Despite scientific consensus that COVID-19 vaccines are safe and effective [1], there is still widely circulated controversial information on social media, with statements such as “while vaccinations offer good protection, they do not provide full immunity, and the extent to which they would be effective against new variants of the virus remains uncertain,” which damages public confidence [2]. This misinformation leads to vaccine hesitancy, which has been recognized by the World Health Organization as a major global health threat [3]. Social media platforms such as Twitter, Facebook, TikTok, and YouTube provide an ideal medium for spreading and reinforcing antivaccine ideas [4-7]. First, the information-filtering mechanism based on algorithmic recommendation technology mediates and facilitates content promotion by considering users’ interest and attitudes [8]. Second, online users have access to a wealth of information and narratives. Affected by individual and social factors such as selective exposure and group pressure, users prefer to select information that fits their belief system, while ignoring dissident information. Gradually, echo chambers emerge, in which like-minded people continue to frame and strengthen shared narratives [9]. In the vaccine promotion campaign, the trend of simplification of users’ vaccine-information sources is strengthened and the flow of information between groups with different ideologies toward vaccines is blocked, which widens the knowledge gap and assimilates value cognition [10]. The accompanying group polarization and social fragmentation blind the public to preconceived misconceptions and undermine authorities’ efforts to improve the public’s information literacy [11], causing inefficiency in the vaccine campaign [12,13].

Users in a social network can be divided into three roles: opinion leaders, structural hole spanners, and ordinary users [14]. Lou and Tang [15] pointed out that the top 1% of users acting as structural hole spanners control almost 80% of information diffusion between communities and 25% of information diffusion on Twitter. Wu et al [16] revealed that 50% of URLs were posted by 1% of users serving as opinion leaders. Further, Cossard et al [17] identified key users in echo chambers, while Jeon et al [18] evaluated the characteristics of users who broke the echo chamber.

To avoid or break an echo chamber, it is critical to characterize these key users and determine their impact on topic dissemination and opinion evolution, which could facilitate the communication within and between pro- and antivaccine groups, and thereby eliminate vaccine hesitancy. Toward this end, in this study, we developed a framework to evaluate and compare the degree of the effect of different forms of echo chambers on users’ interactive behavior using quantitative measurements. We further explored the hidden mechanisms of an echo chamber’s formation and its strengthening or disintegration by detecting key users who occupy critical network positions, analyzing the relationship between their behavior pattern, network location, and function both inside and outside of echo chambers. Although this framework was designed based on online debates of COVID-19 vaccine hesitancy as the background to offer insights for public health administrators, it could also be applied and expanded to other controversial theme discussions to serve as a reference for public opinion managers.

Online echo chambers have been studied in the context of users’ interactions (eg, posting, retweeting, commenting, mentioning), focusing on rumor spread and management [42-44], political debates [29,45], and news consumption [46,47]; however, related studies on vaccine hesitancy are rare, especially during public health emergencies. To assess whether users on Sina Weibo, the most popular microblogging platform in China (with a structure similar to Twitter), exhibited an echo chamber effect when discussing COVID-19 vaccines and vaccinations, and to further understand the formation mechanism or to design strategies to break it, we sought to identify the key users, how their behavior patterns relate to their online positions, and how they cooperate or compete to promote (prevent) the formation or strengthening (breaking) of the echo chamber. Toward this end, we established the following research questions (RQs):

RQ1: Is there an echo chamber effect in topic selection and opinion contagion of users on Weibo when discussing COVID-19 vaccines and vaccinations? Does it differ between users’ retweeting and commenting behaviors?

RQ2: Do users with different behavior patterns on Weibo tend to be regarded as opinion leaders or structural hole spanners?

RQ3a: Do online opinion leaders and structural hole spanners tend to act as echoers or bridgers in topic dissemination?

RQ3b: Do online opinion leaders and structural hole spanners tend to act as echoers or bridgers in attitude interaction?

RQ4a: Do these key users acting as echoers in topic dissemination tend to play the same role in attitude interaction?

RQ4b: Do these key users acting as bridgers in topic dissemination tend to play the same role in attitude interaction?

Figure 1 outlines the research framework. Note that although Weibo posts were analyzed in this study, we use the terms “tweet” and “retweet” throughout the manuscript to refer to activities on the platform, equivalent to activities on Twitter, for the sake of convenience. An original “tweet” refers to posts created by a registered Weibo user. A “retweet” refers to users’ forwarding behavior on Weibo. “Comments” refer to replies to an original post on Weibo.

