Background: Messages on one’s stance toward vaccination on microblogging sites may affect the reader’s decision on whether to receive a vaccine. Understanding the dissemination of provaccine and antivaccine messages relating to COVID-19 on social media is crucial; however, studies on this topic have remained limited.
Objective: This study applies the elaboration likelihood model (ELM) to explore the characteristics of vaccine stance messages that may appeal to Twitter users. First, we examined the associations between the characteristics of vaccine stance tweets and the likelihood and number of retweets. Second, we identified the relative importance of the central and peripheral routes in decision-making on sharing a message.
Methods: English-language tweets from the United States that contained provaccine and antivaccine hashtags (N=150,338) were analyzed between April 26 and August 26, 2021. Logistic and generalized negative binomial regressions were conducted to predict retweet outcomes. The content-related central-route predictors were measured using the numbers of hashtags and mentions, emotional valence, emotional intensity, and concreteness. The content-unrelated peripheral-route predictors were measured using the numbers of likes and followers and whether the source was a verified user.
Results: Content-related characteristics played a prominent role in shaping decisions regarding whether to retweet antivaccine messages. Particularly, positive valence (incidence rate ratio [IRR]=1.32, P=.03) and concreteness (odds ratio [OR]=1.17, P=.01) were associated with higher numbers and likelihood of retweets of antivaccine messages, respectively; emotional intensity (subjectivity) was associated with fewer retweets of antivaccine messages (OR=0.78, P=.03; IRR=0.80, P=.04). However, these factors had either no or only small effects on the sharing of provaccine tweets. Retweets of provaccine messages were primarily determined by content-unrelated characteristics, such as the numbers of likes (OR=2.55, IRR=2.24, P<.001) and followers (OR=1.31, IRR=1.28, P<.001).
Conclusions: The dissemination of antivaccine messages is associated with both content-related and content-unrelated characteristics. By contrast, the dissemination of provaccine messages is primarily driven by content-unrelated characteristics. These findings signify the importance of leveraging the peripheral route to promote the dissemination of provaccine messages. Because antivaccine tweets with positive emotions, objective content, and concrete words are more likely to be disseminated, policymakers should pay attention to antivaccine messages with such characteristics.
Vaccination against COVID-19 has been promoted by governments as a key strategy to prevent infections and fatalities. The wide spread of the highly contagious omicron variant has made vaccination coverage more imperative than ever. However, an overabundance of information has prevented people from protecting themselves against COVID-19 . Scholars have discovered that people are easily influenced by vaccine-related opinion pieces published on microblogging sites. For example, vaccine hesitancy is closely related to antivaccination campaigns on social media [ , ]. Therefore, understanding the dissemination of provaccine and antivaccine messages on social media websites is crucial. The World Health Organization has called for a greater focus on infodemiology, the area of science research dedicated to understanding the distribution of information through electronic mediums [ - ]. This study examined what characteristics of vaccine stance messages are likely to result in dissemination and whether those characteristics differ between provaccine and antivaccine messages. Answers to these questions will help governments proactively engage in disseminating provaccine messages and identify potentially influential antivaccine messages.
We selected Twitter as the data source because it is the most popular microblogging site, with 397 million active global users as of January 2022 . Microblogging sites have proven their effectiveness in public information adoption and decision-making when used to promote a government vaccination policy [ ]. Twitter allows users to retweet another user’s text to disseminate information among their followers, thus enabling widespread information diffusion.
Prior to the COVID-19 pandemic, researchers have used Twitter data to examine public opinion on vaccinations through text analysis, image analysis, topic modeling, and community detection [, ]. More recently, studies have analyzed the sentiments, opinions, topics, and persuasion techniques related to COVID-19 vaccination on Twitter [ - ]. Furthermore, much effort has been devoted to identifying the determinants of attitudes toward COVID-19 vaccines [ - ], the origin of vaccine misinformation, and its negative effect on vaccine acceptance [ ]. One paper by Germani and Biller-Andorno [ ] reported that compared with provaxxers, antivaxxers tweet less but are more engaged in discussions (through replies or retweets) on Twitter.
Another line of the literature focused on persuasive message appeals, including logos (fact/logic of the argument), pathos (emotion of the argument), and ethos (credibility of the author) . Those rhetoric appeals have been applied to political campaigns, health issues, fund raising, promotion of technological products, and vaccination intake [ - ]. In the Gazette of Australia, logos appeal has been widely utilized for vaccination strategy [ ]. Utilization of pathos on antivaccine websites has been found to provide the functionality of social interactivity [ ]. During the COVID-19 pandemic, New York City Department of Health and Mental Hygiene and the official Twitter account of the US government have extensively utilized rhetoric appeals for vaccine communication and to promote COVID-19 vaccination [ , ].
The existing literature suggests that research on the dissemination of provaccine and antivaccine messages during the COVID-19 pandemic has remained limited. This study applied a theoretical framework called the elaboration likelihood model (ELM) to explore message characteristics that may appeal to Twitter users. Specifically, the aims are (1) to examine the associations between message characteristics and the likelihood and number of retweets and (2) to identify the relative importance of the central and peripheral routes in decision-making on sharing a message. Because vaccine discourse on social media is polarized between groups of provaccine and antivaccine communities , and since provaxxers and antivaxxers hardly interact with each other on Twitter [ ], we conjectured that provaccine messages were predominately shared by provaxxers and antivaccine messages predominately shared by antivaxxers. As a result, we used a common set of message characteristics and tested them separately on provaccine and antivaccine messages. We then explored the role of each route in 2 different groups and compared whether the decision-making on retweets is the same for provaxxers and antivaxxers. To the best of our knowledge, this study is the first to examine the association between the dissemination and characteristics of COVID-19 vaccine stance tweets. The results will facilitate the design of effective messages by scientists, clinicians, and policymakers to promote vaccination.
Theoretical Framework: The Elaboration Likelihood Model
The ELM was developed by Petty and Cacioppo in 1986  and is 1 of the most popular persuasion models in consumer research and social psychology. The ELM proposes that attitude changes and consequent behavior changes among individuals may be caused by 2 processing approaches: the central route and the peripheral route. The central route requires an individual to think deeply about relevant arguments in a message and reflect on the relative merits and relevance of those arguments before developing an informed decision about the target behavior. In the context of decisions to retweet on Twitter, such arguments refer to the message content, such as information richness, argument sentiment, and concreteness, of the tweet. The peripheral route, however, involves less cognitive effort. A message is accepted or rejected without any critical thinking or conscious thought. Recipients simply rely on general criteria or content-unrelated characteristics, such as the information source, to make quick decisions [ ]. In the context of making a decision to retweet, such cues include the number of likes received by the tweet and whether the tweet was posted by a verified user. The ELM predicts that decisions made through central-route processing will be more difficult to alter than those formed through peripheral-route processing.
