolkits intervention shows significant effects on advanced mental health knowledge construction, which involves expressions of agreement or disagreement, clarification questions, reinterpretation of knowledge, or its application to other contexts. We further examined whether videos containing toolkit-related themes, such as the idea that climate change can negatively impact mental health, were more likely to induce audience mental health knowledge construction. This would further suggest that how content creators integrate toolkit information into their content is responsible for the treatment effects we observe. Among 30,030 comments on 296 videos that featured themes from the toolkits-only training toolkits, the intervention increased the probability of foundational mental health knowledge construction by 8.59% compared to the control group (P = 0.026). Videos in the toolkits-only group featuring toolkit-related themes showed an estimated 5.91% (95% CI: 0.85–10.97%) increase in foundational mental health knowledge construction (Fig. 2 , Panel A) from 33.45–39.36% (Table A5). This finding further supports the effectiveness of the toolkit intervention. Videos that featured themes from the toolkit suggest that the content creator likely engaged with and applied the toolkit in their content creation 20 . Among these videos, the improvement in knowledge construction was greater than in other groups, reinforcing the effectiveness of the toolkit intervention. Because user commenting behavior may be influenced by how engaging a video is, we conducted subgroup analyses based on video engagement levels, measured by comment-to-view, like-to-view, and share-to-view ratios. We find that the toolkits-only intervention increased foundational mental health knowledge construction by 5.56% (P = 0.024) and 6.46% (P = 0.014) among videos with high engagement in comment (comment-to-view ratio > median of 0.97%) and like (like-to-view ratio > median of 0.02%), respectively. For videos in the toolkits-only group with high comment engagement, foundational mental health knowledge construction increased by an estimated 2.35% (95% CI: -0.66 to 5.36%; Fig. 2 , Panel B), from 32.45–34.79% (Table A5). Among the toolkits-only group, videos with high like engagement, the estimated increase in foundational mental health knowledge construction was 2.91% (95% CI: -0.80 to 6.63%; Fig. 2 , Panel C), rising from 31.65–34.56% (Table A5). Videos with higher commenting and liking engagement appear to elicit more user comments, resulting in a higher level of revealed foundational knowledge construction captured by our analysis. Interestingly, we find that the toolkits-only intervention increased the probability of foundational mental health knowledge construction by 7.86% (P = 0.013) in comments on videos with lower, but not higher, share-to-view ratios. Videos in the toolkits-only group with high sharing engagement showed an estimated 2.68 percentage point (95% CI: 0.30 to 5.07%; Fig. 2 , Panel D) increase in foundational knowledge construction, from 37.17–39.86% (Table A5). Videos with lower share-to-view ratios may resonate more intimately with users' personal mental health experiences, making them highly engaging but less shareable due to privacy and sensitivity. Consequently, these videos might more effectively promote foundational mental health knowledge construction among viewers. To better direct creator-training resources, we further assessed the heterogeneous effects of creator toolkits on user foundational mental health knowledge construction across content creators with different characteristics, including their engagement rate, popularity (number of followers on TikTok), and whether the creator is a licensed mental health professional. The creator engagement rate was calculated as the ratio of total comments and likes to the number of views among videos published between March and May 2023. We found that providing evidence-based mental health toolkits increased foundational knowledge construction by 10.03% (P = 0.005) among creators with engagement rates over the median of 1.20%, compared to the control group (Table 2 ). Among the toolkits-only group creators with high engagement, the estimated increase in foundational knowledge construction was 2.01% (95% CI: -0.77 to 4.79%; Fig. 2 , Panel E), rising from 29.63–31.65% (Table A5). No significant effect was observed among creators with lower engagement rates. Furthermore, we found that the effects of evidence-based mental health toolkits on foundational mental health knowledge construction are unaffected by creators’ follower size or professional licensing status. Following Brambor et al. (2006) 34 , we do not estimate marginal effects or predicted values for these non-significant moderator variables, as the results may be misleading. Institutions aiming to engage creators for health promotion should consider focusing training resources on those most likely to benefit, such as creators with higher engagement rates (above 1.20%) as identified in our study. Study 2 – Off-platform, survey-based RTC Effect of emotional support training on emotional support competency In the second survey-based RCT, we examine whether exposure to videos developed with evidence-based mental health communication training, compared to standard creator content, affects emotional support competencies among a nationally representative sample of U.S. youth aged 14–22. Manipulation check We begin our analysis by first determining whether respondents viewed the study’s treatment video as comparatively more informative than the control. We have reason to believe (see: Figure A3) that the treatment video was more likely to feature evidence-based content than the control. However, respondents must then recognize these differences for our treatments to have the hypothesized effects that we pre-registered. Table 3 tests this possibility by presenting the predicted percentage of respondents assigned to view the control (column 2) vs. treatment (column 3) videos who indicated that they found the video they watched to be helpful, informative, and/or any of the other qualities listed in column 1. The difference between the treatment vs. control group (∆ T-C) is listed in column 4, as well as a corresponding significance test of whether these differences are significantly different from zero (p) in column 5. If our theoretical expectations are supported, we should expect to see large and positive differences in column 4, which attain statistical significance at the p < 0.05 level (two-tailed) in column 5. All predictions are derived from logistic regression models regressing each indicator listed in column 1 on dichotomous indicators of treatment group assignment, and an indicator of whether respondents were recruited into the study’s adult (aged 18–22) or teen (aged 14–17) samples (see Methods). Note that items presented in the final three rows (denoted by “R”) are reverse coded, for consistency with the items presented in the first seven rows. The results strongly suggest that respondents viewed the treatment video as significantly more helpful, informative, evidence-based, interesting, and engaging than the control. At times, these differences were quite large. For example, the likelihood that treatment group respondents rated the video they viewed as informative was more than 16 percentage points greater than those in the study’s control group. Respondents were also significantly less likely to report that the treatment video was unclear, or boring. Only two estimates (trustworthiness and inaccuracy) failed to attain statistical significance between the two groups. Thus, as we intended, Table 3 provides strong evidence that respondents drew meaningful distinctions between the two videos to which they could have been exposed. Table 3 Manipulation check summary % of Respondents who Found the Video to Be… Control Treatment ∆ (T-C) p Helpful 58.96 75.43 16.47 < 0.001 Informative 60.22 76.97 16.75 < 0.001 Evidence-Based 32.13 37.94 5.81 0.054 Trustworthy 54.48 59.20 4.72 0.132 Interesting 58.75 65.29 6.54 0.033 Engaging 52.41 60.59 8.18 0.009 Unclear (R) 62.47 71.90 9.43 < 0.001 Boring (R) 45.62 56.13 10.51 < 0.001 Inaccurate (R) 58.39 60.56 2.17 0.483 Note: Predicted probabilities are derived from logistic regression models. Predictions hold all covariates at their observed sample means. Full model output is available on Table A7. Perceived emotional support competencies Next, we test our pre-registered expectation (H1) that exposure to the study’s treatment videos will be associated with elevated confidence in one’s ability to provide emotional support to one’s peers. This includes respondents’ perceived abilities to (a) provide emotional support to one’s peers, and (b) employ the evidence-based pillars of the A.S.K. method (which we described in full in the survey question; please see Table A8 for additional information.) Figure 3 plots the predicted probability that respondents select each of the response options to the questions listed in Table A8 Row 2 (see: Perceived Emotional Support Capabilities) across exposure to the treatment (in red) vs. control (in blue) videos. 95% confidence intervals extend out from each estimate, with predicted values printed next to each one, for reference. Predictions are derived from ordered logistic regression models that control for differences in sample selection (Table A9), and again hold all covariates at their sample means. The results provide support for our theoretical expectations. In both cases, exposure to the study’s treatment video is positively and significantly associated with increased confidence (A.S.K. Method: β = 0.24, p = 0.05; Emotional Support: β = 0.34, p < 0.01). However, predicted differences between groups are relatively modest, and only approach conventional levels of two-tailed significance (as designated by non-overlapping 90% confidence intervals) in one case: whether respondents feel “very prepared” to provide emotional support. Overall, the results suggest that exposure to the study’s treatment video is associated with increased confidence in respondents’ emotional support capabilities; although we caution that the substantive and statistical magnitude of these effects is relatively modest. Objective emotional support competencies. We next test whether respondents are more likely to take actions that reflect evidence-based best practices for providing emotional support in the study’s treatment condition, relative to the control. We measure respondents’ objective competencies by asking them to report how they would act in a series of hypothetical scenarios; described in detail below (see: Methods). The results are displayed in Table 4 . If our pre-registered theoretical predictions (H2) are correct, we should expect to see positive (column 4, denoted by “∆(T-C)”) and statistically significant (column 5, denoted by “p”) differences between the treatment and control in the proportion of respondents providing a correct answer to the hypothetical scenarios testing each of the different emotional support competencies listed in Table A8. Results are again derived from logistic regression models that regress dichotomous indicators of whether respondents provided a correct answer to each competency question (bolded entries in Table A8) on treatment group assignment and a sample selection indicator. Table 4 The effect of treatment video assignment on objective emotional support competency % Answering Question Correctly Control Treatment ∆ (T-C) p Competency: Active Listening #1 28.11 43.71 15.60 < 0.001 Competency: Validation #1 33.05 47.40 14.35 < 0.001 Competency: Emotional Skill #1 20.97 38.39 17.42 < 0.001 Competency: Emotional Skill #2 68.27 68.32 0.05 0.987 Competency: Validation #2 30.18 45.83 15.65 < 0.001 Competency: Active Listening #3 41.45 48.19 6.74 0.032 Note. Predicted probabilities are derived from logistic regression models. Predictions hold all covariates at their observed sample means. Full model output is available on Table A10. Moderation tests of results in Table A10 are presented on Table A12. The results strongly suggest that exposure to the study’s treatment video is positively and significantly associated with increased competencies in every substantive area that we tested. In all but one case—Emotional Skill Question #2 (which most respondents answered correctly irrespective of experimental assignment)—treatment group respondents were considerably more likely to demonstrate the competencies tested in our hypothetical vignettes. We caution, however, that overall levels of competency in each of the areas we tested were relatively low; even though they are significantly higher in the treatment group. Aside from the second Emotional Skill question, most respondents answered each question incorrectly. Further demonstrative of this tension, we constructed a negative binomial regression model (Table A10) regressing a count of correctly answered items on treatment video exposure and sample selection indicators. Results suggest that treatment group respondents were predicted to answer 2.92 questions correctly, versus just 2.22 in the control (β = 0.27, p < 0.01). While treatment video exposure significantly elevated the number of correct answers we expect to observe, we nevertheless note that overall performance on the hypothetical competency vignettes is relatively low. Discussion The current study underscores the potential of mental health content creators as influential community leaders for promoting mental health in online social networks. The network-based "influencing the influencers" strategy 5 , providing training to creators to enhance mental health-related behaviors among social media users, has demonstrated effectiveness in both an on-platform RCT (providing evidence-based mental health communication training to creators improved TikTok commenters’ foundational mental health knowledge construction) and an off-platform survey experiment (providing emotional support training to creators improved young viewer’s perceived and objective emotional support competencies). To our knowledge, this is the first study to assess the impact of large-scale content creator training on not just the content they produce, but social media user outcomes . Our study advances the previous findings–that creator training programs boost evidence-based content creation 5 –by showing that this enhancement also translates to changes in audience attitudes and behaviors. The benefits of the creator training program are significant. In the on-platform RCT, based on our March to May 2023 creator training program, an estimated 4.15% increase in foundational mental health knowledge construction in comments from over 2 million views accumulated by the 42 creators in the treatment groups could potentially enhance informal learning for up to 83,000 viewers, though this is an upper estimate since individuals may account for multiple views. In the off-platform survey experiment, exposure to TikTok-style videos created with support of the training increase young people’s perceived and objective emotional support competencies. These effects are similar for both regular and infrequent social media users, thereby implying that our approach is unlikely to be limited in its effectiveness to just those most familiar with social-video driven platforms like TikTok or Instagram. Our study also combines the external validity strengths of an on-platform RCT with the internal validity of a survey-based experiment, jointly supporting the robustness of the identified effects of creator training toolkits on audience mental health behavioral outcomes. In the on-platform RCT, embedding video interventions within real-life social media interactions supports the idea that exposure to evidence-based content is associated with attitude change in naturalistic settings. In the survey-based experiment, partnering with a lifestyle influencer to produce both treatment and control videos enhances internal validity, as key message features, such as the messenger, style, diction, and camera work, are held constant across conditions, and allows us to directly manipulate exposure to evidence-based (vs. control) videos. Thus, in the off-platform survey experiment, the effect is assessed across all video viewers, not limited to those who left comments. In the on-platform RCT, the finding that only foundational knowledge construction improved, while advanced stages did not, aligns with the nature of TikTok videos: they are typically short (less than 60 seconds) and involve only one round of interaction between creators and viewers. Advanced knowledge construction may require repeated exposure to more iterative discussions. Future studies should explore these repeated exposures. Similar to the on-platform RCT, the off-platform survey experiment asked participants to watch a single short video (< 60 seconds), and we examined how one video influenced user outcomes. An important direction for future research is to investigate the effects of repeated exposure and reinforcement, as continued engagement with a creator may lead to stronger effects. In addition to identifying mental health content creators as influential online-community leaders in online social networks, our study proposes delivering easy-to-navigate toolkits directly to them as an effective way to intervene. In the on-platform RCT, we found that creators who only received training toolkits produced videos that were more effective at improving their audiences’ mental health knowledge construction, compared to their peers who received the toolkits and also attended online training. This may be because the online training sessions, a series of hourlong health professional-led briefings and networking opportunities, left creators with less time to make content using the toolkits during the study period. More importantly, delivering easy-to-navigate toolkits to content creators is not only effective but also scalable, enabling engagement with a broader range of mental and general health creators across platforms in evidence-based health communication. We believe this is the case for two reasons. First, our results suggest that creator training need not be resource intensive, as we found that the development and delivery of easy-to-scale toolkits tailored to creators' needs can meaningfully impact users’ attitudes and behaviors. Consequently, health organizations could create and distribute these toolkits exponentially to more mental health creators and to creators across different health domains. Second, many creators we've worked with produce content on multiple digital platforms, including TikTok, Instagram, and YouTube 20 . Some of these platforms have adopted algorithmically personalized, short-form video feeds in recent years, such as Instagram Reels and YouTube Shorts 35 – 38 . Given these similarities, we believe that the effects identified in the current study on TikTok could be generalized to other platforms featuring algorithmically personalized and short-form video content, particularly Instagram and YouTube. The current study also points to the efficacy of health professionals engaging with content creators to promote public health messages. Public health practitioners have utilized social media for health promotion, with organizations disseminating information via official social media accounts 39 – 41 and researchers initiating social media campaigns on health topics such as vaccination or cancer screening 42 , 43 . One challenge of the existing health campaigns on social media is that they sometimes fail to fully leverage the platforms’ interactive potential, often focusing more on one-way dissemination than on actively engaging the audience 40 , 44 . A solution is to collaborate with social media content creators, who possess skills in producing appealing content, understanding their audiences’ needs, and building trust and community 45 , 46 . A recent survey of creators reveals that they are most likely to use personal experience for content creation, highlighting an urgent need to enhance their media and information literacy skills, including the ability to identify and use reliable fact-checking resources 47 . The domain knowledge that health institutions can provide through creator training programs fills this content creation gap. Health institutions, platforms, and other organizations that wish to enhance online health communication may want to make such training materials available to more creators. Limitations of the current study Our study has several limitations. In the on-platform RCT, our analysis of mental health knowledge construction in user comments represents just one of the many ways that creator-engaged programs can impact users. Future research could explore other outcomes, such as increased social support or reduced self-diagnosis of mental health problems 48 . Moreover, in the on-platform RCT, our study did not examine if the impact of the training conducted in April 2023 would be sustained over the long term. Future research should aim to understand the longer-term effects of creator programs. Furthermore, our on-platform RCT was solely focused on TikTok. As such, we cannot empirically generalize our findings to other platforms. However, as discussed above, we hypothesize that similar findings could be observed on other platforms featuring personalized short-form videos like Instagram or YouTube. In the on-platform RCT, the finding that only foundational knowledge construction improved, while advanced stages did not, aligns with the nature of TikTok videos: they are typically short (less than 60 seconds) and involve only one round of interaction between creators and viewers. Advanced knowledge construction may require repeated exposure to more iterative discussions. Future studies should explore the impact of repeated exposures to evidence-based mental health content on user behaviors. Similar to the on-platform RCT, the off-platform survey experiment asked participants to watch a single short video (< 60 seconds), and we examined how one video influenced user outcomes. Whether treatment effects are enhanced in situations where social media users are exposed to messages reinforcing similar themes (“dosage effects”) or diluted in situations where they are exposed to competing messages (“competitive framing environments”) are also outside the purview of our study and remain a fruitful avenue for research. An important direction for future research is to investigate the benefits of sustained and reinforced training for content creators, as well as the cumulative effects of repeated exposure on audiences. Besides, in the off-platform survey experiment, we study only a single (albeit important) aspect of mental health attitudes and behavior in the present research; emotional support competencies. Whether findings generated from this research generalize to other settings, such as communication about suicide, eating disorders, and other mental health topics, remains an important and potentially fruitful avenue for future research. Additionally, in the off-platform survey experiment, we partnered with just a single creator to measure the effects of evidence-based mental health communication. Our decision to partner with a lifestyle influencer was strategic; as lifestyle influencers often cultivate large followings that may exceed those of influencers focusing specifically on mental health, likely have less experience with emotional support training than creators who are mental health professionals, and have been shown to shape