Our research did not require ethical board approval because it did not involve human or animal trials. The research data were derived from open access data available on social media, mainly through voluntary contributions from users. Our data and analysis of data were conducted in an unbiased and transparent manner, and the data were used only for scientific research without any ethical violations. To be specific, we anonymized key identifiable information, including the nickname field provided by each user and the ID number assigned to each user by the platform when they registered their unique account. We represented these two fields as nonrepeating consecutive integers incremented from 1 to uniquely identify each user, thus hiding the users’ personal information, which had no influence on the study results.

From January 23, 2020, to February 11, 2021, there were numerous messages posted about the outbreak and cessation of the COVID-19 epidemic, as well as the initial exploration of vaccine development and vaccination on Weibo [48]. As a medical innovation, the vaccine was widely debated in its early diffusion stage [49]. We first used the Sina Weibo application programming interface to collect original tweets containing keywords (“COVID-19 vaccine [新冠疫苗]” or “COVID-19 vaccination [新冠疫苗接种]”). Considering that the interactive data (ie, retweet, comment, and like) of an original tweet could become stable approximately 1 week after it was posted [50,51], we crawled the following-week interactive data for each original tweet, involving likes, retweets, and comments, and the information of posters, retweeting, and commenting users. There were initially 29,218 original tweets, 50,693 retweets, and 50,796 comments.

For data preprocessing, we deleted the original tweets without any text (eg, only pictures, videos, or audio) or those that were duplicated or contained the above keywords but did not include any meaningful content. In addition, blank or meaningless comments and their subcomments were also eliminated. After excluding the corresponding retweets and comments as well as the user information, there were 26,788 original tweets, 48,231 retweets, and 46,224 comments from 77,625 users retained for analysis.

To answer RQ1, we used Python’s NetworkX package to calculate each network’s assortativity coefficient r (–1≤r≤1) based on the nodes’ attributes (topic in information networks, attitude in user networks) and their interaction, which measures the network’s homophily [5,44]. An r>0 indicates that the node generally tends to connect with other nodes with similar properties, and the network is referred to as an assortative network. A larger r value indicates more prominent assortativity. If r≤0, assortativity does not hold [61].

Tweets about domestic, foreign status, and conspiracy accounted for 24.46% (n=29,653), 20.40% (n=24,734), and 16.46% (n=19,955) of total tweets (N=121,243), respectively. Overall, 42.51% (51,544/121,243) of tweets expressed a positive attitude toward vaccines and 13.12% (15,907/121,243) of tweets held a negative attitude. Figure 2 shows that discussions about domestic status were the least controversial, whereas discussions related to counterfeit vaccines and fraudulent information were the most divisive.

Retweet, comment, and global information networks were all sparse, with a density of 0.003, 0.003, and 0.0002, respectively. In Figure 3, the outer ring of 13 different colors represents a collection of 13 different topics of original tweets, the arc length represents the total connection volume for all of the original tweets belonging to this topic, and the inner colored connecting bands indicate the flow direction and magnitude of the data relationship. The top four topics that interacted most frequently with others were “Foreign status,” “Domestic status,” “Conspiracy,” and “Means.” “Foreign status” was often retweeted by users with topics such as “Domestic status,” “Conspiracy,” and “Means” at the same time, along with “Effectiveness, “Severity,” and “Risk.” The assortativity coefficients of retweet, comment, and global information networks were 0.060, 0.022, and 0.048, respectively, indicating low topic-based homogeneity and that the retweet information network displayed more obvious homogeneity compared with the comment information network.

Retweet, comment, and global user networks were also sparse, with densities lower than those of information networks. Compared with those of the retweet user network (0.003, 0.0003, 0.011), the comment user network had a higher clustering coefficient, transitivity, and reciprocity (0.007, 0.055, 0.025), indicating that the network built on comment relationships was more cohesive and stable, where users were closely connected and relatively stable [69], while retweeting was mostly used for a one-way flow of information [70]. As shown in Figure 4, in the three user networks, clusters brought together people who were confident about vaccines and people with uncertainty [71]. The more common edges were found between users holding a positive attitude and between users with a neutral attitude to users with a positive attitude. Users with a clear attitude hardly retweeted posts from users without a determined attitude. Compared with the retweet user network, the tendency of users with a negative/neutral attitude to comment on posts of other users with the same attitude was more obvious, whereas this tendency was less obvious for users with a positive attitude. The assortativity coefficients of the retweet, comment, and global user networks were 0.031, 0.042, and 0.055, respectively, indicating low attitude-based homogeneity and that the comment user network displayed more obvious homogeneity compared with the retweet user network.

