We propose and define 6 constructs in our proposed model: active participation, passive participation, anxiety, self-efficacy, protective behavior, and trust.

First and second, we propose “active participation” and “passive participation” to qualify the engagement of social media users, respectively. The former refers to the publication of COVID-19 content on social media, whereas the latter to its reception and consumption. These constructs follow the discussion in Chen et al [13] and Yoo et al [31]. Users are likely to blend different forms of engagement, from more active ones such as forwarding, replying to, and creating posts to more passive ones such as reading or “liking” a post.

Third, we consider “anxiety,” defined as the feeling of tension, worried thoughts, and physical changes aroused when reading COVID-19 materials on social media.

Fourth, we define self-efficacy as an individual’s belief in his/her ability to protect herself/himself against COVID-19 [25].

Fifth, modifying Yoo et al [31], protective behavior refers to the adoption of behaviors deemed beneficial to COVID-19 prevention.

Finally, we define trust as a firm belief in the integrity of social media platforms and their content.

The 6 constructs were measured in terms of reflective indicators as captured in responses to questions and are listed in Multimedia Appendix 1. We defined the measurement items according to Chen et al [13], Chan et al [28], Mahmood et al [29], Yoo et al [31], Plohl and Musil [40], Liu and Liu [44], Padidar et al [45], and World Health Organization [46].

All responses were collected on a Likert scale. Responses to questions meant to quantify active participation and passive participation used a Likert scale with the following subjective frequency levels: 5=very frequently, 4=frequently, 3=sometimes, 2=rarely, and 1=never. Questions quantifying the 4 other constructs used Likert scales with the following agreement levels: 5=strongly agree, 4=agree, 3=neutral, 2=disagree, and 1=strongly disagree.

We then used structural equation modeling to test the hypotheses [47,48]. Structural equation modeling is a technique that simultaneously models the relationships between composite items and their constructs and the relationships between constructs. All hypothesis relationships are modeled as linear regressions. Structural equation techniques differ according to the estimation procedure [49]. This study used partial least square path modeling to fit the model to the data, which uses a variance-based optimization criterion with limited information [47]. In essence, it consists of iteratively fitting linear regressions until the coefficients of each of the regressions converge between iterations.

We used an online survey approach driven by the COVID-19 pandemic, as this is expected to reach wider respondents. The questionnaire was prepared in Indonesian language with the expectation that the respondents will easily understand the questions. Prior to the survey distribution, we carried out a pilot study involving 30 respondents who have actively used social media. The Cronbach α value for each variable was above .7. Moreover, we conducted an online survey from February 28, 2022, to March 28 2022. The respondents were all Indonesian residents. The survey instrument was prepared in Indonesian and contained 50 questions organized into 4 sections corresponding to demographics, frequency of social media participation, perceptions of COVID-19 social media content, and additional details on social media behavior. Links to the survey were distributed through popular social media platforms in Indonesia, such as Instagram, Facebook, Twitter, WhatsApp, Telegram, and Line (NHN Japan [now Line Corporation]). Respondents were encouraged to share the links to the survey with their social networks; thus, we used snowball sampling to make it easier to find trusted respondents from the network of friends that each respondent has. The age of users ranged from 18 to 34 years, and comprised approximately 65% (~110 million) of the total active users of social media [10]. In total, we obtained 416 complete responses, 2 of which were discarded as they had been submitted by respondents not residing in Indonesia in the period of interest.

This study has received approval from the Faculty of Computer Science, University of Indonesia and all respondents have agreed to participate in this study. No ethics board review was sought as all respondent data are anonymous and the data can only be used for the purposes of this research.

The survey responses were evenly distributed across gender. There was a 2-peaked income distribution with mass at both extremes (<IDR 3 million and >IDR 7 million; IDR 1=US $0.000066); 7 out of 10 respondents resided in Jakarta (the capital city of Indonesia). More than one-half of the respondents were under the age of 30 and undergraduate students. According to Kemp [10], 93.8% of Indonesians view/post content on YouTube, 86.6% use Instagram, and 85.5% post on Facebook. While most respondents used Instagram and YouTube, Facebook, TikTok, and Twitter were much less popular; 68.4% (283/414) of respondents followed accounts from news organizations, and only a minority followed accounts from government or private organizations. A vast majority sought influencers’ content and health information. Table 1 presents a summary of the respondents’ demographics.

The number of items k used for each construct and their internal consistency as measured by Dijkstra and Henseler ρA are displayed in Table 2. All constructs are internally consistent, attaining ρA well above the recommended value of 0.7 [47]. Further, all pairs of constructs displayed a heterotrait-monotrait ratio below the recommended amount of 0.85 as shown in Table 3. The heterotrait-monotrait ratio is a measure of discriminant validity to check whether constructs are not equivalent. It is constructed as the ratio between the mean correlation of constructs across indicators and the mean correlation of constructs within indicators [47].

