This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in the Journal of Medical Internet Research, is properly cited. The complete bibliographic information, a link to the original publication on https://www.jmir.org/, as well as this copyright and license information must be included.
Social media provide an ideal medium for breeding and reinforcing vaccine hesitancy, especially during public health emergencies. Algorithmic recommendation–based technology along with users’ selective exposure and group pressure lead to online echo chambers, causing inefficiency in vaccination promotion. Avoiding or breaking echo chambers largely relies on key users’ behavior.
With the ultimate goal of eliminating the impact of echo chambers related to vaccine hesitancy on social media during public health emergencies, the aim of this study was to develop a framework to quantify the echo chamber effect in users’ topic selection and attitude contagion about COVID-19 vaccines or vaccinations; detect online opinion leaders and structural hole spanners based on network attributes; and explore the relationships of their behavior patterns and network locations, as well as the relationships of network locations and impact on topic-based and attitude-based echo chambers.
We called the Sina Weibo application programming interface to crawl tweets related to the COVID-19 vaccine or vaccination and user information on Weibo, a Chinese social media platform. Adopting social network analysis, we examined the low echo chamber effect based on topics in representational networks of information, according to attitude in communication flow networks of users under different interactive mechanisms (retweeting, commenting). Statistical and visual analyses were used to characterize behavior patterns of key users (opinion leaders, structural hole spanners), and to explore their function in avoiding or breaking topic-based and attitude-based echo chambers.
Users showed a low echo chamber effect in vaccine-related topic selection and attitude interaction. For the former, the homophily was more obvious in retweeting than in commenting, whereas the opposite trend was found for the latter. Speakers, replicators, and monologists tended to be opinion leaders, whereas common users, retweeters, and networkers tended to be structural hole spanners. Both leaders and spanners tended to be “bridgers” to disseminate diverse topics and communicate with users holding cross-cutting attitudes toward COVID-19 vaccines. Moreover, users who tended to echo a single topic could bridge multiple attitudes, while users who focused on diverse topics also tended to serve as bridgers for different attitudes.
This study not only revealed a low echo chamber effect in vaccine hesitancy, but further elucidated the underlying reasons from the perspective of users, offering insights for research about the form, degree, and formation of echo chambers, along with depolarization, social capital, stakeholder theory, user portraits, dissemination pattern of topic, and sentiment. Therefore, this work can help to provide strategies for public health and public opinion managers to cooperate toward avoiding or correcting echo chamber chaos and effectively promoting online vaccine campaigns.
Despite scientific consensus that COVID-19 vaccines are safe and effective [
Users in a social network can be divided into three roles: opinion leaders, structural hole spanners, and ordinary users [
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.
Most studies in this field have concentrated on the presence, form, and degree of echo chambers, whereas limited research has aimed to develop efficient strategies to address the echo chamber effect. Schmidt et al [
Social capital, as a set of resources embedded in relationships, results from holding certain locations in a social structure [
Burt [
The lack of connection among communities forms structural holes in social structures [
Research about opinion leaders’ impact on echo chambers has resulted in contradictory conclusions with respect to different social issues. Limited research has focused on the impact of structural hole spanners on echo chambers. Rather, research in this field has mainly focused on opinion-based echo chambers under specific topics, while ignoring users’ topic selection prior to opinion contagion. Despite these advancements, a gap remains in the literature: if both bonding and bridging arguments are valid depending on the context, under which conditions should they be complementary or otherwise?
Online echo chambers have been studied in the context of users’ interactions (eg, posting, retweeting, commenting, mentioning), focusing on rumor spread and management [
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
Research framework.
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 [
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 RQ1-4, we constructed interactive networks. Information representational and user communication flow networks are commonly used as the basis to measure polarization [
First, we marked the topic for each original tweet. To cover all aspects of the vaccine, we performed this process based on the Health Belief Model, which indicates that the perceived susceptibility, perceived severity, perceived benefits, perceived barriers, cues to action, and self-efficacy have impacts on individuals’ motivation to carry out preventive health behaviors [
Retweeting or commenting on an original tweet indicates that the users are interested in the tweet’s topic [
Topic-coding scheme based on the Health Belief Model.
Construct | Topics |
Perceived susceptibility | Risk of getting COVID-19 infection |
Perceived severity | Severity of getting COVID-19 infection or refusing COVID-19 vaccination |
Perceived benefits | Effectiveness of COVID-19 vaccination |
Perceived barriers | Adverse effects of COVID-19 vaccination; cost of COVID-19 vaccination; fake (eg, counterfeit) vaccines, fraudulent information; safety (eg novelty), infectivity of vaccines, and standardization of vaccination process; conspiracy theory |
Cues to action | Means to get vaccination; dos and don’ts of vaccination; domestic vaccine development, production, and vaccination; foreign vaccine development, production, and vaccination; personal vaccination experience |
Many studies adopted the sentiment expressed in tweets created/retweeted/commented by users to represent their attitudes toward vaccines [
To answer RQ1, we used Python’s NetworkX package to calculate each network’s assortativity coefficient
To answer RQ2, we characterized online users’ behavior patterns and detected opinion leaders and structural hole spanners based on their network locations. To answer RQ3-4, we defined two types of mediators to represent the above key users’ contributions to echo chambers. After coding users from these three perspectives, statistical tests were used to examine the relationships.
Villodre and Criado [
User behavior taxonomy.