As shown in Figure 5, most users were coded as “common user,” sending 86.9% of retweets and 93.9% of comments. Only 1.0% of users were speakers, but receiving 81.3% of retweets and 62.8% of comments. Most of the original tweets were created by isolators. Retweeters not only often created tweets but also frequently retweeted others’ tweets.

As shown in Figure 6, with respect to the weighted degree, most of the speakers’ weighted in-degree centrality was much higher than their weighted out-degree centrality. A different situation was found for retweeters. Speakers had a relatively higher average clustering coefficient (high speakers: 0.012, medium speakers: 0.013, low speakers: 0.007), whereas the average clustering coefficient of retweeters (0.0002) was the lowest among users (apart from isolators). Compared with that of retweeters, replicators had a higher average clustering coefficient (0.004), with in-degrees and out-degrees of similar size.

Figure 7 and Figure 8 show that influentials had more obvious structural hole properties than broadcasters. Among influentials, speakers had a higher effective size, efficiency, and lower constraint than networkers. High speakers performed in the same manner but with a much greater effect. However, some networkers had higher hierarchy than speakers, which indicated that a networker’s constraint was more concentrated on this actor and was more important. Among broadcasters, half of the monologists’ constraint values were lower than 0.500 and half of them had hierarchy values lower than 0.092. Although most of the replicators’ constraint values were lower than 0.333, they largely showed hierarchy values lower than 0.278. Compared with that of replicators, retweeters’ constraint was more concentrated in a single contact.

Given the massive network size, we considered the top 5% of users in weighted in-degree centrality and local clustering coefficient as opinion leaders (n=386, 0.5% of all users), and the other users in the bottom 5% in constraint were considered as structural hole spanners (n=3123, 4.0% of all users). These two types of users were considered key users. As shown in Figure 9, opinion leaders were responsible for 38.4% of all of the information flow, while structural hole spanners were responsible for 50.2% of the information flow. Compared with the former, the latter tended to create, retweet, and comment on more tweets.

As shown in Figure 10, common users, speakers, replicators, and networkers accounted for 44.0%, 36.3%, 10.6%, and 4.1% of opinion leaders, respectively. Common users, speakers, networkers, and retweeters accounted for 59.3%, 19.7%, 8.7%, and 8.7% of structural hole spanners, respectively. The χ² tests showed a significant difference in the distribution of categories of users between opinion leaders and structural hole spanners (χ²₇=184.650, P<.001). Posthoc testing further showed that speakers, replicators, and monologists tended to be opinion leaders, whereas common users, retweeters, and networkers tended to be structural hole spanners.

Isolators did not become opinion leaders or structural hole spanners, whereas 89.2% of isolaters were topic-based echoers and all of them were attitude-based echoers. The results of χ² tests showed that the proportion of structural hole spanners acting as topic-based bridgers (74.2%) was significantly higher than that of opinion leaders (64.2%) (χ²₁=17.148, P<.001). The opposite result (χ²₁=13.193, P<.001) was found when considering attitude-based bridgers (structural hole spanners: 88.1%; opinion leaders: 94.3%). Hence, compared with being echoers, both opinion leaders and structural hole spanners tended to act as bridgers. Structural hole spanners were more likely to become bridgers than opinion leaders in topic-based echo chambers, whereas structural hole spanners were less likely to become bridgers than opinion leaders in attitude-based echo chambers.

To address RQ4, the support and the confidence of the rule were calculated. As shown in Table 4, RQ4a was declined, whereas RQ4b was supported, with 62.8% of all key users acting as both topic-based and attitude-based bridgers. Approximately 85.9% of topic-based bridgers also acted as attitude-based bridgers. Specifically, 60.6% of opinion leaders (32.9% government accounts, 30.3% We-media, 19.2% traditional media) and 63.0% of structural hole spanners (39.4% common personnel, 28.4% We-media, 17.3% traditional media) acted as both topic-based and attitude-based bridgers.

Users showed an overall low echo chamber effect in vaccine-related topic selection and they tended to comment on more diverse topics than retweeting them. Discussions about the status of vaccine development, and vaccination at home and abroad, mostly mixed with conspiracy, largely caught users’ attention [72]. The risk of contracting COVID-19 and the serious consequences of refusing to be vaccinated were cocommented with claims of vaccine effectiveness.