The ELM has been adopted to study the effects of persuasive communication on attitude and behavioral changes with respect to online reviews , health information [ ], and false reviews [ ]. Drawing on the ELM, Guo et al [ ] investigated patients’ continual usage intentions of mobile health services and Ju and Zhang [ ] investigated the factors influencing patients’ continual use of web-based diagnosis and treatment. The ELM has also been applied to explain users’ decisions to share online reviews of consumer products [ ] and information on social networking sites [ ]. In the field of health communication, the ELM model has helped understand the effectiveness of tobacco package warning labels [ ] and designing of peripheral messages to prevent drunkorexia [ ]. Despite the various empirical studies, applications of the ELM for dissemination of COVID-19 vaccine stance messages are still limited.
Other researchers have explored the effects of message content on users’ retweeting decisions without applying the ELM. Their findings have revealed the impact of argument sentiment [, ] and hashtags [ ]. Studies that did not apply the ELM and focused on content-unrelated factors have also reported positive results. Source trustworthiness, source attractiveness, and favorite counts [ ] affect retweeting decisions.
In practice, to explore the central route, this study used a natural language processing (NLP) technique to construct content-related variables. Content analysis was useful for this study because it exploited timely and real-world messages collected from Twitter and allowed us to identify the actual response (retweet decisions) to specific content. Furthermore, the use of algorithmic content analysis in this study helped the analysis of big data from online discourse faster compared to traditional content analysis methods (where researchers need to formulate a coding scheme and train coders to analyze the text manually) and at scale . Alternatively, experiment and survey methods can be used to discover message characteristics that appeal to users. However, concerns have been raised that convenient sampling widely used by those methods could result in sample selection bias and that the survey methodology captures self-reported behavior rather than the actual behavior/response [ ].
This study is original in a number of ways. We extended the scope of the ELM to vaccine communication and clarified the relative importance of 2 psychological routes in sharing pro- and antivaccine messages. We discovered that both the central and the peripheral route play key roles in the decision-making on whether to share an antivaccine message, whereas dissemination of a provaccine message was mostly determined by the peripheral route. These findings are useful for devising effective messages to promote COVID-19 vaccination and to reach out to different communities on social media. Furthermore, we included a new variable in the central route, called concreteness, that has not been explicitly considered by ELM studies before. We borrowed the concreteness construct from construal level theory (CLT) , which states that concrete words help individuals understand psychological proximity to the respective object or event. Originally, CLT was developed to explain how people think about an event at a concrete or abstract level [ , ]. CLT studies have demonstrated the ability of natural language to prime concrete or abstract mindsets [ , ]—association between lexical concreteness and psychological proximity [ ]. By incorporating concreteness, we have not only enriched the ELM but also extended applications of CLT to vaccine stance message dissemination.
The Elaboration Likelihood Model
Our empirical analysis focused on the 2 routes of the ELM, as presented in. We expected that when users processed information through the central route, message content would be a key predictor of dissemination, whereas when users processed information through a peripheral route, content-unrelated characteristics would be more important predictors. The central route is composed of variables for information richness, argument sentiment (emotional valence and emotional intensity), and concreteness. The peripheral route is composed of variables for informational social influence, source trustworthiness, and source attractiveness.
Study Design, Outcome Variables, and Data Collection
To investigate retweeting behavior, a cross-sectional study design was applied to United States data. The outcome variables were (1) whether a provaccine or antivaccine tweet (collectively termed “vaccine stance tweets”) was retweeted and (2) the number of times a vaccine stance tweet was retweeted.
We used the R library (R Core Team and the R Foundation for Statistical Computing) package rtweet  to access the Twitter application programming interface (API) service to collect provaccine- and antivaccine-related tweets between April 26 and August 26, 2021. We excluded non-English tweets and tweets with a geolocation outside the United States. The provaccine search term hashtags were as follows: #GetVaccinated, #GetVaxxex, #Immunization, #Jab, #Vaccinate, #Vaccinated, #VaccinateNY, #Vaccinesafety, #vaccineswork, and #vaxxed. The following terms were used to target antivaccine tweets: #antivaxx, #antivaxxer, #naturalimmunity, #novaccinepassports, #vaccinefailure, #vaccineinjury, #vaccinemurder, #vaccinesarepoison, #vaccinedontwork, and #vaccinekill. Additionally, we looked into user IDs associated with individual tweets and excluded users who tweeted both pro- and antivaccine messages. This reduced approximately 8.8% of vaccine stance tweets identified in the original data set. The inclusion of only users whose vaccine stances remained consistent during the study period ensured that the tweets analyzed conveyed a clear stance. The final sample was composed of 141,782 provaccine and 8556 antivaccine tweets posted by 57,067 and 4308 distinct users (authors), respectively. The flowchart of Twitter data collection is presented in .
Predictors: Central Route
We operationalized the information richness of a tweet by using 2 measures: the number of hashtags and mentions. A hashtag is a word beginning with the # symbol, which is added to posts to aggregate messages of the same topic. A mention references another user in a microblog with the @ symbol and represents an active user interaction . In the literature, the number of mentions is operationalized as a subdimension of information richness [ ], and we adopted a similar method in this work.
Emotional Valence and Emotional Intensity
In psychology, emotional valence indicates the emotional value expressed on a continuum from unpleasant to pleasant or from negative to positive . Emotional intensity is the expression of emotion in content, indicating the level of subjectivity from no emotion (objective) to highly emotional [ ]. We operationalized these 2 dimensions of argument strength [ ] by using TextBlob [ , ], which generated scores for these dimensions. The values of emotional valence range from −1 to 1, where −1 is extremely negative, 1 is extremely positive, and 0 is neutral. The emotional intensity values range from 0 to 1, where 0 is highly objective and 1 is highly subjective. provides examples of emotional valence and emotional intensity. For example, tweets with positive emotions/valance often contained positive words, such as “natural,” “granted,” “better,” “fine,” “good,” and “healthy.” In contrast, tweets with negative valence used negative words, including “bad,” “evil,” “terrible,” “criminal,” “sick,” “illegal,” and “painful.” TextBlob assigns individual scores to all the words in a set of predefined dictionaries and takes an average of all the sentiments in a sentence to generate the final valence score. Studies have suggested that positive emotions are significantly related to retweets [ ].