opinion in a wide range of social and health-related domains 49 , 50 . We encourage researchers in the future to replicate studies like this one to study whether other types of creators might be similarly powerful in motivating attitudinal and behavioral change. Conclusion This study demonstrates the promise of engaging mental health content creators as influential community leaders for advancing mental health in online social networks. By leveraging a network-based strategy of "influencing the influencers," we show that providing creators with training can meaningfully improve mental health behavioral outcomes in audiences. Evidence from both an on-platform RCT, analyzing 188,169 comments from 1,882 videos created by 49 mental health content creators on TikTok, and an off-platform survey experiment conducted with a large, demographically representative sample of 1,000 U.S. youth aged 14–22 supports the effectiveness of this approach. In the on-platform RCT, evidence-based mental health communication toolkits lead to an over 4% increase in foundational mental health knowledge construction among TikTok commenters, while in the off-platform survey experiment, training creators improved both perceived and objectively assessed emotional support competency among video viewers. Taken together, these findings highlight the potential of creator-focused interventions to promote positive mental health behavioral outcomes on video-based social media platforms. They also suggest that relatively simple, scalable resources, such as training toolkits, can be a powerful means of achieving public health impact through social media. Method This study evaluates the effectiveness of creator-focused interventions on audiences through an on-platform RCT and an off-platform, survey-based RCT. On-platform RCT From March to May 2023, the Harvard Chan School’s Center for Health Communication conducted an RCT testing the effects of training toolkits and training programs for identified mental health content creators (MHCCs) (see Method-Creator Training Program). After the RCT, we collected TikTok videos created by MHCCs and their associated user comments (see Method-Comment Data Collection). Based on the experimental assignment protocol embedded in our previous RCT research 5 , we requested all comments (n = 188,169) from all videos (n = 1,882) created by the 49 mental health content creators enrolled in the RCT in March, April, and May 2023 using TikTok’s Research Tools API 29 . We developed a codebook to measure the study outcome, mental health knowledge construction, through content analysis (see Method – Outcome Measurement). We randomly sampled 54 TikTok videos, yielding 4,152 comments. Three research assistants coded the comments, achieving an average percent agreement of 0.81 and a Gwet’s AC of 0.80. To ensure a balanced number of comments with and without knowledge construction for LLM fine-tuning, we applied text augmentation that maintained a cosine similarity of 0.85 between the augmented and original texts (see Method – Text Augmentation). We then fine-tuned LLMs to classify foundational and advanced mental health knowledge construction, achieving over 90% accuracy and F1 scores for both (see Method – LLM Fine-Tuning). These models were then used to assess mental health knowledge construction across all comments in the dataset. Finally, we fitted statistical models to evaluate the effects of the creator training toolkits and programs on user knowledge construction (see Method – Analytical Strategy). The fine-tuned LLMs are available on HuggingFace: https://huggingface.co/chc-harvard . The code and data for replicating the on-platform RCT analysis are available on The Open Science Framework at: https://osf.io/vbkh8/ . The Harvard Longwood Campus Institutional Review Board determined that this study is not research as defined by DHHS regulations or FDA regulations. All procedures complied with applicable guidelines and regulations. Creator training program Our team built on a previous intervention by Harvard Center for Health Communication that identified a sampling frame of MHCCs and conducted a randomized control trial. The inclusion criteria of the MHCCs are: aged 18 or over, English-language mental health content, have at least 10,000 followers across TikTok or Instagram social media platforms, posted videos on the platform at least 4 times per month from December 2022 - February 2023, and have been active on the site since February 2022. MHCCs were randomly selected to participate in interventions held in April 2023, including a series of virtual summits and an online content creation toolkit around evidence-based mental health communication. Details of the randomization process and intervention are documented in our previous work 5 . The trial comprised three groups: one provided with an asynchronous online content creation toolkit (“Toolkits-Only” Condition, N = 17 creators), another receiving the toolkit plus a synchronous virtual summit component (“Sessions Plus Toolkits” Condition, N = 25 creators), and a control group (N = 20 creators). Comment data collection To evaluate the intervention's impact on user behavior, we obtained all videos and related comments from 62 participating MHCCs via the TikTok research API 29 in August 2023, using Python version 3.9. From this group, we successfully retrieved 3,465 videos posted by 58 MHCCs in March, April, and May 2023. Additionally, we collected user information, such as the number of followers, for these MHCCs through the same API. Four creators were excluded from the analysis because they either did not produce videos during the study period or had deleted their videos at the time of the request. Videos without comments were also excluded from the study. We accessed comments from 2,858 videos using the TikTok research API. Note that the TikTok research API imposes data filters, excluding, for instance, “public data from users under 18 and data from Canada.” For videos not captured by the API, we manually collected comments, ultimately compiling comments from 2,901 videos. To align with our research focus on online mental health communication, we applied a filter to exclude videos not relevant to mental health discussions. This filter was based on the content analysis by research assistants in our previous study that determined whether each video is relevant to mental health communication 5 . Applying the germane-to-mental health filter resulted in a final sample of 1,882 videos and 188,169 comments from 49 MHCCs (toolkits-only: N = 14 creators, sessions + toolkits: N = 20 creators, control: N = 15 creators). The flow chart in Figure A1 presents the process of video and comments collection. Outcome measurement We hypothesized that creator training programs enhance mental health knowledge construction in video comments. To measure knowledge construction, we utilized the Interaction Analysis Model (IAM) 51 , 52 , a framework proven effective in exploring informal learning in social media interactions 28 , 53 . IAM outlines learning progresses as users exchange ideas, confront dissonance, negotiate meanings, evaluate emerging understandings, and apply new knowledge to novel contexts, with each stage growing in complexity. Based on IAM and prior research, we crafted a measure for knowledge construction in TikTok video comments for mental health communication across six dimensions: personal reflection, expressing agreement, expressing disagreement, asking for clarifications, reinterpreting knowledge, and applying knowledge to new areas, with each dimension quantified as a binary variable. From this, we derived two binary measures: foundational mental health knowledge construction, coded as 1 if the comment expressing personal experiences or opinions, and advanced mental health knowledge construction, coded as 1 if the comment contains any expression of agreement or disagreement, a clarification question, reinterpretation of knowledge, or application to other contexts. Definitions and examples for these measures are detailed in Table A2. Content analysis We conducted a content analysis of comments from a sample of TikTok videos to measure mental health knowledge construction. Given the hierarchical nature of the creator-video-comment data, we employed clustered randomized sampling to select comments for content analysis, resulting in 54 videos and 4,152 comments. Three research assistants from Harvard University were trained and served as coders for the project. After an initial code training session, which contained a series of iterative pilot coding and feedback sessions, each assistant was assigned to code approximately 9% of all sampled comments (376 comments from 5 randomly selected videos). In the content analysis, each assistant first watched the TikTok video, then started coding the comments. The RAs were not informed about whether the videos and comments are from creators in the treatment or control group in the experimental study. Inter-coder reliability (ICR) was assessed for these comments. The ICR results on the “triple assigned” comments highlighted just three outcome variables with percent agreement scores below 0.70: knowledge construction through personal reflection, expressing agreement, and knowledge reinterpretation. Based on the ICR results, a second training session was held to focus on improving coder agreement in the three outcomes. Coders subsequently re-coded a subset of 120 comments where discrepancies had been noted for the three outcomes. Following the two-round training sessions, ICR exceeded 80% agreement for all variables in the filtering session, and surpassed 70% for most knowledge construction variables—except for expressing agreement and knowledge reinterpretation, which both achieved a 68% agreement rate (Table A3). In general, coders achieved an average percent agreement of 0.81, with Gwet's AC of 0.80 (Table A3) across a random sample of 376 double-assigned comments from the total 4,152 comments. Upon completing the two training sessions, coders formally started the coding of video comments, with each coder handling approximately one-third of the selected comments from the training set. Text augmentation Using the true labels derived from this content analysis, we initially performed character, word-level, and LLM-based text augmentation to ensure a sufficiently large and balanced sample for model fine-tuning. We employed the textattack library to conduct character and word-level text augmentation 54 . Specifically, at the character level, two substituted sentences were generated using transformations including WordSwapRandomCharacterDeletion, WordSwapRandomCharacterInsertion, WordSwapRandomCharacterSubstitution. At the word level, five augmented sentences were generated using transformations including WordInsertionRandomSynonym, WordSwapChangeLocation, WordSwapChangeName, WordSwapChangeNumber, WordInnerSwapRandom, WordSwapEmbedding, and WordSwapQWERTY. Both sets of transformations were subject to two constraints: Repeat Modification and Stopword Modification. For LLM-based text augmentation, five augmented sentences were generated by applying the model “gpt-4-turbo” in the Chat Completions API from OpenAI with temperature set as 0 55 . The prompts we used are presented in Appendix Section 1. All text augmentation analyses were conducted in Python 3.9. We apply cosine similarity to evaluate the performance of the augmented text. The original and augmented texts were embedded using the sentence_transformer library with the "bert-base-nli-mean-tokens" model, and the cosine similarity between the embeddings of the original and augmented texts was calculated. Summary statistics for the cosine similarity are presented in Table A6. The overall mean and median cosine similarity between the original and all augmented texts were 0.85 and 0.90, respectively (SD = 0.16). Specifically, the mean and median cosine similarity between the original text and character-level augmented texts were 0.88 and 0.91, respectively (SD = 0.10); for word-level augmented texts, the mean and median were 0.92 and 0.94 (SD = 0.08); and for LLM-based augmented texts, they were 0.76 and 0.83 (SD = 0.20). Overall, the results indicate that the augmented texts closely resemble the original texts, ensuring that the texts used in the downstream fine-tuning of the LLM are representative of the original content. LLM fine-tuning Subsequently, we fine-tuned LLMs for text classification tasks to measure mental health knowledge construction outcomes. In total, we fine-tuned 8 LLMs to facilitate text classification, with all outcomes treated as binary classification problems. These included 2 outcomes related to filtering questions (i.e., whether a comment only tagged another user or contained only emojis) and 6 outcomes related to knowledge construction (see: "Method-Outcome Measurement"). Fine-tuning was performed on the un-augmented text for the two filtering outcomes and on the augmented text for the remaining 6 outcomes. We ensured a balanced class distribution to avoid issues arising from unbalanced samples when using the augmented text by maintaining a roughly equal ratio of comments from both classes. For each outcome, we fine-tuned the following models: BERT (bert-base-uncased) 56 , RoBERTa (roberta-base) 57 , MentalBERT (mental/mental-bert-base-uncased) 58 , and MentalRoBERTa (mental/mental-roberta-base) 58 . The models were trained with a batch size of 32, for 20 epochs, and a learning rate of 2e-5, following the recommendations from the BERT developers 56 . Among the models, the BERT-based LLM consistently achieved the highest accuracy for each outcome, leading us to use the fine-tuned BERT models for downstream analyses and predictions. The performance of the BERT-based LLMs for each of the 8 outcomes, as well as the specific epoch at which each model achieved its highest accuracy, is detailed in Table A1. Finally, we applied the fine-tuned LLMs to label all comments from the content creators included in the creator training program. We conducted text preprocessing by removing comments that consist solely of a single word (ie, ‘fine’), only emojis (ie, 😎), or only tagging users (ie, @userid_happypuppy), following the approach in previous studies 59 , 60 , as these comments provide no relevant information on our outcomes of interest. After the preprocessing, a total of 23,314 comments (12.39%) were excluded, resulting in 164,855 comments from 1,863 videos created by the 49 creators remaining in the following analysis, as shown in Figure A1. Analytical strategy For the toolkits-only condition, the pre period includes videos published from March 1, 2023, to before 9:30 a.m. ET on May 2, 2023. The post period includes videos published after this time through May 31, 2023. This cutoff reflects the time the toolkits were delivered to MHCCs via email. For the sessions + toolkits condition, the intervention, which consisted of online training sessions, was delivered at 10 a.m. ET on April 8, 2023. Videos published before this time are coded as pre, and those after as post. We aimed to assess whether comments on videos produced by MHCCs across the toolkits-only, sessions + toolkits, and control groups were more likely to demonstrate mental health knowledge construction. To achieve this, we employed three-level multilevel Linear Probability Models (LPM) with random effects at the influencer and video levels to account for potential systematic video-level differences in commenting behavior and/or creators’ differential responsiveness to our experimental treatments. The analysis included 164,855 comments nested within 1,863 videos created by 49 influencers. These models (Table 1 ) assessed the impact of our study interventions on the outcomes (foundational and advanced mental health knowledge construction) by interacting dichotomous fixed effect indicators of treatment group assignment (toolkits-only vs. sessions + toolkits, with the control group as the reference) with the fixed effect indicator of whether each video was produced pre- or post-intervention. The estimated prevalence of knowledge construction outcomes in treatment and control groups at the pre and post period are presented in Table A4. To examine how the toolkit and online training influenced user comments on evidence-based mental health content, we included four video-level moderators: (1) whether the video featured themes from the training toolkits (26.95% yes, 73.05% no, based on coding in our prior research 5 ), (2) comment engagement rate (comments per view, dichotomized at the median of 0.97%), (3) like engagement rate (likes per view, dichotomized at the median of 0.02%), and (4) share engagement rate (shares per view, dichotomized at the median of 0.18%). We also included three creator-level moderators: (1) number of followers (38.78% with more than 500,000; 61.22% with fewer), (2) engagement rate, calculated as (comments + likes) per view from March to May 2023 (dichotomized at the median of 1.20%), and (3) whether the creator is a licensed mental health professional (63.27% yes, 36.73% no). Note that a creator's number of followers, indicating content reach and popularity, does not directly correspond to engagement rate, which reflects audience involvement on online platforms 61 . We dichotomized the number of followers at a cutoff of 500k, a threshold at which creators typically begin to employ managers for account and content management. Furthermore, the mental health professional licensing status of each content creator was collected based on their self disclosure on their social media accounts or websites. We conducted subgroup analyses based on the four video-level and three creator-level indicators to examine whether the effect of the toolkits in the toolkits-only group was stronger in any subgroup. These analyses used the same three-level multilevel linear probability model. The parameter estimation of the moderated multilevel LPMs are shown in Table 2 . The estimated prevalence of knowledge construction outcomes in treatment and control groups at the pre and post period in different subgroups are presented in Table A5. The analyses were conducted using the lme4 package in R version 4.2.1 62 . We also estimated the average marginal effect (AME) of interventions in outcomes among different subgroups of videos and creators. Visualized in Fig. 2 . We presented the distribution of post–pre changes in the outcome variable—the proportion of comments exhibiting foundational mental health knowledge construction—across four groups, defined by intervention condition (toolkits-only vs. control) and each subgroup. AME was estimated by the margins package in R 63 . Off-platform, survey-based RCT Pre-registered Study Design: Assessing the Effects of EBMHC via a “Creator-Engaged” Survey-Based RCT Experimental protocol Our pre-registered survey-based RCT randomly exposed respondents to one of two video messages about how to emotionally support friends. Figure A3 presents a side-by-side comparison of the text featured in the control—developed prior to the administration of interventional training materials (described below)—and treatment videos—developed after the provision of those materials—assembled by a popular social media lifestyle influencer enlisted as a confidential collaborator in our study. Our “creator-engaged” experimental procedures are described in detail below. Registration materials are available at the following Open Science Framework page: https://osf.io/jsg9t . In the first video, which serves as the study’s control, lifestyle content creator Kenzie Brenna, who has a cumulative following across TikTok and Instagram of 430,000, talks frankly and direct-to-camera about how she goes about supporting friends who are “going through something.” The second treatment video resembles the first, with the exception that it has been informed by our training materials and makes much more of an effort to feature evidence-based mental health content therein. For example, Brenna calls out the importance of both validating the friend’s feelings and asking them open-ended questions—critical evidence-backed elements of high-quality emotional support 31 . We created these messages in partnership with Brenna. The hundreds of millions of mental health videos that a TikTok user might encounter are made by a diverse array of creators. Some are lifestyle creators like Brenna who also aim to reduce stigma by sharing their personal mental health experiences; others are licensed mental health providers like Shahem McLaurin (@5hahem) who aim to use their professional expertise to democratize access to high-quality mental health information. We hypothesized that our training might have the biggest impact when offered to creators who satisfy all three of the following criteria: 1) make the type of lifestyle content that TikTok and Instagram users are most likely to come across on these platforms 2) have not trained as a therapist or other licensed mental health provider 3) have demonstrated in past content a consistent interest in educating and serving their community. As a consequence, we chose to expose participants in our survey-based RCT to messages made in partnership with Brenna. Note that all recruitment messages are available as supplementary materials alongside this manuscript. After agreeing to assist in our study, we asked Brenna to develop a video about “emotionally supporting friends” using conventional media production strategies and content elements that they otherwise would use when creating for and posting on the TikTok platform (which hosts videos of a similar style and length to other social networking sites, like Instagram). This video served as the study’s control condition. After doing this, we provided Brenna with a series of materials we co-created with the nonprofit Active Minds. The materials were designed to help creators like Brenna spread evidence-based information on (1) how today’s youth mental health crisis has put more pressure on kids to provide emotional support to friends 2) why emotional support makes a difference and 3) the 3 key characteristics of high-quality emotional support: Acknowledge the person’s feelings. Support their feelings by validating their emotions and asking what they need. Keep in touch, and check in regularly. Upon completion of the training sessions, we asked Brenna to make a second video using the evidence-based mental health information they learned from the training materials, and comport with best practices for effective mental health communication, while sticking as closely as possible to the video they created originally. Storyboards for both videos are presented below. Stable links to access both videos can be accessed at the following webpages: for the control: https://drive.google.com/file/d/1lSYaP8kmCpdUi97bXMMxS0eQ7L8hTsXF/view?usp=sharing for the treatment https://drive.google.com/file/d/1PfI6raNz1eluO6IZcK-uDIntQfVs0syw/view?usp=sharing Analytical strategy We assess the effectiveness of our experimental treatment by constructing multivariate regression models that regress dichotomous indicators (see: Method-Measures) of perceived helpfulness, informativeness, and the evidentiary basis of the videos respondents watched (Manipulation Check); measures of both perceived (H1) and objective (H2) emotional support competencies, and toward mental health content created on TikTok (RQ1), on a dichotomous indicator of experimental treatment assignment, as well as an indicator of how respondents were recruited to participate in the survey (see: Method-Data). Additional