In contrast to the findings of Mønsted and Lehmann [5] and Schmidt et al [6], users showed a low echo chamber effect in attitude interaction. Because the COVID-19 vaccine represents a medical innovation directly related to the safety of human life, a rational public, threatened by the public health emergency, was less bound by herd mentality [73]. As the dominant opinion, a positive attitude appealed to following of a neutral crowd, which helped to weaken the echo chamber. These findings are inconsistent with those of Rathje et al [19], possibly because national cultural backgrounds influence the cognition, decision-making, and interactive behavior of people belonging to different parties in the United States and United Kingdom. In addition, in contrast to the findings of Tsai et al [45], we found that the overall homophily was more obvious in commenting than in retweeting. Specifically, users approving vaccines showed a more significant tendency to interact with like-minded neighbors by retweeting than by commenting [43], while users against vaccines or with a neutral attitude acted more significantly by commenting than by retweeting, which suggested that the commenting mechanism might serve as an “anti-spiral of silence” to compete with a “silence spiral” in retweeting to form the global opinion climate [74]. Retweeting amplifies the visibility of individuals’ opinions [75], influenced by selective psychology, and opinions contrary to mainstream opinions are silenced. While the commenting network was more modularized and cohesive, users were under greater pressure from within their own communities.

The most common behaviors were helpful in spreading information (high percentages of common users and retweeters), while few users frequently participated in two-way dialogs (low percentage of replicators) [62]. Speakers were relatively scarce, but they created content provoking responses of others, which contributed to their popularity in the network, so as to be regarded as information centers within their communities. Networkers who demonstrated a balance between creating content, sharing content, and being retransmitted were more likely to fill structural holes to link otherwise less-connected communities. The commenting mechanism offered more chances to create cohesive communities, and hence nominate replicators in each cluster as opinion leaders, while the opposite situation was found for the retweeting mechanism and retweeters.

Consistent with Yang et al [14], opinion leaders and structural hole spanners tended to have a stronger influence than ordinary users. These two jointly played an important role in making the information propagate over a wider scale; the former affected their entire communities of the network, while the latter connecting to different communities affected the entire network [76]. Specifically, spanners initiated interactions proactively [77].

Tan et al [66] found that degree centrality and structural holes were complementary at enhancing an organization’s innovation performance in low-density networks. Similarly, we found that both opinion leaders and structural hole spanners played a positive role in breaking the echo chamber for topic dissemination and attitude contagion about COVID-19 vaccines. Opinion leaders insulated others against rather than exacerbated the echo chamber [35,36], which contradicts with the findings of Cossard et al [17]. As gatekeepers, because of social pressure and social support based in part on interpersonal trust [78], they were responsible for filtering, curating, and disseminating information they deemed relevant to their social circle to prevent their followers from being trapped in echo chambers. Structural hole spanners diffused information from one group to another, negotiated and synthesized different topics and standpoints, and promoted cooperation in diverse knowledge and ideological fields [79]. An et al [80] found that the same topic could breed multiple emotions and stakeholders with high topic influence that might not necessarily have high sentiment influence, which, to some extent, explained why users as topic-based echoers might not necessarily act as attitude-based echoers, while users as topic-based bridgers also tended to act as attitude-based bridgers in this study. Aware of the negative impact of echo chambers on crisis management and vaccine promotion, despite different cultural backgrounds, government and We-media positively promoted heterogeneity, and the traditional media’s agenda-setting power was also evident in both topic and opinion spread [29]. Moreover, Wagner and Reifegerste [81] declared that although isolators were rarely found in their interviews, since participants reported communicating about pandemic-related media coverage “with basically everyone,” some participants might turn into isolators during the trajectory of the pandemic. Our findings certainly confirmed this prediction. We found many isolators, which meant that they did not contribute to increasing the scale of information dissemination, which is distinct from the phenomenon noted in disaster-information diffusion [62]. However, the isolators were potential echoers by creating homogenized information and constantly reiterating a single point of view. Although no other users interacted with them, the words of isolators might invisibly reinforce the thoughts of others who saw or read their tweets.

The main theoretical contributions of this study are as follows. First, echo chambers in vaccine debates during a crisis differ from those related to general social issues. This study not only examined the echo chamber effect in different information-dissemination dimensions (topic, attitude) and based on different interactive mechanisms (retweeting, commenting), but also dug out the reasons for a low echo chamber effect from the perspective of the relationship of users’ network location and their function in preventing or breaking echo chambers. This offers a powerful complement to existing research focusing on echo chambers’ form, degree, formation, and depolarization.