TextBlob is a Python library for processing textual data. It provides a simple API for examining common NLP tasks, such as part-of-speech tagging, noun phrase extraction, sentiment analysis, classification, and translation. To extract emotional valence (polarity) and emotional intensity (subjectivity) data, we processed the data set pooled from the final analysis corpus of each vaccine stance. Initially, we used the Python function NeatText, a simple NLP package for cleaning textual data and text preprocessing; we removed user handlers, Universal Resource Locators (URLs), punctuation, non–American Standard Code for Information Interchange characters, numbers, hypertext markup language (html) tags, stopwords, special characters, emojis, and multiple spaces. We then used TextBlob to calculate the value of emotional valence and emotional intensity. Generally, the data are supplied as a bag-of-words, and after assigning individual scores to each word, the final sentiment is represented through a sum pooling of all the sentiments. TextBlob has semantic labels that facilitate fine-grained sentiment analysis. The workflow for calculating emotional valence and emotional intensity is presented in.
Concreteness is an aspect of communication in which the information provided in a message is highly descriptive, specific, and vivid; users generally rely more on concrete wording to make their decisions . Studies have suggested that individuals recall concrete words more effectively than abstract words [ ] and that concrete words are more persuasive in affecting user behavior [ ]; thus, we expected language concreteness to play a role in users’ decision to disseminate vaccine stance tweets. Examples of concreteness are presented in .
To measure the content concreteness of cleaned tweets, we relied on the R package doc2concrete , which uses a dictionary of 40,000 common English words and expressions [ ]. The concrete score has a range of 0-5, where 0 is abstract and 5 is concrete. The validity and reliability of this dictionary have been confirmed in the medical setting [ ] and in online reviews [ ]. Furthermore, this dictionary includes words from the medical domain. For example, “virus” has a concreteness rating of 3.48, whereas “vaccination” has a concreteness rating of 4.24. We calculated concreteness through the workflow presented in .
Predictors: Peripheral Route
Informational Social Influence
We measured informational social influence by using the “favorite” count (ie, number of likes) of a tweet. Researchers have studied informational social influence under the bandwagon effect and related concepts, such as herd behavior and social proof . In practice, we took the square root of the favorite count to resolve convergence problems caused by its large scale (from 0 to nearly 30,000) in regression analysis. This approach has been used by researchers to normalize a skewed distribution. The resultant scale for the favorite count was from 0 to 173.1 for provaccine and from 0 to 100.5 for antivaccine tweets. We also used other normalization techniques, including the z score and min-max normalization; however, for the current model, these methods performed less well in the iterative procedure of maximum likelihood estimation.
The trustworthiness of a tweet is determined by whether the tweet is from a user whose status has been verified . Twitter uses an authentication mechanism to ensure the authenticity of user identity, and a verified user is signified by a blue tick next to the screen name. Therefore, this variable is binary, with 1 indicating a trustworthy user and 0 reflecting a nontrustworthy user. Researchers have noted that tweets from verified users disseminate more rapidly than those from nonverified users [ ].
A Twitter user can follow any other user, and the number of followers reflects the likeability of the user’s real-world status. We measured source attractiveness as the number of followers. Studies that have utilized source attractiveness have identified a substantial effect of a user’s number of followers on the retweetability of a tweet [, ]. We log-transformed the variable to render its scale comparable to other predictors.
Regression Analysis and Sensitivity Analysis
We performed logistic regression and generalized negative binomial (NB) regression on the binary retweet outcome and the number of retweets, respectively. The generalized NB extends the NB mean dispersion model by providing flexibility in parameterizing the dispersion parameter α. We specified that the log of α is a linear function of the same covariates used in the main model. The chi-squared test rejected the null hypothesis that none of the covariates in the dispersion function have predictive power (P<.001). Akaike and Bayesian information criteria also indicated that the generalized NB is preferable to the NB model (). The user-clustered sandwich variance estimator, which accommodates intragroup correlation of observations, was used to improve statistical inferences about regression coefficients. Because vaccine stance tweets were posted across multiple points in time, accounting for various exposures in generalized NB regressions was necessary. We included the log-transformed exposure variable, defined as the number of days from the tweet date to the last day of the study period, August 26, 2021. The correlation coefficient matrix ( ) indicated that the correlation between predictors were generally low, except for the 3 peripheral-route variables, which were moderately correlated (0.3-0.4). All regressions were performed using Stata 16 software (Stata Corp Inc.).
We conducted sensitivity analyses to verify whether the results were robust for various model specifications. First, to capture common trends that may affect decisions of retweets, we included monthly binary variables for June, July, and August in logistic regression models. Data for April and May were combined to serve as the reference group; the results are summarized in. Second, to avoid results being driven by outliers, we excluded tweets that had an exceptionally high number of retweets, using the top 0.5% as a cut-off point. As a result, provaccine tweets that had more than 83 retweets and antivaccine tweets with more than 325 retweets were excluded; see the results in . All analyses revealed that our regression results remained consistent across various model specifications.
Informed consent cannot be obtained to analyze Twitter postings as Twitter posts are publicly available information.
The summary statistics of the model variables are presented in. For provaccine and antivaccine tweets, 28% and 32% were retweeted and the average number of retweets was 3.16 and 8.86, respectively. These findings are consistent with the existing evidence that antivaxxers are more active in message sharing on Twitter [ ]. The average number of hashtags was higher for antivaccine (3.18, SD 2.83) than for provaccine (2.82, SD 2.50) tweets. The mean emotional valence score was 0.07 (SD 0.30) for provaccine and 0.03 (SD 0.28) for antivaccine tweets, indicating that provaccine tweets had more positive emotions than antivaccine tweets. The mean emotional intensity score was similar (0.37, SD 0.34, and 0.35, SD 0.33) for the 2 groups. The mean concreteness score was 2.12 (SD 0.68) for provaccine and 1.92 (SD 0.66) for antivaccine tweets. The mean square root of the number of “likes” was 1.55 (SD 3.41) for provaccine and 1.76 (SD 4.56) for antivaccine tweets. Approximately 6% and 1% of provaccine and antivaccine messages, respectively, were tweeted by a verified user, which was considerably low. This finding accords with research that antivaccine messages are led by nonverified Twitter users [ ]. The mean log number of followers was 6.82 (SD 2.06) for provaccine and 5.94 (SD 1.98) for antivaccine tweets.