estimation information can be found throughout the results and Supplementary Materials. Note that supplemental randomization checks (Table A11) demonstrate that assignment to our experimental treatments was well-balanced (i.e., we document no significant differences across treatment and control groups). Correspondingly, we do not account for any additional socio-demographic controls in our analysis. Finally, we assess the possibility that more-regular social media users may be comparatively more (or less) receptive to our video treatments (RQ2) by amending these models to interact the dichotomous treatment assignment indicator with an ordinal measure of social media use frequency when assessing RQ2. Please refer to Table A8 for additional information about these questions. Data Data for this study were derived from a nationally representative survey of N = 1,000 American youth aged 14–22, administered via YouGov. YouGov worked with an external data provider to recruit a large online opt-in sampling frame of teens (“teen sample:” N = 500, aged 14–17), and relied on its own online opt-in panel to recruit college aged adults (“young adult sample:” N = 500, aged 18–22) to participate in this study. Note that teen participants were recruited via their parents, and that both teens and parents provided written consent in order to participate in this study. Institutional Review approval for this study was granted by Harvard University’s Longwood Campus Office of Regulatory Affairs & Research Compliance. YouGov ensures national representativeness in the samples it draws by using propensity score matching techniques. To do this, YouGov first drew a simple random sample of teens and young adults from US Census data. This serves as a nationally representative cross-section of each subpopulation of interest. YouGov then used its proprietary propensity score matching algorithm to search for analogues from each online opt-in sampling frame that most closely match respondents drawn from the US Census on a wide range of demographic observables: including respondents’ age, racial identity, and gender identity. Additional demographic information about the composition of both the adult and teen samples can be found in the Supplementary Materials. Table A13 presents the demographic composition of the teen and adult samples. Measures Table A8 summarizes the primary outcome and moderating variables used to test our theoretical expectations (outlined above). Column 1 provides information about the general concepts we hope to capture via the questions outlined in Column 3, with Column 2 serving as a reference regarding each measure’s role in our analysis. Declarations Acknowledgements We wish to thank Q Garcia, Amy Gatto, and Rita DeBateat Active Minds for their help in developing the training materials and survey questions used in the survey-based RCT. The survey-based RCT was supported by a gift to the Center for Health Communication from Showtime/MTV Entertainment Studios. The funders had no role in study design, data collection, data analysis, data interpretation, decision to publish, or preparation of the manuscript. Author contributions A.Y., M.M., and Y.L. contributed to conceptualization and study design; Y.L., M.M., A.Y., and E.E. contributed to data collection in the on-platform RCT; A.Y., M.M., K.B., S.M., K.S., and E.E. contributed to data collection in the off-platform survey experiment; Y.L. and M.M. contributed to data analysis and visualization; Y.L. and M.M. drafted manuscript; A.Y., M.M., and Y.L. interpreted the results and contribute to the theoretical framing; all authors conducted critical reviews. All authors have read and approved the final version of the manuscript. Competing interests The authors declare no competing interests, but provide the following information in the interests of transparency and full disclosure. AY is an unpaid advisor to the nonprofit Science To People. AY and the Center for Health Communication have received grant funding for other ongoing research projects from YouTube Health and the WellWithAll Foundation. Materials & Correspondence. Correspondence and material requests should be addressed with Yuning Liu. References Korda, H. & Itani, Z. Harnessing Social Media for Health Promotion and Behavior Change. Health Promotion Practice 14 , 15–23 (2013). Stellefson, M., Paige, S. R., Chaney, B. H. & Chaney, J. D. Evolving Role of Social Media in Health Promotion: Updated Responsibilities for Health Education Specialists. International Journal of Environmental Research and Public Health 17 , 1153 (2020). Alex Montero, Grace Sparks, Marley Presiado, & Liz Hamel. KFF Health Misinformation Tracking Poll: Health and Election Issues on TikTok. (2024). Edelman Trust Institute. 2025 Edelman Trust Barometer: Special Report - Trust and Health. Edelman https://www.edelman.com/trust/2025/trust-barometer/special-report-health. Motta, M., Liu, Y. & Yarnell, A. “Influencing the influencers:” a field experimental approach to promoting effective mental health communication on TikTok. Sci Rep 14 , 5864 (2024). Valente, T. W. Network Interventions. Science 337 , 49–53 (2012). Latkin, C. A., Edwards, C., Davey-Rothwell, M. A. & Tobin, K. E. The relationship between social desirability bias and self-reports of health, substance use, and social network factors among urban substance users in Baltimore, Maryland. Addict Behav 73 , 133–136 (2017). Hunter, R. F. et al. Social network interventions for health behaviours and outcomes: A systematic review and meta-analysis. PLOS Medicine 16 , e1002890 (2019). Choi, B., Kim, H. & Huh-Yoo, J. Seeking Mental Health Support Among College Students in Video-Based Social Media: Content and Statistical Analysis of YouTube Videos. JMIR Formative Research 5 , e31944 (2021). Lee, N., Buchanan, K. & Yu, M. Each post matters: a content analysis of #mentalhealth images on Instagram. Journal of Visual Communication in Medicine 43 , 128–138 (2020). Wang, X., & Cohen, R. A. Health Information Technology Use Among Adults: United States, July–December 2022. (2023). Basch, C. H., Donelle, L., Fera, J. & Jaime, C. Deconstructing TikTok Videos on Mental Health: Cross-sectional, Descriptive Content Analysis. JMIR Formative Research 6 , e38340 (2022). #MentalHealth. TikTok https://www.tiktok.com/tag/mentalhealth. Ooi, K.-B. et al. Social media influencers: An effective marketing approach? Journal of Business Research 160 , 113773 (2023). Casaló, L. V., Flavián, C. & Ibáñez-Sánchez, S. Influencers on Instagram: Antecedents and consequences of opinion leadership. Journal of Business Research 117 , 510–519 (2020). Powell, J. & Pring, T. The impact of social media influencers on health outcomes: Systematic review. Social Science & Medicine 340 , 116472 (2024). Eddy, K. Republicans, young adults now nearly as likely to trust info from social media as from national news outlets. Pew Research Center https://www.pewresearch.org/short-reads/2024/10/16/republicans-young-adults-now-nearly-as-likely-to-trust-info-from-social-media-as-from-national-news-outlets/ (2024). YouTube. THE-IQ Creator Program. https://health.youtube/features-application/#:~:text=THE%2DIQ%20Creator%20Program%20supports,YouTube%20%E2%80%93%20with%20messages%20that%20resonate. USC. Mental Health Storytelling Initiative. https://annenberg.usc.edu/research/annenberg-inclusion-initiative/mental-health-storytelling-initiative (2024). Xu, M. M. et al. ‘Influencing the influencers’: Exploring the impact of an in-person summit on TikTokers’ mental health communication habits and beliefs. Preprint at https://doi.org/10.31235/osf.io/v83ht (2024). Janz, N. K. & Becker, M. H. The Health Belief Model: a decade later. Health Educ Q 11 , 1–47 (1984). Limbu, Y. B., Gautam, R. K. & Pham, L. The Health Belief Model Applied to COVID-19 Vaccine Hesitancy: A Systematic Review. Vaccines (Basel) 10 , 973 (2022). Nguyen, H. TikTok as Learning Analytics Data: Framing Climate Change and Data Practices. in LAK23: 13th International Learning Analytics and Knowledge Conference 33–43 (Association for Computing Machinery, New York, NY, USA, 2023). doi:10.1145/3576050.3576055. Beautemps, J. & Bresges, A. What Comprises a Successful Educational Science YouTube Video? A Five-Thousand User Survey on Viewing Behaviors and Self-Perceived Importance of Various Variables Controlled by Content Creators. Front. Commun. 5 , (2021). Zawacki, E. E., Bohon, W., Johnson, S. & Charlevoix, D. J. Exploring TikTok as a promising platform for geoscience communication. Geoscience Communication 5 , 363–380 (2022). Latha, K., Meena, K. S., Pravitha, M. R., Dasgupta, M. & Chaturvedi, S. K. Effective use of social media platforms for promotion of mental health awareness. Journal of Education and Health Promotion 9 , 124 (2020). Shahbaznezhad, H., Dolan, R. & Rashidirad, M. The Role of Social Media Content Format and Platform in Users’ Engagement Behavior. Journal of Interactive Marketing 53 , 47–65 (2021). Nguyen, H. & Diederich, M. Facilitating knowledge construction in informal learning: A study of TikTok scientific, educational videos. Computers & Education 205 , 104896 (2023). TikTok. Research API | TikTok for Developers. https://developers.tiktok.com/products/research-api/. Motta, M. et al. Assessing the Effect of Evidence-Based Mental Health TikTok Videos on Youth Emotional Support Competencies: A “Creator-Engaged” Approach. Preprint at https://doi.org/10.31235/osf.io/kydf4 (2024). High, A. C. & and Dillard, J. P. A Review and Meta-Analysis of Person-Centered Messages and Social Support Outcomes. Communication Studies 63 , 99–118 (2012). Cannava, K. & Bodie, G. D. Language use and style matching in supportive conversations between strangers and friends. Journal of Social and Personal Relationships 34 , 467–485 (2017). Priem, J. S. & Solomon, D. H. What Is Supportive About Supportive Conversation? Qualities of Interaction That Predict Emotional and Physiological Outcomes. Communication Research 45 , 443–473 (2018). Brambor, T., Clark, W. R. & Golder, M. Understanding Interaction Models: Improving Empirical Analyses. Political Analysis 14 , 63–82 (2006). Murariu, S. From Vine to TikTok: The Evolution of Short-Form Video and the Rise of a Social Media Juggernaut. Bootcamp https://medium.com/design-bootcamp/from-vine-to-tiktok-the-evolution-of-short-form-video-and-the-rise-of-a-social-media-juggernaut-98967a7d8d7e (2023). Gerbaudo, P. TikTok and the algorithmic transformation of social media publics: From social networks to social interest clusters. New Media & Society 14614448241304106 (2024) doi:10.1177/14614448241304106. Dodds, D. Council Post: Short-Form Video Content: Capturing Attention In The Digital Age. Forbes https://www.forbes.com/councils/forbesagencycouncil/2024/03/19/short-form-video-content-capturing-attention-in-the-digital-age/ (2024). Duan, Y. et al. Comparing Climate Change Content and Comments across Instagram Reels, TikTok, and YouTube Shorts and Long Videos. Proceedings of the Association for Information Science and Technology 61 , 103–114 (2024). Brownson, R. C., Eyler, A. A., Harris, J. K., Moore, J. B. & Tabak, R. G. Getting the Word Out: New Approaches for Disseminating Public Health Science. Journal of Public Health Management and Practice 24 , 102 (2018). Heldman, A. B., Schindelar, J. & Weaver, J. B. Social Media Engagement and Public Health Communication: Implications for Public Health Organizations Being Truly “Social”. Public Health Rev 35 , 1–18 (2013). Hunt, I. de V. & Linos, E. Social Media for Public Health: Framework for Social Media–Based Public Health Campaigns. Journal of Medical Internet Research 24 , e42179 (2022). Faus, M., Alonso, F., Javadinejad, A. & Useche, S. A. Are social networks effective in promoting healthy behaviors? A systematic review of evaluations of public health campaigns broadcast on Twitter. Front. Public Health 10 , (2022). Shi, J., Poorisat, T. & Salmon, C. T. The Use of Social Networking Sites (SNSs) in Health Communication Campaigns: Review and Recommendations. Health Communication 33 , 49–56 (2018). Herrera-Peco, I., Jiménez-Gómez, B., Peña Deudero, J. J., Benitez De Gracia, E. & Ruiz-Núñez, C. Healthcare Professionals’ Role in Social Media Public Health Campaigns: Analysis of Spanish Pro Vaccination Campaign on Twitter. Healthcare 9 , 662 (2021). Glotfelter, A. Algorithmic Circulation: How Content Creators Navigate the Effects of Algorithms on Their Work. Computers and Composition 54 , 102521 (2019). Si Willmore. Becoming a successful social media content creator. https://memberful.com/blog/social-media-content-creator/ (2024). Ha, Louisa. Behind the Screens: Insights from Digital Content Creators; Understanding Their Intentions, Practices and Challenges - UNESCO Digital Library . https://unesdoc.unesco.org/ark:/48223/pf0000392006 (2024). Foulkes, L. The problem with mental health awareness. The British Journal of Psychiatry 225 , 337–338 (2024). Hasell, A. & Chinn, S. The Political Influence of Lifestyle Influencers? Examining the Relationship Between Aspirational Social Media Use and Anti-Expert Attitudes and Beliefs. Social Media + Society 9 , 20563051231211945 (2023). Chinn, S. & and Hasell, A. How Different Uses of Social Media Inform Perceptions of Offline Social Norms and Changes in Vaccine Intentions. Health Communication 39 , 1198–1208 (2024). Gunawardena, C. N., Lowe, C. A. & Anderson, T. Analysis of a Global Online Debate and the Development of an Interaction Analysis Model for Examining Social Construction of Knowledge in Computer Conferencing. Journal of Educational Computing Research 17 , 397–431 (1997). Gunawardena, C., Flor, N., Gomez, D. & Sanchez, D. Analyzing Social Construction of Knowledge Online by Employing Interaction Analysis, Learning Analytics, and Social Network Analysis. The Quarterly Review of Distance Education 17 , 35–60 (2016). Haythornthwaite, C. et al. Learning in the wild: coding for learning and practice on Reddit. Learning, Media and Technology 43 , 219–235 (2018). Morris, J. X. et al. TextAttack: A Framework for Adversarial Attacks, Data Augmentation, and Adversarial Training in NLP. Preprint at https://doi.org/10.48550/arXiv.2005.05909 (2020). OpenAI et al. GPT-4 Technical Report. Preprint at https://doi.org/10.48550/arXiv.2303.08774 (2024). Devlin, J., Chang, M.-W., Lee, K. & Toutanova, K. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. Preprint at https://doi.org/10.48550/arXiv.1810.04805 (2019). Liu, Y. et al. RoBERTa: A Robustly Optimized BERT Pretraining Approach. Preprint at https://doi.org/10.48550/arXiv.1907.11692 (2019). Ji, S. et al. MentalBERT: Publicly Available Pretrained Language Models for Mental Healthcare. Preprint at https://doi.org/10.48550/arXiv.2110.15621 (2021). Müller, M., Salathé, M. & Kummervold, P. E. COVID-Twitter-BERT: A natural language processing model to analyse COVID-19 content on Twitter. Front. Artif. Intell. 6 , (2023). Pota, M., Ventura, M., Catelli, R. & Esposito, M. An Effective BERT-Based Pipeline for Twitter Sentiment Analysis: A Case Study in Italian. Sensors 21 , 133 (2021). Pourazad, N., Stocchi, L. & Narsey, S. A Comparison of Social Media Influencers’ KPI Patterns across Platforms: Exploring Differences in Followers and Engagement On Facebook, Instagram, YouTube, TikTok, and Twitter. Journal of Advertising Research 63 , 139–159 (2023). Bates, D. et al. lme4: Linear Mixed-Effects Models using ‘Eigen’ and S4. (2024). Ben Bolker. An Introduction to ‘Margins’ . https://cran.r-project.org/web/packages/margins/vignettes/Introduction.html (2024). Additional Declarations There is NO Competing Interest. The authors declare no competing interests, but provide the following information in the interests of transparency and full disclosure. AY is an unpaid advisor to the nonprofit Science To People. AY and the Center for Health Communication have received grant funding for other ongoing research projects from YouTube Health and the WellWithAll Foundation. Supplementary Files ttcommentappendixvfmay25.docx Appendix of the main article Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6787693","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Social Sciences - Article","associatedPublications":[],"authors":[{"id":466043616,"identity":"12e2b3d7-b121-42a8-8a6e-24f49ebf9c10","order_by":0,"name":"Yuning Liu","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABAElEQVRIiWNgGAWjYJCCA0Asw8bAfICBsQEiIsHAwExQCw8bA1sC8VpAgAeIDIjTYnAj9+Dhgl8MPHwSOZ9fMO6wyzc4fvbhDYYK68QGnFryEg7P7AM6TCJ3mwXjmWTLDWfSjS0YzqTj0ZJjcJi3B6LFgLGN2UCyIY1NgrHtMDFacp4BtdQbSPY/A2r5R0ALzw+wFuYHQMMN+CVAtjTg1iJ55g3QlgagFp5nZgyJbceBWp4xWyQcSzfGpYXveI7xZ54/DHLy7cmPP3xsqzZg409jvPGhxloWlxaFA0CCse0/iM0mkQATTsCuGgzkwWb9AbOZP+BROApGwSgYBSMYAAD/KlUhXJCIMQAAAABJRU5ErkJggg==","orcid":"","institution":"The Harvard Kenneth C. Griffin Graduate School of Arts and Sciences, Harvard T.H. Chan School of Public Health","correspondingAuthor":true,"prefix":"","firstName":"Yuning","middleName":"","lastName":"Liu","suffix":""},{"id":466043617,"identity":"2e9dc016-91b0-47ba-bf3e-28f17d9e3dca","order_by":1,"name":"Matt Motta","email":"","orcid":"","institution":"Dept. of Health Law, Policy, \u0026 Management, Boston University School of Public Health","correspondingAuthor":false,"prefix":"","firstName":"Matt","middleName":"","lastName":"Motta","suffix":""},{"id":466043618,"identity":"1a74a8b1-3713-4c7f-966a-6a7217356c64","order_by":2,"name":"Kenzie Brenna","email":"","orcid":"","institution":"Content Creator","correspondingAuthor":false,"prefix":"","firstName":"Kenzie","middleName":"","lastName":"Brenna","suffix":""},{"id":466043619,"identity":"f370ab2a-1db1-4386-b134-81b844583315","order_by":3,"name":"Shahem McLaurin","email":"","orcid":"","institution":"Therapist and Creator @5hahem","correspondingAuthor":false,"prefix":"","firstName":"Shahem","middleName":"","lastName":"McLaurin","suffix":""},{"id":466043620,"identity":"92dfaec7-7acd-48b2-ac8c-970efd3e1f55","order_by":4,"name":"Katharine Speer","email":"","orcid":"","institution":"Center for Health Communication, Harvard T.H. Chan School of Public Health","correspondingAuthor":false,"prefix":"","firstName":"Katharine","middleName":"","lastName":"Speer","suffix":""},{"id":466043621,"identity":"77cecbe4-3de8-4f2a-9407-c4e7e468d6a1","order_by":5,"name":"Elissa Scherer","email":"","orcid":"","institution":"Center for Health Communication, Harvard T.H. Chan School of Public Health","correspondingAuthor":false,"prefix":"","firstName":"Elissa","middleName":"","lastName":"Scherer","suffix":""},{"id":466043622,"identity":"88c04f1a-e72b-4295-a482-98f3a6ed43e1","order_by":6,"name":"Maurice Smith","email":"","orcid":"https://orcid.org/0000-0003-4214-1277","institution":"Harvard University","correspondingAuthor":false,"prefix":"","firstName":"Maurice","middleName":"","lastName":"Smith","suffix":""},{"id":466043623,"identity":"dc6eb41e-89e5-4116-a26d-b6e408408a2b","order_by":7,"name":"Amanda Yarnell","email":"","orcid":"","institution":"Center for Health Communication, Department of Social and Behavioral Sciences, Harvard T.H. Chan School of Public Health","correspondingAuthor":false,"prefix":"","firstName":"Amanda","middleName":"","lastName":"Yarnell","suffix":""}],"badges":[],"createdAt":"2025-05-30 22:11:06","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6787693/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6787693/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":84205087,"identity":"413b2394-e901-4a76-93c7-85c7d91eeea9","added_by":"auto","created_at":"2025-06-09 08:57:30","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":807387,"visible":true,"origin":"","legend":"\u003cp\u003eOverview of the on-platform RCT\u003c/p\u003e\n\u003cp\u003eTraining toolkit: https://hsph.harvard.edu/research/health-communication/creator-program/creator-resources/\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-6787693/v1/249fb99a9a4cbdea464b7d1f.png"},{"id":84204153,"identity":"ebf94284-a3eb-45a6-9a06-7c1ac4a9159e","added_by":"auto","created_at":"2025-06-09 08:49:30","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":885955,"visible":true,"origin":"","legend":"\u003cp\u003ePre/Post-treatment changes in foundational mental health knowledge construction (outcome) in observed data and from model estimates by intervention group and study moderators\u003c/p\u003e\n\u003cp\u003e* Video comment engagement level: Defined by the comment-to-view ratio for each video. Videos with a ratio greater than or equal to the median (≥ 0.97%) were categorized as having high comment engagement, while those below the median were categorized as low comment engagement.\u003c/p\u003e\n\u003cp\u003e** Video like engagement level: Defined by the like-to-view ratio for each video. Videos with a ratio greater than or equal to the median (≥ 0.02%) were categorized as having high like engagement, while those below the median were categorized as low like engagement.\u003c/p\u003e\n\u003cp\u003e*** Video share engagement level: Defined by the share-to-view ratio for each video. Videos with a ratio greater than or equal to the median (≥ 0.18%) were categorized as having high share engagement, while those below the median were categorized as low share engagement.\u003c/p\u003e\n\u003cp\u003e**** Creator engagement level: Calculated as the average comment-to-view ratio across all videos posted by each creator during the study period. Creators with an average ratio above the median (1.20%) were classified as having high engagement, and those below the median as having low engagement.\u003c/p\u003e\n\u003cp\u003eNote: This plot visualizes the distribution of post–pre changes in the outcome variable—the proportion of comments exhibiting foundational mental health knowledge construction—across four groups, defined by intervention condition (toolkits-only vs. Control) and subgroups. The vertical ridgelines represent bootstrapped distributions of the raw, video or creator-level change in the outcome between post- and pre-study period. Superimposed on each ridgeline are the points and error bars, indicating the model-based marginal estimate of the post-pre changes in the prevalence of foundational mental health knowledge construction (outcome), along with its 95% confidence interval derived from a linear probability model: P(outcome=1)=Treatment*Prepost — among different subgroups. Marginal effects were estimated using the margins package in R.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-6787693/v1/af1f3555a9709caf9f5b9599.png"},{"id":84204152,"identity":"bd70dc08-8531-445c-9c74-61ff3e7158c3","added_by":"auto","created_at":"2025-06-09 08:49:30","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":428559,"visible":true,"origin":"","legend":"\u003cp\u003eThe effect of treatment video assignment on perceived emotional support capabilities\u003c/p\u003e\n\u003cp\u003eNote. Predicted probabilities are derived from ordered logistic regression models. Predictions hold all covariates at their observed sample means. ”Prep.” is an abbreviation for “Prepared.” Full model output is available in Table A9. Moderation tests of results in Table A9 are presented in Table A12.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-6787693/v1/74a2a70265fdf16fdb46c93f.png"},{"id":84205414,"identity":"4840772d-d96c-4853-84fe-a073781b0985","added_by":"auto","created_at":"2025-06-09 09:05:32","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3581542,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6787693/v1/4160bb6b-faf0-459f-813e-4265f900578d.pdf"},{"id":84204155,"identity":"a4dcc25d-67b5-4ee2-a6de-e6d4a38aa2f6","added_by":"auto","created_at":"2025-06-09 08:49:30","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":2057343,"visible":true,"origin":"","legend":"Appendix of the main article","description":"","filename":"ttcommentappendixvfmay25.docx","url":"https://assets-eu.researchsquare.com/files/rs-6787693/v1/47b327fc1208f39a6d2ec2bf.docx"}],"financialInterests":"There is \u003cb\u003eNO\u003c/b\u003e Competing Interest.