Second, we focused on two types of key users, namely opinion leaders and structural hole spanners, and characterized their behavioral patterns, which could be a supplement for feature engineering of these key users’ detection or prediction. In addition, referring to the bonding and bridging relationship of social capital, this study proposes two new types of social mediators, namely echoers and bridgers, to quantify key users’ impact on echo chambers, thereby enriching the application scope of social capital theories. Hence, users could be classified based on their behavior, network location, impact on echo chambers, and stakeholder theory [63], offering insights for the construction of user portraits.

Third, previous studies about online key users either focused only on their antecedents (factors contributing to individuals occupying a central location/filling a structural hole [77,82,83]) or only on outcome variables (such as the impact of their locations on knowledge management and innovation performance [84,85], information diffusion [14,76], and emotion contagion [86]). This work linked key users’ antecedents and outcomes at the same time, which could be used to explore hidden behavioral paradigms.

Fourth, we analyzed the relationship of users’ roles in topic-based and attitude-based echo chambers, providing a new research perspective for the dissemination pattern of topic and sentiment.

Finally, most previous studies excluded users who did not interact with others in the data preprocessing step, ignoring their large-scale presence and potential influence on public opinion evolution. This study is thus the first to explore the impact of such users on echo chambers, which could offer a reference for further research about isolators.

First, although a low echo chamber effect existed in users’ selection of topics about vaccines, users tended to focus on some specific topics, namely the status of vaccine development, vaccination at home and abroad, and conspiracies. Health medical and public opinion managers should be aware of the emergence of echo chambers centered on these topics, which might damage international cooperation for vaccinations and epidemic control [87].

Second, users with neutral attitudes toward vaccines were easily influenced by others with determined standpoints. The interaction between opposing viewpoints remained limited. Managers should invite online opinion leaders and structural hole spanners who act as bridgers to offer multiple aspects of vaccine knowledge to correct opponents’ misunderstanding and improve their health literacy. At the same time, although provaccine sentiment, as the mainstream opinion, was largely spread and echoed in retweeting, managers should monitor the evolution of other opinions in commenting to prevent the wrong view from turning defeat into victory.

Third, echo chambers have been a major concern of the government, traditional media, and We-media. To obtain better effectiveness, these stakeholders should try to become opinion leaders or structural hole spanners according to their aims by adjusting their own usage behavior on social media. Our results showed that, compared with opinion leaders, structural hole spanners performed better in diffusing diversified topics, whereas opinion leaders performed better in bridging heterogeneous views.

Finally, online isolators should not be ignored. Although these users were reluctant to interact with others and did not receive any feedback from others, they showed interest in creating messages. They were also immersed in personal echo chambers. Managers should take specific measures to break these isolators’ echo chambers.

First, we simply divided users into two categories, namely echoers and bridgers, according to the rule as to whether the user spread more than one topic or interacted with cross-cutting neighbors, rather than quantifying the extent to which they acted as echoers/bridgers using continuous values. Further exploration is therefore warranted. Second, we did not manually find bot accounts in our data set, which was part of the strategy of Villodre and Criado [62] in their study on Twitter data. To date, no tool has been developed for robot account identification for Weibo. In future research, it will be important to develop automated detection algorithms for larger-scale data [88]. Bot accounts could be classified based on their behaviors such as posting repeatedly to appeal for attention or posting maliciously to damage credibility [89], which might have different impacts on echo chambers. Finally, our data were limited to the early stage of vaccine promotion, and we did not consider the impact of subsequent virus variants on public perceptions of vaccines. Updated data should be supplemented in follow-up studies. Moreover, this research should be extended to other social media platforms (eg, Zhihu), users with higher information literacy [90], and in discussions about different controversial social issues to evaluate the consistency or differences from our results.

By adopting network analysis, this study evaluated and compared the echo chamber effect in users’ topic selection and attitude interaction based on different social media mechanisms (retweeting, commenting) in the vaccine debate during the public health emergency of COVID-19. We further used statistical and visual analyses to characterize behavioral patterns of key users (opinion leaders, structural hole spanners), and explored their function in avoiding/breaking or preventing/strengthening topic-based and attitude-based echo chambers. These findings could provide meaningful inspiration for health medical and public opinion managers to break online echo chambers and eliminate vaccine hesitancy.