User category | Criterion | Behavior description | |
|
|||
|
Speaker | Number of retweets received was three times higher (low speakers), 10 times higher (medium speakers), or 100 times higher (high speakers) than that of tweets they had posted | Users create widely shared content. They show less content-sharing behavior |
|
Networker | Number of tweets≥total mean; number of retweets received≥total mean; number of retweets received/number of retweets sent≥0.5 | Users show equilibrium between creating content, sharing content, and being retransmitted |
|
|||
|
Monologist | Number of tweets≥total mean; number of retweets received/own tweets≤0.3 | Users create original content that is not widely shared |
|
Retweeter | Number of tweets≥total mean; number of retweets sent/own tweets≥0.5 | Users mostly share others’ content |
|
Replicator | Number of comments sent/own tweets≥0.6 | Users mostly comment on others’ content |
|
Isolator | Number of retweets sent=0; number of retweets received=0; number of comments sent=0; number of comments received=0 | Users never share/comment on others’ content and they create some content that is never shared/commented by others |
|
Automatic | Send same comments multiple times under one tweet; personal information is blank | Users seem to act with automatization |
Common user | None of the above | Not applicable |
Stakeholder types and related keywords.
Stakeholder types | Keywords (partial) |
Government | government, police, court, judicial bureau, judicial office, procuratorate, commission for discipline inspection, political and legal committee |
Hospital | hospital |
Traditional media | newspaper, radio, TV station, news, magazine, broadcast, daily, timely, weekly, monthly, morning post, evening post, channel |
We-media | We-media, author, writer, reporter, editor, blogger, commentator, critic |
Platform account | Sina Weibo, Weibo medical and health operation, Weibo secretary, Weibo administrator, Weibo rumor rebuttal, Weibo politics |
Social organization | association, public welfare |
Medical company | vaccine manufacturer (“SINOVAC BIOTECH CO., LTD”. [“科兴”], “CanSino Biologics Inc.” [“康希诺”], “Hualan,” “Zhifei,” “Kangtai”), medical enterprise |
Common company | company, enterprise |
Educational institution | middle school, high school, campus, technical school |
Medical personnel | doctor, nurse |
Common personnel | None of the above |
To measure the extent of each user being regarded as an opinion leader, we adopted in-degree centrality [
To uncover the mechanisms of intra- and intergroup communication among holders of different interests and viewpoints, we conceptualized two types of social mediators. One was the “echoer,” who only initiated interactions with peers whose interests and viewpoints were highly homogenous, thereby contributing to the formation and even consolidation of echo chambers [
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.
The distribution of attitudes expressed in tweets (original tweets, retweets, comments) about different topics.
Retweet, comment, and global information networks were all sparse, with a density of 0.003, 0.003, and 0.0002, respectively. In
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 [
Chord diagram representation of the retweet information network (a), comment information network (b), and global information network (c) colored by topic.
Communication flow network of users in the (a) retweet user network (b) comment user network, and (c) global user network. The size and color of each node represent its in-degree and user’s attitude (red=“positive”, blue=“negative”, orange=“neutral”), respectively. The color of the edge is explained in the corresponding legends in the figure.
As shown in
As shown in
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
As shown in
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
To address RQ4, the support and the confidence of the rule were calculated. As shown in
Percentage of user categories based on their behavior. No automatics were detected in the data set.
The distribution of users’ weighted in-degree centrality, weighted out-degree centrality, and local clustering coefficient (the size of the circle). The depth of the shadow represents the number of users with corresponding centrality and clustering coefficients.
Distribution of users’ structural hole indices (speakers and networks). The white dot, and upper and lower lines of the thick black line represent the index’s median, third quantile, and first quantile, respectively. The width of the red shadow represents the percentage of specific-category users whose index took on that value.
Distribution of users’ structural hole indices (monologists, retweeters, replicators, and common users). The white dot, and upper and lower lines of the thick black line represent the index’s median, third quantile, and first quantile, respectively. The width of the red shadow represents the percentage of specific-category users whose index took on that value.
Percentages of tweets from key users.
Composition of user categories in opinion leaders/structural hole spanners and their role in the topic-based echo chamber (left) and attitude-based echo chamber (right).
Support and confidence of research question 4 (RQ4).
User category | Number of users | Number of users as topic-based echoers/bridgers | Number of users as both topic-based and attitude-based echoers/bridgers | Supporta | Confidenceb | ||||||
|
|||||||||||
|
Opinion leader | 386 | 138 | 8 | 0.021 | 0.058 | |||||
|
Structural hole spanner | 3123 | 807 | 23 | 0.007 | 0.029 | |||||
|
Total | 3509 | 945 | 31 | 0.009 | 0.033 | |||||
|
|||||||||||
|
Opinion leader | 386 | 248 | 234 | 0.606 | 0.944 | |||||
|
Structural hole spanner | 3123 | 2316 | 1968 | 0.630 | 0.850 | |||||
|
Total | 3509 | 2564 | 2202 | 0.628 | 0.859 |
aSupport equals the number of users as both topic-based and attitude-based echoers (RQ4a)/bridgers (RQ4b) divided by the total number of users.
bConfidence equals the number of users as both topic-based and attitude-based echoers (RQ4a)/bridgers (RQ4b) divided by the number of users as topic-based echoers/bridgers.
cRQ4a: Do key users acting as echoers in topic dissemination tend to play the same role in attitude interaction?
dRQ4b
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 [
In contrast to the findings of Mønsted and Lehmann [
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) [
Consistent with Yang et al [
Tan et al [
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 [
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 [
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 [
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 [
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.
research question
Funding for this study was provided by the National Natural Science Foundation of China (71661167007, 71420107026) and by the National Key Research and Development Program of China (2018YFC0806904-03).
None declared.