\nThe authors declare no competing interests, but provide the following information in the interests of transparency and full disclosure. AY is an unpaid advisor to the nonprofit Science To People. AY and the Center for Health Communication have received grant funding for other ongoing research projects from YouTube Health and the WellWithAll Foundation.","formattedTitle":"Training TikTok creators in mental health communication can benefit their audiences","fulltext":[{"header":"Introduction","content":"\u003cp\u003eSocial media has created new opportunities for implementing network interventions to promote health outcomes\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e,\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e. Network intervention refers to a set of strategies that leverage social network structures to promote behavior change\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e. These strategies have been successfully applied in interventions targeting various health outcomes, including exercise, smoking control, and mental health\u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e,\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e. Video-based platforms such as TikTok, Instagram, and YouTube host online social networks for mental health communication, where users seek information, share personal experiences, and access community support\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e,\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e,\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e. According to a national survey by the U.S. CDC, 58.5% of adults used the internet for health or medical information in 2022\u003csup\u003e11\u003c/sup\u003e. On TikTok, views of the hashtag #MentalHealth surged from 25.3\u0026nbsp;billion in 2022 to nearly 44\u0026nbsp;billion in 2023, with over 22.1\u0026nbsp;million posts in 2025\u003csup\u003e12,13\u003c/sup\u003e. Notably, 66% of all TikTok users\u0026mdash;and 91% of those aged 18\u0026ndash;29\u0026mdash;report using the platform to access mental health information or advice\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eMental health content creators can serve as influential online community leaders in promoting mental health within online social networks. Content creators have been shown to influence user behavior in fields such as fashion and lifestyle\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e,\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e, as well as promote hygiene practices during COVID-19\u003csup\u003e16\u003c/sup\u003e. Mental health content creators reach broad audiences\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e on platforms such as TikTok. More importantly, their content has gained trust from audiences. For example, a recent poll found that 53% of U.S. adults aged 18\u0026ndash;29 trust health information found on TikTok\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e, and young adults trust social media health content nearly as much as information from national news outlets\u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e. Moreover, younger users are more likely than older adults to act on this information\u0026mdash;for example, by consulting a doctor (19% for ages 18\u0026ndash;29, compared to 14% for ages 30\u0026ndash;49 and 5% for 50+), or seeking mental health treatment (26% for ages 18\u0026ndash;29, versus 9% and 2%, respectively)\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e. A 2025 study across 16 countries found 33% of participants aged 18 to 34 made health decisions influenced by content creators without medical training\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eTo promote evidence-based online mental health communication via content creators, organizations have initiated creator training programs. For example, Harvard T.H. Chan School\u0026rsquo;s Center for Health Communication (CHC), the World Health Organization (WHO), and YouTube Health have each launched programs for mental-health creators\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e,\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e. These programs aim to improve the quality of mental health communication on social media by providing creators with resources such as expert-vetted mental health information and training in evidence-based content practices. Meanwhile, organizations like the Mental Health Storytelling Coalition and CHC maintain publicly available resource libraries for mental health content creators\u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eEmpirical \u003cem\u003ecreator-engaged\u003c/em\u003e studies have recently shown that training programs can effectively change social media content. A large-scale randomized controlled trial conducted on TikTok\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e demonstrated that providing creators with toolkits containing evidence-based mental health talking points increases the prevalence of such content in the videos creators produced. Creators also reported increasing motivation to incorporate evidence-based research into their videos after attending an in-person training\u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e. Moreover, the creator training program significantly boosted views of trained creators' content\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e, demonstrating the effectiveness of promoting evidence-based content to social media users through a network-based strategy of \"influencing the influencers.\"\u003c/p\u003e \u003cp\u003eHowever, it remains unclear whether the effects of creator training influence mental health attitudes and behaviors of the content consumers \u0026ndash; which is the primary goal of network interventions. According to health behavior change theories\u003csup\u003e\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e,\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e, exposure to evidence-based mental health information can shift beliefs and perceptions, improve health literacy, and potentially prompt protective behaviors. Yet, no previous studies have empirically examined these training programs' effects on viewer-level mental health outcomes. Our research aims to address this gap.\u003c/p\u003e \u003cp\u003eWe examine the potential of training content creators to improve mental health communication within online social networks through two randomized controlled trials (RCTs). The first is an on-platform RCT on TikTok that evaluates how training creators in evidence-based content influences mental health knowledge construction\u0026mdash;the process of understanding concepts, phenomena, and situations\u0026mdash;as evidenced by user comments left on each video. The second is an off-platform, nationally representative survey-based RCT of U.S. youth aged 14\u0026ndash;22, assessing how training creators in how to offer emotional support to friends affects the audience\u0026rsquo;s ability to provide such support. In so doing, our approach combines the \u003cem\u003eexternal validity\u003c/em\u003e benefits associated with on-platform experiments (i.e., the fact that video interventions were embedded in \u0026ldquo;real life\u0026rdquo; social media interactions) with the \u003cem\u003einternal validity\u003c/em\u003e associated with survey-based experiments (i.e., standardization of the content to which people were exposed; ability to measure attitudinal and behavioral data from even those participants who may be reluctant to publicly comment on a video).\u003c/p\u003e \u003cp\u003eOur first study, an on-platform RCT, tests the hypothesis that increases in the supply of evidence-based mental health communication content on TikTok, driven by creator training programs\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e, can enhance mental health knowledge construction in video comments. Knowledge construction is the consequence of informal learning among social media users\u0026mdash;a non-didactic, socially collaborative process driven by individual interests\u0026mdash;across various domains, including climate change\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e, education\u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e, geoscience communication\u003csup\u003e\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e, and mental health\u003csup\u003e\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e. Users\u0026rsquo; comments offer real-time insights into their online communication and activities\u003csup\u003e\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e, as well as their processes of socialization and idea sharing\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e,\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u003c/sup\u003e. For example, a study of comments on TikTok mental health videos found that 66% offered validation or support of peers and 56% discussed other mental health issues\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eAn overview of the on-platform RCT is presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. We assess mental health knowledge construction in TikTok video comments, by using the Interaction Analysis Model (IAM). IAM isa widely applied framework that conceptualizes knowledge construction as a cascading process of increasing complexity. Specifically, we examine whether evidence-based mental health content enhances both the foundational mental health knowledge construction\u0026mdash;the initial stage of reflection on personal experiences or opinions\u0026mdash;and the advanced stages of mental health knowledge construction, which are characterized by expressions of agreement or disagreement, clarification questions, reinterpretation of knowledge, or its application to other contexts. We apply Large Language Model (LLM)-based content analysis to identify knowledge construction stages in \u0026gt;\u0026thinsp;180,000 user comments. Based on the experimental assignment protocol embedded in our previous RCT research\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e, we requested all comments (n\u0026thinsp;=\u0026thinsp;188,169) from all videos (n\u0026thinsp;=\u0026thinsp;1,882) created by the 49 mental health content creators enrolled in the RCT in March, April, and May 2023 using TikTok\u0026rsquo;s Research Tools API\u003csup\u003e\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e (Figure A1). We fine-tuned BERT-based LLMs (Table A1) to predict the study outcomes (Table A2)\u0026mdash;foundational and advanced mental health knowledge construction\u0026mdash;using 4,152 human-annotated comments and augmented texts (Table A3) (see Methods: Outcome Measurement).\u003c/p\u003e \u003cp\u003eOur second study, an off-platform survey experiment, addresses the selection bias inherent in the on-platform RCT, where individuals who choose to comment on videos are an unrepresentative subset of the audience. To overcome this, we conducted a pre-registered\u003csup\u003e\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/sup\u003e, survey-based RCT with a demographically representative sample (N\u0026thinsp;=\u0026thinsp;1,000) of U.S. youth aged 14\u0026ndash;22. This study examines whether exposure to videos created using evidence-based mental health communication toolkits influences perceived and objective emotional support competencies; i.e., the process by which people offer guidance and understanding to peers experiencing hardship.\u003c/p\u003e \u003cp\u003eWe partnered with a popular social media content creator to produce a short, TikTok-style video about \u003cem\u003eemotionally supporting friends\u003c/em\u003e using their conventional media production strategies (control, prior to engaging with any interventional materials). The creator then completed training analogous to that featured in our on-platform experiment that provided guidance on how to create evidence-based emotional support information for young people. After the session, the creator made a second video on the same topic (treatment), incorporating as they saw fit the evidence-based content while keeping the format as similar as possible to the first video. Respondents were randomly assigned to view either the treatment or control video and then asked questions about (1) their perceived efficacy in providing emotional support (perceived support) and (2) hypothetical vignette scenarios that put respondents in a position to offer emotional support, pursuant with best practice recommendations (objective support).\u003c/p\u003e \u003cp\u003eThe training on emotional support was designed based on key competencies identified in prior research\u003csup\u003e\u003cspan additionalcitationids=\"CR32\" citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u003c/sup\u003e. These competencies include: (a) validating emotions by acknowledging what the person is feeling and why it is understandable, (b) active listening through asking questions, paraphrasing, and avoiding unsolicited problem-solving, and (c) supporting peers' emotions by refraining from minimizing, challenging, or dismissing their concerns. Correspondingly, we expected that treated respondents would be more likely to both express perceived competencies in providing emotional support to friends experiencing mental health challenges and exhibit greater competencies in providing such support, in line with evidence-based best practices.\u003c/p\u003e "},{"header":"Results","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eStudy 1 \u0026ndash; On-platform RCT:\u003c/h2\u003e \u003cdiv id=\"Sec4\" class=\"Section3\"\u003e \u003ch2\u003eEffect of creator training on mental health knowledge construction\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eEffect of toolkits-only and sessions\u0026thinsp;+\u0026thinsp;toolkits interventions assignment on mental health knowledge construction in video comments\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"8\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colspan=\"3\" nameend=\"c4\" namest=\"c2\"\u003e \u003cp\u003eModel result\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colspan=\"3\" nameend=\"c8\" namest=\"c6\"\u003e \u003cp\u003eSample size (N)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e% changes in comments with knowledge construction due to treatment\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSE\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eP value\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eComments\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eVideos\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eCreators\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"8\" nameend=\"c8\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eToolkits-Only group (ref\u0026thinsp;=\u0026thinsp;control)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFoundational\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003e+\u0026thinsp;4.15%\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e1.85\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.025\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\" morerows=\"1\" rowspan=\"2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e123323\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e1207\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e29\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAdvanced\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e-0.86%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.29\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.502\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"8\" nameend=\"c8\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eSessions\u0026thinsp;+\u0026thinsp;Toolkits group (ref\u0026thinsp;=\u0026thinsp;control)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFoundational\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e+\u0026thinsp;2.29%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.86\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.219\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\" morerows=\"1\" rowspan=\"2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e77910\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e1171\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e35\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAdvanced\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e+\u0026thinsp;0.64%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.46\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.662\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e\u003cp\u003eNote:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eParameter estimates are standardized and from three-level multilevel linear probability models (LPM), with comments nested in videos that are created by creators engaged in the study (sessions\u0026thinsp;+\u0026thinsp;toolkits group N\u0026thinsp;=\u0026thinsp;20, toolkits-only group N\u0026thinsp;=\u0026thinsp;14, control group N\u0026thinsp;=\u0026thinsp;15). Multilevel LPMs were separately fitted for sessions\u0026thinsp;+\u0026thinsp;toolkits (compared to the control) and toolkits-only (compared to the control).\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eThe study outcomes are foundational mental health knowledge construction\u0026mdash;the initial stage involving personal reflection\u0026mdash;and advanced mental health knowledge construction\u0026mdash;the stages with deeper cognitive engagement and higher-order thinking through agreeing, disagreeing, clarifying, reconceptualization, and application\u0026mdash;in mental health measured following Interactive Analysis Model (IAM) by LLM-assisted content analysis. Foundational and advanced knowledge construction are binary variables.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eThe models identify the effect of change in the outcomes attributable to exposure to our study\u0026rsquo;s interventions by interacting dichotomous fixed effect indicators of treatment group assignment (sessions\u0026thinsp;+\u0026thinsp;toolkits vs. toolkits-only, with the control serving as a reference group), with the fixed effect indicator of whether or not each video was produced pre or post intervention. The pre- and post-intervention periods were defined based on the timing of the study intervention. For the toolkits-only group, the pre-intervention period spanned from March 1, 2023, at 12:00 AM ET to May 2, 2023, at 9:30 AM ET, while the post-intervention period covered May 2, 2023, at 9:30 AM ET to May 31, 2023, at 11:59 PM ET. For the sessions\u0026thinsp;+\u0026thinsp;toolkits group, the pre-intervention period lasted from March 1, 2023, at 12:00 AM ET to April 7, 2023, at 11:59 AM ET, and the post-intervention period extended from April 8, 2023, at 12:00 AM ET to May 31, 2023, at 11:59 PM ET.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eThe estimated prevalence of knowledge construction outcomes in treatment and control groups at the pre and post period are presented in Table A4.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eEffect of creator training toolkits (the toolkits-only intervention) on foundational mental health knowledge construction in video comments by video and creator characteristics\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"10\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c10\" colnum=\"10\"\u003e\u003c/div\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colspan=\"3\" nameend=\"c4\" namest=\"c2\"\u003e \u003cp\u003eModel result\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\" morerows=\"1\" rowspan=\"2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eLevel\u003c/p\u003e \u003cp\u003eof analysis\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colspan=\"3\" nameend=\"c10\" namest=\"c8\"\u003e \u003cp\u003eSample size (N)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e% changes in comments\u003c/p\u003e \u003cp\u003ewith outcome* due to training toolkits\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSE\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eP value\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eComments\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003eVideos\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003eCreators\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"4\" nameend=\"c4\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eVideo featured themes in training toolkits\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eVideo\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eYes, featured\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e+\u0026thinsp;\u003cb\u003e8.59%\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e3.84\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.026\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e30030\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e296\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e24\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNo, not featured\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e+\u0026thinsp;2.14%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.318\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e93293\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e911\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e28\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"4\" nameend=\"c4\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eVideo comment engagement\u003c/b\u003e \u003cem\u003e(comment-to-view ratio)\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eVideo\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u0026ge;0.97% (median)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e+\u0026thinsp;\u003cb\u003e5.56%\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e2.46\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.024\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e81322\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e652\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e27\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u0026lt;0.97%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e+\u0026thinsp;0.59%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.84\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.837\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e42001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e555\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e26\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"4\" nameend=\"c4\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eVideo like engagement\u003c/b\u003e \u003cem\u003e(like-to-view ratio)\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eVideo\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u0026ge;0.02% (median)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e+\u0026thinsp;\u003cb\u003e6.46%\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e2.61\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.014\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e42202\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e603\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e26\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u0026lt;0.02%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e-1.24%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3.10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.688\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e81121\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e604\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e28\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"4\" nameend=\"c4\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eVideo