|Model variables||Provaccine tweets (N=141,782)||Antivaccine tweets (N=8556)|
|Mean (SD)||Minimum||Maximum||Mean (SD)||Minimum||Maximum|
|Whether retweeted (0/1)||0.28 (0.45)||0||1||0.32 (0.47)||0||1|
|Retweet count||3.16 (60.87)||0||12,500||8.86 (99.72)||0||5141|
|Number of hashtags||2.82 (2.50)||1||32||3.18 (2.83)||1||35|
|Number of mentions||0.72 (1.47)||0||24||0.70 (1.23)||0||14|
|Emotional valence score (–1 to 1)||0.07 (0.30)||–1||1||0.03 (0.28)||–1||1|
|Emotional intensity score (0–1)||0.37 (0.34)||0||1||0.35 (0.33)||0||1|
|Concreteness score (0–5)||2.12 (0.68)||0||4.59||1.92 (0.66)||0||3.74|
|Informational social influence: number of likes (square root)||1.55 (3.41)||0||173.12||1.76 (4.56)||0||100.5|
|Source trustworthiness: a verified user (0/1)||0.06 (0.24)||0||1||0.01 (0.12)||0||1|
|Source attractiveness: number of followers (log)||6.82 (2.06)||0||16.55||5.94 (1.98)||0||12.83|
|Exposurea (log days)||3.20 (1.01)||0||4.81||3.39 (1.16)||0||4.81|
aExposure is defined as the number of days from the tweet date to the last day of the study period, August 26, 2021.
The results from the logistic and generalized NB regressions are presented inand , respectively. All regressions were run separately for provaccine (green color) and antivaccine (red color) tweets to examine the characteristics of messages that may determine the likelihood and number of retweets.
An additional hashtag increased the odds of sharing by 13.3% (95% CI 1.12-1.15, P<.001) and 9.1% (95% CI 1.06-1.12, P<.001) for provaccine and antivaccine tweets, respectively. An additional mention (of another user) increased the odds of sharing provaccine tweets by 3.1% (95% CI 1.01-1.06, P=.02) but reduced the odds of retweeting antivaccine tweets by 10.2% (95% CI 0.84-0.96, P=.002). A 1-point increase in the emotional intensity (subjectivity) score reduced the odds of sharing an antivaccine tweet substantially by 21.6% (95% CI 0.63-0.97, P=.03). Finally, a 1-point increase in concreteness scores increased the odds of sharing an antivaccine tweet substantially by 16.9% (95% CI 1.05-1.30, P=.01).
When the outcome variable was the number of retweets, we obtained similar results to those of the likelihood of retweets. The number of hashtags increased the retweet rate for both provaccine (incidence rate ratio [IRR]=1.07, 95% CI 1.06-1.09, P<.001) and antivaccine (IRR=1.08, 95% CI 1.05-1.11, P<.001) tweets. For antivaccine tweets, the number of mentions decreased the retweet rate by 12% (IRR=0.88, 95% CI 0.83-0.93, P<.001); positive valence increased the retweet rate substantially by 31.8% (IRR=1.32, 95% CI 1.03-1.69, P=.03), and emotional intensity decreased the retweet rate substantially by 20.5% (IRR=0.80, 95% CI 0.64-0.99, P=.04). With respect to provaccine tweets, a 1-point increase in the concreteness score increased the incidence rate of retweets marginally (IRR=1.06, 95% CI 1.00-1.12, P=.046).
The results associated with peripheral routes are presented inand for the likelihood and number of retweets, respectively. An additional square-root number of likes increased the odds of retweets for provaccine and antivaccine messages by a factor of 2.55 (95% CI 2.48-2.61, P<.001) and 4.23 (95% CI 3.84-4.67, P<.001), respectively. Verification of user status increased the odds of retweets for provaccine messages substantially by 45% (95% CI 1.26-1.67, P<.001). A 1% increase in the number of followers increased the odds of retweeting provaccine and antivaccine messages by 30.7% (95% CI 1.27-1.34, P<.001) and 12.1% (95% CI 1.07-1.17, P<.001), respectively.
The generalized NB model indicated that provaccine and antivaccine tweets that had 1 more square-root number of likes had 2.24 (95% CI 2.15-2.34, P<.001) and 2.36 (95% CI 2.08-2.68, P<.001) times more retweets, respectively. When the author was a verified user, the rate of retweeting decreased for both groups (IRR=0.85 [pro], IRR=0.46 [anti], P=.03 [pro], P=.001 [anti]). In contrast, the number of followers increased the incidence rate of retweeting for both groups (IRR=1.28 [pro], IRR=1.18 [anti], P<.001).
This study applied the ELM to investigate characteristics of COVID-19 vaccine stance–related tweets that were associated with the likelihood and number of retweets on Twitter. The key finding is that content-related (central-route) predictors are strongly associated with retweets of antivaccine messages. Specifically, for antivaccine messages, the number of hashtags was positively associated with (the likelihood and number of) retweets; positive valence was associated with a higher number of retweets, concreteness was positively associated with the likelihood of retweets, whereas the number of mentions and emotional intensity were negatively associated with (the likelihood and number of) retweets. Regarding provaccine messages, only the number of hashtags was strongly and positively associated with (the likelihood and number of) retweets; the number of mentions and concreteness were positively but weakly associated with the likelihood of retweets. Among the content-unrelated (peripheral-route) predictors, the number of likes and followers were strongly and positively associated with (the likelihood and number of) retweets of provaccine and antivaccine messages.
Central-Route Predictors Predominantly Associated with Dissemination of Antivaccine Tweets
The ELM predicts that if recipients have a high desire or ability to process a message, they will use the central route and spend more time deliberating on their decision. In this context, if antivaccine messages were mostly shared by antivaxxers, our finding of strong associations between central-route predictors and dissemination of antivaccine messages may imply that antivaxxers have relied more on cognitive cues than provaxxers to make retweeting decisions. Particularly, having positive emotions, low emotional intensity (objective content), and concrete words considerably increased the dissemination of antivaccine tweets. These results warrant attention because they conflict with the general perception that antivaxxers are irrational and attracted by negative emotions and abstract slogans [- ].
The positive associations between concreteness and dissemination of antivaccine messages may be explained by the strategy used by antivaccine message creators to specify the harm caused by COVID-19 vaccines. Specifically, if antivaccine messages include concrete words, then it is likely to motivate the reader to share a descriptive, specific, and factual vaccine stance message. However, the same cannot be said when it comes to utilizing concrete words in provaccine messages, where there is little impact on readers' sharing of vaccine stance messages in this study.