Hypotheses H1-H8 are represented in terms of the direct acyclic graph (Figure 1). Each node represents a dependent variable in a linear regression in which the sources of its incoming edges are the independent variables. To estimate the model, we standardized each variable such that its mean is equal to 0 and the SD is equal to 1. The estimated coefficients are displayed in Table 3. The table also displays SEs that are computed via bootstrapping with 5000 repetitions; the significance level presented is based on the assumption that the t-statistics coefficient divided by SE approximates a standard normal distribution.

The average variance extracted (AVE) for all fitted constructs is above the recommended level of 0.5 [47]. The AVE can be interpreted as the average loading on each construct. Lower AVEs suggest that some constructs might have weak item loading. Table 3 shows that “protective behavior” has a loading of 0.581, which might indicate a lower agreement between its composite items. However, due to the disparate nature of protective measures ranging from frequent handwashing to vaccination, it is reasonable to expect a lower AVE.

Concerning the fitted coefficients, we observed an agreement with hypothesis H1a, which posits a positive relationship between passive social media participation and anxiety. By contrast, hypotheses H1b (P=.45) and H2 (P=.001) do not agree with the data as there is no statistically significant relationship between social media contribution and self-efficacy and there is a negative relationship with protective behavior. Next, hypothesis H3b is satisfied, whereas H3a (P=.67) and H4 (P=.15) are not. There is a positive association of passive social media participation with self-efficacy, but no significant association with either anxiety or protective behavior. We also observed an agreement with hypothesis H5b, which posits an association between self-efficacy and protective behavior. Taken together, the previous 2 results suggest that self-efficacy works as a mediator between social media consumption and protective behavior. By contrast, we observed no statistically significant relationship between anxiety and protective behavior invalidating hypothesis H5a (P=.15).

Moreover, we wanted to uncover the roles that trust in social media could play on the other variables. Indeed, we found that trust is positively and significantly associated with all of our constructs in agreement with hypotheses H6a (P<.001), H6b (P<.001), H7a (P=.02), H7b (P<.001), and H8 (P<.001). However, there are notable differences. In particular, the coefficients of “trust” on “passive participation” and “self-efficacy” are close to 0.4 of an SD, which indicates that higher degrees of trust have a strong association with the use of social media as a medium of information acquisition. The coefficients of “trust” on “active participation” and “protective behavior” are smaller compared with the previous 2. By contrast, trust is only weakly associated with anxiety.

As we observed a lack of relationship between the constructs “active participation” and “self-efficacy,” we fitted an alternative model in which the causality between those variables is reversed, and there is a direct path between “passive participation” and “active participation.” In this alternative model, we found a statistically significant coefficient between “passive participation” and “active participation” equal to 0.354 (P<.001). The relationship between “active participation” and “self-efficacy” remained insignificant (P=.47) and there were no changes to the fit of the other dependent constructs.

With 6 constructs under investigation, there are just above 3 million possible networks. It is thus computationally feasible to evaluate the fit of all possible models defined as a direct acyclic graph. As a robustness check, we investigated whether our proposed model could capture most of the variation in the data as compared with alternatives based on the Akaike information criterion (AIC). Although seldom used for exploratory analysis, structural equation modeling is suited for this sort of exercise [49]. Neither the AIC nor any other similar measure was used as a selection criterion, and this exercise was only meant to compare the performance of our model with all other alternatives to provide some perspective.

The AIC for our model is presented in Table 3. The main target of our research is explaining “protective behavior” with regard to social media usage. We find that our proposed model for “protective behavior” falls in the 95th percentile of the AIC distribution for all models for this variable, which obtains a minimum of –75.1. The AIC penalizes models with a larger number of variables. However, given that most of the covariates for “protective behavior” are significant as displayed in Table 3, we obtain a strong fit for this variable. Besides, the AIC for self-efficacy and protective behavior is in the 75th percentile. Conversely, the fit for active and social media consumption falls in the 25th percentile, which is expected because we did not focus on explaining these variables.

This study shows that passive and active participation in social media has diametrically opposite effects. On the one hand, the active contribution of COVID-19 content in social media is detrimental to public health outcomes because it is associated positively with anxiety, negatively with protective behavior, and does not affect self-efficacy. Our results align with Yoo et al [31], who also found a negative direct relationship between active participation and protective behavior in South Korea. Although Liu and Liu [44] modeled the reverse directionality between active participation and anxiety, they also found a positive relationship between anxiety and active participation in China. Anxious participants in Belgium used social media more often as a coping mechanism for COVID-19 [50], and Belgium adolescents used social media to deal with anxious feelings during the COVID-19 quarantine period [50]. O’Day and Heimberg [5] found that people with social anxiety search social support from social media due to the lack of in-person support. In addition, based on Tsao et al [1], positive social media content linked to positive behavioral shifts of social media users, such as stay home and social distancing content messages, in this study.