share engagement\u003c/b\u003e \u003cem\u003e(share-to-view ratio)\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eVideo\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u0026ge;0.18% (median)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e+\u0026thinsp;1.18%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.25\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.601\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e93555\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e640\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e26\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u0026lt;0.18%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e+\u0026thinsp;\u003cb\u003e7.86%\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e3.15\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.013\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e29768\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e567\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e28\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"4\" nameend=\"c4\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eCreator engagement level**\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eComment\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u0026ge;1.20% (median)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e+\u0026thinsp;\u003cb\u003e10.03%\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e3.56\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.005\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e68541\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e557\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e15\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u0026lt;1.20%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e+\u0026thinsp;2.65%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.36\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.261\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e54782\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e650\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e14\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"4\" nameend=\"c4\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eCreator follower count\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eComment\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u0026ge;500,000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e+\u0026thinsp;3.87%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3.04\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.204\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e82774\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e570\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e9\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u0026lt;500,000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e+\u0026thinsp;4.75%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.81\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.092\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e40549\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e637\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e20\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"4\" nameend=\"c4\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eCreator is licensed health professional\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eComment\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLicensed\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e+\u0026thinsp;4.70%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.42\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.053\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e50481\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e696\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e19\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNot licensed\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e+\u0026thinsp;4.29%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3.22\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.184\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e72842\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e511\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e10\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"10\"\u003e* Outcome: foundational mental health knowledge construction\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e** Creator engagement level is calculated by: \u003cem\u003e[(comment\u0026thinsp;+\u0026thinsp;like)/view] using data in Mar 2023-May 2023.\u003c/em\u003e\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eNote\u003c/strong\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e Parameter estimates are standardized and derived from three-level multilevel linear probability models (LPMs), with comments nested within videos, which are nested within creators (toolkits-only group N\u0026thinsp;=\u0026thinsp;14, control group N\u0026thinsp;=\u0026thinsp;15).\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eSeparate LPMs were fitted for comment subsets based on video or creator characteristics. As indicated in the \u0026ldquo;level of analysis\u0026rdquo; column, video-level subgroup analyses include all comments from videos in a given subgroup (e.g., videos featuring toolkit-related themes), with one LPM per subgroup. The number of videos in each subgroup defined by a single characteristic (e.g., whether the video features a toolkit-related theme or not) sums to the total of 1,207 videos across the toolkits-only and control groups. Similarly, creator-level subgroup analyses include comments on videos by creators in each subgroup (e.g., \u0026ge;\u0026thinsp;500,000 followers), with one LPM per subgroup. The number of creators in each subgroup defined by a single characteristic (e.g., follower count\u0026thinsp;\u0026ge;\u0026thinsp;500,000 vs. \u0026lt;500,000) sums to the total of 29 across the toolkits-only and control groups.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eThe study outcome is the foundational mental health knowledge construction\u0026mdash;a binary variable presenting the initial stage involving personal reflection in mental health measured following Interactive Analysis Model (IAM) by LLM-assisted content analysis.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eThe models identify the effect of change in the outcomes attributable to exposure to our study\u0026rsquo;s interventions by interacting dichotomous fixed effect indicators of treatment group assignment (toolkits-only, with the Control serving as a reference group), with the fixed effect indicator of whether or not each video was produced pre or post intervention. The pre- and post-intervention periods were defined based on the timing of the study intervention. For the toolkits-only group, the pre-intervention period spanned from March 1, 2023, at 12:00 AM ET to May 2, 2023, at 9:30 AM ET, while the post-intervention period covered May 2, 2023, at 9:30 AM ET to May 31, 2023, at 11:59 PM ET.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eThe estimated prevalence of knowledge construction outcomes in treatment and control groups at the pre and post period for different subgroups are presented in Table A5.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eWith the LLM-assisted content analysis results on 188,169 comments, we first explored the pre- vs. post-treatment change in mental health knowledge construction. We assess mental health knowledge construction in video comments across content creators assigned to a \u0026ldquo;Toolkits-Only\u0026rdquo; training condition (N\u0026thinsp;=\u0026thinsp;14 creators, N\u0026thinsp;=\u0026thinsp;692 videos) that exposed TikTok creators to a series of asynchronous digital training toolkits, a \u0026ldquo;Sessions Plus Toolkits\u0026rdquo; condition (N\u0026thinsp;=\u0026thinsp;20 creators, N\u0026thinsp;=\u0026thinsp;656 videos) which exposed creators to both digital toolkits and a series of synchronous, virtual training sessions (see: Methods\u0026ndash;Creator training program); all compared to a randomly-assigned control group (N\u0026thinsp;=\u0026thinsp;15 creators, N\u0026thinsp;=\u0026thinsp;515 videos). Figure A2 presents the estimated prevalence of foundational mental health knowledge construction in video comments across intervention groups over the study period, allowing us to examine the parallel trends assumption and assess the distribution of the raw data.\u003c/p\u003e\n\u003cp\u003eShown in Table \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e (Note that two-tailed \u003cem\u003ep\u003c/em\u003e values are presented throughout), we find that exposure to videos by creators in the toolkits-only intervention significantly increased the likelihood that comments featured foundational mental health knowledge construction\u0026mdash;the initial stage of reflection on personal experiences or opinions\u0026mdash;by 4.15% (P\u0026thinsp;=\u0026thinsp;0.025), relative to the control group; following exposure to our interventional materials. In the toolkits-only group, the prevalence of foundational mental health knowledge construction increased from 32.86% in the pre period to 35.02% in the post period, while in the control group, it decreased from 31.57\u0026ndash;29.58% over the same period (Table A4). Additionally, neither the toolkits-only nor sessions\u0026thinsp;+\u0026thinsp;toolkits intervention shows significant effects on advanced mental health knowledge construction, which involves expressions of agreement or disagreement, clarification questions, reinterpretation of knowledge, or its application to other contexts.\u003c/p\u003e\n\u003cp\u003eWe further examined whether videos containing toolkit-related themes, such as the idea that climate change can negatively impact mental health, were more likely to induce audience mental health knowledge construction. This would further suggest that how content creators integrate toolkit information into their content is responsible for the treatment effects we observe. Among 30,030 comments on 296 videos that featured themes from the toolkits-only training toolkits, the intervention increased the probability of foundational mental health knowledge construction by 8.59% compared to the control group (P\u0026thinsp;=\u0026thinsp;0.026). Videos in the toolkits-only group featuring toolkit-related themes showed an estimated 5.91% (95% CI: 0.85\u0026ndash;10.97%) increase in foundational mental health knowledge construction (Fig. \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e, Panel A) from 33.45\u0026ndash;39.36% (Table A5). This finding further supports the effectiveness of the toolkit intervention. Videos that featured themes from the toolkit suggest that the content creator likely engaged with and applied the toolkit in their content creation\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e. Among these videos, the improvement in knowledge construction was greater than in other groups, reinforcing the effectiveness of the toolkit intervention.\u003c/p\u003e\n\u003cp\u003eBecause user commenting behavior may be influenced by how engaging a video is, we conducted subgroup analyses based on video engagement levels, measured by comment-to-view, like-to-view, and share-to-view ratios. We find that the toolkits-only intervention increased foundational mental health knowledge construction by 5.56% (P\u0026thinsp;=\u0026thinsp;0.024) and 6.46% (P\u0026thinsp;=\u0026thinsp;0.014) among videos with high engagement in comment (comment-to-view ratio\u0026thinsp;\u0026gt;\u0026thinsp;median of 0.97%) and like (like-to-view ratio\u0026thinsp;\u0026gt;\u0026thinsp;median of 0.02%), respectively. For videos in the toolkits-only group with high comment engagement, foundational mental health knowledge construction increased by an estimated 2.35% (95% CI: -0.66 to 5.36%; Fig. \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e, Panel B), from 32.45\u0026ndash;34.79% (Table A5). Among the toolkits-only group, videos with high like engagement, the estimated increase in foundational mental health knowledge construction was 2.91% (95% CI: -0.80 to 6.63%; Fig. \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e, Panel C), rising from 31.65\u0026ndash;34.56% (Table A5). Videos with higher commenting and liking engagement appear to elicit more user comments, resulting in a higher level of revealed foundational knowledge construction captured by our analysis.\u003c/p\u003e\n\u003cp\u003eInterestingly, we find that the toolkits-only intervention increased the probability of foundational mental health knowledge construction by 7.86% (P\u0026thinsp;=\u0026thinsp;0.013) in comments on videos with lower, but not higher, share-to-view ratios. Videos in the toolkits-only group with high sharing engagement showed an estimated 2.68 percentage point (95% CI: 0.30 to 5.07%; Fig. \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e, Panel D) increase in foundational knowledge construction, from 37.17\u0026ndash;39.86% (Table A5). Videos with lower share-to-view ratios may resonate more intimately with users\u0026apos; personal mental health experiences, making them highly engaging but less shareable due to privacy and sensitivity. Consequently, these videos might more effectively promote foundational mental health knowledge construction among viewers.\u003c/p\u003e\n\u003cp\u003eTo better direct creator-training resources, we further assessed the heterogeneous effects of creator toolkits on user foundational mental health knowledge construction across content creators with different characteristics, including their engagement rate, popularity (number of followers on TikTok), and whether the creator is a licensed mental health professional. The creator engagement rate was calculated as the ratio of total comments and likes to the number of views among videos published between March and May 2023. We found that providing evidence-based mental health toolkits increased foundational knowledge construction by 10.03% (P\u0026thinsp;=\u0026thinsp;0.005) among creators with engagement rates over the median of 1.20%, compared to the control group (Table \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e). Among the toolkits-only group creators with high engagement, the estimated increase in foundational knowledge construction was 2.01% (95% CI: -0.77 to 4.79%; Fig. \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e, Panel E), rising from 29.63\u0026ndash;31.65% (Table A5). No significant effect was observed among creators with lower engagement rates.\u003c/p\u003e\n\u003cp\u003eFurthermore, we found that the effects of evidence-based mental health toolkits on foundational mental health knowledge construction are unaffected by creators\u0026rsquo; follower size or professional licensing status. Following Brambor et al. (2006)\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e34\u003c/span\u003e\u003c/sup\u003e, we do not estimate marginal effects or predicted values for these non-significant moderator variables, as the results may be misleading. Institutions aiming to engage creators for health promotion should consider focusing training resources on those most likely to benefit, such as creators with higher engagement rates (above 1.20%) as identified in our study.\u003c/p\u003e\n\u003ch3\u003eStudy 2 \u0026ndash; Off-platform, survey-based RTC\u003c/h3\u003e\n\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\n \u003ch2\u003eEffect of emotional support training on emotional support competency\u003c/h2\u003e\n \u003cp\u003eIn the second survey-based RCT, we examine whether exposure to videos developed with evidence-based mental health communication training, compared to standard creator content, affects emotional support competencies among a nationally representative sample of U.S. youth aged 14\u0026ndash;22.\u003c/p\u003e\n\u003c/div\u003e\n\u003ch3\u003eManipulation check\u003c/h3\u003e\n\u003cp\u003eWe begin our analysis by first determining whether respondents viewed the study\u0026rsquo;s treatment video as comparatively more informative than the control. We have reason to believe (see: Figure A3) that the treatment video was more likely to feature evidence-based content than the control. However, respondents must then recognize these differences for our treatments to have the hypothesized effects that we pre-registered.\u003c/p\u003e\n\u003cp\u003eTable \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e tests this possibility by presenting the predicted percentage of respondents assigned to view the control (column 2) vs. treatment (column 3) videos who indicated that they found the video they watched to be helpful, informative, and/or any of the other qualities listed in column 1. The difference between the treatment vs. control group (∆ T-C) is listed in column 4, as well as a corresponding significance test of whether these differences are significantly different from zero (p) in column 5. If our theoretical expectations are supported, we should expect to see large and positive differences in column 4, which attain statistical significance at the p\u0026thinsp;\u0026lt;\u0026thinsp;0.05 level (two-tailed) in column 5.\u003c/p\u003e\n\u003cp\u003eAll predictions are derived from logistic regression models regressing each indicator listed in column 1 on dichotomous indicators of treatment group assignment, and an indicator of whether respondents were recruited into the study\u0026rsquo;s adult (aged 18\u0026ndash;22) or teen (aged 14\u0026ndash;17) samples (see Methods). Note that items presented in the final three rows (denoted by \u0026ldquo;R\u0026rdquo;) are reverse coded, for consistency with the items presented in the first seven rows.\u003c/p\u003e\n\u003cp\u003eThe results strongly suggest that respondents viewed the treatment video as significantly more helpful, informative, evidence-based, interesting, and engaging than the control. At times, these differences were quite large. For example, the likelihood that treatment group respondents rated the video they viewed as informative was more than 16 percentage points greater than those in the study\u0026rsquo;s control group. Respondents were also significantly less likely to report that the treatment video was unclear, or boring. Only two estimates (trustworthiness and inaccuracy) failed to attain statistical significance between the two groups.\u003c/p\u003e\n\u003cp\u003eThus, as we intended, Table \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e provides strong evidence that respondents drew meaningful distinctions between the two videos to which they could have been exposed.\u003c/p\u003e\u003ctable id=\"Tab3\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eManipulation check summary\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e% of Respondents who Found the Video to Be\u0026hellip;\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eControl\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eTreatment\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e∆ (T-C)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003ep\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eHelpful\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e58.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e75.43\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e16.47\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eInformative\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e60.22\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e76.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e16.75\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eEvidence-Based\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e32.13\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e37.94\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e5.81\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.054\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eTrustworthy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e54.48\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e59.20\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e4.72\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.132\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eInteresting\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e58.75\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e65.29\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e6.54\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.033\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eEngaging\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e52.41\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e60.59\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e8.18\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.009\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eUnclear (R)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e62.47\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e71.90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e9.43\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eBoring (R)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e45.62\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e56.13\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e10.51\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eInaccurate (R)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e58.39\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e60.56\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e2.17\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.483\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003ctfoot\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"5\"\u003eNote: Predicted probabilities are derived from logistic regression models. Predictions hold all covariates at their observed sample means. Full model output is available on Table A7.\u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tfoot\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003c/p\u003e\n\u003cp\u003e\u003c/p\u003e\n\u003ch2\u003ePerceived emotional support competencies\u003c/h2\u003e\n\u003cp\u003eNext, we test our pre-registered expectation (H1) that exposure to the study\u0026rsquo;s treatment videos will be associated with elevated confidence in one\u0026rsquo;s ability to provide emotional support to one\u0026rsquo;s peers. This includes respondents\u0026rsquo; perceived abilities to (a) provide emotional support to one\u0026rsquo;s peers, and (b) employ the evidence-based pillars of the A.S.K. method (which we described in full in the survey question; please see Table A8 for additional information.)\u003c/p\u003e\n\u003cp\u003eFigure \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e plots the predicted probability that respondents select each of the response options to the questions listed in Table A8 Row 2 (see: Perceived Emotional Support Capabilities) across exposure to the treatment (in red) vs. control (in blue) videos. 95% confidence intervals extend out from each estimate, with predicted values printed next to each one, for reference. Predictions are derived from ordered logistic regression models that control for differences in sample selection (Table A9), and again hold all covariates at their sample means.