The information systems literature contains inconsistent findings on valence (positive and negative) in electronic word-of-mouth studies [, , ]. A study demonstrated that negative valence has more influence on sharing online reviews of consumer products than positive valence [ ]. Our work provides additional evidence that emotional valence predominantly has positive effects on retweeting antivaccine messages. Furthermore, the negative association between emotional intensity (subjectivity) and dissemination of antivaccine messages supports the existing research [ ] that also indicates a negative effect of emotional intensity on the sharing of information behavior.
With respect to information richness, we discovered that hashtags increase the dissemination of both provaccine and antivaccine tweets, which is consistent with the findings of prior research . Mentioning another user had a small negative effect on the dissemination of antivaccine messages, which is consistent with results that indicate mentions have a negative effect on information sharing [ ]. One possible explanation for this is that in antivaccine messages, mentions are used to cite provaccine users, which is not welcomed by the antivaxxer community.
Peripheral-Route Predictors Associated with Dissemination of Both Provaccine and Antivaccine Tweets
The number of likes (favorite count) measures social influence. It consistently demonstrated a positive association with dissemination of vaccine stance tweets in all models. The finding can be explained by the bandwagon effect, where people follow a trend regardless of the underlying evidence. This trend was stronger for antivaccine users than for provaccine users probably because of their desire to fit into the antivaxxers’ groups . Existing research has revealed that a strong sense of community is a key factor contributing to the success of the antivaccination movement [ ].
In the provaccine models, the association between the verified user status and retweets was inconsistent; in the antivaccine models, the verified user status was negatively associated with the number of retweets. This contradicts our hypothesis that tweets from verified users are more likely to be retweeted. One possible explanation for this trend is that the percentage of verified users was low in both groups (6% and 1% in the provaccine and antivaccine groups, respectively). The predictor varied little, which made fitting the regression line difficult. Moreover, the data revealed that the verified users received more likes and had more followers compared to the nonverified users; the 3 variables were correlated (correlation coefficients=0.3-0.4). When we excluded either the favorite count or the number of followers, the verified user status was positively associated with retweets in all models for provaccine tweets and in 1 model for antivaccine tweets ().
Source attractiveness (number of followers) had positive associations with disseminating both provaccine and antivaccine messages. The literature indicates that having many followers leads to a higher probability of information dissemination .
Recommendations for COVID-19 Vaccination Campaigns Using Social Media
This study provides several insights into how COVID-19 vaccination campaigns can be strengthened. First, to promote the dissemination of provaccine messages, policymakers may consider focusing on peripheral-route predictors (content-unrelated characteristics), such as increasing the likeability of their tweets, engaging with provaxxers who have many followers, and gaining more followers on Twitter. Moreover, to leverage central-route predictors, policymakers may use more hashtags in their messages. Using concrete words in a provaccine message may also increase the number of retweets the message receives, although the effect of doing so was small in this study.
Second, because antivaccine tweets with positive emotions, objective content, and concrete words are more likely to be disseminated, policymakers should pay attention to antivaccine messages with such characteristics. Additionally, paying attention to antivaccine tweets with many likes and followers could be crucial because those tweets are likely to be widely circulated. Research has demonstrated that dissemination of antivaccine messages is driven by strong influencers .
This study has several limitations. First, despite the popularity of Twitter, its users are a selected population and may not be representative of the general United States population. The identification of tweets may be incomplete because of a limited use of hashtags. Second, because Twitter has a strict policy of removing vaccine misinformation tweets from its platform, our data set may have been limited. Third, we examined a user’s retweeting decision when confronting a particular tweet. We were not able to identify people who retweeted those vaccine stance messages, and thus we could not be sure of their vaccine stances. Although most people retweet messages that are consistent with their own principles, some may retweet information that contradicts their beliefs. This limitation has been discussed in another Twitter studies  and should be considered when interpreting the results. Fourth, this study adopted content analysis and did not incorporate the effects of images or emoticons. In a literature review, we found that researchers removed emojis during preprocessing and cleaning of vaccine message text data to study multiple topics in Twitter, such as online vaccination debates [ ], childhood vaccination opinions [ ], COVID-19 vaccine sentiment in the United States [ ], and key themes and topics on COVID-19 vaccines [ ]. On similar lines of the literature, we removed emojis from the Twitter text corpus to analyze our dissemination model. However, emojis can enrich our findings by providing useful information alongside text tweets. Future research may consider including emojis in empirical analysis. Finally, 1 study utilized the data from Twitter posts and compared the sentiment outcomes of TextBlob, VADER, and Word2Vec–bidirectional long short-term memory (Word2Vec-BiLSTM) models. The results showed that TextBlob provides fewer positive sentiments compared to Word2Vec-BiLSTM but provides more positive sentiments compared to VADER [ ]. Despite the wide applications of TextBlob on Twitter data for sentiment analysis [ , ], using different tools to validate emotional valence will help confirm the main findings of this study.
This study identified the characteristics of COVID-19 vaccine stance tweets that are associated with the likelihood and number of retweets. This was performed by applying the ELM and examining 2 psychological routes involved in retweet decisions. A major finding of this study is that the dissemination of antivaccine messages is strongly associated with characteristics related to message content (central-route processing), including emotional valence and intensity. However, message content exhibited a much weaker association with dissemination of provaccine messages. We discovered that dissemination of provaccine messages is predominately determined by content-unrelated characteristics, such as the numbers of likes and followers.
We thank the anonymous reviewers for their careful reading of our manuscript and their insightful comments. The research was conducted through collaboration between the Research Center for Epidemic Prevention, National Yang Ming Chiao Tung University (RCEP-NYCU), and the Intelligent Electronic Commerce Research Center, National Sun Yat-sen University (ECRC-NSYSU), Taiwan. We acknowledge the grant support from the Ministry of Science & Technology, RCEP-NYCU and ECRC-NSYSU.
In addition, the author Y-CY acknowledges the financial support of the Intelligent Electronic Commerce Research Center from the Featured Areas Research Center Program within the framework of the Higher Education Sprout Project by the Ministry of Education (MoE) in Taiwan.