On the other hand, this study finds that passive participation in social media was positively associated with public health outcomes. We identified a strong positive relationship between passive participation and self-efficacy, but no statistically significant relationship between anxiety and protective behavior. Appropriate practices toward COVID-19 are influenced by user’s good COVID-19 knowledge [51]. Thus, users who remained up to date with COVID-19 information through social media and other online channels will protect themselves [45]. Liu and Liu [44] described social media as a vital information channel that might positively influence people’s preventive behaviors. According to Shiloh et al [52], providing coping information and increasing behavior efficacy beliefs can effectively mobilize the public’s adoption of protective behaviors during COVID-19. However, this study presented results different from Zhang et al [27] in China and Mahmood et al [29] in Pakistan, which found that passive participation in social media influences anxiety and protective behavior.

Finally, this study shows that social media trust is a crucial antecedent where trust in social media is positively associated with active contribution and passive consumption of COVID-19 content in social media, users’ anxiety, self-efficacy, and protective behavior. Plohl and Musil [40] described that users’ trust in the entire social network and a member of the network is critical to their decisions to share knowledge in social networks. Therefore, users will choose social media where they find a network of trusted friends.

When trusted friends post messages concerning the efficacy or impact of various COVID-19 vaccines, personally connected people consider those messages with care. According to our study, the more trust users have in COVID-19 content on social media, the more protective behaviors they deploy. Therefore, social media trust indeed influences users’ compliance with COVID-19 protocol guidance [40]. Trust in others or social media peers (ie, governments, citizens, and news organizations) has been identified as a significant predictor of the willingness to cooperate, and trust toward fellow citizens is associated with prosocial behavioral intentions [53].

This study analyzes the effects of social media usage on user’s protective behavior against the COVID-19 epidemic in Indonesia, the fourth most populous country in the world. This study also investigated both active and passive participation in social media, shedding light on cofounding effects of these different forms of engagement. Moreover, this study analyzed the role of trust in social media platforms and its effect on public health outcomes. Thus, this study enriches the study of Wijayanti et al [17], Sujarwoto et al [18], and Maurizka et al [19]. This study also adds to the research study of Wang et al [16], which only focused on pregnant women searching for COVID-19 information. This study also considered the perspective of social media users, thus enriching the study of Huesch et al [11], Ahmed and Rasul [12], Chen et al [13], Lwin et al [14], and Buchanan et al [15], all of which only analyzed the secondary data from social media posts.

This study provides practical implications for public health campaigns and social media users. Public health campaigns should use social media to post health promotion content to make social media users more aware of protecting themselves from infectious diseases. Social media users should also be aware of their active participation on social media with the hope of releasing their anxieties. Passive participation on social media was positively associated with public health outcomes. Therefore, social media users should consume health information from credible sources such as news, governments, or peers’ social media accounts. Social media providers should also filter their content to increase their users’ trust. Social media providers, together with public health campaigns, should provide awareness and knowledge about how to filter credible messages to social media users. Currently, the number of COVID-19–positive cases is starting to decrease; accordingly, the results of this study can potentially be used for other infectious diseases that have handling characteristics like COVID-19, for example, for dengue fever and others that can be transmitted by air and have handling protocols similar to COVID-19.

The respondents in this study are mostly located in the greater Jakarta region and may not faithfully reflect the diversity of the population of Indonesia. Then, although there likely exist feedback loops between some investigated constructs, we refrain from modeling such loops because of limitations in structural equation modeling. For instance, higher levels of active participation may lead to higher levels of passive social media participation because those users that post content are likely to spend more time reading reactions to their material. The challenge with feedback loops is that they can be hard to identify statistically. Dijkstra and Henseler [54] proposed the consistent partial least square estimator, which uses 2-stage least squares to identify endogenous effects. Identifying feedback loops using this methodology requires additional assumptions on the coefficients of the structural model, which we leave for future work.

Opposite effects for active and passive participation in social media are found in this study. When social media is used passively, much like traditional media, we observe more positive public health outcomes. By contrast, active participation is associated with worse health outcomes. Thus, social media could be used as a medium for health promotion. However, public health campaigns must be aware of this reality when engaging with social media users. Future studies should further investigate social media interventions to make social media users more active in posting health-related information on social media. Moreover, they can analyze the tendency from social media consumption that could have an impact on social media users’ anxiety and fear.