\u003c/p\u003e\n\u003cp\u003eThe results provide support for our theoretical expectations. In both cases, exposure to the study\u0026rsquo;s treatment video is positively and significantly associated with increased confidence (A.S.K. Method: \u0026beta;\u0026thinsp;=\u0026thinsp;0.24, p\u0026thinsp;=\u0026thinsp;0.05; Emotional Support: \u0026beta;\u0026thinsp;=\u0026thinsp;0.34, p\u0026thinsp;\u0026lt;\u0026thinsp;0.01). However, predicted differences between groups are relatively modest, and only approach conventional levels of two-tailed significance (as designated by non-overlapping 90% confidence intervals) in one case: whether respondents feel \u0026ldquo;very prepared\u0026rdquo; to provide emotional support.\u003c/p\u003e\n\u003cp\u003eOverall, the results suggest that exposure to the study\u0026rsquo;s treatment video is associated with increased confidence in respondents\u0026rsquo; emotional support capabilities; although we caution that the substantive and statistical magnitude of these effects is relatively modest.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eObjective emotional support competencies.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe next test whether respondents are more likely to take actions that reflect evidence-based best practices for providing emotional support in the study\u0026rsquo;s treatment condition, relative to the control. We measure respondents\u0026rsquo; objective competencies by asking them to report how they would act in a series of hypothetical scenarios; described in detail below (see: Methods). The results are displayed in Table \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e. If our pre-registered theoretical predictions (H2) are correct, we should expect to see positive (column 4, denoted by \u0026ldquo;∆(T-C)\u0026rdquo;) and statistically significant (column 5, denoted by \u0026ldquo;p\u0026rdquo;) differences between the treatment and control in the proportion of respondents providing a correct answer to the hypothetical scenarios testing each of the different emotional support competencies listed in Table A8.\u003c/p\u003e\n\u003cp\u003eResults are again derived from logistic regression models that regress dichotomous indicators of whether respondents provided a correct answer to each competency question (bolded entries in Table A8) on treatment group assignment and a sample selection indicator. \u0026nbsp;\u003c/p\u003e\n\u003ctable id=\"Tab4\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eThe effect of treatment video assignment on objective emotional support competency\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e% Answering Question Correctly\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eControl\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eTreatment\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e∆ (T-C)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003ep\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCompetency: Active Listening #1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e28.11\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e43.71\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e15.60\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCompetency: Validation #1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e33.05\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e47.40\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e14.35\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCompetency: Emotional Skill #1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e20.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e38.39\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e17.42\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCompetency: Emotional Skill #2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e68.27\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e68.32\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.05\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.987\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCompetency: Validation #2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e30.18\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e45.83\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e15.65\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCompetency: Active Listening #3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e41.45\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e48.19\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e6.74\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.032\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003ctfoot\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"5\"\u003eNote. Predicted probabilities are derived from logistic regression models. Predictions hold all covariates at their observed sample means. Full model output is available on Table A10. Moderation tests of results in Table A10 are presented on Table A12.\u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tfoot\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003c/p\u003e\n\u003cp\u003eThe results strongly suggest that exposure to the study\u0026rsquo;s treatment video is positively and significantly associated with increased competencies in every substantive area that we tested. In all but one case\u0026mdash;Emotional Skill Question #2 (which most respondents answered correctly irrespective of experimental assignment)\u0026mdash;treatment group respondents were considerably more likely to demonstrate the competencies tested in our hypothetical vignettes.\u003c/p\u003e\n\u003cp\u003eWe caution, however, that overall levels of competency in each of the areas we tested were relatively low; even though they are significantly higher in the treatment group. Aside from the second Emotional Skill question, most respondents answered each question incorrectly. Further demonstrative of this tension, we constructed a negative binomial regression model (Table A10) regressing a count of correctly answered items on treatment video exposure and sample selection indicators. Results suggest that treatment group respondents were predicted to answer 2.92 questions correctly, versus just 2.22 in the control (\u0026beta;\u0026thinsp;=\u0026thinsp;0.27, p\u0026thinsp;\u0026lt;\u0026thinsp;0.01). While treatment video exposure significantly elevated the number of correct answers we expect to observe, we nevertheless note that overall performance on the hypothetical competency vignettes is relatively low.\u003c/p\u003e\n\u003cp\u003e\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThe current study underscores the potential of mental health content creators as influential community leaders for promoting mental health in online social networks. The network-based \"influencing the influencers\" strategy\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e, providing training to creators to enhance mental health-related behaviors among social media users, has demonstrated effectiveness in both an on-platform RCT (providing evidence-based mental health communication training to creators improved TikTok commenters\u0026rsquo; foundational mental health knowledge construction) and an off-platform survey experiment (providing emotional support training to creators improved young viewer\u0026rsquo;s perceived and objective emotional support competencies).\u003c/p\u003e \u003cp\u003eTo our knowledge, this is the first study to assess the impact of large-scale content creator training on not just the content they produce, but \u003cem\u003esocial media user outcomes\u003c/em\u003e. Our study advances the previous findings\u0026ndash;that creator training programs boost evidence-based content creation\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e\u0026ndash;by showing that this enhancement also translates to changes in audience attitudes and behaviors. The benefits of the creator training program are significant. In the on-platform RCT, based on our March to May 2023 creator training program, an estimated 4.15% increase in foundational mental health knowledge construction in comments from over 2\u0026nbsp;million views accumulated by the 42 creators in the treatment groups could potentially enhance informal learning for up to 83,000 viewers, though this is an upper estimate since individuals may account for multiple views. In the off-platform survey experiment, exposure to TikTok-style videos created with support of the training increase young people\u0026rsquo;s perceived and objective emotional support competencies. These effects are similar for both regular and infrequent social media users, thereby implying that our approach is unlikely to be limited in its effectiveness to just those most familiar with social-video driven platforms like TikTok or Instagram.\u003c/p\u003e \u003cp\u003eOur study also combines the external validity strengths of an on-platform RCT with the internal validity of a survey-based experiment, jointly supporting the robustness of the identified effects of creator training toolkits on audience mental health behavioral outcomes. In the on-platform RCT, embedding video interventions within real-life social media interactions supports the idea that exposure to evidence-based content is associated with attitude change in naturalistic settings. In the survey-based experiment, partnering with a lifestyle influencer to produce both treatment and control videos enhances internal validity, as key message features, such as the messenger, style, diction, and camera work, are held constant across conditions, and allows us to directly manipulate exposure to evidence-based (vs. control) videos. Thus, in the off-platform survey experiment, the effect is assessed across all video viewers, not limited to those who left comments.\u003c/p\u003e \u003cp\u003eIn the on-platform RCT, the finding that only foundational knowledge construction improved, while advanced stages did not, aligns with the nature of TikTok videos: they are typically short (less than 60 seconds) and involve only one round of interaction between creators and viewers. Advanced knowledge construction may require repeated exposure to more iterative discussions. Future studies should explore these repeated exposures. Similar to the on-platform RCT, the off-platform survey experiment asked participants to watch a single short video (\u0026lt;\u0026thinsp;60 seconds), and we examined how one video influenced user outcomes. An important direction for future research is to investigate the effects of repeated exposure and reinforcement, as continued engagement with a creator may lead to stronger effects.\u003c/p\u003e \u003cp\u003eIn addition to identifying mental health content creators as influential online-community leaders in online social networks, our study proposes delivering easy-to-navigate toolkits directly to them as an effective way to intervene. In the on-platform RCT, we found that creators who only received training toolkits produced videos that were more effective at improving their audiences\u0026rsquo; mental health knowledge construction, compared to their peers who received the toolkits and also attended online training. This may be because the online training sessions, a series of hourlong health professional-led briefings and networking opportunities, left creators with less time to make content using the toolkits during the study period.\u003c/p\u003e \u003cp\u003eMore importantly, delivering easy-to-navigate toolkits to content creators is not only effective but also scalable, enabling engagement with a broader range of mental and general health creators across platforms in evidence-based health communication. We believe this is the case for two reasons. First, our results suggest that creator training need not be resource intensive, as we found that the development and delivery of easy-to-scale toolkits tailored to creators' needs can meaningfully impact users\u0026rsquo; attitudes and behaviors. Consequently, health organizations could create and distribute these toolkits exponentially to more mental health creators and to creators across different health domains. Second, many creators we've worked with produce content on multiple digital platforms, including TikTok, Instagram, and YouTube\u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e. Some of these platforms have adopted algorithmically personalized, short-form video feeds in recent years, such as Instagram Reels and YouTube Shorts\u003csup\u003e\u003cspan additionalcitationids=\"CR36 CR37\" citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e\u003c/sup\u003e. Given these similarities, we believe that the effects identified in the current study on TikTok could be generalized to other platforms featuring algorithmically personalized and short-form video content, particularly Instagram and YouTube.\u003c/p\u003e \u003cp\u003eThe current study also points to the efficacy of health professionals engaging with content creators to promote public health messages. Public health practitioners have utilized social media for health promotion, with organizations disseminating information via official social media accounts\u003csup\u003e\u003cspan additionalcitationids=\"CR40\" citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e\u003c/sup\u003e and researchers initiating social media campaigns on health topics such as vaccination or cancer screening\u003csup\u003e\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e,\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e\u003c/sup\u003e. One challenge of the existing health campaigns on social media is that they sometimes fail to fully leverage the platforms\u0026rsquo; interactive potential, often focusing more on one-way dissemination than on actively engaging the audience\u003csup\u003e\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e,\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e\u003c/sup\u003e. A solution is to collaborate with social media content creators, who possess skills in producing appealing content, understanding their audiences\u0026rsquo; needs, and building trust and community\u003csup\u003e\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e,\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e\u003c/sup\u003e. A recent survey of creators reveals that they are most likely to use personal experience for content creation, highlighting an urgent need to enhance their media and information literacy skills, including the ability to identify and use reliable fact-checking resources\u003csup\u003e\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e\u003c/sup\u003e. The domain knowledge that health institutions can provide through creator training programs fills this content creation gap. Health institutions, platforms, and other organizations that wish to enhance online health communication may want to make such training materials available to more creators.\u003c/p\u003e\n\u003ch3\u003eLimitations of the current study\u003c/h3\u003e\n\u003cp\u003eOur study has several limitations. In the on-platform RCT, our analysis of mental health knowledge construction in user comments represents just one of the many ways that creator-engaged programs can impact users. Future research could explore other outcomes, such as increased social support or reduced self-diagnosis of mental health problems\u003csup\u003e\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e\u003c/sup\u003e. Moreover, in the on-platform RCT, our study did not examine if the impact of the training conducted in April 2023 would be sustained over the long term. Future research should aim to understand the longer-term effects of creator programs. Furthermore, our on-platform RCT was solely focused on TikTok. As such, we cannot empirically generalize our findings to other platforms. However, as discussed above, we hypothesize that similar findings could be observed on other platforms featuring personalized short-form videos like Instagram or YouTube.\u003c/p\u003e \u003cp\u003eIn the on-platform RCT, the finding that only foundational knowledge construction improved, while advanced stages did not, aligns with the nature of TikTok videos: they are typically short (less than 60 seconds) and involve only one round of interaction between creators and viewers. Advanced knowledge construction may require repeated exposure to more iterative discussions. Future studies should explore the impact of repeated exposures to evidence-based mental health content on user behaviors.\u003c/p\u003e \u003cp\u003eSimilar to the on-platform RCT, the off-platform survey experiment asked participants to watch a single short video (\u0026lt;\u0026thinsp;60 seconds), and we examined how one video influenced user outcomes. Whether treatment effects are enhanced in situations where social media users are exposed to messages reinforcing similar themes (\u0026ldquo;dosage effects\u0026rdquo;) or diluted in situations where they are exposed to competing messages (\u0026ldquo;competitive framing environments\u0026rdquo;) are also outside the purview of our study and remain a fruitful avenue for research. An important direction for future research is to investigate the benefits of sustained and reinforced training for content creators, as well as the cumulative effects of repeated exposure on audiences.\u003c/p\u003e \u003cp\u003eBesides, in the off-platform survey experiment, we study only a single (albeit important) aspect of mental health attitudes and behavior in the present research; emotional support competencies. Whether findings generated from this research generalize to other settings, such as communication about suicide, eating disorders, and other mental health topics, remains an important and potentially fruitful avenue for future research.\u003c/p\u003e \u003cp\u003eAdditionally, in the off-platform survey experiment, we partnered with just a single creator to measure the effects of evidence-based mental health communication. Our decision to partner with a lifestyle influencer was strategic; as lifestyle influencers often cultivate large followings that may exceed those of influencers focusing specifically on mental health, likely have less experience with emotional support training than creators who are mental health professionals, and have been shown to shape opinion in a wide range of social and health-related domains\u003csup\u003e\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e,\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e\u003c/sup\u003e. We encourage researchers in the future to replicate studies like this one to study whether other types of creators might be similarly powerful in motivating attitudinal and behavioral change.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThis study demonstrates the promise of engaging mental health content creators as influential community leaders for advancing mental health in online social networks. By leveraging a network-based strategy of \"influencing the influencers,\" we show that providing creators with training can meaningfully improve mental health behavioral outcomes in audiences. Evidence from both an on-platform RCT, analyzing 188,169 comments from 1,882 videos created by 49 mental health content creators on TikTok, and an off-platform survey experiment conducted with a large, demographically representative sample of 1,000 U.S. youth aged 14–22 supports the effectiveness of this approach. In the on-platform RCT, evidence-based mental health communication toolkits lead to an over 4% increase in foundational mental health knowledge construction among TikTok commenters, while in the off-platform survey experiment, training creators improved both perceived and objectively assessed emotional support competency among video viewers. Taken together, these findings highlight the potential of creator-focused interventions to promote positive mental health behavioral outcomes on video-based social media platforms. They also suggest that relatively simple, scalable resources, such as training toolkits, can be a powerful means of achieving public health impact through social media.\u003c/p\u003e "},{"header":"Method","content":"\u003cp\u003eThis study evaluates the effectiveness of creator-focused interventions on audiences through an on-platform RCT and an off-platform, survey-based RCT.\u003c/p\u003e\u003ch2\u003eOn-platform RCT\u003c/h2\u003e\u003cp\u003eFrom March to May 2023, the Harvard Chan School’s Center for Health Communication conducted an RCT testing the effects of training toolkits and training programs for identified mental health content creators (MHCCs) (see Method-Creator Training Program). After the RCT, we collected TikTok videos created by MHCCs and their associated user comments (see Method-Comment Data Collection). Based on the experimental assignment protocol embedded in our previous RCT research\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e, we requested all comments (n = 188,169) from all videos (n = 1,882) created by the 49 mental health content creators enrolled in the RCT in March, April, and May 2023 using TikTok’s Research Tools API\u003csup\u003e\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e. We developed a codebook to measure the study outcome, mental health knowledge construction, through content analysis (see Method – Outcome Measurement). We randomly sampled 54 TikTok videos, yielding 4,152 comments. Three research assistants coded the comments, achieving an average percent agreement of 0.81 and a Gwet’s AC of 0.80. To ensure a balanced number of comments with and without knowledge construction for LLM fine-tuning, we applied text augmentation that maintained a cosine similarity of 0.85 between the augmented and original texts (see Method – Text Augmentation). We then fine-tuned LLMs to classify foundational and advanced mental health knowledge construction, achieving over 90% accuracy and F1 scores for both (see Method – LLM Fine-Tuning). These models were then used to assess mental health knowledge construction across all comments in the dataset. Finally, we fitted statistical models to evaluate the effects of the creator training toolkits and programs on user knowledge construction (see Method – Analytical Strategy).\u003c/p\u003e\u003cp\u003eThe fine-tuned LLMs are available on HuggingFace: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://huggingface.co/chc-harvard\u003c/span\u003e\u003cspan address=\"https://huggingface.co/chc-harvard\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. The code and data for replicating the on-platform RCT analysis are available on The Open Science Framework at: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://osf.io/vbkh8/\u003c/span\u003e\u003cspan address=\"https://osf.io/vbkh8/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. The Harvard Longwood Campus Institutional Review Board determined that this study is not research as defined by DHHS regulations or FDA regulations. All procedures complied with applicable guidelines and regulations.