Conflicts of Interest
Examples of emotional valence, emotional intensity, and concreteness in pro- and antivaccine tweets.DOCX File , 22 KB
Workflow for calculating emotional valence, emotional intensity, and concreteness.DOCX File , 35 KB
Akaike information criterion (AIC) and Bayesian information criterion (BIC) for model selection.DOCX File , 16 KB
Correlation coefficient matrices for variables in the pro- and the antivaccine model.DOCX File , 22 KB
Logistic regressions of whether a message was retweeted, with month indicators included.DOCX File , 16 KB
Excluding vaccine stance–related tweets that had an exceptionally high number of retweets, using the top 0.5% as a cut-off point.DOCX File , 17 KB
Regression results from the pro- and antivaccine models, excluding either the number of likes or the number of followers.DOCX File , 18 KB
- World Health Organization. 1st WHO Infodemiology Conference, WHO Infodemic Management: How Infodemics Affect the World & How They Can Be Managed. URL: https://www.who.int/news-room/events/detail/2020/06/30/default-calendar/1st-who-infodemiology-conference [accessed 2021-11-09]
- Wilson SL, Wiysonge C. Social media and vaccine hesitancy. BMJ Glob Health 2020 Oct;5(10):e004206 [FREE Full text] [CrossRef] [Medline]
- Megget K. Even covid-19 can't kill the anti-vaccination movement. BMJ 2020 Jun 04;369:m2184. [CrossRef] [Medline]
- Eysenbach G. Infodemiology and infoveillance: framework for an emerging set of public health informatics methods to analyze search, communication and publication behavior on the Internet. J Med Internet Res 2009 Mar 27;11(1):e11 [FREE Full text] [CrossRef] [Medline]
- Tangcharoensathien V, Calleja N, Nguyen T, Purnat T, D'Agostino M, Garcia-Saiso S, et al. Framework for managing the COVID-19 infodemic: methods and results of an online, crowdsourced WHO technical consultation. J Med Internet Res 2020 Jun 26;22(6):e19659 [FREE Full text] [CrossRef] [Medline]
- Calleja N, AbdAllah A, Abad N, Ahmed N, Albarracin D, Altieri E, et al. A public health research agenda for managing infodemics: methods and results of the first WHO infodemiology conference. JMIR Infodemiol 2021 Sep 15;1(1):e30979 [FREE Full text] [CrossRef] [Medline]
- Statista. Most Popular Social Networks Worldwide as of January 2022, Ranked by Number of Monthly Active Users. URL: https://www.statista.com/statistics/272014/global-social-networks-ranked-by-number-of-users/ [accessed 2022-06-22]
- Langley D, Wijn R, Epskamp S, van BR, Bork V, van R. Should I Get That Jab? Exploring Influence to Encourage Vaccination via Online Social Media. URL: http://aisel.aisnet.org/ecis2015_rip/64 [accessed 2022-06-22]
- Chen T, Dredze M. Vaccine images on Twitter: analysis of what images are shared. J Med Internet Res 2018 Apr 03;20(4):e130 [FREE Full text] [CrossRef] [Medline]
- Tavoschi L, Quattrone F, D'Andrea E, Ducange P, Vabanesi M, Marcelloni F, et al. Twitter as a sentinel tool to monitor public opinion on vaccination: an opinion mining analysis from September 2016 to August 2017 in Italy. Hum Vaccin Immunother 2020 May 03;16(5):1062-1069 [FREE Full text] [CrossRef] [Medline]
- Kwok SWH, Vadde SK, Wang G. Tweet topics and sentiments relating to COVID-19 vaccination among Australian Twitter users: machine learning analysis. J Med Internet Res 2021 May 19;23(5):e26953 [FREE Full text] [CrossRef] [Medline]
- Lyu JC, Han EL, Luli GK. COVID-19 vaccine-related discussion on Twitter: topic modeling and sentiment analysis. J Med Internet Res 2021 Jun 29;23(6):e24435 [FREE Full text] [CrossRef] [Medline]
- Scannell D, Desens L, Guadagno M, Tra Y, Acker E, Sheridan K, et al. COVID-19 vaccine discourse on Twitter: a content analysis of persuasion techniques, sentiment and mis/disinformation. J Health Commun 2021 Jul 03;26(7):443-459. [CrossRef] [Medline]
- Yousefinaghani S, Dara R, Mubareka S, Papadopoulos A, Sharif S. An analysis of COVID-19 vaccine sentiments and opinions on Twitter. Int J Infect Dis 2021 Jul;108:256-262 [FREE Full text] [CrossRef] [Medline]
- Kreps S, Prasad S, Brownstein JS, Hswen Y, Garibaldi BT, Zhang B, et al. Factors associated with US adults' likelihood of accepting COVID-19 vaccination. JAMA Netw Open 2020 Oct 01;3(10):e2025594 [FREE Full text] [CrossRef] [Medline]
- Machingaidze S, Wiysonge CS. Understanding COVID-19 vaccine hesitancy. Nat Med 2021 Aug 16;27(8):1338-1339. [CrossRef] [Medline]
- Razai MS, Chaudhry UAR, Doerholt K, Bauld L, Majeed A. Covid-19 vaccination hesitancy. BMJ 2021 May 20;373:n1138. [CrossRef] [Medline]
- Solís AJS, Warren SS, Meriggi NEA. COVID-19 vaccine acceptance and hesitancy in low and middle income countries, and implications for messaging. medRxiv Preprint posted online May 12, 2021. [CrossRef]
- Jamison A, Broniatowski DA, Smith MC, Parikh KS, Malik A, Dredze M, et al. Adapting and extending a typology to identify vaccine misinformation on Twitter. Am J Public Health 2020 Oct;110(S3):S331-S339. [CrossRef]
- Germani F, Biller-Andorno N. The anti-vaccination infodemic on social media: a behavioral analysis. PLoS One 2021;16(3):e0247642 [FREE Full text] [CrossRef] [Medline]
- Kennedy G. Aristotle, On Rhetoric: A Theory of Civic Discourse. New York and Oxford: Oxford University Press; 2007.
- Johnson J. Twitter bites and Romney: examining the rhetorical situation of the 2012 presidential election in 140 characters. J Contemp Rhetor 2012;2(3/4):54-64.
- McNeill A, Briggs P. Understanding Twitter influence in the health domain: a social-psychological contribution. 2014 Presented at: WWW 2014 Companion - Proceedings of the 23rd International Conference on World Wide Web; April 7-11, 2014; Seoul, Korea p. 673-678. [CrossRef]
- Tirdatov I. Web-based crowd funding: rhetoric of success. Tech Commun 2014;61(1):3-24.
- Cudmore A, Slattery D. An analysis of physical and rhetorical characteristics of videos used to promote technology projects, on the Kickstarter crowdfunding platform. Tech Commun 2019;66(4):319-346.