\u003c/p\u003e\u003ch2\u003eCreator training program\u003c/h2\u003e\u003cp\u003eOur team built on a previous intervention by Harvard Center for Health Communication that identified a sampling frame of MHCCs and conducted a randomized control trial. The inclusion criteria of the MHCCs are: aged 18 or over, English-language mental health content, have at least 10,000 followers across TikTok or Instagram social media platforms, posted videos on the platform at least 4 times per month from December 2022 - February 2023, and have been active on the site since February 2022. MHCCs were randomly selected to participate in interventions held in April 2023, including a series of virtual summits and an online content creation toolkit around evidence-based mental health communication. Details of the randomization process and intervention are documented in our previous work\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e. The trial comprised three groups: one provided with an asynchronous online content creation toolkit (“Toolkits-Only” Condition, N = 17 creators), another receiving the toolkit plus a synchronous virtual summit component (“Sessions Plus Toolkits” Condition, N = 25 creators), and a control group (N = 20 creators).\u003c/p\u003e\u003ch2\u003eComment data collection\u003c/h2\u003e\u003cp\u003eTo evaluate the intervention's impact on user behavior, we obtained all videos and related comments from 62 participating MHCCs via the TikTok research API\u003csup\u003e\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e in August 2023, using Python version 3.9. From this group, we successfully retrieved 3,465 videos posted by 58 MHCCs in March, April, and May 2023. Additionally, we collected user information, such as the number of followers, for these MHCCs through the same API. Four creators were excluded from the analysis because they either did not produce videos during the study period or had deleted their videos at the time of the request. Videos without comments were also excluded from the study. We accessed comments from 2,858 videos using the TikTok research API. Note that the TikTok research API imposes data filters, excluding, for instance, “public data from users under 18 and data from Canada.” For videos not captured by the API, we manually collected comments, ultimately compiling comments from 2,901 videos. To align with our research focus on online mental health communication, we applied a filter to exclude videos not relevant to mental health discussions. This filter was based on the content analysis by research assistants in our previous study that determined whether each video is relevant to mental health communication\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e. Applying the germane-to-mental health filter resulted in a final sample of 1,882 videos and 188,169 comments from 49 MHCCs (toolkits-only: N = 14 creators, sessions + toolkits: N = 20 creators, control: N = 15 creators). The flow chart in Figure A1 presents the process of video and comments collection.\u003c/p\u003e\u003ch2\u003eOutcome measurement\u003c/h2\u003e\u003cp\u003eWe hypothesized that creator training programs enhance mental health knowledge construction in video comments. To measure knowledge construction, we utilized the Interaction Analysis Model (IAM)\u003csup\u003e\u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e,\u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e52\u003c/span\u003e\u003c/sup\u003e, a framework proven effective in exploring informal learning in social media interactions\u003csup\u003e\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e,\u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e\u003c/sup\u003e. IAM outlines learning progresses as users exchange ideas, confront dissonance, negotiate meanings, evaluate emerging understandings, and apply new knowledge to novel contexts, with each stage growing in complexity. Based on IAM and prior research, we crafted a measure for knowledge construction in TikTok video comments for mental health communication across six dimensions: personal reflection, expressing agreement, expressing disagreement, asking for clarifications, reinterpreting knowledge, and applying knowledge to new areas, with each dimension quantified as a binary variable. From this, we derived two binary measures: foundational mental health knowledge construction, coded as 1 if the comment expressing personal experiences or opinions, and advanced mental health knowledge construction, coded as 1 if the comment contains any expression of agreement or disagreement, a clarification question, reinterpretation of knowledge, or application to other contexts. Definitions and examples for these measures are detailed in Table A2.\u003c/p\u003e\u003ch2\u003eContent analysis\u003c/h2\u003e\u003cp\u003eWe conducted a content analysis of comments from a sample of TikTok videos to measure mental health knowledge construction. Given the hierarchical nature of the creator-video-comment data, we employed clustered randomized sampling to select comments for content analysis, resulting in 54 videos and 4,152 comments. Three research assistants from Harvard University were trained and served as coders for the project. After an initial code training session, which contained a series of iterative pilot coding and feedback sessions, each assistant was assigned to code approximately 9% of all sampled comments (376 comments from 5 randomly selected videos). In the content analysis, each assistant first watched the TikTok video, then started coding the comments. The RAs were not informed about whether the videos and comments are from creators in the treatment or control group in the experimental study.\u003c/p\u003e\u003cp\u003eInter-coder reliability (ICR) was assessed for these comments. The ICR results on the “triple assigned” comments highlighted just three outcome variables with percent agreement scores below 0.70: knowledge construction through personal reflection, expressing agreement, and knowledge reinterpretation. Based on the ICR results, a second training session was held to focus on improving coder agreement in the three outcomes. Coders subsequently re-coded a subset of 120 comments where discrepancies had been noted for the three outcomes. Following the two-round training sessions, ICR exceeded 80% agreement for all variables in the filtering session, and surpassed 70% for most knowledge construction variables—except for expressing agreement and knowledge reinterpretation, which both achieved a 68% agreement rate (Table A3). In general, coders achieved an average percent agreement of 0.81, with Gwet's AC of 0.80 (Table A3) across a random sample of 376 double-assigned comments from the total 4,152 comments. Upon completing the two training sessions, coders formally started the coding of video comments, with each coder handling approximately one-third of the selected comments from the training set.\u003c/p\u003e\u003ch2\u003eText augmentation\u003c/h2\u003e\u003cp\u003eUsing the true labels derived from this content analysis, we initially performed character, word-level, and LLM-based text augmentation to ensure a sufficiently large and balanced sample for model fine-tuning. We employed the textattack library to conduct character and word-level text augmentation\u003csup\u003e\u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e\u003c/sup\u003e. Specifically, at the character level, two substituted sentences were generated using transformations including WordSwapRandomCharacterDeletion, WordSwapRandomCharacterInsertion, WordSwapRandomCharacterSubstitution. At the word level, five augmented sentences were generated using transformations including WordInsertionRandomSynonym, WordSwapChangeLocation, WordSwapChangeName, WordSwapChangeNumber, WordInnerSwapRandom, WordSwapEmbedding, and WordSwapQWERTY. Both sets of transformations were subject to two constraints: Repeat Modification and Stopword Modification. For LLM-based text augmentation, five augmented sentences were generated by applying the model “gpt-4-turbo” in the Chat Completions API from OpenAI with temperature set as 0\u003csup\u003e55\u003c/sup\u003e. The prompts we used are presented in Appendix Section 1. All text augmentation analyses were conducted in Python 3.9.\u003c/p\u003e\u003cp\u003eWe apply cosine similarity to evaluate the performance of the augmented text. The original and augmented texts were embedded using the sentence_transformer library with the \"bert-base-nli-mean-tokens\" model, and the cosine similarity between the embeddings of the original and augmented texts was calculated. Summary statistics for the cosine similarity are presented in Table A6. The overall mean and median cosine similarity between the original and all augmented texts were 0.85 and 0.90, respectively (SD = 0.16). Specifically, the mean and median cosine similarity between the original text and character-level augmented texts were 0.88 and 0.91, respectively (SD = 0.10); for word-level augmented texts, the mean and median were 0.92 and 0.94 (SD = 0.08); and for LLM-based augmented texts, they were 0.76 and 0.83 (SD = 0.20). Overall, the results indicate that the augmented texts closely resemble the original texts, ensuring that the texts used in the downstream fine-tuning of the LLM are representative of the original content.\u003c/p\u003e\u003ch2\u003eLLM fine-tuning\u003c/h2\u003e\u003cp\u003eSubsequently, we fine-tuned LLMs for text classification tasks to measure mental health knowledge construction outcomes. In total, we fine-tuned 8 LLMs to facilitate text classification, with all outcomes treated as binary classification problems. These included 2 outcomes related to filtering questions (i.e., whether a comment only tagged another user or contained only emojis) and 6 outcomes related to knowledge construction (see: \"Method-Outcome Measurement\"). Fine-tuning was performed on the un-augmented text for the two filtering outcomes and on the augmented text for the remaining 6 outcomes. We ensured a balanced class distribution to avoid issues arising from unbalanced samples when using the augmented text by maintaining a roughly equal ratio of comments from both classes.\u003c/p\u003e\u003cp\u003eFor each outcome, we fine-tuned the following models: BERT (bert-base-uncased)\u003csup\u003e\u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e56\u003c/span\u003e\u003c/sup\u003e, RoBERTa (roberta-base)\u003csup\u003e\u003cspan citationid=\"CR57\" class=\"CitationRef\"\u003e57\u003c/span\u003e\u003c/sup\u003e, MentalBERT (mental/mental-bert-base-uncased)\u003csup\u003e58\u003c/sup\u003e, and MentalRoBERTa (mental/mental-roberta-base)\u003csup\u003e58\u003c/sup\u003e. The models were trained with a batch size of 32, for 20 epochs, and a learning rate of 2e-5, following the recommendations from the BERT developers\u003csup\u003e\u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e56\u003c/span\u003e\u003c/sup\u003e. Among the models, the BERT-based LLM consistently achieved the highest accuracy for each outcome, leading us to use the fine-tuned BERT models for downstream analyses and predictions. The performance of the BERT-based LLMs for each of the 8 outcomes, as well as the specific epoch at which each model achieved its highest accuracy, is detailed in Table A1.\u003c/p\u003e\u003cp\u003eFinally, we applied the fine-tuned LLMs to label all comments from the content creators included in the creator training program. We conducted text preprocessing by removing comments that consist solely of a single word (ie, ‘fine’), only emojis (ie, 😎), or only tagging users (ie, @userid_happypuppy), following the approach in previous studies\u003csup\u003e\u003cspan citationid=\"CR59\" class=\"CitationRef\"\u003e59\u003c/span\u003e,\u003cspan citationid=\"CR60\" class=\"CitationRef\"\u003e60\u003c/span\u003e\u003c/sup\u003e, as these comments provide no relevant information on our outcomes of interest. After the preprocessing, a total of 23,314 comments (12.39%) were excluded, resulting in 164,855 comments from 1,863 videos created by the 49 creators remaining in the following analysis, as shown in Figure A1.\u003c/p\u003e\u003ch2\u003eAnalytical strategy\u003c/h2\u003e\u003cp\u003eFor the toolkits-only condition, the pre period includes videos published from March 1, 2023, to before 9:30 a.m. ET on May 2, 2023. The post period includes videos published after this time through May 31, 2023. This cutoff reflects the time the toolkits were delivered to MHCCs via email. For the sessions + toolkits condition, the intervention, which consisted of online training sessions, was delivered at 10 a.m. ET on April 8, 2023. Videos published before this time are coded as pre, and those after as post.\u003c/p\u003e\u003cp\u003eWe aimed to assess whether comments on videos produced by MHCCs across the toolkits-only, sessions + toolkits, and control groups were more likely to demonstrate mental health knowledge construction. To achieve this, we employed three-level multilevel Linear Probability Models (LPM) with random effects at the influencer and video levels to account for potential systematic video-level differences in commenting behavior and/or creators’ differential responsiveness to our experimental treatments. The analysis included 164,855 comments nested within 1,863 videos created by 49 influencers. These models (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e) assessed the impact of our study interventions on the outcomes (foundational and advanced mental health knowledge construction) by interacting dichotomous fixed effect indicators of treatment group assignment (toolkits-only vs. sessions + toolkits, with the control group as the reference) with the fixed effect indicator of whether each video was produced pre- or post-intervention. The estimated prevalence of knowledge construction outcomes in treatment and control groups at the pre and post period are presented in Table A4.\u003c/p\u003e\u003cp\u003eTo examine how the toolkit and online training influenced user comments on evidence-based mental health content, we included four video-level moderators: (1) whether the video featured themes from the training toolkits (26.95% yes, 73.05% no, based on coding in our prior research\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e), (2) comment engagement rate (comments per view, dichotomized at the median of 0.97%), (3) like engagement rate (likes per view, dichotomized at the median of 0.02%), and (4) share engagement rate (shares per view, dichotomized at the median of 0.18%). We also included three creator-level moderators: (1) number of followers (38.78% with more than 500,000; 61.22% with fewer), (2) engagement rate, calculated as (comments + likes) per view from March to May 2023 (dichotomized at the median of 1.20%), and (3) whether the creator is a licensed mental health professional (63.27% yes, 36.73% no).\u003c/p\u003e\u003cp\u003eNote that a creator's number of followers, indicating content reach and popularity, does not directly correspond to engagement rate, which reflects audience involvement on online platforms\u003csup\u003e\u003cspan citationid=\"CR61\" class=\"CitationRef\"\u003e61\u003c/span\u003e\u003c/sup\u003e. We dichotomized the number of followers at a cutoff of 500k, a threshold at which creators typically begin to employ managers for account and content management. Furthermore, the mental health professional licensing status of each content creator was collected based on their self disclosure on their social media accounts or websites.\u003c/p\u003e\u003cp\u003eWe conducted subgroup analyses based on the four video-level and three creator-level indicators to examine whether the effect of the toolkits in the toolkits-only group was stronger in any subgroup. These analyses used the same three-level multilevel linear probability model. The parameter estimation of the moderated multilevel LPMs are shown in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. The estimated prevalence of knowledge construction outcomes in treatment and control groups at the pre and post period in different subgroups are presented in Table A5. The analyses were conducted using the lme4 package in R version 4.2.1 \u003csup\u003e62\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eWe also estimated the average marginal effect (AME) of interventions in outcomes among different subgroups of videos and creators. Visualized in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. We presented the distribution of post–pre changes in the outcome variable—the proportion of comments exhibiting foundational mental health knowledge construction—across four groups, defined by intervention condition (toolkits-only vs. control) and each subgroup. AME was estimated by the margins package in R\u003csup\u003e\u003cspan citationid=\"CR63\" class=\"CitationRef\"\u003e63\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003ch2\u003eOff-platform, survey-based RCT\u003c/h2\u003e\u003ch2\u003ePre-registered Study Design: Assessing the Effects of EBMHC via a “Creator-Engaged” Survey-Based RCT\u003c/h2\u003e\u003ch2\u003eExperimental protocol\u003c/h2\u003e\u003cp\u003eOur pre-registered survey-based RCT randomly exposed respondents to one of two video messages about how to emotionally support friends. Figure A3 presents a side-by-side comparison of the text featured in the control—developed prior to the administration of interventional training materials (described below)—and treatment videos—developed after the provision of those materials—assembled by a popular social media lifestyle influencer enlisted as a confidential collaborator in our study. Our “creator-engaged” experimental procedures are described in detail below. Registration materials are available at the following Open Science Framework page: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://osf.io/jsg9t\u003c/span\u003e\u003cspan address=\"https://osf.io/jsg9t\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/p\u003e\u003cp\u003eIn the first video, which serves as the study’s control, lifestyle content creator Kenzie Brenna, who has a cumulative following across TikTok and Instagram of 430,000, talks frankly and direct-to-camera about how she goes about supporting friends who are “going through something.”\u003c/p\u003e\u003cp\u003eThe second treatment video resembles the first, with the exception that it has been informed by our training materials and makes much more of an effort to feature evidence-based mental health content therein. For example, Brenna calls out the importance of both validating the friend’s feelings and asking them open-ended questions—critical evidence-backed elements of high-quality emotional support\u003csup\u003e\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eWe created these messages in partnership with Brenna. The hundreds of millions of mental health videos that a TikTok user might encounter are made by a diverse array of creators. Some are lifestyle creators like Brenna who also aim to reduce stigma by sharing their personal mental health experiences; others are licensed mental health providers like Shahem McLaurin (@5hahem) who aim to use their professional expertise to democratize access to high-quality mental health information. We hypothesized that our training might have the biggest impact when offered to creators who satisfy all three of the following criteria: 1) make the type of lifestyle content that TikTok and Instagram users are most likely to come across on these platforms 2) have not trained as a therapist or other licensed mental health provider 3) have demonstrated in past content a consistent interest in educating and serving their community. As a consequence, we chose to expose participants in our survey-based RCT to messages made in partnership with Brenna. Note that all recruitment messages are available as supplementary materials alongside this manuscript.\u003c/p\u003e\u003cp\u003eAfter agreeing to assist in our study, we asked Brenna to develop a video about “emotionally supporting friends” using conventional media production strategies and content elements that they otherwise would use when creating for and posting on the TikTok platform (which hosts videos of a similar style and length to other social networking sites, like Instagram). This video served as the study’s control condition.\u003c/p\u003e\u003cp\u003e After doing this, we provided Brenna with a series of materials we co-created with the nonprofit Active Minds. The materials were designed to help creators like Brenna spread evidence-based information on (1) how today’s youth mental health crisis has put more pressure on kids to provide emotional support to friends 2) why emotional support makes a difference and 3) the 3 key characteristics of high-quality emotional support: Acknowledge the person’s feelings. Support their feelings by validating their emotions and asking what they need. Keep in touch, and check in regularly.\u003c/p\u003e\u003cp\u003eUpon completion of the training sessions, we asked Brenna to make a second video using the evidence-based mental health information they learned from the training materials, and comport with best practices for effective mental health communication, while sticking as closely as possible to the video they created originally. Storyboards for both videos are presented below.\u003c/p\u003e\u003cp\u003eStable links to access both videos can be accessed at the following webpages:\u003c/p\u003e\u003cul\u003e \u003cli\u003e \u003cp\u003efor the control:\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e\u003cspan class=\"RefSource\"\u003ehttps://drive.google.com/file/d/1lSYaP8kmCpdUi97bXMMxS0eQ7L8hTsXF/view?usp=sharing\u003c/span\u003e\u003cul\u003e \u003cli\u003e \u003cp\u003efor the treatment\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e\u003cspan class=\"RefSource\"\u003ehttps://drive.google.com/file/d/1PfI6raNz1eluO6IZcK-uDIntQfVs0syw/view?usp=sharing\u003c/span\u003e\u003ch2\u003eAnalytical strategy\u003c/h2\u003e\u003cp\u003eWe assess the effectiveness of our experimental treatment by constructing multivariate regression models that regress dichotomous indicators (see: Method-Measures) of perceived helpfulness, informativeness, and the evidentiary basis of the videos respondents watched (Manipulation Check); measures of both perceived (H1) and objective (H2) emotional support competencies, and toward mental health content created on TikTok (RQ1), on a dichotomous indicator of experimental treatment assignment, as well as an indicator of how respondents were recruited to participate in the survey (see: Method-Data). Additional estimation information can be found throughout the results and Supplementary Materials.\u003c/p\u003e\u003cp\u003eNote that supplemental randomization checks (Table A11) demonstrate that assignment to our experimental treatments was well-balanced (i.e., we document no significant differences across treatment and control groups). Correspondingly, we do not account for any additional socio-demographic controls in our analysis.\u003c/p\u003e\u003cp\u003eFinally, we assess the possibility that more-regular social media users may be comparatively more (or less) receptive to our video treatments (RQ2) by amending these models to interact the dichotomous treatment assignment indicator with an ordinal measure of social media use frequency when assessing RQ2. Please refer to Table A8 for additional information about these questions.\u003c/p\u003e\u003ch2\u003eData\u003c/h2\u003e\u003cp\u003eData for this study were derived from a nationally representative survey of N = 1,000 American youth aged 14–22, administered via YouGov. YouGov worked with an external data provider to recruit a large online opt-in sampling frame of teens (“teen sample:” N = 500, aged 14–17), and relied on its own online opt-in panel to recruit college aged adults (“young adult sample:” N = 500, aged 18–22) to participate in this study. Note that teen participants were recruited via their parents, and that both teens and parents provided written consent in order to participate in this study. Institutional Review approval for this study was granted by Harvard University’s Longwood Campus Office of Regulatory Affairs \u0026amp; Research Compliance.\u003c/p\u003e\u003cp\u003eYouGov ensures national representativeness in the samples it draws by using propensity score matching techniques. To do this, YouGov first drew a simple random sample of teens and young adults from US Census data. This serves as a nationally representative cross-section of each subpopulation of interest. YouGov then used its proprietary propensity score matching algorithm to search for analogues from each online opt-in sampling frame that most closely match respondents drawn from the US Census on a wide range of demographic observables: including respondents’ age, racial identity, and gender identity. Additional demographic information about the composition of both the adult and teen samples can be found in the Supplementary Materials. Table A13 presents the demographic composition of the teen and adult samples.\u003c/p\u003e\u003ch2\u003eMeasures\u003c/h2\u003e\u003cp\u003eTable A8 summarizes the primary outcome and moderating variables used to test our theoretical expectations (outlined above). Column 1 provides information about the general concepts we hope to capture via the questions outlined in Column 3, with Column 2 serving as a reference regarding each measure’s role in our analysis.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe wish to thank Q Garcia, Amy Gatto, and Rita DeBateat Active Minds for their help in developing the training materials and survey questions used in the survey-based RCT.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;The survey-based RCT was supported by a gift to the Center for Health Communication from Showtime/MTV Entertainment Studios. The funders had no role in study design, data collection, data analysis, data interpretation, decision to publish, or preparation of the manuscript.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003cstrong\u003eAuthor contributions\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA.Y., M.M., and Y.L. contributed to conceptualization and study design; Y.L., M.M., A.Y., and E.E. contributed to data collection in the on-platform RCT; A.Y., M.M., K.B., S.M., K.S., and E.E. contributed to data collection in the off-platform survey experiment; Y.L. and M.M. contributed to data analysis and visualization; Y.L. and M.M. drafted manuscript; A.Y., M.M., and Y.L. interpreted the results and contribute to the theoretical framing; all authors conducted critical reviews. All authors have read and approved the final version of the manuscript.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare no competing interests, but provide the following information in the interests of transparency and full disclosure. AY is an unpaid advisor to the nonprofit Science To People. AY and the Center for Health Communication have received grant funding for other ongoing research projects from YouTube Health and the WellWithAll Foundation.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003cstrong\u003eMaterials \u0026amp; Correspondence.\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eCorrespondence and material requests should be addressed with Yuning Liu.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eKorda, H. \u0026amp; Itani, Z. Harnessing Social Media for Health Promotion and Behavior Change. \u003cem\u003eHealth Promotion Practice \u003c/em\u003e\u003cstrong\u003e14\u003c/strong\u003e, 15\u0026ndash;23 (2013).\u003c/li\u003e\n\u003cli\u003eStellefson, M., Paige, S. R., Chaney, B. H. \u0026amp; Chaney, J. D. Evolving Role of Social Media in Health Promotion: Updated Responsibilities for Health Education Specialists. \u003cem\u003eInternational Journal of Environmental Research and Public Health \u003c/em\u003e\u003cstrong\u003e17\u003c/strong\u003e, 1153 (2020).\u003c/li\u003e\n\u003cli\u003eAlex Montero, Grace Sparks, Marley Presiado, \u0026amp; Liz Hamel. KFF Health Misinformation Tracking Poll: Health and Election Issues on TikTok. (2024).\u003c/li\u003e\n\u003cli\u003eEdelman Trust Institute. 2025 Edelman Trust Barometer: Special Report - Trust and Health. \u003cem\u003eEdelman \u003c/em\u003ehttps://www.edelman.com/trust/2025/trust-barometer/special-report-health.\u003c/li\u003e\n\u003cli\u003eMotta, M., Liu, Y. \u0026amp; Yarnell, A. \u0026ldquo;Influencing the influencers:\u0026rdquo; a field experimental approach to promoting effective mental health communication on TikTok. \u003cem\u003eSci Rep \u003c/em\u003e\u003cstrong\u003e14\u003c/strong\u003e, 5864 (2024).\u003c/li\u003e\n\u003cli\u003eValente, T. W. Network Interventions. \u003cem\u003eScience \u003c/em\u003e\u003cstrong\u003e337\u003c/strong\u003e, 49\u0026ndash;53 (2012).\u003c/li\u003e\n\u003cli\u003eLatkin, C. A., Edwards, C., Davey-Rothwell, M. A. \u0026amp; Tobin, K. E. The relationship between social desirability bias and self-reports of health, substance use, and social network factors among urban substance users in Baltimore, Maryland. \u003cem\u003eAddict Behav \u003c/em\u003e\u003cstrong\u003e73\u003c/strong\u003e, 133\u0026ndash;136 (2017).\u003c/li\u003e\n\u003cli\u003eHunter, R. F. \u003cem\u003eet al.\u003c/em\u003e Social network interventions for health behaviours and outcomes: A systematic review and meta-analysis. \u003cem\u003ePLOS Medicine \u003c/em\u003e\u003cstrong\u003e16\u003c/strong\u003e, e1002890 (2019).\u003c/li\u003e\n\u003cli\u003eChoi, B., Kim, H. \u0026amp; Huh-Yoo, J. Seeking Mental Health Support Among College Students in Video-Based Social Media: Content and Statistical Analysis of YouTube Videos. \u003cem\u003eJMIR Formative Research \u003c/em\u003e\u003cstrong\u003e5\u003c/strong\u003e, e31944 (2021).\u003c/li\u003e\n\u003cli\u003eLee, N., Buchanan, K. \u0026amp; Yu, M. Each post matters: a content analysis of #mentalhealth images on Instagram. \u003cem\u003eJournal of Visual Communication in Medicine \u003c/em\u003e\u003cstrong\u003e43\u003c/strong\u003e, 128\u0026ndash;138 (2020).\u003c/li\u003e\n\u003cli\u003eWang, X., \u0026amp; Cohen, R. A. Health Information Technology Use Among Adults: United States, July\u0026ndash;December 2022. (2023).\u003c/li\u003e\n\u003cli\u003eBasch, C. H., Donelle, L., Fera, J. \u0026amp; Jaime, C. Deconstructing TikTok Videos on Mental Health: Cross-sectional, Descriptive Content Analysis. \u003cem\u003eJMIR Formative Research \u003c/em\u003e\u003cstrong\u003e6\u003c/strong\u003e, e38340 (2022).\u003c/li\u003e\n\u003cli\u003e#MentalHealth. \u003cem\u003eTikTok \u003c/em\u003ehttps://www.tiktok.com/tag/mentalhealth.\u003c/li\u003e\n\u003cli\u003eOoi, K.-B. \u003cem\u003eet al.\u003c/em\u003e Social media influencers: An effective marketing approach? \u003cem\u003eJournal of Business Research \u003c/em\u003e\u003cstrong\u003e160\u003c/strong\u003e, 113773 (2023).\u003c/li\u003e\n\u003cli\u003eCasal\u0026oacute;, L. V., Flavi\u0026aacute;n, C. \u0026amp; Ib\u0026aacute;\u0026ntilde;ez-S\u0026aacute;nchez, S. Influencers on Instagram: Antecedents and consequences of opinion leadership. \u003cem\u003eJournal of Business Research \u003c/em\u003e\u003cstrong\u003e117\u003c/strong\u003e, 510\u0026ndash;519 (2020).\u003c/li\u003e\n\u003cli\u003ePowell, J. \u0026amp; Pring, T. The impact of social media influencers on health outcomes: Systematic review. \u003cem\u003eSocial Science \u0026amp; Medicine \u003c/em\u003e\u003cstrong\u003e340\u003c/strong\u003e, 116472 (2024).\u003c/li\u003e\n\u003cli\u003eEddy, K. Republicans, young adults now nearly as likely to trust info from social media as from national news outlets. \u003cem\u003ePew Research Center \u003c/em\u003ehttps://www.pewresearch.org/short-reads/2024/10/16/republicans-young-adults-now-nearly-as-likely-to-trust-info-from-social-media-as-from-national-news-outlets/ (2024).\u003c/li\u003e\n\u003cli\u003eYouTube. THE-IQ Creator Program. https://health.youtube/features-application/#:~:text=THE%2DIQ%20Creator%20Program%20supports,YouTube%20%E2%80%93%20with%20messages%20that%20resonate.\u003c/li\u003e\n\u003cli\u003eUSC. Mental Health Storytelling Initiative. https://annenberg.usc.edu/research/annenberg-inclusion-initiative/mental-health-storytelling-initiative (2024).\u003c/li\u003e\n\u003cli\u003eXu, M. M. \u003cem\u003eet al.\u003c/em\u003e \u0026lsquo;Influencing the influencers\u0026rsquo;: Exploring the impact of an in-person summit on TikTokers\u0026rsquo; mental health communication habits and beliefs. Preprint at https://doi.org/10.31235/osf.io/v83ht (2024).\u003c/li\u003e\n\u003cli\u003eJanz, N. K. \u0026amp; Becker, M. H. The Health Belief Model: a decade later. \u003cem\u003eHealth Educ Q \u003c/em\u003e\u003cstrong\u003e11\u003c/strong\u003e, 1\u0026ndash;47 (1984).\u003c/li\u003e\n\u003cli\u003eLimbu, Y. B., Gautam, R. K. \u0026amp; Pham, L. The Health Belief Model Applied to COVID-19 Vaccine Hesitancy: A Systematic Review. \u003cem\u003eVaccines (Basel) \u003c/em\u003e\u003cstrong\u003e10\u003c/strong\u003e, 973 (2022).\u003c/li\u003e\n\u003cli\u003eNguyen, H. TikTok as Learning Analytics Data: Framing Climate Change and Data Practices. in \u003cem\u003eLAK23: 13th International Learning Analytics and Knowledge Conference\u003c/em\u003e 33\u0026ndash;43 (Association for Computing Machinery, New York, NY, USA, 2023). doi:10.1145/3576050.3576055.\u003c/li\u003e\n\u003cli\u003eBeautemps, J. \u0026amp; Bresges, A. What Comprises a Successful Educational Science YouTube Video? A Five-Thousand User Survey on Viewing Behaviors and Self-Perceived Importance of Various Variables Controlled by Content Creators. \u003cem\u003eFront. Commun. \u003c/em\u003e\u003cstrong\u003e5\u003c/strong\u003e, (2021).\u003c/li\u003e\n\u003cli\u003eZawacki, E. E., Bohon, W., Johnson, S. \u0026amp; Charlevoix, D. J. Exploring TikTok as a promising platform for geoscience communication. \u003cem\u003eGeoscience Communication \u003c/em\u003e\u003cstrong\u003e5\u003c/strong\u003e, 363\u0026ndash;380 (2022).\u003c/li\u003e\n\u003cli\u003eLatha, K., Meena, K. S., Pravitha, M. R., Dasgupta, M. \u0026amp; Chaturvedi, S. K. Effective use of social media platforms for promotion of mental health awareness. \u003cem\u003eJournal of Education and Health Promotion \u003c/em\u003e\u003cstrong\u003e9\u003c/strong\u003e, 124 (2020).\u003c/li\u003e\n\u003cli\u003eShahbaznezhad, H., Dolan, R. \u0026amp; Rashidirad, M. The Role of Social Media Content Format and Platform in Users\u0026rsquo; Engagement Behavior. \u003cem\u003eJournal of Interactive Marketing \u003c/em\u003e\u003cstrong\u003e53\u003c/strong\u003e, 47\u0026ndash;65 (2021).\u003c/li\u003e\n\u003cli\u003eNguyen, H. \u0026amp; Diederich, M. Facilitating knowledge construction in informal learning: A study of TikTok scientific, educational videos. \u003cem\u003eComputers \u0026amp; Education \u003c/em\u003e\u003cstrong\u003e205\u003c/strong\u003e, 104896 (2023).\u003c/li\u003e\n\u003cli\u003eTikTok. Research API | TikTok for Developers. https://developers.tiktok.com/products/research-api/.\u003c/li\u003e\n\u003cli\u003eMotta, M. \u003cem\u003eet al.\u003c/em\u003e Assessing the Effect of Evidence-Based Mental Health TikTok Videos on Youth Emotional Support Competencies: A \u0026ldquo;Creator-Engaged\u0026rdquo; Approach. Preprint at https://doi.org/10.31235/osf.io/kydf4 (2024).\u003c/li\u003e\n\u003cli\u003eHigh, A. C. \u0026amp; and Dillard, J. P. A Review and Meta-Analysis of Person-Centered Messages and Social Support Outcomes. \u003cem\u003eCommunication Studies \u003c/em\u003e\u003cstrong\u003e63\u003c/strong\u003e, 99\u0026ndash;118 (2012).\u003c/li\u003e\n\u003cli\u003eCannava, K. \u0026amp; Bodie, G. D. Language use and style matching in supportive conversations between strangers and friends. \u003cem\u003eJournal of Social and Personal Relationships \u003c/em\u003e\u003cstrong\u003e34\u003c/strong\u003e, 467\u0026ndash;485 (2017).\u003c/li\u003e\n\u003cli\u003ePriem, J. S. \u0026amp; Solomon, D. H. What Is Supportive About Supportive Conversation? Qualities of Interaction That Predict Emotional and Physiological Outcomes. \u003cem\u003eCommunication Research \u003c/em\u003e\u003cstrong\u003e45\u003c/strong\u003e, 443\u0026ndash;473 (2018).\u003c/li\u003e\n\u003cli\u003eBrambor, T., Clark, W. R. \u0026amp; Golder, M. Understanding Interaction Models: Improving Empirical Analyses. \u003cem\u003ePolitical Analysis \u003c/em\u003e\u003cstrong\u003e14\u003c/strong\u003e, 63\u0026ndash;82 (2006).\u003c/li\u003e\n\u003cli\u003eMurariu, S. From Vine to TikTok: The Evolution of Short-Form Video and the Rise of a Social Media Juggernaut. \u003cem\u003eBootcamp \u003c/em\u003ehttps://medium.com/design-bootcamp/from-vine-to-tiktok-the-evolution-of-short-form-video-and-the-rise-of-a-social-media-juggernaut-98967a7d8d7e (2023).\u003c/li\u003e\n\u003cli\u003eGerbaudo, P. TikTok and the algorithmic transformation of social media publics: From social networks to social interest clusters. \u003cem\u003eNew Media \u0026amp; Society\u003c/em\u003e 14614448241304106 (2024) doi:10.1177/14614448241304106.\u003c/li\u003e\n\u003cli\u003eDodds, D. Council Post: Short-Form Video Content: Capturing Attention In The Digital Age. \u003cem\u003eForbes \u003c/em\u003ehttps://www.forbes.com/councils/forbesagencycouncil/2024/03/19/short-form-video-content-capturing-attention-in-the-digital-age/ (2024).\u003c/li\u003e\n\u003cli\u003eDuan, Y. \u003cem\u003eet al.\u003c/em\u003e Comparing Climate Change Content and Comments across Instagram Reels, TikTok, and YouTube Shorts and Long Videos. \u003cem\u003eProceedings of the Association for Information Science and Technology \u003c/em\u003e\u003cstrong\u003e61\u003c/strong\u003e, 103\u0026ndash;114 (2024).\u003c/li\u003e\n\u003cli\u003eBrownson, R. C., Eyler, A. A., Harris, J. K., Moore, J. B. \u0026amp; Tabak, R. G. Getting the Word Out: New Approaches for Disseminating Public Health Science. \u003cem\u003eJournal of Public Health Management and Practice \u003c/em\u003e\u003cstrong\u003e24\u003c/strong\u003e, 102 (2018).\u003c/li\u003e\n\u003cli\u003eHeldman, A. B., Schindelar, J. \u0026amp; Weaver, J. B. Social Media Engagement and Public Health Communication: Implications for Public Health Organizations Being Truly \u0026ldquo;Social\u0026rdquo;. \u003cem\u003ePublic Health Rev \u003c/em\u003e\u003cstrong\u003e35\u003c/strong\u003e, 1\u0026ndash;18 (2013).\u003c/li\u003e\n\u003cli\u003eHunt, I. de V. \u0026amp; Linos, E. Social Media for Public Health: Framework for Social Media\u0026ndash;Based Public Health Campaigns. \u003cem\u003eJournal of Medical Internet Research \u003c/em\u003e\u003cstrong\u003e24\u003c/strong\u003e, e42179 (2022).\u003c/li\u003e\n\u003cli\u003eFaus, M., Alonso, F., Javadinejad, A. \u0026amp; Useche, S. A. Are social networks effective in promoting healthy behaviors? A systematic review of evaluations of public health campaigns broadcast on Twitter. \u003cem\u003eFront. Public Health \u003c/em\u003e\u003cstrong\u003e10\u003c/strong\u003e, (2022).\u003c/li\u003e\n\u003cli\u003eShi, J., Poorisat, T. \u0026amp; Salmon, C. T. The Use of Social Networking Sites (SNSs) in Health Communication Campaigns: Review and Recommendations. \u003cem\u003eHealth Communication \u003c/em\u003e\u003cstrong\u003e33\u003c/strong\u003e, 49\u0026ndash;56 (2018).\u003c/li\u003e\n\u003cli\u003eHerrera-Peco, I., Jim\u0026eacute;nez-G\u0026oacute;mez, B., Pe\u0026ntilde;a Deudero, J. J., Benitez De Gracia, E. \u0026amp; Ruiz-N\u0026uacute;\u0026ntilde;ez, C. Healthcare Professionals\u0026rsquo; Role in Social Media Public Health Campaigns: Analysis of Spanish Pro Vaccination Campaign on Twitter. \u003cem\u003eHealthcare \u003c/em\u003e\u003cstrong\u003e9\u003c/strong\u003e, 662 (2021).\u003c/li\u003e\n\u003cli\u003eGlotfelter, A. Algorithmic Circulation: How Content Creators Navigate the Effects of Algorithms on Their Work. \u003cem\u003eComputers and Composition \u003c/em\u003e\u003cstrong\u003e54\u003c/strong\u003e, 102521 (2019).\u003c/li\u003e\n\u003cli\u003eSi Willmore. Becoming a successful social media content creator. https://memberful.com/blog/social-media-content-creator/ (2024).\u003c/li\u003e\n\u003cli\u003eHa, Louisa. \u003cem\u003eBehind the Screens: Insights from Digital Content Creators; Understanding Their Intentions, Practices and Challenges - UNESCO Digital Library\u003c/em\u003e. https://unesdoc.unesco.org/ark:/48223/pf0000392006 (2024).\u003c/li\u003e\n\u003cli\u003eFoulkes, L. The problem with mental health awareness. \u003cem\u003eThe British Journal of Psychiatry \u003c/em\u003e\u003cstrong\u003e225\u003c/strong\u003e, 337\u0026ndash;338 (2024).\u003c/li\u003e\n\u003cli\u003eHasell, A. \u0026amp; Chinn, S. The Political Influence of Lifestyle Influencers? Examining the Relationship Between Aspirational Social Media Use and Anti-Expert Attitudes and Beliefs. \u003cem\u003eSocial Media + Society \u003c/em\u003e\u003cstrong\u003e9\u003c/strong\u003e, 20563051231211945 (2023).\u003c/li\u003e\n\u003cli\u003eChinn, S. \u0026amp; and Hasell, A. How Different Uses of Social Media Inform Perceptions of Offline Social Norms and Changes in Vaccine Intentions. \u003cem\u003eHealth Communication \u003c/em\u003e\u003cstrong\u003e39\u003c/strong\u003e, 1198\u0026ndash;1208 (2024).\u003c/li\u003e\n\u003cli\u003eGunawardena, C. N., Lowe, C. A. \u0026amp; Anderson, T. Analysis of a Global Online Debate and the Development of an Interaction Analysis Model for Examining Social Construction of Knowledge in Computer Conferencing. \u003cem\u003eJournal of Educational Computing Research \u003c/em\u003e\u003cstrong\u003e17\u003c/strong\u003e, 397\u0026ndash;431 (1997).\u003c/li\u003e\n\u003cli\u003eGunawardena, C., Flor, N., Gomez, D. \u0026amp; Sanchez, D. Analyzing Social Construction of Knowledge Online by Employing Interaction Analysis, Learning Analytics, and Social Network Analysis. \u003cem\u003eThe Quarterly Review of Distance Education \u003c/em\u003e\u003cstrong\u003e17\u003c/strong\u003e, 35\u0026ndash;60 (2016).\u003c/li\u003e\n\u003cli\u003eHaythornthwaite, C. \u003cem\u003eet al.\u003c/em\u003e Learning in the wild: coding for learning and practice on Reddit. \u003cem\u003eLearning, Media and Technology \u003c/em\u003e\u003cstrong\u003e43\u003c/strong\u003e, 219\u0026ndash;235 (2018).\u003c/li\u003e\n\u003cli\u003eMorris, J. X. \u003cem\u003eet al.\u003c/em\u003e TextAttack: A Framework for Adversarial Attacks, Data Augmentation, and Adversarial Training in NLP. Preprint at https://doi.org/10.48550/arXiv.2005.05909 (2020).\u003c/li\u003e\n\u003cli\u003eOpenAI \u003cem\u003eet al.\u003c/em\u003e GPT-4 Technical Report. Preprint at https://doi.org/10.48550/arXiv.2303.08774 (2024).\u003c/li\u003e\n\u003cli\u003eDevlin, J., Chang, M.-W., Lee, K. \u0026amp; Toutanova, K. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. Preprint at https://doi.org/10.48550/arXiv.1810.04805 (2019).\u003c/li\u003e\n\u003cli\u003eLiu, Y. \u003cem\u003eet al.\u003c/em\u003e RoBERTa: A Robustly Optimized BERT Pretraining Approach. Preprint at https://doi.org/10.48550/arXiv.1907.11692 (2019).\u003c/li\u003e\n\u003cli\u003eJi, S. \u003cem\u003eet al.\u003c/em\u003e MentalBERT: Publicly Available Pretrained Language Models for Mental Healthcare. Preprint at https://doi.org/10.48550/arXiv.2110.15621 (2021).\u003c/li\u003e\n\u003cli\u003eM\u0026uuml;ller, M., Salath\u0026eacute;, M. \u0026amp; Kummervold, P. E. COVID-Twitter-BERT: A natural language processing model to analyse COVID-19 content on Twitter. \u003cem\u003eFront. Artif. Intell. \u003c/em\u003e\u003cstrong\u003e6\u003c/strong\u003e, (2023).\u003c/li\u003e\n\u003cli\u003ePota, M., Ventura, M., Catelli, R. \u0026amp; Esposito, M. An Effective BERT-Based Pipeline for Twitter Sentiment Analysis: A Case Study in Italian. \u003cem\u003eSensors \u003c/em\u003e\u003cstrong\u003e21\u003c/strong\u003e, 133 (2021).\u003c/li\u003e\n\u003cli\u003ePourazad, N., Stocchi, L. \u0026amp; Narsey, S. A Comparison of Social Media Influencers\u0026rsquo; KPI Patterns across Platforms: Exploring Differences in Followers and Engagement On Facebook, Instagram, YouTube, TikTok, and Twitter. \u003cem\u003eJournal of Advertising Research \u003c/em\u003e\u003cstrong\u003e63\u003c/strong\u003e, 139\u0026ndash;159 (2023).\u003c/li\u003e\n\u003cli\u003eBates, D. \u003cem\u003eet al.\u003c/em\u003e lme4: Linear Mixed-Effects Models using \u0026lsquo;Eigen\u0026rsquo; and S4. (2024).\u003c/li\u003e\n\u003cli\u003eBen Bolker. \u003cem\u003eAn Introduction to \u0026lsquo;Margins\u0026rsquo;\u003c/em\u003e. https://cran.r-project.org/web/packages/margins/vignettes/Introduction.html (2024).\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-6787693/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6787693/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThe rapid rise of social media presents new opportunities for implementing network interventions to improve health outcomes\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e,\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e. In the mental health domain, content creators can serve as influential community leaders\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e,\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e, yet it remains unclear whether creator-focused training impacts the mental health attitudes and behaviors of their audiences. This study evaluates the effectiveness of creator-focused interventions on the mental health knowledge and knowledge-based skill bases of their audiences using two experiments. First, an on-platform randomized controlled trial (RCT) provided evidence-based mental health communication toolkits and training to TikTok creators\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e. Analyzing 188,169 comments from 1,882 videos by 49 creators (March\u0026ndash;May 2023), we found a 4% increase in foundational mental health knowledge construction among viewers\u0026mdash;defined as the initial stage of learning through reflection on personal experiences or opinions. Second, an off-platform survey experiment exposed a nationally representative sample of 1,000 U.S. youth (aged 14\u0026ndash;22) to pre- and post-training videos created by a lifestyle influencer. Participants who viewed the post-training video showed significant improvements in both perceived and objectively assessed emotional support skills. By combining the external validity of the on-platform RCT with the internal validity of the survey experiment, our study provides robust evidence that creator training toolkits can enhance mental health knowledge and competencies among audiences, supporting the promise of scalable, evidence-based communication strategies on social media.\u003c/p\u003e","manuscriptTitle":"Training TikTok creators in mental health communication can benefit their audiences","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-06-09 08:49:25","doi":"10.21203/rs.3.rs-6787693/v1","editorialEvents":[],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"2842a0f8-17bd-40e2-801b-91d75527029c","owner":[],"postedDate":"June 9th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":49479149,"name":"Scientific community and society/Social sciences/Interdisciplinary studies"},{"id":49479150,"name":"Scientific community and society/Social sciences/Communication"},{"id":49479151,"name":"Health sciences/Risk factors"}],"tags":[],"updatedAt":"2025-06-09T08:49:25+00:00","versionOfRecord":[],"versionCreatedAt":"2025-06-09 08:49:25","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6787693","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6787693","identity":"rs-6787693","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}