- McKinnon M, Orthia L. Vaccination communication strategies: what have we learned, and lost, in 200 years? J Sci Commun 2017;16(03):A08. [CrossRef]
- Grant L, Hausman BL, Cashion M, Lucchesi N, Patel K, Roberts J. Vaccination persuasion online: a qualitative study of two provaccine and two vaccine-skeptical websites. J Med Internet Res 2015 May 29;17(5):e133 [FREE Full text] [CrossRef] [Medline]
- Jensen RL. Vaccine Communication During the COVID-19 Pandemic as Vaccine Communication During the COVID-19 Pandemic as Displayed by NYC Health Displayed by NYC Health. URL: https://tinyurl.com/yckstpf9 [accessed 2022-06-24]
- Su M. US Government’s Engagement in Social Media: A Case Study of How Government Promote COVID-19 Vaccination. URL: http://download.bcpub.org/proceedings/2021/ECSS2021/ECSS023.pdf [accessed 2022-06-22]
- Gunaratne K, Coomes EA, Haghbayan H. Temporal trends in anti-vaccine discourse on Twitter. Vaccine 2019 Aug 14;37(35):4867-4871. [CrossRef] [Medline]
- Milani E, Weitkamp E, Webb P. The visual vaccine debate on Twitter: a social network analysis. Media Commun 2020 Jun 25;8(2):364-375. [CrossRef]
- Petty R, Cacioppo J. The elaboration likelihood model of persuasion. Adv Exp Soc Psychol (C) 1986;19:123-205. [CrossRef]
- Guo X, Chen S, Zhang X, Ju X, Wang X. Exploring patients' intentions for continuous usage of mHealth services: elaboration-likelihood perspective study. JMIR Mhealth Uhealth 2020 Apr 06;8(4):e17258 [FREE Full text] [CrossRef] [Medline]
- Cheung C, Sia C, Kuan K. Is this review believable? A study of factors affecting the credibility of online consumer reviews from an ELM perspective. J Assoc Inf Syst 2012 Aug;13(8):618-635. [CrossRef]
- Hanmei F, Reeva L, Stephen S, Shanton C. How online health forum users assess usergenerated content: mixed-method research. 2013 Presented at: Proceedings of the 21st European Conference on Information Systems (ECIS); August 2013; Utrecht, the Netherlands.
- Munzel A. Assisting consumers in detecting fake reviews: the role of identity information disclosure and consensus. J Retail Consum Serv 2016 Sep;32:96-108. [CrossRef]
- Ju C, Zhang S. Influencing factors of continuous use of web-based diagnosis and treatment by patients with diabetes: model development and data analysis. J Med Internet Res 2020 Sep 28;22(9):e18737 [FREE Full text] [CrossRef] [Medline]
- Teng S, Khong KW, Chong AY, Lin B. Examining the impacts of electronic word-of-mouth message on consumers’ attitude. J Comput Inf Syst 2016 Sep 09;57(3):238-251. [CrossRef]
- Shi J, Hu P, Lai KK, Chen G. Determinants of users’ information dissemination behavior on social networking sites: an elaboration likelihood model perspective. Internet Res 2018 Apr 04;28(2):393-418. [CrossRef]
- Strahan E, White K, Fong G, Fabrigar L, Zanna M, Cameron R. Enhancing the effectiveness of tobacco package warning labels: a social psychological perspective. Tob Control 2002 Sep;11(3):183-190 [FREE Full text] [CrossRef] [Medline]
- Glassman T, Paprzycki P, Castor T, Wotring A, Wagner-Greene V, Ritzman M, et al. Using the elaboration likelihood model to address drunkorexia among college students. Subst Use Misuse 2018 Jul 29;53(9):1411-1418. [CrossRef] [Medline]
- Berger J. What makes online content viral? Strateg Dir 2012;28(8):90-91. [CrossRef]
- Stieglitz S, Dang-Xuan L. Emotions and information diffusion in social media—sentiment of microblogs and sharing behavior. J Manag Inf Syst 2014 Dec 08;29(4):217-248. [CrossRef]
- Suh B, Hong L, Pirolli P, Chi E. Want to be retweeted? Large scale analytics on factors impacting retweet in twitter network. 2010 Presented at: SocialCom 2010: the Second IEEE International Conference on Social Computing, PASSAT nd IEEE International Conference on Privacy, Security, Risk and Trust. Published online; August 20-22, 2010; Minneapolis, Minnesota, USA. [CrossRef]
- Rudat A, Buder J. Making retweeting social: the influence of content and context information on sharing news in Twitter. Comput Human Behav 2015 May;46:75-84. [CrossRef]
- Lai L, To W. Content analysis of social media: a grounded theory approach. J Electron Commer Res 2015;16(2):138.
- Melki J, Tamim H, Hadid D, Makki M, El Amine J, Hitti E. Mitigating infodemics: the relationship between news exposure and trust and belief in COVID-19 fake news and social media spreading. PLoS One 2021 Jun 4;16(6):e0252830 [FREE Full text] [CrossRef] [Medline]
- Trope Y, Liberman N. Construal-level theory of psychological distance. Psychol Rev 2010 Apr;117(2):440-463 [FREE Full text] [CrossRef] [Medline]
- Liberman N, Trope Y. The role of feasibility and desirability considerations in near and distant future decisions: a test of temporal construal theory. J Pers Soc Psychol 1998;75(1):5-18. [CrossRef]
- Trope Y, Liberman N. Temporal construal. Psychol Rev 2003 Jul;110(3):403-421. [CrossRef] [Medline]
- Lermer E, Streicher B, Sachs R, Raue M, Frey D. The effect of abstract and concrete thinking on risk-taking behavior in women and men. SAGE Open 2016 Aug 26;6(3):215824401666612. [CrossRef]
- Lermer E, Streicher B, Sachs R, Raue M, Frey D. Thinking concretely increases the perceived likelihood of risks: the effect of construal level on risk estimation. Risk Anal 2016 Mar 26;36(3):623-637. [CrossRef] [Medline]
- Kirkpatrick AW. The spread of fake science: lexical concreteness, proximity, misinformation sharing, and the moderating role of subjective knowledge. Public Underst Sci 2021 Jan 24;30(1):55-74. [CrossRef] [Medline]
- Package ‘rtweet’. URL: https://cran.r-project.org/web/packages/rtweet/rtweet.pdf [accessed 2021-11-10]
- Yang J, Counts S. Predicting the speed, scale, and range of information diffusion in Twitter. 2010 Presented at: Fourth International AAAI Conference on Weblogs and Social Media; May 16, 2010; Washington, DC.
- Liu Z, Liu L, Li H. Determinants of information retweeting in microblogging. Internet Res 2012 Aug 10;22(4):443-466. [CrossRef]
- Eschrich S, Münte TF, Altenmüller EO. Unforgettable film music: the role of emotion in episodic long-term memory for music. BMC Neurosci 2008 May 28;9(1):48 [FREE Full text] [CrossRef] [Medline]
- Peng C, Yin D, Zhang H. More than words in medical question-and-answer sites: a content-context congruence perspective. Inf Syst Res 2020 Sep;31(3):913-928. [CrossRef]
- Sattar NS, Arifuzzaman S. COVID-19 vaccination awareness and aftermath: public sentiment analysis on Twitter data and vaccinated population prediction in the USA. Appl Sci 2021 Jun 30;11(13):6128. [CrossRef]
- Gbashi S, Adebo OA, Doorsamy W, Njobeh PB. Systematic delineation of media polarity on COVID-19 vaccines in Africa: computational linguistic modeling study. JMIR Med Inform 2021 Mar 16;9(3):e22916 [FREE Full text] [CrossRef] [Medline]
- Larrimore L, Jiang L, Larrimore J, Markowitz D, Gorski S. Peer to peer lending: the relationship between language features, trustworthiness, and persuasion success. J Appl Commun Res 2011 Feb;39(1):19-37. [CrossRef]
- Tse C, Altarriba J. The word concreteness effect occurs for positive, but not negative, emotion words in immediate serial recall. Br J Psychol 2010 Dec 31;100(1):91-109. [CrossRef]
- Yeomans M. A concrete example of construct construction in natural language. Organ Behav Human Decis Process 2021 Jan;162:81-94. [CrossRef]
- Brysbaert M, Warriner AB, Kuperman V. Concreteness ratings for 40 thousand generally known English word lemmas. Behav Res 2013 Oct 19;46(3):904-911. [CrossRef]
- Cousins KAQ, Ash S, Olm CA, Grossman M. Longitudinal changes in semantic concreteness in semantic variant primary progressive aphasia (svPPA). eNeuro 2018 Dec 26;5(6):1-10. [CrossRef]
- Ransbotham S, Lurie NH, Liu H. Creation and consumption of mobile word of mouth: how are mobile reviews different? Mark Sci 2019 Sep;38(5):773-792. [CrossRef]
- Zhang L, Xu L, Zhang W. Social media as amplification station: factors that influence the speed of online public response to health emergencies. Asian J Commun 2017 Feb 20;27(3):322-338. [CrossRef]
- Broniatowski DA, Jamison AM, Qi S, AlKulaib L, Chen T, Benton A, et al. Weaponized health communication: Twitter bots and Russian trolls amplify the vaccine debate. Am J Public Health 2018 Oct;108(10):1378-1384. [CrossRef] [Medline]
- Memon S, Tyagi A, Mortensen D, Carley K. Characterizing sociolinguistic variation in the competing vaccination communities. In: Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). 2020;12268 LNCS. Cham: Springer International; 2020:118-129.
- Cohen N. Anti-Vaxxers Posing as Victims Has a History: Look at Andrew Wakefield. URL: https://www.theguardian.com/commentisfree/2021/jan/30/why-would-big-tech-police-anti-vaxxers-when-they-are-so-good-for-business [accessed 2022-06-22]
- Miyazaki K, Uchiba T, Tanaka K, Sasahara K. Characterizing the anti-vaxxers' reply behavior on social media. 2021 Presented at: WI-IAT '21: IEEE/WIC/ACM International Conference on Web Intelligence and Intelligent Agent Technology; December 2021; Melbourne, VIC, Australia p. 83-89. [CrossRef]
- van Herpen E, Pieters R, Zeelenberg M. When demand accelerates demand: trailing the bandwagon. J Consum Psychol 2009 Jul;19(3):302-312. [CrossRef]
- Yuan X, Schuchard RJ, Crooks AT. Examining emergent communities and social bots within the polarized online vaccination debate in Twitter. Soc Media Soc 2019 Sep 04;5(3):205630511986546. [CrossRef]
- Featherstone JD, Ruiz JB, Barnett GA, Millam BJ. Exploring childhood vaccination themes and public opinions on Twitter: a semantic network analysis. Telemat Inform 2020 Nov;54:101474. [CrossRef]
- Bustos VP, Comer CD, Manstein SM, Laikhter E, Shiah E, Xun H, et al. Twitter voices: Twitter users’ sentiments and emotions about COVID-19 vaccination within the United States. Eur J Environ Public Health 2022;6(1):em0096. [CrossRef]
- Chandrasekaran R, Desai R, Shah H, Kumar V, Moustakas E. Examining public sentiments and attitudes toward COVID-19 vaccination: infoveillance study using Twitter posts. JMIR Infodemiol 2022 Apr 15;2(1):e33909 [FREE Full text] [CrossRef] [Medline]
- Hussain A, Tahir A, Hussain Z, Sheikh Z, Gogate M, Dashtipour K, et al. Artificial intelligence-enabled analysis of public attitudes on Facebook and Twitter toward COVID-19 vaccines in the United Kingdom and the United States: observational study. J Med Internet Res 2021 Apr 05;23(4):e26627 [FREE Full text] [CrossRef] [Medline]
- Cresswell K, Tahir A, Sheikh Z, Hussain Z, Domínguez Hernández A, Harrison E, et al. Understanding public perceptions of COVID-19 contact tracing apps: artificial intelligence-enabled social media analysis. J Med Internet Res 2021 May 17;23(5):e26618 [FREE Full text] [CrossRef] [Medline]
|API: application programming interface|
|BiLSTM: bidirectional long short-term memory|
|CLT: construal level theory|
|ELM: elaboration likelihood model|
|IRR: incidence rate ratio|
|NB: negative binomial|
|NLP: natural language processing|
|OR: odds ratio|
Edited by A Mavragani; submitted 11.02.22; peer-reviewed by SF Tsao, A Pal; comments to author 11.04.22; revised version received 28.05.22; accepted 16.06.22; published 27.06.22Copyright
©Vipin Saini, Li-Lin Liang, Yu-Chen Yang, Huong Mai Le, Chun-Ying Wu. Originally published in JMIR Infodemiology (https://infodemiology.jmir.org), 27.06.2022.
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 JMIR Infodemiology, is properly cited. The complete bibliographic information, a link to the original publication on https://infodemiology.jmir.org/, as well as this copyright and license information must be included.